Not impressive on its face. Children do it at school. There are people on TikTok who do it faster than you can finish reading this sentence. What made it interesting was that the hand had never touched a Rubik’s cube. It had never touched anything physical at all. It had been trained, instead, inside a simulator for thirteen thousand years.

Thirteen thousand simulated years. To learn one move.

Real Olympic gymnasts train for fifteen.

The project was called Dactyl, run by Sergey Levine, Wojciech Zaremba, and a team at OpenAI. It was a tour de force and an embarrassment at the same time. Tour de force: you could, in principle, transfer a skill from a simulator into the real world. Embarrassment: doing so required burning a small data centre’s worth of compute to teach a robotic hand a child’s pastime, and even then, the hand dropped the cube most of the time and could not generalise to anything else. The technique was called domain randomisation. The grim joke in the field was that the only thing it reliably randomised was the patience of whoever was paying the cloud bill.

For the next four years, this was the state of robotics. You could simulate. You could not transfer. The gap between what worked in code and what worked in the world had a name. Sim-to-real. Roboticists said it the way oncologists say “metastasis.” It was the thing that killed your project.

Then, very quietly, in the eighteen months between mid-2024 and early 2026, it stopped killing projects.

You did not see a press release. There was no DeepMind blog post called “Sim-to-Real Just Quietly Solved.” But somewhere between NVIDIA opening its Newton simulator under the Linux Foundation, Physical Intelligence shipping π0.5, Unitree pricing a humanoid at sixteen thousand dollars, and Boston Dynamics retrofitting its production line with electric Atlas units, the calculus changed. The thing that used to take thirteen thousand simulated years now takes a weekend. The thing that used to require a team of twenty-five PhDs now ships from a startup with twelve people. The thing that used to be a demo is now a deployment.

This is what people miss when they talk about AI right now. They argue about chatbots and image generators and whether transformer scaling has plateaued. Meanwhile, the world’s most consequential AI breakthrough is happening in the part of the stack that touches atoms. The physical AI stack is evolving faster than anything in our lifetimes. And it is opening up a category of companies that have not been viable since the late 1990s. Companies whose value comes from understanding a piece of the physical world better than anyone else.

The world of interesting technology, the kind that touches the things you can hold, is coming back. And this essay is a map of how it got here, what is now possible, and where we think the next ten companies should be built.

The Forty Years of Falling Over

To understand why the last eighteen months feel like a phase change, you have to understand what came before. Robotics simulation is older than the personal computer.

In 1986, Marc Raibert’s leg lab at MIT built the first dynamic running robot. The control loops were written in LISP. The simulators were hand-coded Newtonian physics, run on minicomputers, and the gap between simulation and reality was so wide that Raibert famously preferred to debug on the real machine. Forty years later, this remained the standard advice in the field. Trust the metal, not the model. The simulator was a toy.

The next decade tried to fix this with better physics. The Open Dynamics Engine, written by Russell Smith in the late 1990s, gave the field a free rigid-body simulator that anyone could plug into a robotics project. Gazebo, launched in 2002, gave the field a simulation environment with sensors. MuJoCo, written by Emo Todorov at the University of Washington in 2012, gave the field a contact-rich physics engine fast enough for control research. None of these closed the gap. They got the physics closer. The reality gap stayed huge. A controller that ran at sixty hertz in Gazebo would, on a real robot, hit the wall, fall over, and look at you accusingly.

The cleverest people in the field gave up trying to make simulators more accurate and instead tried to make them more diverse. Randomise the friction. Randomise the mass. Randomise the lighting. Randomise the texture. Train a controller across thousands of randomised worlds and pray that the real world fell somewhere inside that distribution. This was domain randomisation, the idea Sergey Levine and his collaborators formalised in a 2017 paper that became one of the most cited works in modern robotics. It is the technique that made Dactyl possible. And it is the technique that, by 2022, had been pushed to its conceptual limits.

Domain randomisation could close the gap on simple tasks. It could not close the gap on dexterity, on locomotion across novel terrain, on manipulation of objects the robot had not seen. It was the right idea trapped in the wrong substrate. The compute was too expensive. The simulators were too slow. The neural networks were too small.

Meanwhile, the rest of AI was eating the world.

Between 2017 and 2022, language models scaled from one hundred million parameters to one trillion. Image generators went from blurry rabbits to photorealistic everything. The compute spent on training a single large model went from a few thousand GPU-hours to a few million. And robotics? Robotics watched.

There was a robotics renaissance happening on paper. RT-1 from Google in 2022. RT-2 in 2023. Diffusion policies. Behaviour cloning at scale. Cross-embodiment learning. But on the floor, on the actual floors of actual factories, on the actual roofs of actual warehouses, in the actual control rooms of actual refineries, the robots that worked were the ones that had been there since 1995. Articulated arms doing pick-and-place. AGVs running on pre-mapped routes. Conveyor systems with computer vision. The new stuff lived in YouTube videos. A robotics friend in the Bay Area told me in late 2023 that he had watched more demos in his career than products. And the demos were always two years from production.

He is now CTO of a venture-backed humanoid company. He told me, in March 2026, that he no longer says that.

What Changed (And Why You Didn’t Notice)

In a single eighteen-month window, three things collided.

The first was the simulators. NVIDIA’s Newton, an open-source GPU-native physics engine developed jointly with Google DeepMind, Disney Research, and the Linux Foundation, hit a public release in 2025. Newton runs differentiable physics. Which means a robot’s training loop can pass gradients all the way through the simulated world the way a deep neural network passes gradients through its own layers. Genesis, an academic simulator released in late 2024 by a multi-institution team led out of CMU, claimed forty-three million simulation frames per second on a single RTX 4090. A claim that would have been dismissed as a typo three years ago. And Isaac Lab, NVIDIA’s robotics training stack, started shipping reference implementations that let a researcher iterate on a humanoid locomotion policy in hours instead of weeks.

The second was the world models. Not the physics-engine kind. The video-generation kind. In 2024, DeepMind released Genie 2. In 2025, Genie 3, which produced minutes-long playable worlds with persistent physics and consistent object permanence. NVIDIA released Cosmos in January 2025, a foundation model explicitly trained to generate physically plausible video for robot training. Decart’s Oasis open-sourced a real-time interactive world model the same quarter. These are not toys. A world model trained on internet video can produce training data for a robot that has never existed yet, in environments that have never been simulated, with physics that emerge from data rather than equations. The robot is no longer learning from a coded physics engine. It is learning from a model that learned the physics from watching the world.

The third, and most under-appreciated, was the foundation models for embodiment. Physical Intelligence’s π0, released in October 2024, was the first cross-embodiment robotics foundation model that could be fine-tuned for new robot bodies with a fraction of the data of training from scratch. π0.5, released April 2025, generalised to homes the model had never seen. Google DeepMind’s Gemini Robotics shipped in March 2025. Skild AI announced its “general-purpose brain” in early 2025 and raised at four and a half billion dollars on the back of it. The pattern across all of them is the same. A single model that can drive any robot, that can learn from any data, that generalises across tasks without retraining.

Put them together and what you get is a closed loop. A robot collects data in the real world. The data feeds a world model. The world model generates synthetic experience. The synthetic experience trains the foundation model. The foundation model deploys to the next robot. The next robot collects more real data. Loop. The thing that took thirteen thousand simulated years in 2019 takes a weekend in 2026, because the loop spins faster than anyone in the field predicted.

It is not that the sim-to-real gap closed. It is that the loop that closes the gap got cheap.

What Pixar Already Did to Itself

The closest historical parallel to what is happening in robotics right now is not a robotics parallel. It is a Pixar parallel.

In 1986, Steve Jobs bought a small computer graphics division from Lucasfilm for five million dollars and called it Pixar. Their dream was the first feature-length computer-animated film. The technology to make that film did not exist. To render a single frame of Toy Story at the resolution needed for a cinema screen, the Pixar render farm (a custom-built cluster of Sun SPARCstations) took, on average, between six and twenty-nine hours. The film contained one hundred and ten thousand frames. The render farm ran around the clock for over a year just to finish the pixels. The film cost thirty million dollars to make and earned three hundred and seventy-three million. Pixar went public the next week. The world changed.

What changed at Pixar over the following twenty years is the more interesting story.

By the time Monsters University shipped in 2013, each frame of fur on Sulley took an average of twenty-nine hours to render. Not less. More. Pixar had spent two decades pouring every available efficiency back into higher resolution, more accurate light transport, finer geometry, more particles, more cloth, more hair. The render farm did not get faster in any user-facing sense. It got more honest. Pixar moved from approximating reality to simulating it.

Around 2015, GPUs and machine learning denoisers and real-time path tracing collided. Soul in 2020 used live previews on workstations. Lightyear in 2022 used Unreal Engine for set lighting. By 2024, every major studio was iterating on photorealistic scenes in something close to real time. The render farm did not disappear. The render farm became the substrate that everything else ran on. The bottleneck moved from “can we render this” to “do we have a story worth rendering.”

This is the parallel. Robotics right now is in the Toy Story moment. The render farm in this case is the simulator, the world model, the foundation model, the GPU cluster. It has just gotten good enough and cheap enough to make the impossible possible. The next ten years of robotics will look like the next ten years of Pixar. The simulators will keep getting better. The world models will keep absorbing more data. The foundation models will keep transferring across more bodies. The cost of trying something physical will keep collapsing.

And, exactly as in animation, the bottleneck will move. It will move from “can the robot do the thing” to “do we have a story worth telling, in atoms, with this robot.”

Which is to say: the bottleneck moves to the operator. The person who actually knows what the plant needs done. The person who knows which inspection on which pipe at which interval would actually pay for itself. The person who knows where the work is hidden.

The Bottleneck Moved

Look at where capital is being priced in physical AI right now and you will see the same shift.

Physical Intelligence: founded 2024, last reported valuation around five billion dollars [Fact, public reports late 2024]. Skild AI: four and a half billion [Fact, 2025]. Figure: forty billion in the late 2025 round [Fact, public reports]. Wayve: raised over a billion at a multi-billion valuation [Fact, 2024]. Apptronik: backed by Google and Mercedes [Fact, 2025]. The Bot Company, founded by Kyle Vogt after Cruise: building household robots, valued in the low single-digit billions before shipping a product [Fact, 2025].

The numbers are large. The numbers are not the interesting part. The interesting part is what they are pricing.

None of these companies are priced on hardware. The Unitree G1 humanoid, available since 2024, lists at $16,000 [Fact, manufacturer]. Industrial six-axis arms from Universal Robots cost less than $30k. Mobile bases from Clearpath, Boston Dynamics Spot, AgileX are all commoditising. The cost of an embodied platform that can be programmed to do useful work has collapsed by more than ten times since 2020 [Hypothesis, based on aggregated public pricing]. The hardware is not where the moat lives.

None of these companies are priced on the foundation model itself, either. The models are convergent. Cross-embodiment foundation models from PI, Skild, Google DeepMind, NVIDIA, and a handful of academic groups will follow the same arc as LLMs. They will commoditise. The architectures will leak. The weights will leak. The differentiation will compress. The model is a feature, not a moat.

What the valuations are pricing is the loop.

The loop is the proprietary thing. Who owns the closed circuit between real-world deployment, real-world data collection, real-to-sim ingestion, sim-to-real retraining, and re-deployment. Physical Intelligence’s loop runs across multiple commercial deployments. Skild’s loop runs across academic and partner robots. Figure’s loop runs through BMW Spartanburg and a handful of other plants. Tesla’s loop runs through its own factories. Over a thousand Optimus units inside Fremont and Austin in early 2026 [Hypothesis, based on Musk’s public statements; the verified number is lower]. The loop is what compounds. The loop is what is, finally, defensible.

Three eras of robotics simulation:

Classical robotics (1980s–2010s). Hand-coded physics. Bottleneck: the reality gap. Canonical projects: Honda ASIMO, Boston Dynamics BigDog.

Reinforcement learning renaissance (2017–2022). Domain randomisation. Bottleneck: compute and transfer. Canonical projects: OpenAI Dactyl, Google RT-1.

Foundation model era (2024 onward). Differentiable sim + world models + cross-embodiment foundation models. Bottleneck: loop ownership. Canonical projects: π0.5, GR00T, Helix, Newton.

This shift is why operator-founders are suddenly the right profile for the category. The bottleneck has moved from technology to context. Whoever knows the asset best is now the person whose embedded knowledge is most valuable to wrap a deployment around. The plant. The port. The grid. The well. The warehouse. The pipe. The robot can be commodity. The simulator can be open source. The foundation model can be rented. The thing that cannot be rented is fifteen years of knowing which valve fails first in summer.

This is also why the GCC has a structural advantage that almost nobody outside the region has priced.

Why Qatar Owns a Piece of This

The countries with the most physical infrastructure being built right now are also the countries with the youngest digital substrate underneath that infrastructure.

Qatar is in the middle of an eighty-five percent expansion of its LNG export capacity [Fact, QatarEnergy public filings]. Saudi Arabia is building NEOM. The UAE is laying out Etihad Rail, expanding DP World’s port network, and standing up two of the largest AI data centre footprints outside the United States. India is building airports. Indonesia is building a capital. Nigeria is building rail. None of these projects are running on twenty-year-old SAP installations. None of them have the entrenched, legacy enterprise software layer that paralyses physical AI deployment in the United States and Europe.

The result is that the region’s industrial assets are simultaneously some of the most modern in the world and some of the most software-poor. The LNG trains. The smelters. The desalination plants. The petrochemical complexes. The airports. The rail networks. The new construction sites. The control rooms are full. The data is in spreadsheets. The maintenance schedules are in people’s heads. The simulators do not exist. The digital twins do not exist. The training data for the next generation of physical AI is sitting in proprietary sensor networks that have never been wrapped in a product.

This is the structural advantage. Every plant, every grid, every port in the region is a private dataset. A foundation model trained on that data, fine-tuned for that asset, deployed on that asset, is not commoditisable in the way an LLM is. It is bound to a piece of infrastructure that has a sovereign owner. The data does not leak. The deployment does not generalise to anyone else’s site. The moat is a physical thing.

This is what we mean when we say that Utopia Studio exists for Qatar-first deployment and global commercial scalability at the same time. The Qatar-first deployment is what gives a company its proprietary data loop. The global commercial scalability is what makes that loop a category. Both, always. The fellow who runs a refinery in Mesaieed and builds a live twin of it is sitting on a defensible position in a global market that nobody else in the world can replicate.

