Elon Musk called Terafab “the largest and most valuable building on Earth.” We can agree that he’s right about the size, but whether it’s valuable depends on whether anyone outside three companies can actually build one of these things.

You can also read this, along with other related pieces, on the Orbital Industries website.

Tesla and SpaceX are putting an initial $16.8 billion into Terafab, a semiconductor facility covering 100 million square feet in Grimes County, Texas, designed to combine logic, memory, packaging, and testing for use with robotaxis, humanoid robots, and space-based computing. Most news coverage has focused on the dollar funding amount in question; I want to focus on the building itself, because it’s tricky to grasp what a fab actually has to do, and why it’s so difficult.

What actually happens inside

Let’s start with the wafer. First, a disc of silicon has to be purified to 99.9999% to 99.9999999%, so we’re already starting at incredible constraints. Building a chip on it means repeating a handful of operations an enormous number of times, since material gets added in layers just a few atoms thick. A mask projects the circuit pattern through light onto a light-sensitive coating, and the exposed areas dissolve away. Impurities get implanted to change how the silicon conducts. The wafer is then subsequently heated above 1,000 degrees in order to repair that damage, and finally cooled down slowly to recover its crystal structure. The entire process is then carried out again on the following layer.

A logic chip runs this loop roughly 80 to 120 times, generating 500 to 1,000 process steps and around 90 masks over a three- to four-month cycle, with the wafer cleaned more than 200 times along the way, since any stray particle anywhere in that sequence can ruin the chip. A fab plant is essentially just this one loop involving clean, deposit, pattern, etch, clean, and measure.

Now if we were to compare this to a typical manufacturing plant, most of them have some built-in tolerance or allowance, so a component that’s a little out of specification will still fit, or a slightly impure metal can still be used. But in the case of a modern chip, the features are measured in single-digit nanometres, so the process has to be accurate hundreds of thousands of times tighter than conventional manufacturing, with almost no room for drift. Just a few atoms in the wrong place or a tiny speck of invisible dust is enough to cause a step to fail and stop the chip before the process even fully begins.

This level of meticulousness has resulted in the industry spending decades chasing sources of interference nobody would even think to look for. Early researchers traced mysterious failures to workers who’d touched a copper doorknob, carrying enough stray atoms to contaminate machinery and chips. Intel later found that swapping in a slightly longer cable during construction could quietly degrade yield for months. Their fix was a practice called Copy EXACTLY, the idea being to build every new fab as a complete physical duplicate of one that already worked, right down to the exact paint used on the walls. That’s the level of control a fab must maintain permanently, while still shipping hundreds of millions of chips a year.

The three companies that know how

Chip design is now “fabless”, with companies such as Apple and Nvidia designing their own chips and then outsourcing the manufacturing to a foundry, since only a very small number of companies still attempt leading-edge nodes: the likes of TSMC, Samsung, and Intel. It takes decades to acquire the process knowledge needed to operate one of these facilities at a yield, and money alone doesn’t provide a shortcut.

Even for the three companies that do know how to do it, the price of staying at the frontier keeps climbing. The same week Terafab made headlines, TSMC lifted its 2026 capex to $60–$64 billion, up from $52–$56 billion, and added $100 billion to its Arizona commitment, bringing the total there to $265 billion. All of this on the back of net income that was up 77.4% year over year.

The limitation in chip processing isn’t demand; it’s execution. Even if one had unlimited capital, a semiconductor fabrication plant would still be bottlenecked by a single company from the Netherlands. ASML is the sole maker of the EUV lithography machines that pattern advanced chips, and its newest High-NA systems cost $350 to $400 million each; a modern fab needs 9 to 18 of them. Each weighs 150,000 kilograms, ships in 250 crates, and takes six months and 250 engineers to assemble before running a single wafer. ASML builds only 12 to 15 a year, and yield learning after installation adds another 12 to 18 months before volume production starts. This is less a purchase order than a place in a queue behind every other advanced fab on Earth, for a machine that takes over a year to earn its keep.