This is why the next ten companies will not come from Silicon Valley. They will come from people who can hold a heavy industrial asset in their head, who have a small software team that can build the loop, and who have access to capital that can write a multi-year ticket against a sovereign customer. The studio is in the business of putting those three things together.

What follows is the map of the categories we think are now buildable. Six of them. Each one corresponds to a layer of the new physical AI stack. Each one is a venture shape that did not exist three years ago. Each one is a call.

The Six Spaces We Want To Build

These are the categories the Infrastructure Intelligence pod, with our partner pods in Sovereign Systems and National Resilience, are actively mapping. Each one wants a particular founder profile. Each one is a piece of the stack that is now, for the first time, a viable venture rather than a research project.

Apply to the Co-Build Fellowship

1. The Plant’s Live Twin.

Every heavy industrial plant in the world runs on static physics. A consultancy ran a CFD model of the boiler in 2019. A different consultancy ran the corrosion analysis in 2022. The maintenance schedule was set in a board meeting based on a slide that referenced both reports. None of this is alive. None of these updates. The plant changes. The model does not. The first company that builds a vertical-specific, physics-grounded, continuously-updating digital twin of a single asset class owns the operational layer for that industry. LNG trains. Ammonia plants. District cooling networks. Water treatment. Refineries. Not horizontal. Vertical. Deep.

The founder profile: fifteen-plus years on a specific class of plant, ideally as a reliability engineer, operations director, or chief metallurgist. Paired with an ML engineer who has shipped production models on physical systems. If you have lived inside a control room and watched four different consultancy reports contradict each other on the same boiler, the studio wants to talk to you.

2. The Body in the Hazardous Place.

There are categories of physical work where the cost of human presence is enormous, the dexterity required is real, and the deployment environment is too messy for traditional industrial robotics. Oil and gas inspection. Decommissioning. Offshore wind maintenance. Mining drift inspection. Nuclear handling. Defence applications. The combination of a sub-twenty-thousand-dollar humanoid platform, a foundation model that can be fine-tuned for the specific task, and a simulator that can pre-train the policy on a digital replica of the site is, in 2026, a buildable product. It was not in 2023.

The founder profile: a robotics or controls systems engineer paired with a deep operator in the target vertical. Ideally, someone who has been called to a job site at 3 a.m. and decided which technician went into the hazardous zone. That person knows things no Silicon Valley team can guess at.

3. The Real-World Data Harvester.

The largest barrier to deploying a physical AI system into a piece of industrial infrastructure is not the model. It is the data. Most of the data that matters is locked inside legacy SCADA systems, vendor-proprietary formats, and undocumented historian databases that have been accumulating for 40 years. The company that builds the ingestion layer owns a flywheel position. The ingestion layer takes a plant’s entire historical sensor record, turns it into training data for a simulator, and feeds it back into the loop. It is the gatekeeper between the asset and the AI.

The founder profile: control systems engineer with deep experience in OT and IT integration, paired with a data infrastructure builder. A specific edge if either side has worked in a sovereign infrastructure context, where the data governance constraints are the moat. We are particularly interested in builders who have run SCADA migrations and lived the gap between the spec and the reality.

4. The Sovereign Training Loop.

A national infrastructure operator sits on a proprietary dataset that, properly wrapped, is a foundation model for an entire industry. A port authority. A national oil company. A sovereign rail operator. A national grid. The company that builds the wrapper does not own the data. It builds the technical, legal, and commercial architecture that lets a sovereign operator turn their physical infrastructure into a sovereign AI asset. This is a long-tail bet on Qatar, the UAE, Saudi, Singapore, and a small number of comparable jurisdictions. The product is the loop, but the customer is a country.

The founder profile: ex-operator at a national infrastructure entity, sovereign wealth fund, or strategic government office, paired with a product builder who has worked at the edge of regulated industries. This is the most political of the six categories, and the highest value. It requires someone who can maintain relationships with a minister and a CTO in the same conversation without losing either.

5. The Cross-Body Control Stack.

In any given industrial site today, you will find a mix of legacy fixed industrial robots, AGVs, drones, inspection robots, and increasingly humanoids. Each with its own software stack, its own integration cost, its own training pipeline. The integration cost of running all of these on a single control layer is what kills most multi-robot deployments. The company that builds the unified control stack owns the operating system for an industrial site. It runs policies trained on one body and deploys them on another. It hands off tasks between fixed and mobile platforms. It abstracts the hardware diversity into a coherent runtime.

The founder profile: robotics systems engineer with experience across multiple commercial platforms, paired with a supply chain operator who has lived the integration pain. Or an ex-ROS or NVIDIA Isaac engineer who has watched the same problems recur in every customer engagement.

6. The Capital Wrapper for Robot Capex.

When a robot becomes capable of doing the work of a human in a hazardous setting, or when an autonomous twin becomes capable of retiring a million dollars a year in consultancy spend, the unit economics start to make sense. But the financing layer is not built. Insurance does not price the new risk. Banks do not underwrite the new asset class. Project finance does not have a template for a five-year capex on a fleet of foundation-model-driven inspection robots. The first company that builds the actuarial science, the underwriting playbook, and the financing instruments for embodied AI deployments owns the wedge between the technology and the capital that pays for it.

The founder profile: ex-infrastructure project finance, paired with an actuary and a machine learning engineer. Specific edge for someone who has done capex underwriting at a major insurer, a multilateral, or a sovereign fund.

A Call For Curiosity

The thing that excites me most about the physical AI stack right now is that, for the first time in twenty years, the most interesting technology in the world is, again, the one that touches things you can hold.

The 1990s were a hardware decade. So was the 2000s. We built networks, phones, chips, and sensors. Then, for fifteen years, almost all the venture money in the world went into software. SaaS, marketplaces, infrastructure for other software, social, the cloud. Hardware became a thankless category. Robotics became a YouTube category. The most ambitious technical talent in the world worked on better ways to share photos.

That era is over.

The robot in your hand at OpenAI in 2019 took thirteen thousand simulated years to learn to do nothing useful. The robot on the BMW production line in 2026 was trained in a weekend on a closed loop that did not exist three years ago. The simulator runs on a single GPU. The foundation model fits on a workstation. The hardware costs less than a used car. And the most defensible position in this entire stack is the one held by the person who knows the specific piece of physical infrastructure best.

Most breakthroughs look obvious in hindsight and invisible while they are happening. This one is in the second category. The papers are public. The simulators are open source. The hardware is shipping. The valuations are pricing the loop. And yet most of the operators who have the embedded knowledge to win in this category have not been told that the substrate underneath their work has just changed. The reliability engineers. The chief metallurgists. The LNG schedulers. The port operators. The grid dispatchers. The maintenance directors. The inspection chiefs.

If you are one of them, the studio exists to remove the obstacles. Co-build OS. Technical co-founders. Sovereign capital. Pilots in the region. A structured path from problem to product to company in under six months. The map above is intentionally incomplete. It is expected to be redrawn by the people who walk onto it.

The world of interesting technology is back. The render farm is open. The robots are listening.

oliver@utopia-studio.co

References & Source Notes

OpenAI Dactyl: “Solving Rubik’s Cube with a Robot Hand,” OpenAI, October 2019. The roughly 13,000 simulated years figure is from the original technical paper.

Sergey Levine et al., “Sim-to-Real Transfer of Robotic Control with Dynamics Randomization,” 2017. The seminal domain randomisation paper.

NVIDIA Newton: announced jointly with Google DeepMind and Disney Research at SIGGRAPH 2024; placed under Linux Foundation governance 2025. Differentiable, GPU-native, open source.

Genesis simulator: multi-institution academic release, December 2024. Throughput claims of roughly 43M FPS on RTX 4090.

DeepMind Genie 3: announced August 2025, playable world model with minutes-long persistence.

NVIDIA Cosmos: foundation world model for physical AI, announced CES January 2025.

Physical Intelligence π0 (Oct 2024) and π0.5 (April 2025): cross-embodiment foundation models.

Skild AI: announced general-purpose robotics brain, valuation reported at roughly $4.5B mid-2025.

Figure 02 deployments at BMW Spartanburg, public statements from Figure and BMW, 2024 and 2025.

Unitree G1 humanoid: launched 2024, list price roughly $16,000.

Pixar Toy Story render statistics: drawn from Pixar engineering retrospectives, multiple sources. Average frame render times of six to twenty-nine hours on the original render farm.

QatarEnergy North Field Expansion: 77 mtpa to roughly 126 mtpa, public regulatory filings.

There is a category of machines that did not exist in commercial form three years ago. It does not predict the next word. It predicts the next moment of a physical system — how a fluid will flow through a pipe, how a material will crystallise under a given pressure, how a gravitational field will vary across a landmass, how a battery will age in heat, and how a stress wave will propagate through a weld. It learns the grammar of force.

Subscribe now

I have been watching the edge of this shift for some time, across a career that has moved between venture, product, and heavy industry. What has kept pulling me back is a single observation. The intelligence layer that governs physical assets — refineries, grids, data centers, utilities, ports, buildings, and fabs — has been built almost entirely on the wrong substrate. Text. Documents. Structured records. Spreadsheets. The machines were reading the symbolic description of the asset. They were not reading the asset.

That is the constraint that is now breaking.

The point of this essay is not a career retrospective. It is a map. Five fields where the physics is now learnable. Named breakthroughs, exact numbers, open gaps. And, at the end, a call to the engineers, physicists, operators, and software builders who sit at the seam between physical reality and a new generation of models. The companies that own this category will not be built by a single discipline. They will be built by small teams that can speak all three.

If you are one of those people, this is a letter written to you.

Every decade or so the bottom layer of what machines can do shifts, and a whole category of work bends around it. The current shift has a name most people have not learned yet.

A large language model, stripped to its mechanics, is a function that predicts the next token given the previous ones. It works because natural language has statistical structure — words depend on words, phrases on phrases — and scale across a big enough corpus yields a machine that can generate plausible continuations. It was a narrow insight that turned into a civilisational one the moment the substrate got big enough.

The physical world has its own statistical structure. A fluid flowing through a pipe obeys Navier-Stokes. A field propagating through space obeys Maxwell. A material deforming under load obeys constitutive equations calibrated by a century of mechanical testing. A battery ageing in heat obeys electrochemistry governed by Arrhenius kinetics and diffusion equations. None of that structure is in the written corpus in any recoverable form. It is in simulations, sensor streams, experiments, telemetry. The corpus that describes the physical world is not a library. It is a million years of compute time burned against the partial differential equations that govern how the universe actually behaves.

A physics language model — I'll use the term loosely to cover physics-informed neural networks, neural operators, and the foundation models that have started appearing across weather, materials, and fluid dynamics — is a machine trained on that corpus. It learns the grammar of force. And because force is universal in a way language is not, the implications ripple into every industry whose economics turn on what happens when materials meet the world.

This is the break. Not better chatbots. Not smarter search. Machines that know what the pipe is actually doing.

What follows is how far that break has already travelled.

Beyond Optics: Seeing Through The Earth

The sensor stack of the twentieth century was, overwhelmingly, optical. Cameras, satellites, LiDAR, spectrometers, fibre. It worked because light is easy to manipulate, cheap to deploy, and reaches a lot of places. It also has a fundamental limit. Light does not pass through rock, water, concrete, or steel at useful resolution. Most of what matters in an industrial asset is behind something opaque.

The physics that travels through opacity is gravity and the cosmic muon flux, and both are finally becoming tractable as sensing substrates.

In February 2022 a team at the University of Birmingham's UK National Quantum Technology Hub, led by Kai Bongs and Michael Holynski, published in Nature the first demonstration of a quantum gravity gradiometer operating outside laboratory conditions [1]. They detected a two-metre tunnel half a metre below the surface from ground level, with a signal-to-noise ratio of eight and a sensitivity of around twenty eötvös. They suppressed micro-seismic noise, laser phase noise, thermal drift, and magnetic field variation — all of which had kept the technology in the lab for thirty years. Their result cut gravity survey time from a month to a few days, a tenfold speed-up against a category of work that has barely moved since the 1960s.

NASA JPL is now building the first quantum gravity gradiometer designed to operate from space, backed by the Earth Science Technology Office [2]. If it flies, the same physics that let Birmingham find a tunnel will let us measure groundwater, mineral deposits, hydrocarbon reservoirs, and carbon storage sites at continental scale without drilling anything.

The second opacity-piercing technology is muon tomography. Cosmic rays generate a continuous flux of high-energy muons that travel through kilometres of rock. Differential absorption tells you about density. The Khufu pyramid void was famously mapped this way in 2023 [3]. The commercial application is mining, not archaeology. Ideon Technologies activated the world's first in-mine muon tomography system at Evolution Mining's Northparkes site in Australia in December 2024, and deployed at Rio Tinto's Bingham Canyon — the largest surface mine in the United States — in September 2024, for live four-dimensional monitoring of ore bodies [4]. The value unlock at Rio Tinto scale is material. The technology is not a lab curiosity. It is a platform for the category of physical intelligence I'll call geodetic sensing — the ability to see forces and densities, not surfaces.

Combine quantum gravimetry with muon tomography with physics-informed inversion, and the entire subsurface — groundwater, ore bodies, pipes, tunnels, faults, CO₂ plumes in storage formations — becomes legible at a resolution and speed that the industries built around ignorance of those things have not started to price.

That last point is the one that will move capital. The mining industry, the oil and gas industry, the water utility, the civil engineering firm, the carbon storage developer — all of them operate against a subsurface they cannot see. Their risk models are built on that blindness. When the blindness lifts, the pricing of every project they underwrite moves. Companies that own the translation layer between the new sensors and the existing capital stack will be the MSCI of a sensing category that did not exist in commercial form before 2024.

We are exploring the insurance-and-capital-markets wrapper around this in our Infrastructure Intelligence pod. Subsurface sensing is a sensor-hardware business at the point of origin. It becomes a category-defining software business at the point of translation into underwriting, reserve certification, and development finance. The hardware is coming from Birmingham, NASA, and Ideon. The translation layer is unbuilt.

The Grammar Of Materials

Chemistry has, since its founding, been a search problem. The periodic table is small. The combinatorial space of how those elements can arrange themselves into ordered solids is astronomical. For most of the twentieth century, finding a new useful material — a battery cathode, a catalyst, a superconductor, a coating — took a decade of graduate students and kilogrammes of precursor.

That clock is breaking.

In November 2023, Amil Merchant and colleagues at Google DeepMind published in Nature results from GNoME, a graph neural network trained on materials data that predicted 2.2 million stable crystal structures — the equivalent of about 800 years of prior experimental knowledge — of which 381,000 were novel materials below the convex hull of known stability [5]. Of those, 736 have now been experimentally synthesised and verified by independent laboratories. Among them are fifty-two thousand layered compounds with the structural fingerprints of superconductors.

A month later, the Lawrence Berkeley A-Lab, directed by Gerbrand Ceder, published in the same issue of Nature a robotic laboratory that synthesised forty-one novel materials from fifty-eight GNoME-predicted targets in seventeen days [6]. Seventy-one percent success rate. Twenty-one experiments a day. This is the industrial equivalent of the moment when DNA synthesis stopped being a two-year PhD project and became a service you order online.

Microsoft Research followed with MatterGen in 2025, a diffusion-based generative model that produces novel crystalline structures conditioned on target properties — chemistry, symmetry, mechanical response, magnetism, electronic behaviour — published in Nature [7]. Orbital Materials, founded by Jonathan Godwin out of DeepMind, launched its Orb foundation model for advanced materials in late 2024 and signed a multi-year strategic agreement with AWS focused on data-centre cooling and CO₂ capture [8]. Separately, a National University of Singapore group announced in 2025 the first copper-free high-temperature superconductor operating above 30 K at ambient pressure — a category-breaking result that ended a thirty-plus-year presumption about where superconductivity could exist [9].

This is the second grammar in its most visceral form. The model is not predicting the next word. It is predicting the next crystal. And the crystal is a battery that holds more energy, a catalyst that drops the cost of hydrogen, a membrane that makes desalination thirty percent cheaper, a thermal interface material that lets a GPU run ten degrees cooler.

The venture opportunity is not in building another foundation model. The foundation models exist. The opportunity is in the last mile — taking a GNoME-predicted material from Nature supplementary figure to commercial product. The A-Lab is doing it inside a national lab. Orbital is doing it inside AWS. Neither is a company structure that scales across the industries that will consume these materials. That is a gap we are mapping specifically for desalination, district cooling, battery storage, and carbon capture, where the incumbents run on materials chemistry set in the 1980s and the uplift from a single generation of new compounds is billions per sector.

Chips That Know Themselves

The chip industry has been on a single vector — more transistors per square millimetre — for sixty years. The next five years it will fragment. Security is one of the fragments, and it is the one where physics is entering the substrate most directly.

Three lines are converging.

The first is the physical unclonable function. A PUF exploits the atomic-scale manufacturing variation that makes every silicon die slightly different and turns that variation into a cryptographic key that cannot be cloned because it was never stored. Silicon photonic PUFs demonstrated in Nature Communications in 2024 achieved a cloning probability below 10⁻³⁰ with zero false-positive authentication rate across their test runs [10]. For comparison, no conventional secret-storage scheme can claim cloning probabilities below the probability that the silicon vendor itself is compromised.

The second is photonic chaos. A patterned silicon photonic chip that emits chaotic light waves, covered by SciTechDaily in 2024, produces cryptographic keys from a process that is deterministic to the chip owner and statistically indistinguishable from noise to any attacker [11]. This is not quantum key distribution with its fibre-and-detector infrastructure. It is a chip that, installed in a device, issues one-time pads at line rate.

The third is post-quantum cryptography hardware. NIST finalised ML-KEM, ML-DSA, and SLH-DSA in August 2024 [12]. The lattice mathematics those standards rely on is computationally heavy; software implementations are slow. Dedicated silicon acceleration for these algorithms is a hardware category that did not exist three years ago and will be in every network card, every HSM, every TPM within five. Separately, IonQ has published a roadmap targeting around twenty thousand physical qubits by 2028, and PsiQuantum was selected into the final phase of DARPA's Underexplored Systems for Utility-Scale Quantum Computing programme in February 2025 on a million-qubit photonic architecture targeted for 2033 [13].

Put those three together. Every authenticating device in the world — every phone, every car, every industrial sensor, every card reader — will, by 2030, carry a chip whose security posture is grounded in a physical process rather than a stored secret. And the chips that underpin the banking and defence stacks will need to resist an adversary with several thousand logical qubits at their disposal.

This is a market that did not have a sales call five years ago. Today it has a procurement line in every Tier 1 bank, every defence ministry, and every hyperscaler. The GCC is disproportionate in this category because sovereign systems are a core pod for us and because the region's cybersecurity budgets have grown double digits year on year since 2022. A chip company that owns the physics-grounded security layer for regional critical infrastructure is one of the clearest category bets we are watching.

Forces As Tokens

The most under-discussed development in AI over the last eighteen months is that the simulation stack has moved.

For forty years, computational physics — the workhorse of every industrial sector from aerospace to oil and gas to civil engineering to weather forecasting — ran on finite-element and finite-volume solvers on clusters of CPUs. A single high-fidelity simulation of a turbine blade, a data-centre cooling volume, or a regional weather pattern took hours to days. Pricing any decision that depended on that simulation was bottlenecked by the simulation itself.

In 2023, Google DeepMind published GraphCast in Science [14]. A graph neural network, trained on forty years of ECMWF reanalysis data, produced ten-day global weather forecasts at 0.25-degree resolution in under a minute, outperforming the ECMWF High-Resolution Forecast on over ninety-nine percent of tropospheric variables. In 2024 Microsoft Research released Aurora [15], an atmospheric foundation model achieving approximately five-thousand-times speed-up against the Integrated Forecasting System at equivalent or better accuracy, with variants now shipping for high-resolution weather, air pollution, and ocean waves.

Weather is the canary because weather modellers are the most rigorous, the most adversarial, and the most data-rich community in applied physics. If the foundation-model approach works against ECMWF — the gold standard — it works against every less-rigorous simulation domain in the world. And that is what is now happening. NVIDIA's PhysicsNeMo framework, formerly Modulus, hit production v2.0 in March 2026, with enterprise support shipping inside NVIDIA AI Enterprise [16]. It is the operational substrate for physics-informed neural networks across thermal, structural, electromagnetic, and fluid domains — the same domains that sit under every heavy industry asset class we track.

Combine foundation-model inference speed with physics-informed constraints, and what was a two-day simulation becomes a two-second query. The implications are severe. A corrosion model that used to be run quarterly by a consulting firm becomes a continuously updating field. A flood assessment that used to be a twelve-week engineering study becomes a real-time overlay. A seismic inversion — Shell and SparkCognition reported a ninety-nine-percent reduction in the number of physical seismic shots required per survey through generative-AI techniques — stops being a multi-million-dollar campaign and becomes a weekly refresh [17].

The digital twin, after a decade of disappointment, finally earns the name. The twin matches the plant because the twin runs at the speed of the plant.

That reshapes every vertical. The refining base. The LNG train. The grid dispatch desk. The water utility. The municipal flood manager. The building energy optimiser. Each of them has been running on static physics and quarterly consultancy refreshes. Each of them is about to have a live model instead. The gap between who has that live model in production and who does not will be the primary competitive axis in industrial operations by the end of the decade.

Why This Round Looks Different

A pattern has repeated across every decade of attempts to close the gap between the drawing and the asset. Two rooms, each of which cannot hear the other. In one room, financial models and consultancy reports assume the asset behaves the way its specification says it does. In the other, the operator knows what the asset is actually doing and long ago gave up trying to explain it, because the last three software vendors did not listen.

Every previous attempt to close that gap failed on the same reef. GE Predix consumed around seven billion dollars — four billion in direct spend over six years — before being written down, because it was built top-down by a conglomerate that had sold turbines for a century and thought software was another SKU [18]. Siemens MindSphere pivoted three times. Honeywell Forge survived by narrowing scope to vertical deployments on Microsoft Cloud. The pattern was the same each time. Outside-in product teams building horizontal platforms and asking operators to come to the tool.

The physics LM inverts that failure mode.

The reason is mechanical. A large language model can read a plant's documentation, SCADA logs, turnaround history, and handover notes the way a new hire reads their predecessor's binder, absorbing the institutional knowledge in hours. A physics foundation model can take that context and run the relevant multi-physics simulation at inference speed. The tacit knowledge that used to live in a maintenance director's head can now leverage the same scale of compute that previously required a team of twenty-five PhDs and an SAP contract.

The operator does not need to learn Python. They need to encode their judgment into the model. Which, with the current generation of tooling, is a several-week exercise, not a several-year one.

That is why the interesting founder profile for the next decade is the fifteen-to-twenty-year operator paired with a software builder who has done production ML. Not because the operator understands AI. Because they carry the physics of their asset in a way no outside researcher can reconstruct from public data, and the researcher's tools have finally caught up to what the operator always knew.

1. ERP era (1995–2015) — Intelligence substrate: Relational records. Founder profile: Software vendor. Canonical outcome: GE Predix, Target Canada, Lidl SAP.

2. First-wave IIoT (2010–2020) — Intelligence substrate: Sensor telemetry. Founder profile: Conglomerate spin-out. Canonical outcome: MindSphere pivots, Predix write-down.

3. AI-native infrastructure (2022–) — Intelligence substrate: Physics + language models. Founder profile: Operator-founder. Canonical outcome: Azraq, Barrier Intelligence, the next ten.

The Categories That Are Now Buildable

If the physics is learnable, a class of company becomes obvious. Each of the five fields above maps to a venture shape that did not exist two years ago. These are the categories the studio is actively mapping, and the ones we think most need builders from the intersection.

Subsurface sensing, productised. Quantum gravimetry and muon tomography are hardware businesses at the point of origin. They become category-defining software businesses at the point of translation into carbon storage certification, mineral reserve verification, groundwater accounting, and civil-engineering-grade subsurface mapping. The hardware is coming from Birmingham, NASA JPL, and Ideon. The translation layer — the MSCI of the sensing category — is unbuilt.

Materials, as a service. A GNoME-predicted material is useless until someone takes it from supplementary figure to commercial product. The A-Lab is doing that inside a national lab. Orbital is doing it inside AWS. Neither is a company structure that scales across the verticals — desalination membranes, battery cathodes, thermal interface materials, catalysts for low-carbon hydrogen — that will consume these compounds. Each vertical is a billion-dollar uplift sitting on a chemistry set in the 1980s.

Physics-grounded chip security. PUFs, photonic chaos sources, and post-quantum accelerators are a hardware category that did not have a sales call five years ago and will be in every network card within five more. Regional demand is disproportionate because sovereign cybersecurity budgets have compounded since 2022. The domestic product does not yet exist.

Live twins, not reports. The corrosion model that is a PDF. The cooling loop without a twin. The hydraulic network that gets a consultancy refresh every quarter. Each of these becomes a live, physics-grounded model the moment someone does the integration work. The horizontal platform era failed. The vertical, operator-led, physics-LM-native era is beginning.

The underwriting layer. Every live physics model is a new input to the capital stack. Insurance, project finance, development credit, and reserve certification are priced on assumptions that were set before any of this sensing or simulation existed. Companies that translate physics into underwriting language will set the price of the next generation of infrastructure.

This is the map the studio is building against. Azraq and Barrier Intelligence are the first two fellows in production. The next ten are the point of this essay.

A Call For Curiosity

Most breakthroughs in a technology category look obvious in hindsight and invisible while they are happening. The physics LM is one of the invisible ones. The papers are in Nature and Science. The foundation models are shipping in NVIDIA, Microsoft, and Google stacks. The hardware is deploying at Rio Tinto and through NASA. And almost none of the people who will use it day-to-day — the reliability engineer on a refinery, the chief metallurgist at a desalination plant, the civil engineer mapping a carbon storage reservoir, the chip architect securing a sovereign network — have been told that the substrate underneath their work is changing.

The people who will build the category-defining companies of the next decade will be curious across an unusual number of domains. They will read a graph neural network paper and immediately think about a corrosion inspection dataset they have seen. They will read an earnings call from a mining major and think about how a muon detector array would reprice the reserve. They will read the NIST post-quantum specification and think about the TPM in a sovereign identity card. They will refuse to accept that their discipline is the only one that matters.

If you are one of those people, there are a small number of questions worth asking yourself.

The studio exists to remove the obstacles. The co-build model is designed to get a boiler-suit founder to a contracted pilot in under six months. The capital is there. The pilots are in the region. The map is incomplete on purpose — it expects to be redrawn by the people who walk onto it.

The companies that will own the next decade will be built, at the intersection, by the people who are curious enough to stand in all three rooms at once.

If you are one of them, write to us.

References

[1] B. Stray, A. Lamb, A. Kaushik, J. Vovrosh, A. Rodgers, J. Winch, F. Hayati, D. Boddice, A. Stabrawa, A. Niggebaum, M. Langlois, Y.-H. Lien, S. Lellouch, S. Roshanmanesh, K. Ridley, G. de Villiers, G. Brown, T. Cross, G. Tuckwell, A. Faramarzi, N. Metje, K. Bongs, M. Holynski, "Quantum sensing for gravity cartography," Nature 602, 590–594 (2022). https://www.nature.com/articles/s41586-021-04315-3

[2] NASA JPL press materials on the Quantum Gravity Gradiometer Pathfinder, Earth Science Technology Office, 2025. SciTechDaily coverage: NASA's Quantum Sensor Could Revolutionize Gravity Mapping. https://scitechdaily.com/nasas-quantum-sensor-could-revolutionize-gravity-mapping/

[3] S. Procureur et al., "Precise characterization of a corridor-shaped structure in Khufu's Pyramid by observation of cosmic-ray muons," Nature Communications 14, 1144 (2023).

[4] Ideon Technologies announcements: Evolution Mining Northparkes (December 2024), Rio Tinto Bingham Canyon (September 2024). https://ideon.ai

[5] A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, E. D. Cubuk, "Scaling deep learning for materials discovery," Nature 624, 80–85 (2023). https://www.nature.com/articles/s41586-023-06735-9

[6] N. J. Szymanski, B. Rendy, Y. Fei, R. E. Kumar, T. He, D. Milsted, M. J. McDermott, M. Gallant, E. D. Cubuk, A. Merchant, H. Kim, A. Jain, C. J. Bartel, K. Persson, Y. Zeng, G. Ceder, "An autonomous laboratory for the accelerated synthesis of novel materials," Nature 624, 86–91 (2023). https://www.nature.com/articles/s41586-023-06734-w

[7] C. Zeni et al., "A generative model for inorganic materials design" (MatterGen), Nature (2025). https://www.nature.com/articles/s41586-025-08628-5

[8] Orbital Materials Orb launch (Sep 2024); AWS strategic partnership announcement (Dec 2024). Globe Newswire press release.