Fifty picometers of margin

Once the tools are in, the room around them has to be cleaner and stiller than almost anywhere else humans build. Leading-edge cleanrooms run at ISO Class 4 or 5, permitting as few as 29 particles 0.5 microns or larger per cubic meter, versus over 3.5 million in a much laxer ISO 8 room, held by 300 to 480 full air changes an hour, replacing the room’s entire air volume roughly every eight to twelve seconds. Even that isn’t clean enough for the wafer, which travels between tools sealed inside pods, often entering an even cleaner sealed micro-environment than the room itself.

Underneath the cleanroom sits a sub-fab, one or two floors of pumps, chemical lines, and vacuum equipment that most visible tools depend on but that never appears on a factory tour, with an interstitial level above recirculating the air back down. Running through all of it is vibration control, an equally unforgiving standard: advanced EUV tools need vibration held below 50 picometers, a thousand times stricter than a decade ago, at a scale where a passing truck can ruin a wafer. The fix is entire rooms hung on isolation systems independent of the main structure, floating on elastomeric pads or springs. You’re building a structure, then building several more inside it that aren’t allowed to touch it. This holds even somewhere seismically calm like Texas, Arizona, or upstate New York, all preferred for new fabs for exactly that reason. The tools inside don’t care whether vibration comes from a fault line or a passing train.

The constraints nobody sees

The sub-fab also contains the chemical stores, some of which are genuinely dangerous, such as toxic doping gases, and others that ignite on contact with air, all piped through stainless steel lines polished inside so they can’t shed particles of their own. None of this is visible to a visitor, but a leak in this area can shut the entire factory down just as quickly as a problem in the cleanroom.

Above ground, water decides where a fab can be built. Every rinse needs ultrapure water, thousands of times cleaner than drinking water, with contamination held below 0.5 parts per billion at advanced nodes. A single leading-edge fab can use up to 10 million gallons a day, roughly what 33,000 US households use, and it takes 1,400 to 1,600 gallons of municipal water to produce 1,000 gallons of the ultrapure grade. Terafab’s water will come from the Gibbons Creek Reservoir, built to cool a coal plant that shut in 2018, repurposing capacity that would otherwise sit unused.

Beneath all this is a power supply that cannot fail for even a fraction of a moment. A voltage dip lasting just 10 to 20 milliseconds can ruin an entire in-process wafer lot, since a wafer mid-fabrication can’t be paused and resumed. Fabs aim to counteract this by layering redundant substations, battery storage, and diesel generators under their UPS, but even then, they can still get caught out. TSMC’s own Fab 21 in Arizona lost hours of production to a disruption traced to an external gas supplier, not even its own grid connection. A single unplanned hour of downtime at a leading fab runs $2 to 5 million in lost production, before scrapped lots or recalibration.

The part nobody can predict

By now, the shape and complexity of the problem should be clear. A wafer has to survive a hundred repetitions of a process with almost zero tolerance for error, in a room stiller than the ground beneath it and thousands of times cleaner than the air outside it, fed by chemicals that can ignite on contact with air. All of this topped off by a power supply that can’t blink for a fraction of a second, water purer than anything in nature, and a process using machines only one company on Earth can build.

A fully in-house, end-to-end process is a genuine solution to a real problem: logic, memory, packaging, and testing under one roof means fewer places for one vendor’s delay to become everyone else’s problem. However, owning the loop doesn’t make it smaller. Every discipline in this piece, from lithography and cleanroom control to chemical handling, water supply, and power resilience, still has to hold at the same tolerance, in the same building, for the entire life of the fab.

The part that no one can predict now is whether Terafab can keep to all of those constraints on the same site for a number of years.

For forty years, the boundary where light beats electricity has been closing in; From the ocean floor to the city, the building, and the rack. It’s now reached the inside of the computer itself.

You can also read this, along with other related pieces, on the Orbital Industries website.

In 1988, a consortium led by AT&T switched on TAT-8, the first transatlantic fiber optic cable, carrying ten times the capacity of the best copper cable ever laid across the ocean. Over those thousands of kilometers photons beat electrons so decisively that copper never competed at that distance again.

But the distance at which light beats electricity didn’t stay at the ocean floor. It’s been shrinking ever since. Fiber took the routes between cities, then the links between buildings, then the connections between rows of a data center, then the short hops between racks in the same room. At today’s data rates, an electrical signal is barely survivable over a couple of meters of copper, and the frontier is still moving: the newest networking silicon converts to light inside the machine itself rather than at its edge.

Thinking is cheap, moving is expensive

Everyone carries the same mental model of a computer: there’s a processor, which does the thinking, and there’s the wiring, which does plumbing. The thinking is the hard part, the expensive part, the part that Moore’s Law spent sixty years making better, while the wires just carry the results around.

That model has flipped. On a modern chip, a floating point operation, the basic unit of thinking, costs somewhere in the region of a picojoule, while fetching the two numbers that operation needs from memory a few centimeters away costs on the order of a thousand picojoules. The ratio has been measured at several hundred to one, and it’s getting worse, because transistor physics keeps improving while wire physics doesn’t. When Google published the energy accounting for one of its TPU generations, the arithmetic had gotten three times cheaper since the previous process node, but the cost of reaching into memory hadn’t moved at all.

A modern AI accelerator spends a small minority of its energy doing the actual mathematics, and the majority hauling numbers from A to B. Across real workloads, data movement has been measured at 40 to 60 percent of total system energy. The machine spends more effort passing notes than doing homework, and every joule spent moving a bit arrives as heat that has to be removed from a rack. This is the pressure that has been pushing light inward for forty years: wherever moving a bit electrically costs more than converting it to photons, the photons eventually win.

Distance is the new clock speed

Once movement dominates, a strange thing happens to the definition of a computer: its performance becomes a function of geometry. How fast your machine runs depends on how far apart its parts are, and what medium connects them.

This matters right now because AI stopped fitting on single chips years ago. A frontier training run is tens of thousands of accelerators behaving as one machine, which means the network between them has become part of the processor rather than infrastructure around it. Once the wires became part of the machine, the machine inherited their physics.

Those physics are unforgiving. Electrical signaling over copper degrades with both distance and data rate, and the data rates have gone vertical. At 112 gigabits per second per lane, a copper cable can carry a clean signal about two and a half meters. At 224 gigabits, the generation now rolling into AI clusters, that reach collapses to about one meter. The distance from a chip to the far side of its own rack has become, electrically speaking, too far. Copper, which carried human communication for a hundred and fifty years, now can’t reliably cross a server cabinet at the speeds AI demands.

The industry’s answer for the past decade has been to convert to light at the edge of each switch: pluggable optical transceivers, small modules that take an electrical signal, turn it into laser light, and push it down a fiber. Inside each one sits a digital signal processor whose job is repairing the damage the signal suffered on its 20-centimeter electrical journey from the chip to the faceplate. That repair work alone accounts for roughly half the module’s power, spent undoing what a short copper trip did to the signal rather than making light. It works, and it built the modern data center. But the conversion burns real power, roughly 30 watts for a state of the art 1.6 terabit port, and a large AI cluster needs hundreds of thousands of them. Jensen Huang has said that connecting a million GPUs with pluggable optics would consume around 180 megawatts just for the transceivers, more than the total power draw of most entire data centers. The cost had migrated to the conversion points at either end.