[9] NUS copper-free superconductor above 30 K at ambient pressure, covered by SciTechDaily: 40-Year Barrier Broken: Scientists Discover New High-Temperature Superconductor. https://scitechdaily.com/40-year-barrier-broken-scientists-discover-new-high-temperature-superconductor/

[10] Silicon photonic PUF with 10⁻³⁰ cloning probability, Nature Communications (2024). https://www.nature.com/articles/s41467-024-47479-y

[11] SciTechDaily, Patterned Optical Chips That Emit Chaotic Light Waves Keep Secrets Perfectly Safe. https://scitechdaily.com/patterned-optical-chips-that-emit-chaotic-light-waves-keep-secrets-perfectly-safe/

[12] NIST Post-Quantum Cryptography Standardization, FIPS 203/204/205 finalised August 2024.

[13] IonQ publicly published roadmap targeting ~20,000 physical qubits by 2028. PsiQuantum selected into Phase C of DARPA's Underexplored Systems for Utility-Scale Quantum Computing (US2QC) programme, February 2025.

[14] R. Lam et al., "Learning skillful medium-range global weather forecasting," Science 382, 1416–1421 (2023). https://www.science.org/doi/10.1126/science.adi2336

[15] C. Bodnar et al., "Aurora: A foundation model of the atmosphere," Microsoft Research blog and arXiv preprint, 2024; GitHub open source release 2025. https://www.microsoft.com/en-us/research/blog/introducing-aurora-the-first-large-scale-foundation-model-of-the-atmosphere/

[16] NVIDIA PhysicsNeMo v2.0 release notes, March 2026. https://developer.nvidia.com/physicsnemo

[17] SparkCognition–Shell announcement of deep-sea exploration acceleration via generative AI, reporting 99% reduction in required seismic shots. Offshore Magazine (2023).

[18] Platform Engineering analysis, How General Electric Burned $7 Billion on Their Platform. https://platformengineering.org/blog/how-general-electric-burned-7-billion-on-their-platform

oliver@utopia-studio.co

Somewhere behind him, a camera is recording. Somewhere in the building, a trainee is watching a live feed. Somewhere in the hospital’s archives, a case note will be typed, filed, and, within six months, effectively lost — indexed by date, not by what happened.

Subscribe now

Here is what is not being recorded. Not the video. The video is fine. What isn’t being recorded is the reasoning underneath the video. Why Dr. Haji lifted the lesion at this corner and not that one. Why he paused for seven seconds before the next cut. Why, when the submucosal plane dropped away in the middle of the procedure, he switched his injection solution — and why that switch will be invisible on the video unless you already know what to look for. Why he doesn’t cut here. Why he cut there. Why, when the anaesthetist says something about CO₂ absorption, he steps his pace down a fraction without announcing it.

That reasoning is the operation. The cuts are the easy part. The cuts are what a good general surgeon can learn in a year. The reasoning is what takes twenty years to build, and it is almost entirely unwritten.

Dr. Haji trained under Professor Shin-ei Kudo at Showa University in Yokohama, where pit-pattern classification — the visual taxonomy that lets an endoscopist read a lesion’s malignancy from its surface pattern before cutting — was invented in the early 1990s. He trained under Professor Haruhiro Inoue, who in 2008 pioneered peroral endoscopic myotomy, POEM, for the treatment of achalasia: a procedure that, until a single surgeon in a single hospital in Yokohama decided otherwise, required a full chest incision. Kudo’s pit-pattern classification took two decades to cross the Pacific. Inoue’s POEM took a decade to become standard of care. The knowledge exists. The transmission mechanism is the slowest part of modern medicine.

Dr. Haji wrote more than a hundred papers. He runs King’s Live Endoscopy, an annual event where surgeons from around the world watch him operate via live broadcast. He is directly responsible for training a generation of British and European colorectal endoscopists in techniques that Western centres were telling their fellows, as recently as 2015, could not be safely adopted outside Japan.

He’s fifty-one. Even if he operates until he’s seventy, the arithmetic is unforgiving. A lifetime of surgical knowledge dies with the surgeon unless something structural happens between now and then to preserve it.

He’s watched it happen before. Every master does.

The Architecture of the Apprenticeship

There is a useful exercise if you work in medical education or surgical robotics and you want to understand the shape of the problem. Sit in on a teaching ward round. Count the number of times a senior surgeon says something like, “You can see here that…” or “The thing to notice is…” or “In my experience when you get this appearance, you want to…” Now count how many of those statements get written down anywhere a future trainee could find them. The answer is approximately zero.

This is not a failure of diligence. This is the architecture of the apprenticeship.

Surgical training in the West still runs on a model designed in 1889. William Halsted, at Johns Hopkins, formalised a system in which young surgeons learn by watching, then assisting, then doing under supervision, then finally operating unsupervised — passing, as Halsted put it, from observer to assistant to surgeon-in-charge. The method works. It produces excellent surgeons. It also scales linearly with the number of senior surgeons willing to teach and the number of junior surgeons in their operating room, which is to say it does not scale at all.

The cost of that architecture is legible in the learning curves. A surgeon acquiring proficiency in laparoscopic colorectal surgery requires somewhere between 50 and 152 cases before complication rates drop to expert baseline, depending on the definition and the case mix [Fact]. Anastomotic leak rates in the first 200 cases for a new laparoscopic colorectal surgeon run 6–9 percent, falling below 2 percent only after sustained case volume. Endoscopic submucosal dissection — Dr. Haji’s specialty — requires approximately 80 procedures before an endoscopist can reliably achieve the dissection speed, en bloc resection rate, and perforation avoidance that define expert performance [Fact]. Forty procedures just to stop punching holes through the colon wall.

A surgeon who trains one fellow a year transmits her expertise to perhaps twenty people in a career. At the population level that is a rounding error. The NHS diagnoses a colorectal cancer patient every fifteen minutes. Fewer than two hundred endoscopists in the UK are trained in complex ESD. The supply of expertise is not keeping up with demand, and the pipeline cannot be accelerated without compromising the patient on whom the trainee is learning.

Meanwhile, the data that would compress the learning curve sits in three places. On hospital video servers, unsearchable. In case notes filed by date and patient identifier, not by clinical feature. And in the head of the surgeon who did the case, a few terabytes of pattern recognition with no export format.

Miss the capture window and it’s gone. The institution writes a card. Everyone claps. A successor picks up the baton and starts the twenty-year climb from somewhere near the bottom.

Where AI Currently Stops

There is a specific type of surgical knowledge the field has spent a decade trying to automate, and it is worth understanding the ceiling the automation has hit, because it tells you where the actual opportunity is.

Look at polyp detection.

A colonoscopy performed by an average endoscopist misses somewhere between 20 and 30 percent of adenomas — precancerous polyps. That miss rate is the source of a meaningful fraction of post-colonoscopy colorectal cancers, the ones that appear in the three to five years between a “clear” screening and a diagnosis. For a decade, the field has known that computer-aided detection systems — CADe — can improve adenoma detection rate by around 20 percent in head-to-head trials, with a 55 percent decrease in adenoma miss rate when pooled across dozens of randomised studies [Fact]. The pooled adenoma detection rate in the CADe group of a major 44-trial meta-analysis was 44.7 percent, versus 36.7 percent in the standard colonoscopy group. These are not marginal gains. These are meaningful numbers of cancers prevented.

So why is CADe not yet standard of care in every hospital in the world?

Part of the answer is procurement cycles and variable reimbursement. The underlying technical ceiling is that detection is not the hard part of endoscopy. Detection is the foothill. The mountain is what comes next — characterisation of the lesion, decision to resect, choice of technique, execution, management of complication. That entire sequence is where the cost, the risk, and the patient outcome live. CADe answers “is there something there.” It does not answer “what is it, should I take it out, and how.”

Systems that pattern-match against Kudo’s pit-pattern classification can get partway to characterisation. Systems that track instrument kinematics — the path a needle takes through tissue — can score dexterity. Intuitive Surgical, the company that manufactures the da Vinci robotic system, has accumulated an extraordinary video corpus. Their SurgVU dataset contains more than 840 hours of robotic surgery footage, sampled at 60 frames per second, labelled for tool use and surgical step — roughly 18 million annotated images [Fact]. Research groups at ETH Zurich, MIT, and NVIDIA are training video-language foundation models on progressively larger surgical corpora, using frameworks like VidLPRO and GenSurg+ to pair video with transcript-derived captions [Fact].

This is meaningful work. It is not the ceiling.

The ceiling is that almost all of this data is observational. It shows what happened. It does not show why. The master surgeon who stepped back for seven seconds before the next cut did so for a reason, and until the reason is captured alongside the video, the foundation model trained on it will learn the choreography without learning the judgment.

You end up with a system that can recognise an ESD but cannot perform one.

The kinematics are visible. The decision tree that produced the kinematics is not.

What is missing is the expert annotation layer. The running commentary. Here’s why I’m doing this, here’s what I’m looking for, here’s what would make me stop, here’s what this reminds me of. That annotation layer exists — in the heads of Dr. Haji, Kudo, Inoue, and a few hundred other world-class endoscopists globally. It has simply never been systematically captured.

The Expert Inside the Machine

Amyn Dr. Haji has, by his own count, somewhere north of four thousand hours of operative video sitting on hospital servers, going back fifteen years. He has case notes for every single one of those procedures. He remembers, with startling specificity, the difficult ones. Which tells you something: the mind of a master surgeon is not a filing cabinet. It is a retrieval system. Given the right cue — a lesion appearance, a patient history, a scope angle — he can surface the relevant priors from a fifteen-year career in roughly the time it takes to make a decision mid-procedure. Three to five seconds. That retrieval, more than the hands, is what makes him one of the best in the world.

Now imagine you could capture that retrieval system. Not the video alone. Not the notes alone. The reasoning that connects the two.

This is the insight behind what Mentix is building.

The initial concept, which is where the company is live today, is narrower than the eventual vision but already interesting on its own terms. You take a single surgeon — Dr. Haji, as the founder-operator pilot — and you build a system capable of ingesting the raw material of a career: operative videos, case notes, the published papers, the teaching lectures, the recorded broadcasts from King’s Live Endoscopy. You structure it. You annotate it, in partnership with the surgeon himself, layering the why onto the what. You build an interface that lets a trainee, mid-procedure, query the corpus in natural language.

“I’m looking at a 30mm LST-granular in the rectum with a type IV pit pattern near the anorectal junction. What would you do?”

And the system responds not with a textbook answer but with Dr. Haji’s answer — grounded in a specific case he did in 2019, or a variant he discussed in a paper in 2022, or a teaching moment he recorded in 2024.

That is a personalised AI. It is a specific surgeon, abstracted from the biological constraint that he can only be in one operating room at a time.

If you accept that such a thing is technically feasible — and a decade of progress in multimodal foundation models, expert retrieval systems, and surgical video analysis suggests it is — then the second-order implications become the interesting part.

A personalised AI of one expert surgeon is a tool. A network of personalised AIs, covering every major surgical specialty and every major procedural variant, is an infrastructure. It is the first time in the history of the field that expertise can be queried without the expert being physically present. It is the first time that a 28-year-old registrar in Doha can have, in her ear, the reasoning of a specific Japanese endoscopist she will almost certainly never meet. It is the first time that the Halsted apprenticeship model gets genuinely unbundled — not replaced, but extended, because the senior surgeon’s judgment can be carried into rooms the senior surgeon will never enter.

And that infrastructure, once it exists, is the data layer the next generation of surgical robotics has been missing.

The Robotics Parallel

Here is the robotics parallel, which matters because it shows where the real scale of this opportunity lives.

In the last three years, the frontier of robotics has moved from specialised, hand-engineered systems to generalist foundation models — large neural networks trained on enormous, heterogeneous corpora of physical interaction data. Google DeepMind’s RT-2. NVIDIA’s ORBIT-Surgical. Physical Intelligence’s π-series. These models learn motor skills the way a language model learns grammar — by consuming huge volumes of example data and extracting the underlying patterns. The key lesson from the last decade of AI is that data quality and diversity beat architecture [Fact]. The teams that win are the teams with the best corpora.

Surgical robotics has a corpus problem.

The existing datasets — the SurgVU data, the academic collections, the fragments of kinematic logs released from da Vinci installations — are large but narrow. They skew heavily toward training tasks on porcine models, toward early-career surgeons, toward procedures chosen for their pedagogical cleanliness rather than their clinical complexity. The corpus required to train a robot to perform a complex ESD on a live human colon does not currently exist in a structured, queryable form. It exists as raw video on hospital servers, untagged, unannotated, and bound by the data governance of fifty different health systems.

And the missing layer — the one that would transform that video from an observational record into training data — is exactly the layer Mentix is building. The annotations. The reasoning. The expert commentary. The why.

Think about it structurally. Intuitive’s SurgVU dataset is, at 840 hours, smaller than the body of operative video sitting on Dr. Haji’s King’s College servers for a single subspecialty. Multiply that by every subspecialty, every senior endoscopist and surgeon, every teaching hospital globally. The raw footage exists. What is missing is the connection between the raw footage and the reasoning of the operator. Once that connection is made — once you have a systematic way to extract expert judgment and pair it with the corresponding video — you have something the robotics field does not yet have: a structured dataset of surgical decision-making at expert grade.

This is the ImageNet moment for surgery.

The analogy is deliberate. In 2009, Fei-Fei Li and collaborators released ImageNet — fourteen million hand-labelled images, organised by a taxonomy of real-world objects. The dataset was not conceptually novel. What was novel was the scale, the labelling, and the commitment to making it open to researchers. Within three years, ImageNet had catalysed the deep learning revolution. Not because the neural network architectures were new — most of them dated to the 1980s — but because, for the first time, there was enough labelled data for the architectures to learn on. The corpus was the unlock.

Surgical AI is sitting at the equivalent point. The architectures exist. The compute exists. The raw video exists. What is missing is the expert-annotated, structured, reasoning-laden corpus. Whoever builds it becomes the substrate on which the next generation of surgical AI — training tools, decision support, and eventually autonomous surgical robotics — will be trained.

This is not a small prize. Intuitive Surgical alone is a $180 billion company, built on a single robotic platform that requires a human surgeon in the loop. The companies that build the successor platforms will be built on data they do not currently have. Mentix is positioning to be the supplier.