Co-packaged optics is the culmination of that forty-year iteration: instead of converting to light at a module plugged into the switch faceplate, the optical engines are placed on the same package as the switch silicon itself, millimeters from the transistors. The electrical path shrinks from tens of centimeters to almost nothing, and with it goes most of the amplification, retiming, and signal conditioning that made the old conversion so hungry. Nvidia’s co-packaged photonic switches, the first of which reached the market in 2025 with Ethernet versions following through 2026, cut the power of that 1.6 terabit port from around 30 watts to 9 using a quarter of the lasers. A single switch replaces 72 pluggable transceivers. Microsoft, Meta, CoreWeave, and Oracle are among the first buyers, and TSMC has built a dedicated process platform for stacking photonics onto logic.

The caveat is that these first products change less than the headline numbers suggest. Networking is under a tenth of a cluster’s total power, so even dramatic transceiver savings trim only a few percent off the whole. This generation partly serves as a rehearsal for the supply chain. The deeper shift arrives later this decade in the scale-up fabric, the links that let a rack of GPUs behave as one giant chip, where each GPU moves nearly ten times the data it sends across the wider network and copper’s two-meter reach caps the size of the whole architecture.

What this is really about

This took a long time, and not because the physics was unfavorable. The hard problems were practical: attaching hair-thin fibers to silicon at manufacturing yield, keeping lasers alive next to hot chips, and servicing a switch whose optics can no longer be unplugged when a module dies. Co-packaged optics shipped once the packaging matured, years after the idea did, and the companies that solved fiber attach and laser reliability, (problems with no glamour whatsoever), are the reason the roadmap works.

I find the whole arc clarifying for how to think about compute. The story we tell about progress in AI is a story about chips: transistor counts, process nodes, FLOPS. But the real limit has been drifting away from the arithmetic for years, first to memory, now to the network, always toward the movement of data rather than the transformation of it. The densest racks exist largely to keep dozens of GPUs inside copper’s shrinking radius. The liquid cooling they demand is the price of huddling that close together. Cluster topologies are being redrawn around what distance costs. And photons keep getting pulled deeper into the machine, one crossover at a time, with silicon itself as the only territory left.

The next time someone shows you a chart of AI progress denominated in FLOPS, remember that the FLOPS are nearly free. The expensive thing, the thing entire industries are being rebuilt around, is carrying a number from one side of a building, or one side of a package, to the other. The capability roadmap of AI now runs as much through photonics, packaging, and thermal engineering as through transistor counts. Light shows no sign of stopping at the package edge: the next crossovers are already visible, between chiplets, between processors and memory, each one another retreat of the last meter of copper. From the outside, every one of them will look like the machines simply getting faster, and few will notice that what actually changed is how far a number can travel before the journey costs more than the thought.

In 1900, most of the world’s food supply relied heavily on a single desert in Chile. This limited supply chain seemed like a fact that was outright unavoidable. But thirteen years later, a chemical company made that fact irrelevant. AI infrastructure is now accumulating critical dependencies with the same shape, and the same assumption that they’re permanent. So how do we learn from the lessons of the past?

You can also read this, along with other related pieces, on the Orbital Industries website.

Every era has inputs it treats as facts of geography, and for the last decades of the nineteenth century, fixed nitrogen was one of them. Crops need it, soil exhausts it, and by 1900 the world was farming harder than manure and rotation could replenish. Gunpowder and explosives also competed for the same resource, and nearly all of the supply came out of one place: sodium nitrate deposits in the Atacama Desert, shipped around Cape Horn to the farms and munitions works of Europe and America.

Supply was scarce against demand, and by any measure this wasn’t a minor exposure. Peru’s guano islands had already been mined to depletion in a few decades. Chile fought a war with Peru and Bolivia over the nitrate provinces and won, and by the 1890s it controlled roughly four fifths of the world’s supply of usable nitrogen. By 1890 the duty on saltpeter accounted for around half of all Chilean government revenue, and it stayed there for nearly three decades. Everyone understood this was a major fragility, and yet the solutions proposed were, for many years, to find ways to work with the limited supply.

The default responses were the ones you’d recognize from any modern supply chain review: stockpile, diversify, secure the shipping lanes, negotiate access. The dependency was treated as something to be managed, because managing it was seemingly the only available move.

The breakthrough

In September 1898, William Crookes stood up in front of the British Association for the Advancement of Science and told the audience that this wasn’t good enough, and that the wheat-eating world was heading for starvation on current trends. “It is the chemist who must come to the rescue of the threatened communities,” he said.

On July 2, 1909, Fritz Haber took up that challenge and demonstrated the answer in a lab in Karlsruhe: it was possible to pull nitrogen out of the air. Combine air with hydrogen under heat and pressure over a catalyst, and out comes liquid ammonia. His apparatus produced it drop by drop, at about 125 milliliters an hour.

BASF bought the process and gave it to Carl Bosch, a chemist who had trained as a metallurgist, and the lab result promptly met industrial reality. The first high-pressure reactors burst. Bosch worked out why: hot pressurized hydrogen was leaching the carbon out of the steel, leaving the vessel walls soft and brittle from the inside. He redesigned the reactor with a soft iron lining and vented walls, and in doing so more or less founded high-pressure chemical engineering. In parallel, Alwin Mittasch ran what may be the first great high-throughput screening program in history: roughly 2,500 candidate materials across some 6,500 experiments, searching for a catalyst cheap enough to matter. The eventual answer was iron, promoted with alumina and potassium oxide, and it’s essentially the same catalyst the industry uses today, more than a century later.

With all the problems worked through, the Oppau plant then came online in September 1913, and within a year it was making 40 tons of ammonia a day. With this discovery, the most critical supply chain in the world had stopped being a supply chain issue at all.

Parallels today

The reason I’m writing about this today is because the history feels current. AI infrastructure has quietly assembled its own Atacamas. Advanced packaging capacity concentrated in a handful of Taiwanese fabs, high-bandwidth memory from three major suppliers, power rationed by interconnection queues, and grid transformers on four-year lead times. Each of these gets discussed the way Chilean saltpeter was discussed in 1900: as a fact of the landscape, to be hedged, stockpiled, and negotiated around. The entire discipline of AI supply chain strategy is, at present, old-fashioned saltpeter thinking.

But the nitrogen story challenges us to think differently and consider that these aren’t facts of geography, they’re unsolved chemistry and engineering problems wearing geography’s costume. The atmosphere was always 78% nitrogen; the constraint was never the supply of the element but the absence of a defined engineering process. And the process didn’t come from one invention alone. It came from years of reactor metallurgy and six and a half thousand catalyst experiments, funded by a company that decided the dependency was a problem rather than a condition.

And therein lies the answer, and a promising outlook for the future. If we’re drawing parallels to today, we’re in a much better position than we were in 1913. Mittasch’s search took three years because every candidate had to be physically synthesized and tested. That search loop is exactly what AI models now compress, in some domains by orders of magnitude. The tools for experimenting your way off a dependency have never been cheaper, more efficient, or more effective.

Surviving the last century’s fragile supply chain wasn’t about securing the best access to Chilean nitrates, it came down to who made that access irrelevant. Somewhere in today’s list of managed dependencies, the same opportunity is sitting in plain sight, waiting for someone to treat it as an engineering problem instead of geography.

The UK wants to triple its compute capacity by 2030 and rebuild the energy system that will power it. But the trades that will do that work take four years to train, and the pipeline is producing a fifth of what the industry needs. The scarcest input to sovereign AI may not be chips, energy, or capital, it may simply be people.

You can also read this, along with other related pieces, on the Orbital Industries website.