The Strongest Objection

Now engage the hardest version of the counter-argument, because a piece like this is only credible if it does.

Surgical knowledge cannot be extracted because it is fundamentally tacit. It lives in the hands. It lives in the pattern recognition that a surgeon cannot quite put into words. Attempts to codify it have been tried for decades and produced textbooks that no trainee reads.

This is the strongest objection. It is also incomplete, and it’s worth being precise about why.

The objection assumes the extraction problem is linguistic — that the only way to capture tacit knowledge is to get the expert to write it down or speak it aloud, and that this process loses the essential thing. It also assumes the extraction has to be complete. Neither assumption holds.

Two things have changed in the last five years that should update the prior.

The first is multimodal capture. A senior surgeon narrating mid-procedure, with video of the operation running alongside, kinematic data from the scope, the audio of the operating room, the patient’s vital signs, and the pathology result a week later — that is not a linguistic record. That is a dense, time-aligned dataset in which the reasoning is grounded in the physical reality that produced it. The commentary does not need to be complete. It needs to be paired with what is on the screen at the moment it is spoken. Modern foundation models are very good at learning from aligned multimodal data. They are much worse at learning from text alone.

The second is retrieval-augmented generation. The AI systems being deployed in medicine today do not need to memorise the corpus. They need to retrieve the right fragment at the right time. A trainee mid-procedure does not need a generic summary of colorectal ESD. She needs the three cases in Dr. Haji’s career that most resembled the lesion on her screen, the reasoning he applied to each, and the outcomes. That is a retrieval problem, and the tools to solve it at scale exist now.

Tacit knowledge resists extraction into text. It yields to extraction into structured, multimodal, retrievable datasets. The technology to build those datasets did not exist five years ago. It does now. That’s the update.

The historical parallel worth holding in mind is Go. For decades, the orthodox view among Go professionals was that top-level play required an intuition that could not be taught outside of a decade of direct apprenticeship under a master. The 2016 AlphaGo result did not disprove the existence of that intuition. It showed that the intuition could be induced in a system that had consumed enough high-quality game data. The extraction wasn’t linguistic. It was behavioural. The surgical equivalent is where we are heading.

Mentix’s Three-Phase Roadmap

Mentix’s roadmap, stated plainly, has three phases, each of which is independently valuable and none of which depends on the later phases being right.

Phase 0 - The Mentor Platform, a foundational platform allowing surgeons to connect with mentees and effectively train and transfer knowledge, onboarding them to Mentix and allowing them to explore deeper, long-term knowledge sharing. This phase is live.

Phase One: The Personalised Surgeon AI. Pilot with a single expert — Dr. Haji — operating as both founder and subject. Build the capture, annotation, and retrieval infrastructure on one career, proving that a queryable personalised surgical AI is technically feasible and clinically useful. This phase is in R&D.

Phase Two: The Training Platform
Deploy the personalised AI into NHS trusts, Royal College training programmes, and MedTech partner organisations, starting with the institutions where Dr. Haji already has teaching relationships. The unit of sale is not the model; it is the training license. A training programme director at a teaching hospital signs a contract that gives her fellows access to the Dr. Haji corpus, with the retrieval interface embedded in their training workflow. She measures success by time-to-proficiency — the compression of the learning curve that the field has been waiting for since Halsted.

Phase Three: The Multi-Surgeon Corpus
The multi-surgeon corpus and the data layer for surgical AI broadly. By the time ten or twenty senior surgeons across specialties are on the platform, Mentix holds a structured, expert-annotated surgical corpus with no true competitor. That corpus becomes the substrate for three distinct businesses: continued training platform expansion, decision-support tools for practicing surgeons, and — the largest prize — a licensed data layer for the robotics platforms that will need exactly this data to move from teleoperation toward semi-autonomy. That last market is the one Intuitive, Medtronic, CMR, Asensus, and half a dozen well-funded startups are all racing toward, and the one that will be structurally unreachable without an expert-grade surgical dataset.

The moat across all three phases is the same. Every expert onboarded makes the next one cheaper to onboard. The annotation methodology, built with Dr. Haji, gets refined with the second expert and polished with the fifth. The retrieval infrastructure gets more useful with every additional case. The trainees using the platform generate engagement data that feeds back into which annotations are most valuable. And — crucially — exclusive relationships with a small number of world-class surgeons per specialty, early, create a corpus that a competitor with more capital but a later start cannot easily reconstruct.

This is Bloomberg’s moat. Not the terminal. The settlement record. The data that accumulates as a byproduct of the service and that, over a decade, becomes the thing the industry cannot work without.

The Larger Prize

There is a version of this piece that ends on the commercial arithmetic. The surgical simulation market is around $665 million today and projected to approach $1 billion by 2033 [[Assumption]](market reports vary by source; order of magnitude is consistent). Surgical robotics is meaningfully larger — roughly $7 billion, growing at mid-teens. Adjacent markets — decision support, post-operative quality assurance, medical AI broadly — extend the surface into the tens of billions.

That version is real. It is the smaller version.

The larger version is that, for the first time in modern medicine, there is a credible path to decoupling surgical expertise from the physical presence of the expert. A 32-year-old surgeon in Dar es Salaam, operating on a complex colorectal lesion, should have access to the same pattern recognition, the same priors, the same decision support as a surgeon down the corridor from Dr. Haji at King’s. She does not yet. The gap between her outcomes and his trainees’ is not a function of her capability. It is a function of her geography and the slowness of the transmission mechanism that determines what she has access to.

That gap is what Mentix is built to close. Not immediately. Not for every procedure. But the architecture of the closure — a queryable corpus of expert surgical judgment, built in partnership with the experts themselves, delivered as infrastructure — is the first serious attempt at a structural fix.

Dr. Haji has watched the knowledge leave the room before. Every master has. Every master has also known, somewhere behind the frustration, that it did not have to. The video always existed. The notes always existed. The reasoning always existed. The missing piece was a way to connect them.

That piece exists now.

Interested in working with Dr. Haji or collaborating with Mentix?

Dr. Amyn Haji is the clinical lead and founder of Mentix. If you are a surgical trainer, NHS trust, MedTech company, or research institution interested in the Mentix platform, reach out directly.

Contact Mentix

Mentix is a founding fellow company at Utopia Studio, operating within our Co-Build at The Utopia Studio.

Subscribe now

Further reading: Dr. Haji et al. — ESD outcomes study (Springer Surgical Endoscopy, 2026)

Somewhere in the Persian Gulf right now, a gas detector is mounted to a pipe rack on an LNG train. It was installed in 2014. The engineering study that placed it there assumed a specific wind direction, a specific leak rate from a specific pump seal, and a specific arrangement of surrounding equipment. Since then, a Management of Change moved the detector four metres to accommodate new piping. The detector register got updated. The Fire & Gas Mapping Study — the document that actually proves the detector covers the hazard it was designed to cover — did not.

Subscribe now

Nobody noticed. Nobody was supposed to notice. No system exists to notice.

That detector is still counted as a safety barrier. The safety case still claims credit for it. The regulator still sees a compliant facility. The insurer still prices the risk accordingly. And if that pump seal fails and releases a hydrocarbon cloud, the detector may or may not catch it — because the coverage geometry that justified its position is now fiction.

This is not a rare edge case. This is the default condition of Fire & Gas safety systems across the global oil and gas industry.

The Lie Hiding in Plain Sight

Here is what everyone in process safety knows but nobody says plainly enough: Installed does not mean effective.

An oil and gas facility might have 200 gas detectors. Each one appears on an equipment register. Each one has a tag number, a maintenance schedule, a commissioning certificate. From a compliance standpoint, the facility looks protected. But whether those 200 detectors actually cover the hazard scenarios that engineering studies identified — whether the barriers are earning the safety credit that risk assessments assumed — is a question almost nobody can answer. Not the operator. Not the regulator. Not the insurer. And not the engineer who designed the system a decade ago.

The UK Health and Safety Executive analysed eight years of offshore hydrocarbon release data and found that between 36% and 69% of gas releases went undetected. Let that number land. On facilities bristling with detectors, up to two-thirds of actual gas releases were not caught by the detection systems supposedly guarding against them.

This isn’t a technology failure. The detectors work. The gas dispersion models are sound. The engineering standards — IEC 61511, Shell DEP 34.11-20, ISA TR84.00.07 — are rigorous. The failure is informational. The knowledge that connects “this hazard exists” to “this detector covers it” to “this detector is still working and still in the right place” is shattered across five separate document systems that have never been joined together.

Five Silos, Zero Verification

Dr. Satheesh Kumar Veluswamy has spent more than twenty years inside this problem. He’s worked across the GCC’s largest facilities — the kind of plants where a single process train produces more LNG than some countries consume. He wrote Mastering Fire and Gas Detection Systems, which became an Amazon bestseller, not because it was a gentle introduction to the topic but because it was the first book to say what practitioners already knew: the gap between designed safety and actual safety is growing, and the industry’s tools aren’t keeping pace.

What Satheesh understood — from two decades of walking process plants, reviewing P&IDs, sitting through HAZOP workshops, and watching detectors get moved without anyone updating the safety study — is that the verification problem isn’t technical. It’s architectural. The information needed to verify whether a safety barrier works is scattered across five separate silos, each created by a different engineer, at a different phase of the project, and stored in a different system.

Silo One holds the hazard studies. HAZOP worksheets that define what could go wrong — a gas release from a compressor seal, a fire in a pump bay, a toxic cloud from a chemical spill. These documents identify the scenarios. They live as PDFs in a file share, written years ago, rarely updated.

Silo Two holds the risk assessments. LOPA studies that take each hazard and assign a Safety Integrity Level — SIL 1, SIL 2, SIL 3 — and designate which protective layers must deliver how much risk reduction. The LOPA says: this scenario requires a Fire & Gas detection system as an Independent Protection Layer, credited at SIL 2. The LOPA is archived separately, often by the EPC contractor who built the plant, and it may not be in the operator’s document management system at all.

Silo Three holds the design. P&IDs showing the instrumentation layout, the detector types, the interlocks, the cause-and-effect logic that connects a detection event to an emergency shutdown. These are DWG files in a CAD system, revised through multiple turnarounds, and the current revision may not match what’s physically installed.

Silo Four holds the Fire & Gas Mapping Study — the document that actually translates the safety requirement into a physical detector layout. It defines detection zones, specifies how many detectors of what type are needed where, and calculates coverage targets per hazard scenario. This is the document that should answer the question “does this detector cover this hazard?” It is also the document most likely to be out of date, because it was written once, at design, and almost never revised.

Silo Five holds the operational evidence. As-built drawings. Proof test records. Inhibit logs. Maintenance histories from CMMS. Commissioning certificates. The actual field state of every detector — whether it’s working, degraded, inhibited, or missing. This evidence is scattered across multiple systems. Some of it exists only as entries in a DCS historian that nobody queries. Some of it is a printout in a filing cabinet.

To verify whether a single detector is earning its safety credit, you need to trace a thread through all five silos. You need to connect the hazard scenario (Silo 1) to the SIL requirement (Silo 2) to the design specification (Silo 3) to the spatial coverage calculation (Silo 4) to the current field state (Silo 5). For one detector, this might take a skilled process safety engineer half a day. For 200 detectors across 50 to 100 documents — the typical scope for a single facility — it takes six to eight weeks.

And so it almost never happens.

The Consultant Treadmill

When it does happen — usually triggered by an incident, an insurance requirement, or a regulator asking uncomfortable questions — the workflow looks like this.

A safety consultancy gets engaged. Senior process safety engineers, billing at rates that make the operator wince, spend two weeks just collecting documents. They chase the EPC contractor for archived files. They dig through the operator’s document management system, which is usually organised by project phase rather than by safety function, meaning a HAZOP from 2009 and a detector register from 2022 are in completely different directories with no linking metadata. They receive PDFs, DWGs, scanned paper copies, and Excel spreadsheets with inconsistent column names.

Then they spend three weeks manually reading HAZOP worksheets, cross-referencing each scenario to the LOPA, checking whether F&G was assigned as an IPL, opening the mapping study to find the coverage design, and comparing it to whatever equipment data they can locate. They note discrepancies in their own spreadsheet. When they need to verify spatial coverage, they print layout drawings and use physical rulers, highlighters, and calculators to estimate detection zones. If obstructions exist — columns, equipment, pipe racks — they sketch estimated shadows on the printout by hand.

After eight weeks, the consultant delivers a PDF report. It contains a barrier status table, a list of identified gaps, and recommendations. The traceability is minimal — page references at best, but no structured data, no confidence scoring, no way to update the findings when the next Management of Change moves another detector.

The report is accurate for the day it was written. Within months, it begins to decay. And the next time someone needs to verify barrier effectiveness — three, five, ten years later — the entire process starts from scratch. Different consultant. Different methodology. Different assumptions. No institutional memory.

This is a $150,000-to-$250,000 exercise per facility, per assessment. Most facilities run it once, at commissioning, and never again. The industry’s cumulative investment in process safety engineering — decades of HAZOPs, LOPAs, F&G designs, proof tests, and maintenance records — produces no compound intelligence about how barriers actually perform over time. Every facility is an island. Every assessment reinvents the wheel.

March 18, 2026

Then Ras Laffan happened.

On March 18, 2026, Iranian missile strikes hit Ras Laffan Industrial City, destroying the Main Cryogenic Heat Exchangers on two QatarEnergy LNG trains and severely damaging the Shell Pearl GTL facility. The attack took 17% of Qatar’s LNG export capacity offline — 12.8 million tonnes per annum — with repairs estimated at three to five years. QatarEnergy declared force majeure on long-term contracts to Europe and Asia. Twenty billion dollars in annual revenue was suddenly at risk.

Only two manufacturers in the world — Air Products and Linde Engineering — build cryogenic heat exchangers at this scale. Manufacturing lead time alone is 18 to 24 months, before installation even begins.

But the physical reconstruction, massive as it is, exposed a deeper problem. When you rebuild destroyed process trains, you don’t simply bolt new equipment onto old foundations. The process configuration changes. The hazard scenarios shift. The gas dispersion characteristics that determined where detectors should go no longer match the original engineering study. Every safety barrier surrounding the new cold boxes must be re-verified against whatever reality the rebuild creates.

The five-silo problem becomes acute. The HAZOP is from 2009. The LOPA is from 2011. The P&ID has been revised through three turnarounds. The F&G layout was designed for a process configuration that no longer exists. Operational records are scattered across systems that may themselves be damaged or incomplete.

And it’s not just the destroyed trains. Adjacent trains that weren’t hit experienced thermal radiation, shockwave vibration, cable tray stress, and potential sensor degradation from blast overpressure. They maintained structural integrity. But whether their barrier architecture — every detector, every coverage zone, every safety credit — remained valid after the blast was a question nobody could answer quickly.

The Ras Laffan attack didn’t reveal a new problem. It made an old one impossible to ignore. Every major operator in the GCC — QatarEnergy, ADNOC, Saudi Aramco, Kuwait Oil Company, Petroleum Development Oman — started asking the same question: What else don’t we know about the true state of our safety barriers?

The Shape of What’s Coming

Satheesh saw this coming before the missiles did. Not the attack — nobody predicted that — but the inevitability that the industry’s informational debt would eventually collide with a forcing function that demanded repayment. The question was always when, not if.

What he’s building — and what I can’t fully describe yet, because it’s still in the kind of development phase where showing too much invites the wrong kind of attention — is not another consulting methodology. It’s not a better spreadsheet. It’s not a 3D modelling tool that still requires someone to manually enter detector positions from a PDF.

It’s an intelligence layer.

The concept works like this. You take the five silos — the HAZOPs, the LOPAs, the P&IDs, the mapping studies, the operational evidence — and instead of asking a human to trace threads between them, you build a system that reads all of them, extracts structured data with source traceability, constructs a knowledge graph linking hazard scenarios to protection layers to physical detectors to field evidence, and then asks the question that nobody currently asks at scale: Does this detector earn the safety credit that the risk assessment assumed?

Not “is this detector installed?” That’s an asset management question. Not “has this detector been tested recently?” That’s a maintenance question. The real question — the one that determines whether a safety case is fiction or fact — is whether the full chain holds. Hazard identified. Risk assessed. Protection designed. Detector installed in the right place. Detector still working. Coverage still adequate. Logic chain still intact.

Every link in the chain must hold simultaneously. If any one breaks — detector moved, coverage drifted, proof test overdue, maintenance lapsed, process configuration changed — the barrier credit degrades. And the system should know, automatically, which credits are solid, which are at risk, and which are lost.

Why This Hasn’t Been Built Before

The natural question: if this is so obvious, why hasn’t someone built it already?

Three reasons.

First, the documents are genuinely hard to process. A HAZOP worksheet from 2009 is a PDF scan of a printed table with inconsistent formatting, handwritten annotations, and abbreviations that only make sense to the engineer who wrote them. A layout drawing is a CAD export with detector symbols that vary by contractor, coordinate systems that vary by project, and obstruction geometry that requires spatial reasoning to interpret. Until recently, the AI capable of extracting structured data from these documents at production quality didn’t exist. It does now.

Second, the spatial problem is non-trivial. Verifying detector coverage isn’t a matter of checking a box. It requires georeferenced spatial modelling — placing detectors on a coordinate system, calculating detection zones using industry-standard methodologies like Shell DEP 34.11-20, modelling obstructions that block detection, and comparing the result against the coverage targets from the mapping study. This is computational geometry applied to safety engineering, and it requires domain expertise that generic software platforms don’t have.

Third — and this is the real reason — the problem spans organisational boundaries. The hazard study belongs to the operator. The LOPA belongs to the risk consultant. The design belongs to the EPC contractor. The equipment data belongs to the maintenance team. The proof test records belong to the instrument technicians. No single stakeholder owns the full verification chain. Building a system that integrates all of it requires not just technical capability but a deep understanding of how these documents relate to each other — which fields in the HAZOP map to which fields in the LOPA, which tag numbers in the design match which entries in the detector register, which spatial coordinates in the layout correspond to which coverage zones in the mapping study.

This is knowledge that lives in the heads of senior process safety engineers. People like Satheesh, who have spent decades doing this work manually, building the mental model that connects the silos. The insight is that this mental model can be made explicit, encoded as a verification chain, and automated.

The Compound Effect

Here’s what makes this interesting beyond the immediate problem. Every time you verify a facility, you learn something. You learn what gaps look like at LNG plants of this vintage. You learn which detector types degrade fastest in which process environments. You learn what happens to coverage zones after Management of Change activity. You learn how barrier health actually degrades over time — not in theory, but in documented fact, across real facilities.

No individual facility owns this intelligence. But a platform that accumulates it — facility by facility, assessment by assessment — eventually holds a dataset that no operator, no regulator, no insurer, and no competitor has ever assembled. A structured, georeferenced, evidence-backed map of how safety barriers actually perform across the oil and gas industry.

That dataset is the moat. Not the software. The data that the software generates as a byproduct of delivering value to each customer.

Imagine an insurer who can price a portfolio of LNG facilities not based on self-reported safety claims but based on verified barrier health scores, benchmarked against anonymised peers. Imagine a regulator who can identify systemic weaknesses across an operator’s fleet before an incident reveals them. Imagine an operator who receives a forward-looking alert: “Based on proof-test patterns at similar facilities, this detector has an 82% probability of failing its next test within 90 days.”

That’s not verification anymore. That’s prediction. And prediction is where the real value lives.

The Person Behind the Problem

I want to close on something that matters more than the technology.

Satheesh didn’t come to this problem from a software background. He came from the plants. Twenty-plus years walking process areas, reviewing detection layouts, sitting in HAZOP workshops where the decisions that determine whether people live or die get made in fluorescent-lit conference rooms over stale coffee. He’s worked across the very facilities that were hit on March 18. He knows the people. He knows the equipment. He knows which drawings lie and which tell the truth.

When he wrote Mastering Fire and Gas Detection Systems, it wasn’t an academic exercise. It was an attempt to codify what he’d learned — to take the tribal knowledge that senior F&G engineers carry and make it accessible before it retires with them. The book became a bestseller because it filled a void the industry had tolerated for too long.

What he’s building now is the same impulse at a different scale. Not a book that teaches humans to do the verification, but a system that does the verification — with the same rigour, the same domain knowledge, the same insistence on evidence over assumption — and does it continuously, across every facility, for every barrier, every day.

The oil and gas industry has spent three decades building process safety systems and almost no time verifying that they still work. The data to answer the question exists. It’s sitting in five disconnected silos on every facility in the world. Someone just needs to read it all, connect it, and tell the truth about what it says.

That someone doesn’t need to be a person anymore.

But it took a person — one who spent twenty years learning what the truth looks like — to build the system that tells it.

Find out more about Barrier Intelligence

Dr. Satheesh Kumar Veluswamy is the author of Mastering Fire and Gas Detection Systems and the domain architect behind Barrier Intelligence — the first AI-powered platform for continuous Fire & Gas barrier verification in oil and gas.

Oliver Graham-Yooll & Karan Pinto - Co-building as The Utopia Studio

Subscribe now

Alexandra Coleman was a senior chartered electrical engineer at Arup when she first noticed the pattern.

Across multiple large-scale infrastructure projects, she watched six different consultancies produce six different risk assessments — environmental, structural, electrical, regulatory, financial, insurance — and not one of them talked to the others. Each report was thorough on its own terms. None of them modelled what happened when their risks collided.

The power consultant didn't know about the permit delay. The insurance underwriter didn't see the grid constraint. The environmental assessor didn't talk to the structural engineer about what salt air corrosion would do to the cooling infrastructure over a ten-year debt tenor. The lender's credit team got a stack of PDFs that each said "manageable risk" and signed off on a number that was, in practice, a guess dressed in professional formatting.

Alexandra saw it because she was the engineer sitting between the workstreams, watching risks interact in ways nobody was modelling. A grid approval runs late, so the construction timeline slips. The timeline slips, so the off-taker renegotiates. Revenue projections change. The DSCR covenant gets tight. None of it showed up in the financial model, because the model assumed these risks existed in isolation.

They don't.

She filed the insight away. Then she went and helped build BeZero Carbon, a global carbon ratings agency that raised over $50 million during her tenure as Chief Engineering Officer. The entire company was built on one premise: a market works better when risk is quantified and independently assessed. BeZero took an opaque, qualitative market and made it legible. Capital reorganised around the signal.

Then the AI compute boom happened.

Capital flooded into data centres. $7trn in annual capex projected by 2030, growing at 26 percent a year. And Alexandra recognised the same structural gap she'd been staring at for a decade: an enormous, fast-moving asset class, financed by sophisticated institutional capital, priced on risk data that doesn't exist.

She started Azraq.

What the Models Miss

There's a useful exercise if you work in project finance or credit risk. Pull up the financial model for any data centre deal you've underwritten in the last two years. Look at the depreciation schedule. Look at the equipment useful life assumptions.

Now ask yourself: does this model know the difference between a facility in Frankfurt and a facility in Doha?

It almost certainly doesn't. But the difference is enormous.

A UPS battery rated for five years at 25°C will last two and a half years at 35°C and less than eighteen months at 40°C. That's not a rounding error. That's a 60 percent reduction in battery life that shows up nowhere in the capex replacement schedule of a standard financial model. In the Gulf, where ambient temperatures routinely exceed 45°C and humidity can swing from 20 to 90 percent in the same week, the ASHRAE curves that underpin most equipment lifetime assumptions simply don't hold. Corrosion from salt air in coastal environments — Abu Dhabi, Dammam, Mumbai, Jakarta — can halve the rated lifespan of unprotected switchgear and transformers.

Or look at cooling architecture. Most financial models treat power capacity as fungible — 50 megawatts is 50 megawatts. But in a high-density AI training facility running thousands of GPUs at 700 watts each, 40 percent or more of total power draw goes to cooling alone. If the facility wasn't designed for liquid cooling, you don't have a 50-megawatt data centre. You have a stranded asset with a beautiful power contract it can't fully monetise. Thousands of megawatts of contracted capacity globally are already sitting behind this thermal bottleneck.

Or look at water. A single large data centre in an arid climate can consume five million gallons per day. When Qatar, Saudi Arabia, or Arizona face water allocation decisions, a data centre is not going to win that fight against agriculture or residential supply. Water curtailment risk doesn't appear in any standard credit model. But it is the single fastest path to forced operational downtime in water-stressed regions — and those are precisely the regions where the largest new builds are being planned.

Or look at the grid. Utilities are projecting five to seven year interconnection delays for new high-voltage connections. Data centres need to be operational in eighteen to thirty-six months. In West London, data centres have absorbed so much grid capacity that new housing developments can't get connected until 2037. The financial model says "grid connection: confirmed." The reality is that designed capacity and delivered capacity are two completely different numbers.

These are not edge cases. These are the standard operating conditions of the fastest-growing data centre markets in the world.

A billion-dollar SPV built on assumptions that don't account for how salt air eats switchgear, how desert dust degrades cooling performance, how water scarcity forces shutdowns, how grid delays strand capital — that is not a sophisticated financial instrument. It is an expensive guess. And the people writing the cheques are starting to notice.

How Azraq Works

At the core of Azraq is what Alexandra calls a physics-informed probabilistic AI engine.

This is not a spreadsheet with three scenarios. It's not a consultancy opinion with caveats. It is six AI agents, each trained on a specific risk domain, working across a 600 line-item risk framework calibrated for different data centre types, geographies, and climate conditions. The agents model the interdependencies — how a permit delay compounds a delivery overrun, how a grid constraint interacts with a cooling system choice, how a three-month slip becomes a twelve-month slip becomes a credit default.

The phrase "physics-informed" is doing real work. Traditional financial models treat risk factors as independent variables with linear sensitivities. Stress the revenue line by 10 percent, stress the cost line by 10 percent, see what happens to DSCR. That's fine for a bond. It is inadequate for an infrastructure asset where risks propagate like forces in a physical system — where a flood zone reclassification triggers a site relocation that triggers a six-month delay that triggers an off-taker walkaway that triggers a covenant breach.

The cascade is the risk. And the cascade is what Azraq's engine models.

Data comes from two sources. Open-source intelligence — GIS tools mapping power infrastructure, cooling systems, fibre networks, flood zones, grid availability, climate projections. And proprietary operational data from developers, operators, and lenders who feed real-world performance data into the platform. Azraq gets access because operators need what the platform gives back: if you're a developer trying to close financing, a lender-grade risk report that accounts for site-specific physics is the single most valuable document you can produce.

The baseline assumptions are calibrated by subject matter experts — the kind of knowledge that lives in the heads of people who've spent twenty years watching transformers corrode in coastal environments or cooling systems fail in desert heat. Coleman's engine encodes that expertise into replicable, auditable probabilistic models.

The output is a Value at Risk assessment at the 95th and 99th percentile confidence intervals. It tells a lender: here is what your actual downside exposure looks like, not the one your spreadsheet assumed.

Phase one is live now: a secure web-based workflow producing automated, lender-grade risk reports, with models being calibrated against real pilot projects. Phase two is the network: an interactive dashboard with API-connected live data feeds, portfolio analytics, and marketplace infrastructure connecting sell-side operators with buy-side capital through shared risk intelligence.

Why This Is Hard to Replicate

The technology matters. But the technology is not the moat.

Every project Azraq assesses makes the platform harder to compete with. Every assessment creates a training record — inputs entered, risks flagged, actual outcomes observed. After fifty projects, the model is meaningfully calibrated against reality. After five hundred, it is structurally superior to any new entrant. Not because of better algorithms. Because of better data.

The physics-informed engine learns the actual decision trees between each data point and its impact on value at risk. The accuracy compounds in a way that no competitor can shortcut, because they won't have access to the underlying asset data. Think of Bloomberg's bond settlement record — a live dataset built over decades that cannot be purchased or reconstructed from the outside.

Then there's workflow lock-in. Once a lender builds Azraq outputs into their credit decision process, the methodology becomes embedded in their underwriting chain. Switching isn't a software decision. It's a re-underwriting of every loan in the portfolio. The analogy is MSCI index inclusion — once embedded in the process, the platform becomes the process.

Then there's geography. The GCC, Southeast Asia, and Latin America are experiencing the fastest data centre growth globally, and they face the most severe environmental and infrastructure constraints. These are the markets where traditional risk models break down most completely — and where Azraq's physics-informed approach creates the widest performance gap.

And there's the mandate logic. Once lenders build Azraq into their credit approval process, every developer and operator who needs that capital must produce an Azraq risk score. Buy-side adoption makes sell-side adoption non-optional. The platform creates its own distribution.