When people talk about what it takes to build sovereign AI infrastructure, the conversation tends to settle on a few familiar variables: energy, land, planning permission, capital. These are real constraints, and I have spent a good deal of time thinking about them. But spend any time on the construction side of the buildout, whether data centers, grid connections, or the factories that feed them, and a different bottleneck comes into focus, one that’s harder to fix because you can’t buy it, legislate it, or build it. That constraint is the skilled trades: electricians, mechanical engineers, fiber-optic technicians, high-voltage specialists, and the people who commission and maintain these systems once they’re built.

The UK government has set a target of at least 6 GW of AI-capable data center capacity by 2030, roughly triple the current installed base, and the capital is following: Microsoft has committed around £22 billion to UK AI infrastructure, and Google has pledged £5 billion more. Yet the data centers are only the most visible piece of what that capital buys, because every gigawatt of compute drags behind it substations, transmission upgrades, and generation projects that draw on the same limited pool of trades. For once, the money is abundant, and the harder task lies in finding enough people to build it all.

Counting the electricians

To see the problem clearly, it helps to start with electricians because they are arguably the most critical trade in the entire buildout. The IBEW estimates that electrical systems account for 45 to 70 percent of total data center construction costs, which makes sense once you consider that every rack of GPUs needs power distribution, backup generation, cooling controls, and the cabling that connects them, all of it installed by certified electrical workers. The shortage is already biting: Microsoft’s president, Brad Smith, has publicly identified the lack of electricians as the single biggest obstacle to the company’s data center expansion, with electricians commuting from as far as 75 miles away, or relocating temporarily, just to keep projects on schedule.

At home, the UK needs approximately 104,000 new electricians by 2032 to meet demand from renewable energy and infrastructure projects, a figure that does not fully account for the data center surge, and the training provider JTL has warned that the number of trainees could fall by a third by 2038 without intervention, stalling both decarbonization and growth. As Randstad CEO Sander van ‘t Noordende put it: “Ultimately, the real constraint on global tech growth isn’t solely related to a shortage of microchips, energy or capital; it’s the severe scarcity of the specialized talent required to build it.”

Britain is far from alone in this problem. In Germany, electrical contracting firms are carrying around 96,000 unfilled vacancies, and France is expected to need an estimated 200,000 electrical workers filled by 2030. America, where the buildout is most intense, is no better placed: the Bureau of Labor Statistics projects roughly 81,000 electrician openings every year between 2024 and 2034, while McKinsey puts the gap at 130,000 electricians by 2030.

And this is only one trade. The same shortfall runs through almost all of the other crafts a data center depends on, which is why the workforce gap, not chips or capital, keeps surfacing as a major constraint.

Training the people, and the robots

If the UK is serious about 6 GW, and about sovereign AI more broadly, the workforce question needs to be treated with the same urgency as the energy question. Thankfully, that urgency is starting to appear in some places.

The CITB committed £267 million in 2024 to addressing the skills crisis, and the government is rolling out various schemes like the homebuilding skills hubs designed to fast-track training and add 5,000 construction apprenticeship places a year.

But these efforts, encouraging as they are, still run into the fundamental constraint, which is time. It takes four years to train a certified electrician, so a crash program started today would deliver its first graduates in 2030, the same year the 6 GW target is supposed to be met. And that assumes you can also fix dropout rates, and the competition from every other sector that needs exactly the same people.

It’s clear then, that training has some difficult constraints. For that reason, there’s a tempting narrative here to say that automation is the alternative that will solve this. After all, there’s an enormous productivity overhang in UK manufacturing especially, where the technology to automate exists and the investment simply hasn’t been made. That’s a capital problem, and in principle it’s solvable.

But construction sites are a different problem entirely to manufacturing. A factory robot operates in a structured, repetitive environment, while an electrician on site works in an unstructured, constantly changing one, making judgment calls that depend on building codes, physical layout, the work of other trades, and conditions that differ on every project. Humanoid robots capable of that kind of work are, by any credible assessment, beyond the 2030 horizon. Current prototypes that achieve near 90 percent success rates in laboratory simulations succeed at just 12 percent of everyday tasks in the real world, and few commercially available platforms can cover a full eight-hour shift on a single charge.

For these reasons, no serious technologist expects to see a robotic electrician wiring a substation in Northumberland within the next ten years. There is, however, a more productive way to read the constraint: if automation works best in structured environments, then the place to look for leverage is in the most structured part of the whole process, which is the engineering work itself.

Multiplying the expertise

The more immediate leverage, and the area where I have an obvious professional interest, is to multiply the expertise you already have, because the shortage doesn’t stop at the site gate. A data center project draws on mechanical, electrical, thermal, structural, and increasingly chemical expertise, and each of those disciplines is its own scarce hiring pool with its own multi-year pipeline. This is where AI can be genuinely useful today: if robotics fails on construction sites because the environment is unstructured, engineering design work is the opposite case, structured, digital, and full of the simulation and optimization problems that machine learning handles well. AI engineering tools can now run the thermal modeling, layout optimization, and compliance checks that used to occupy teams of specialists for weeks. Because they compile expertise across disciplines, one experienced engineer supported by the right software can cover ground that once required a mechanical engineer, a chemical engineer, and a construction engineer working in sequence. The engineers remain as necessary as ever; what the software changes is how many projects each of them can carry, and that’s the number that matters when the workforce itself can’t grow fast enough.

A skilled workforce is the product of decades of sustained investment in training, career pathways, and institutional knowledge. A single spending announcement won’t buy it, and automation can’t yet stand in for it, which leaves multiplying the expertise of the people already available as the nearest source of leverage.

The sovereign AI debate in the UK is, rightly, focused on ensuring the country is not dependent on others for its critical technology. But sovereignty extends beyond owning the models and controlling the data to the ability to physically build and maintain the infrastructure that everything else runs on. Right now, Britain is planning for a future its workforce is not yet equipped to deliver. But the gap is closable, and the fastest way to start closing it is to put AI to work where it already excels today.

Britain sits on enough always-on clean power to cover most of its electricity, and enough heat to warm its buildings for a thousand years. In February 2026 it switched on its first geothermal power station to reach that heat, half a century after it first went looking.

You can also read this, along with other related pieces, on the Orbital Industries website.

Near Redruth in Cornwall stands the United Downs power station. It came online in February 2026, and it’s the first plant in Britain to make electricity from the heat of the earth. Its production well runs to 5,275 meters, the deepest hole ever drilled on British soil, and it took the better part of a decade to complete. By any ordinary measure it’s production is still relatively small: 3 MW or so, enough energy for perhaps ten thousand homes, which is about 0.01% of national demand. But that solitary output is just the beginning.

The granite under Cornwall is unusually radioactive, rich enough in thorium, uranium, and potassium to generate its own heat, which leaves it running roughly twice as hot as the British average at any given depth. The most recent national survey, published by Project InnerSpace in February, puts the country’s technically recoverable geothermal electricity at around 25 GWe, close to three-quarters of what Britain currently uses, alongside a heat resource large enough to cover national demand for something like a thousand years. The heat is real, it’s enormous, and unlike almost everything else we file under “renewable” it never stops: it remains unaffected by whether the wind is blowing or the sun is shining, and it will still be there, at a steady couple of hundred degrees, long after the last gas turbine has been switched off for good.

So if the resource is genuinely that large and that dependable, the obvious question is why almost none of it has reached the grid in the fifty years we’ve known it was there.

The price of the dig

Cornwall has understood its own heat for a long time: in the late 1970s, when an earlier energy shock had Britain hunting for homegrown power, engineers from the Camborne School of Mines began one of the world’s first Hot Dry Rock experiments at Rosemanowes quarry near Penryn, fracturing the granite and circulating water through it to prove the principle that United Downs would eventually turn into a business. The geology has since been surveyed, the faults mapped, and the temperatures logged. What has never fallen far enough is the price of reaching them, and that price is set almost entirely by drilling, which accounts for somewhere between 30% and 57% of the cost of a geothermal plant and pushes toward the upper end of that range in hard rock.