Why This Matters Now

SoftBank, Amazon, Google, and others are taking on significant debt to finance AI infrastructure at a moment when geopolitical instability is raising borrowing costs and pushing capital toward safer assets. Private credit funds exposed to tech are seeing redemption pressure. AI compute economics remain largely unproven at scale.

When debt-fuelled technology booms unravel, they unravel fast.

Independent, forward-looking risk assessment that identifies hidden leverage, flags refinancing risk early, and prices actual project viability rather than growth narratives — that isn't a nice-to-have. It's the missing piece of the capital stack.

When risk is mispriced, capital flows to the wrong places. Projects that should get funded don't, because lenders can't differentiate between good sites and bad ones. Projects that shouldn't get funded do, because the models said the risk was manageable and nobody checked whether the physics agreed.

Learn more about Azraq

Alexandra Coleman, Founders of Azraq

Ollie Graham-Yooll & Karan Pinto , The Utopia Studio Co-Builders

Subscribe now

We have more surgical techniques than ever before. More specialized procedures. More demand for surgery from aging populations worldwide. Yet the system that trains the humans who perform these procedures hasn’t fundamentally changed since William Halsted formalized surgical residency at Johns Hopkins in the 1890s.

And the crisis isn’t about producing more surgical graduates. It’s about a training model designed for a world that no longer exists, colliding with the realities of modern medicine.

The Halstedian Hangover

Surgical training was built on apprenticeship. “See one, do one, teach one” became the mantra—a resident would observe a procedure, perform it under supervision, then teach it to the next generation. This model assumed something that was true for most of the twentieth century: trainees would spend enormous amounts of time in the operating room, working alongside experienced surgeons who knew them well enough to gradually hand over autonomy.

That world has disappeared.

European surgical trainees now work a maximum of 48 hours per week. In the United States, the cap is 80 hours—down from the 100+ hour weeks that previous generations endured. The curriculum hasn’t shrunk to match. Same procedures to learn, dramatically less time to learn them.

Subspecialization has fragmented the mentor pool. Seventy-five to eighty percent of new general surgery graduates now pursue fellowship training in narrower fields—trauma surgery, surgical oncology, minimally invasive surgery. The broad generalist who could teach across the full scope of operative care has become an endangered species. A trainee rotating through a service may never encounter the same attending surgeon often enough to build the relationship that enables real teaching.

This matters because surgical skill transfer depends on trust. An attending surgeon won’t hand progressively more responsibility to a resident they barely know. The educational literature calls this the “autonomy crisis”—residents completing training without the operative experience and independence required to practice confidently. Shortened rotations, service pressures, and the sheer volume of trainees competing for OR time mean the deep mentor-mentee relationships that defined surgical education have become increasingly rare.

Meanwhile, expertise concentrates geographically. The best surgeons cluster at tertiary academic centers in wealthy cities. A trainee at a rural hospital in Kansas or a teaching institution in Lagos faces the same fundamental problem: the mentor who could accelerate their learning isn’t there.

COVID made all of this worse. A global survey of surgical trainees found that the pandemic created “major disruption in all aspects of training,” with hands-on courses and conferences hit hardest. When researchers asked what would help recovery, mentorship by senior surgeons ranked near the top—but the same structural barriers that existed before the pandemic remained. The system had no mechanism to deliver mentorship at scale.

The Telementoring Paradox

Remote surgical guidance isn’t a new idea. In 1965, Michael DeBakey transmitted guidance on an open heart surgery from the United States to surgeons in Europe via satellite. Since then, decades of research have validated the concept. A systematic review of 66 studies found that 58 percent showed no difference in outcomes between telementoring and in-person supervision. No study found telementoring resulted in poorer postoperative outcomes. The technology works.

So why, sixty years later, does surgical training still depend almost entirely on physical proximity?

The answer lies in what telementoring studies validated versus what the real world requires. Research demonstrated that a surgeon in one location can guide a procedure happening somewhere else. It didn’t solve the harder problems: How does a trainee find the right mentor for their specific learning need? How does that relationship persist beyond a single procedure? How does feedback get structured, tracked, and verified? How do institutions trust that remote training meets their standards?

Existing tools are fragmented. Video conferencing platforms weren’t designed for surgical workflows. Recording systems capture procedures but don’t integrate with feedback mechanisms. Proctoring happens ad hoc, arranged through personal networks rather than systematic matching. There’s no audit trail, no credentialing pathway, no standardized assessment that travels with a trainee from one stage of their career to the next.

The surgical world has accumulated impressive hardware—robotic systems, augmented reality headsets, cameras that can stream in 4K from inside a patient’s abdomen. What it lacks is an operating system. A layer that connects the trainee who needs guidance with the expert who can provide it, captures the learning that happens, and creates accountability across the entire competency journey.

The Mentix Thesis

This is the gap Mentix was built to fill.

The platform connects trainee surgeons with leading experts for live case guidance, structured feedback, and secure global skills training. Not one-off consultations, but longitudinal relationships. Not informal advice, but documented competency development.

The founding insight came from Dr. Amyn Haji, a colorectal surgeon and surgical educator who spent years watching talented trainees struggle to access the mentorship that would accelerate their development. The problem wasn’t motivation or ability. It was infrastructure. The expert who could teach a specific technique existed somewhere in the world—but there was no reliable way to connect trainee to mentor, no framework to structure the learning, no system to verify what had been taught.

Mentix starts with procedures where remote guidance has the strongest evidence base: laparoscopy, endoscopy, and minimally invasive techniques where visualization already happens through cameras rather than direct line of sight. The platform enables an experienced surgeon to watch a procedure in real time, communicate with the operating surgeon, and provide the kind of moment-by-moment guidance that typically requires physical presence in the OR.

But the deeper play is what happens around the live guidance. Structured feedback forms capture what was taught. Competency tracking follows a trainee across procedures and mentors. Assessment data accumulates into a portable record of skill development. For the first time, a surgical trainee can demonstrate not just that they completed a rotation, but what they actually learned and from whom.

This creates value in multiple directions. Trainees get access to world-class expertise regardless of where they’re physically located. Expert surgeons can teach globally without leaving their institutions—monetizing knowledge and extending impact in ways that weren’t previously possible. Training programs gain visibility into competency development that was historically opaque.

The Device Company Wedge

Every new surgical device requires training. When a company like Creo Medical introduces an advanced energy platform or a novel endoscopic tool, surgeons need to learn how to use it safely and effectively. Historically, this meant flying expert proctors to hospitals around the world—an expensive, unscalable proposition that created bottlenecks in adoption.

Mentix offers a different model. Remote proctoring through the platform lets device manufacturers train surgeons globally without the logistics of constant travel. The expert demonstrates technique, watches the trainee perform, provides real-time correction—all the elements of effective proctoring, delivered over secure video with appropriate documentation.

This isn’t theoretical. Mentix has already established trial agreements with device companies seeking scalable training infrastructure. The economics are compelling: faster surgeon onboarding means faster device adoption means faster revenue recognition. For manufacturers, Mentix becomes the embedded training layer for their entire product portfolio.

The strategic value extends beyond cost savings. Regulatory pressure for documented competency is increasing. European medical device regulations increasingly require evidence that surgeons using new technologies have been appropriately trained. A platform that captures this training systematically—with timestamps, assessment data, and credential verification—solves a compliance problem that device companies currently manage with spreadsheets and emails.

The Global South Imperative

The surgical workforce crisis hits hardest where healthcare systems are already stretched thin. Sub-Saharan Africa has roughly 0.7 surgeons per 100,000 population—a fraction of what’s needed for basic coverage. The World Federation of Societies of Anaesthesiologists estimates that over 136,000 physician anesthesia providers would need to be trained globally just to reach a minimal density threshold.

These regions have training programs. They have motivated trainees. What they lack, disproportionately, is access to experienced mentors who can guide them through the learning curve.

Traditional solutions have struggled with dynamics that aren’t purely technical. Short-term surgical missions from high-income countries—sometimes called “voluntourism”—can undermine local expertise and create dependency rather than building sustainable capacity. What’s needed is bilateral partnership: structured relationships where expertise flows across borders without displacing the surgeons who will ultimately serve these populations.

Mentix’s model addresses this directly. A trainee in Nairobi can access mentorship from a specialist in London without either party relocating. The mentor relationship persists across months and years, building the kind of longitudinal development that one-off training trips cannot provide. The platform enables same-gender mentorship matching—data shows women and underrepresented minorities benefit most from mentors who share their background, but finding such mentors locally may be impossible.

Qatar’s positioning as Mentix’s base matters here. Neutral ground between established medical centers and emerging healthcare markets. Increasingly a crossroads for health technology investment. Well-resourced enough to support the infrastructure that global surgical training requires, without the complicated dynamics of colonial-era medical relationships.

Building the Moat

Platform businesses in healthcare face a consistent challenge: what prevents a well-resourced competitor from replicating the model? For Mentix, defensibility emerges from several reinforcing dynamics.

Network effects compound over time. More mentors on the platform attract more trainees, which generates more procedures logged, which produces better matching algorithms, which attracts more mentors. The data generated by each interaction—what was taught, how it was assessed, what outcomes resulted—creates proprietary competency benchmarks that no newcomer can replicate without years of accumulated history.

Device manufacturer relationships create lock-in on the supply side. Once a company has standardized its proctoring on Mentix, switching means retraining proctors, migrating compliance documentation, and disrupting established workflows. The same logic applies to hospital systems that build competency tracking into their residency programs.

Perhaps most importantly, whoever establishes credentialing standards for remote surgical training first sets the rules that others must follow. Regulatory bodies looking for documented evidence of competency will defer to whatever framework exists. Mentix is building that framework—not just a platform for connecting mentors and trainees, but a system for verifying what surgical education actually accomplished.

The Path Forward

Surgical mentorship has been locked in a 130-year-old model not because no one wanted to change it, but because the technological and institutional pieces weren’t in place. Reliable high-bandwidth video to operating rooms worldwide. Acceptance of remote training models, accelerated by COVID. Device manufacturers actively seeking scalable proctoring solutions. Regulatory frameworks that increasingly require documented competency.

Those pieces are now in place.

For surgeons, the implication is that training quality no longer needs to depend on the accident of geography—which hospital you matched at, which attendings happened to be on service during your rotations. Access to expertise can be systematic rather than serendipitous.

For device companies, the opportunity is to accelerate adoption without the cost and complexity of deploying proctors globally. Every week that a surgeon waits for training is a week of delayed revenue and a patient who might benefit from a new technology but can’t access it.

For healthcare systems facing surgeon shortages, the math is clear: you cannot hire your way out of this problem. There will not be enough surgeons. The only solution is multiplying the impact of every expert you already have—letting one experienced surgeon mentor ten trainees across three continents instead of teaching whoever happens to be in the same operating room.

Mentix is building the infrastructure for that world. A world where a trainee in any hospital can learn from the best surgeon for that specific procedure, live, securely, with structured feedback that accumulates into verified competency.

That’s not incremental improvement to surgical education. That’s the foundation for how the next generation of surgeons will be trained.

Learn more about Mentix

Learn more about The Utopia Studio

Dr. Amyn Haji, Founder of Mentix

Ollie Graham-Yooll & Karan Pinto , The Utopia Studio Co-Builders

At A-typical Ventures, we back startups solving urgent, structural problems in markets where timing and local expertise matter. Governata stood out not only for what the team has built, but for the regulatory moment they are capturing – helping enterprises and government entities in Saudi Arabia comply with national data mandates while transforming compliance into a strategic data advantage that enables institutional AI adoption.

The Opportunity: A Large Market Driven by Regulatory Mandate

Saudi Arabia is positioning data as a strategic national asset under Vision 2030. The enforcement of the Personal Data Protection Law (PDPL) and National Data Management Office (NDMO) standards has transformed data governance from a discretionary IT initiative into a national obligation. Thousands of private sector entities and hundreds of government bodies fall within the scope of these mandates. Non-compliance carries regulatory penalties and reputational risks, making data governance a priority for organizations across the Kingdom.

Yet most organizations still operate with siloed systems, manual reporting, and inconsistent data quality. Global incumbents lack the localization, Arabic-native interfaces, and deployment flexibility that local entities require. This creates a clear opening for a homegrown solution built specifically for the Kingdom’s regulatory and cultural context, with potential to expand across the wider region.

Why Governata Stands Out

Governata is a Saudi-built, SDAIA-aligned data governance platform, purpose-built to align with Saudi Arabia’s National Data Management Office (NDMO), National Data Index (NDI), and Personal Data Protection Law (PDPL) requirements. What sets them apart is not just compliance coverage, but how they have built it:

• Arabic-Native from Day One. The platform is built natively in Arabic and English, with data dictionaries and interfaces fully aligned with local standards. This is more than translation; it is true localization across user experience, data structures, and regulatory workflows.

• Deployment Flexibility. Governata supports cloud, hybrid, and on-premise deployments, satisfying the stringent data sovereignty and security requirements of ministries and regulated enterprises. For government clients handling sensitive data, this is non-negotiable.

• Ease of Use. Unlike legacy enterprise tools that require significant training, Governata’s interface is optimized for functional managers rather than data engineers. This accelerates adoption and drives organization-wide participation, which is essential for sustained compliance.

• Modular Architecture. The platform includes numerous interoperable modules covering data cataloging, classification, lineage, quality profiling, and personal data management. Organizations can adopt incrementally or deploy as a unified stack.

• AI-Powered Foundation. AI already powers key functions including data classification, anomaly detection, and compliance monitoring. The roadmap extends this into a decision intelligence layer that will transform governed data into actionable insights for leadership teams.

The Team: Domain Expertise Meets Execution

The founding team combines exactly the right mix of skills for this challenge:

• Djamal Mohand (CEO) brings over 15 years of experience scaling tech ventures across MENA. His previous roles include COO at Foodics, one of the largest restaurant technology companies in the region, as well as leadership positions at Kitopi, Rocket Internet, and Jenny. His operational expertise in building high-growth startups and navigating complex regional markets is directly applicable here.

• Jehad Senan (COO) has over 20 years in data governance and analytics, previously working at Deloitte and PwC Middle East where he supported data strategy and governance initiatives. He has built and scaled data offices across multiple organizations in KSA and understands intimately what enterprises need to achieve compliance.

• Khalid Almudayfir (Cofounder) brings 10+ years of experience in data governance and AI, with deep government relationships developed through multiple roles. His credibility and access within public sector circles have been instrumental in opening doors that would otherwise take years to access.

Strategic Vision: From Compliance to Intelligence

What excites us most is where Governata is headed. While the immediate wedge is compliance, the long-term vision is to become the decision-making layer for government and enterprise clients.

The progression is logical: once you have governed, classified, and connected an organization’s data assets, you have built the foundation for something much more valuable. Governata’s roadmap includes embedding AI to automate decision analytics, essentially creating an intelligent assistant that sits atop trusted data to guide strategic and operational decisions in real time.

This evolution from compliance tool to intelligence platform is where the real value creation lies. In markets where data sovereignty and localization are paramount, this positions Governata to become critical infrastructure for how Saudi and regional institutions make data-driven decisions.

Market Timing and Expansion Strategy

The timing could not be better. PDPL enforcement is now active. Government entities are under direct mandate to comply, and private sector requirements are following. This creates sustained, non-discretionary demand for exactly what Governata offers. The company is building momentum across the public and private sectors. Looking ahead, similar regulatory frameworks are emerging across the Middle East, presenting natural expansion opportunities. Success in KSA serves as proof of concept for replication in other regional markets facing comparable data governance requirements.

Why We Are Confident

Several factors give us conviction in Governata’s potential:

1. First-mover advantage in a compliance-driven market where regulatory pressure is accelerating, not slowing.

2. Deep localization that global competitors cannot easily replicate.

3. A founding team that combines operational excellence, domain expertise, and go-to-market access.

4. Early traction that that validates both product readiness and institutional trust.

5. A clear path from compliance revenue to higher-value intelligence offerings.

At A-typical Ventures, we back founders solving real problems with differentiated solutions in markets they understand deeply. Governata represents exactly this: a strong team, market tailwinds, product differentiation, and the potential to become essential infrastructure for data governance across the Middle East.

We are excited to support Djamal, Jehad, Khalid, and the entire Governata team as they build the data governance standard for Saudi Arabia and beyond.

The Path Forward

The regulatory tailwinds are real and the early traction is promising. Now comes execution. Governata is focused on expanding the team, pushing forward on AI-powered capabilities, and broadening its reach across sectors where governed, trustworthy data is becoming a prerequisite for growth.

About Governata

Launched in 2025, Governata is Saudi Arabia’s first enterprise data-management platform that provides AI-driven solutions in data governance and decision-making for both the public and private sectors.

The platform also has global aspirations and has plans to release new products on top of its foundational data governance software, including a unified AI-driven enterprise decision-making platform for the MENA region.

For more information, visit governata.com

About A-typical Ventures

A-typical Ventures is an early-stage venture capital fund backing exceptional founders from pre-seed through Series A across the Middle East. As part of UTOPIA’s integrated venture platform, A-typical Ventures combines capital and AI-native venture building through The Studio to remove friction from company creation and regional scaling. Rooted in the Global South, the fund pairs deep regional connectivity with global venture expertise, enabling founders to build faster, scale across markets, and focus on solving real, high-impact problems.

For more information, visit a-typicalventures.com

Subscribe now

Share

Modern tech feels like mist: dense from a distance, gone when you walk into it.

Usefulness isn’t a vibe. It’s a force. Category-defining companies compress messy, expensive, high-friction reality into something that feels like a cheat code. That compression creates new behaviour, new budgets, new expectations—a new “obvious” way to operate.

To spot real hidden problems, you need to learn where usefulness is being rationed.

Hidden in Plain Sight

A hidden problem is rarely hidden because it’s small. It’s hidden because it’s normal.

The reconciliation takes three days every month. The inspection cycle was done “because that’s how it’s done.” The compliance process has grown like coral reef over ten years. The maintenance schedule that’s half science, half superstition. The inbox that acts as the company’s API.

These are boring on purpose. Boring is how systems protect themselves.

People don’t describe them as “pain.” They describe them as “operations.” If you’re listening for complaints, you miss them. If you’re listening for resignation, you find them everywhere.

The usefulness lens flips the question: not “what could we build?” but “what do people do every week that they would delete from reality if they could?”

Five Signals of Category-Grade Problems

1. Humans as glue between systems

Smart people copying and pasting between tools, reformatting spreadsheets, translating emails into tickets, chasing approvals across chat threads. This is usefulness leakage. Software failed to cross the last mile: where information becomes a decision.

Most AI products generate text but don’t close loops. They don’t change what gets approved, scheduled, dispatched, paid, logged, or audited. They create output, not outcomes.

Category products close loops. They turn glue work into a system.

2. Metrics are obvious but improvement is socially blocked

“Everyone can see the downtime, but nobody can agree on why.” “Everyone wants fewer incidents, but incentives punish anyone who slows down production.” “Everyone wants faster onboarding, but legal and security are set up to say no.”

Usefulness isn’t limited by knowledge. It’s limited by coordination.

The opportunity is a new operating layer that makes the right action safe, legible, and easy to adopt. Category companies change behavior, not just sell software.

3. Costs are huge but diffused

Maintenance done too early “just in case” and too late “because we didn’t know.” Errors that don’t kill you, just tax you daily. Revenue leakage split across teams. Compliance effort invisible until you fail an audit.

Hidden problems love diffusion. Diffused pain is how organizations stay stable.

The category move: concentrate it, name it, measure it, make a single owner possible.

4. Shadow systems

Entire parallel universes of spreadsheets, Slack rituals, and recurring meetings to compensate for something. Excel sheets treated like sacred objects. The one operator who “just knows” how it works.

Shadow systems are expensive. They only persist when the alternative is worse.

That’s not chaos. That’s product-market fit for the wrong solution.

5. Insider advantage

The best opportunities are inaccessible to outsiders—not because they’re secret, but because you can’t see real data and workflows from the internet. You need tacit context: what counts as a good decision, what gets someone blamed, what the real bottleneck is, what data is trusted versus performative.

Outsiders build for the story. Insiders build for the scar tissue.

The Usefulness Filter

Does the problem have scar tissue? If yes, there are usually workarounds, budgets, and willingness to try something credibly safer and measurably better.

Category creation is not primarily invention. It’s recognition.

A category emerges when you take a recurring, expensive, accepted reality and make it feel optional. When people suddenly realize: we don’t have to live like this.

The Test

Stop asking “is this a big market?” Start asking:

• What recurring decision does this make easier or faster?

• Who gets blamed today when this goes wrong?

• What’s the minimum measurable improvement that would make someone look good?

• What’s the “before” workflow, including all handoffs and exceptions?

• When the system fails, who do they call? What spreadsheet do they open?

• If this disappeared tomorrow, what breaks first?

• If this worked perfectly, what new behavior becomes possible?

Category companies answer these cleanly. They speak in verbs: schedule, approve, reconcile, dispatch, underwrite, inspect, validate, remediate.

A useful product is a verb with a measurable result.

The Unsexy Truth

Usefulness is often upstream of the “sexy” thing. Everyone wants the interface. Nobody wants the plumbing. But categories are built by people willing to touch: data rights, workflow integration, security, edge cases, audit trails, human override, procurement, change management, incentives.

The real game isn’t “build a demo.” It’s “make it safe for an enterprise to let this run.”

Sell before you build. Not as hustle—as a truth machine. If you can’t get a serious operator to commit to an LOI, a pilot, a paid experiment, or recurring working sessions with real workflow detail, you probably don’t have a hidden problem. You have an interesting idea.

AI Changes Everything

AI makes novelty abundant, which means novelty stops being valuable. Usefulness becomes the only defensible currency.

If a model can generate an answer, the product isn’t the answer. The product is the system that turns the answer into an action the organization will actually take.

That’s where categories will form in the next decade, especially in high-barrier verticals where the cost of being wrong is high, workflows are complex, and data is messy.

The Final Filter

Does the problem create a permanent appetite for improvement?

Some pains vanish when patched. Others are structural—they regenerate. Maintenance, compliance, energy optimisation, asset integrity, claims, fraud, security, procurement, quality, workforce scheduling. Anything where the world changes faster than the organisation can update its rules.

Those are usefulness wells. Reality keeps producing new edge cases.

Pair this with insider access to workflows and data that actually matter, and you’re not just building a product. You’re laying tracks. Once tracks exist, a category has somewhere to run.

The usefulness problem in tech isn’t that people don’t want to be useful. It’s that usefulness is hard to see from outside, hard to prove early, and hard to ship because it forces you into real constraints: politics, legacy systems, procurement, liability, trust.

But if you can get comfortable there—if you can learn to love the ugly middle—you start to notice something.

The biggest opportunities don’t announce themselves.

They’re quietly draining time, money, and attention from the most important systems in the world.

They’re waiting for someone to make them optional…

Join us as a Co-Build Fellow

Everywhere we went, we saw the same thing. Thousands of domain and industry experts sitting on latent breakthroughs. Not ideas in the abstract, but hard-earned insight into where systems fail, where money leaks, and where outcomes could change materially if someone built the right thing.

These people are different from generalist founders. They carry proprietary problem insight that outsiders simply don’t have. They understand the workflows, the constraints, the politics, the physics. What they often lack is product translation, AI leverage, venture discipline, and a path from insight to company. That’s exactly the gap a Studio is meant to fill.

Subscribe now

When we formed The Utopia Studio, this was the leverage point we deliberately chose. Generalist founders don’t need us; the market already serves them well. In a world where building is faster and cheaper than ever, a Studio only makes sense if it concentrates on the scarcest input.

That input is domain context.

So we went looking for the people in boiler suits, lab coats, and back rooms. The engineers, scientists, operators, and boffins who keep critical systems running and know precisely what’s broken.

They are the ones capable of unleashing systematic change across entire industries—if partnered with the right co-builders.

Why we chose co-build Fellows

We chose a co-build fellowship model for a very deliberate reason: we want to operate as co-founders, not farmers.

Most studios and accelerators optimise for throughput. They batch founders, water them lightly, and hope a few grow. That works if your raw material is already startup-shaped.

Ours isn’t.

The people we work with often don’t see themselves as founders yet. Others know exactly what’s broken in their industry but recognise that getting from insight to company requires a partner with complementary strengths. They don’t need motivation. They need leverage.

The fellowship is the mechanism that lets us embed properly: shared risk, shared build, shared outcomes. We don’t advise from the edge. We step in and build alongside them until there is something real enough to stand on its own.

How co-build actually works in practice

Every co-build is tailored to the Fellow’s context, not forced through a generic playbook.

We start with their lived workflow, data, and constraints. Then we layer in our build tools: AI-native product design, rapid prototyping, and venture discipline. The result is a strong bias toward selling and testing early, with domain context baked in from day one.

For example, an asset integrity engineer doesn’t start with a dashboard. They start by mapping where inspections fail, where downtime is created, and who owns the budget. The first prototype is tested directly with operators they already trust. Sales conversations happen in parallel with product build, because the buyer and the problem are already known.

This isn’t customer discovery theatre. It’s validation at operational depth.

What we’re actually trying to build

We’re not interested in novelty software or incremental optimisation.

Our ambition is to create useful companies that drive system-level change: improvements measured in billions or trillions of dollars of impact across energy, infrastructure, healthcare, industry, and governance.

That’s why Qatar matters. It’s why the Global South matters. These regions run some of the world’s most critical systems and are undergoing rapid transformation. They offer real scale, real complexity, and real urgency.

We build locally with global ambition: companies born in local environments, designed to scale across regions where the same systemic problems repeat.

Why exits matter more than people admit

We believe in exits. Not as a vanity outcome, but as a test of usefulness.

We’ve seen domain experts build from such deep problem fit that their companies become genuinely valuable to acquirers. Not bolt-on features, but capabilities that matter: data they don’t have, workflows they can’t replicate, insight they can’t buy off the shelf.

This is where domain-led companies win. They make sense inside larger systems. They integrate cleanly. They solve specific, expensive problems.

In contrast, many generalist startups end up as awkward acquisitions: broad, confusing, and hard to place inside a real organisation.

When you start with deep domain truth, exits stop being theoretical. They become logical.

So what’s next

We’re building. Quietly. Seriously.

If you’re a domain or industry expert, the only thing that matters is whether this sounds familiar:

• You’ve accumulated real scar tissue inside an industry.

• You know how work actually gets done, not how it’s described.

• You can unlock data, systems, or pilots without permission theatre.

• People trust you because you’ve delivered.

• You’ve driven change where failure was visible and costly.

• You’re obsessed with fixing a specific, recurring problem/ unlocking hidden solutions.

• You take ownership end-to-end and think in outcomes.

• You understand where money leaks and what triggers buying.

• You know why past solutions failed and build for reality, not hype.

• You care more about things working than sounding clever.

If that resonates, you don’t need a pitch deck.

You already know what’s broken and the promise of the solution.

The question is whether you want the right partners to finally build it properly.

Learn more… here

If you care about the systems that keep economies running — and how AI will remake them — you’ll want to be here.

Season 1: Intersections — A 10-Part Exploration of the Systems That Will Define the Next Decade

1. The Model and the Origin Story

   How we built a venture studio designed for the AI era, why traditional venture models break in high-stakes domains, and the operator-first philosophy that powers everything we do.

2. Introducing the First Pods

   A walkthrough of the opportunity spaces shaping our early build program: infrastructure intelligence, decarb systems, programmable flow rails, sovereign compute, and vertical domain AI.

3. Inside the Studio

   Unfiltered views from our Studio Directors on what it takes to turn insider knowledge into investable AI-native companies, and why this model works where others don’t.

4. Pod 1: Infrastructure Intelligence

   Why modern infrastructure runs blind, where the value is leaking, and how AI-native control layers will rewrite uptime, resilience, and operational economics.

5. Pod 2: Decarb Industry

   A deep look at the industrial decarbonisation problem set, the real bottlenecks, and the intelligent operating systems needed to decarbonise without slowing growth.

6. Pod 3: Sovereign Systems

   How emerging economies build digital, compute, and energy sovereignty — and the strategic advantage of owning the intelligence layers that run nations.

7. Pod 4: Flow Rails

   Programmable money movement, economic coordination, and the next generation of financial plumbing designed for high-velocity economies.

8. Pod 5: Vertical Domain AI

   Why domain experts will build the most valuable AI companies of the next decade — and what it takes to create purpose-built intelligence for high-stakes environments.

9. Building with Insiders

   What we’ve learned working with operators, engineers, analysts, and domain specialists — the people who live the problems and hold the key telemetry to solve them.

10. What’s Next

    A preview of the next wave of Pods, the markets we’re exploring, and the systemic opportunities we believe will create the next generation of enduring companies.