To see why, it helps to picture what the drill is up against. Granite is one of the least accommodating materials in the crust, ten to twenty times harder than the concrete of a pavement, with a compressive strength that runs toward 250 MPa against perhaps 40 for the soft sedimentary shales the oil industry has spent a century learning to drill quickly. In those shales a modern rig can advance at hundreds of meters an hour, while in hot Cornish granite the same rig manages closer to just five. The problem compounds with depth, because at the 200 degrees Celsius and beyond that make the heat worth having, even polycrystalline diamond, the hardest thing we routinely manufacture, blunts and fails under the combined heat and load. Each failure means tripping five kilometers of steel out of the ground to change the bit and running it all the way back down, with the rig and crew billing at full day rate throughout.

Unfortunately, all that accumulated cost ends up in the price of the power. United Downs sells its electricity under a contract worth £119/MWh in 2012 prices, while offshore wind has cleared Britain’s recent auctions at about half that, and the government’s own reviewers have put geothermal’s premium down to the expense of drilling safely.

Reduced to its essentials, the obstacle is a surprisingly small one: what stands between Britain and a large, clean, always-on power source is how long a few centimeters of engineered material can keep cutting hard rock, kilometers down, before heat and abrasion wear it away.

How it gets cheaper

Some of the cost of drilling erodes through sheer repetition. The clearest demonstration of this comes from Fervo Energy in the United States, which has been drilling hot granite in Utah and treating every well as a rehearsal for the next. Across the first four horizontal wells at its Cape Station project, the cost of a single well fell from $9.4m to $4.8m, nearly halving, and by the fourth the rate of penetration had beaten the official American projections for 2035. Fervo managed this by importing, wholesale, the drill bits and the accumulated field habits of the shale industry next door, moving two decades of hard-won oil-and-gas practice into geothermal in the space of a single campaign. This clever structuring of procurement is something we’re seeing other industries like data center construction adopt too.

That kind of learning has a floor, though, and the floor is the material itself. You can optimize the schedule, the drilling mud, the well path, and the crew, but at the bottom of it all you’re still asking an engineered cutting edge to work through the hardest rock in the country while the surrounding heat degrades it almost as fast as it cuts. What’s left at the root is a materials problem, and there are three quite different ways to try and solve it.

The first is a design-engineering challenge; making the cutting edge last longer. It’s incremental work, but far from marginal: engineers at the drilling firm NOV recently redesigned a diamond-composite bit that cut its time on the bottom of the hole by 63% and drilled 500 feet of granite at better than a hundred feet an hour without once coming up for repair.

The second is to reach past the drill-bit materials we already have toward genuinely new ones; harder and tougher and more stable when hot. The American Department of Energy is funding work at Argonne on superhard nanocomposite bit materials meant to double the rate of penetration outright. This could become more interesting as materials research becomes integrated with AI infrastructure platforms. Searching the space of possible hard, heat-stable composites is exactly the sort of vast combinatorial problem that machine-learned models of materials have started to make tractable, collapsing discovery timelines in things like coolants, semi-conductors, and catalysts over the past few years.

The third route is the most radical, which is to rethink the engineering process itself and to stop grinding at the rock at all. Quaise, a company spun out of MIT in 2018, drills with millimeter waves generated by a gyrotron, vaporizing the rock with a beam and leaving no cutting surface in the hole to wear out. Last year, at a granite quarry in Texas, it bored a hundred meters this way, ten times faster than any previous attempt, and it’s aiming for a pilot power plant by 2028. Its ambition runs beyond cost, because with no bit to destroy, Quaise believes it can push far deeper and hotter than conventional drilling allows, into rock at 400 degrees and above where water turns supercritical and carries several times the energy of ordinary steam.

All three approaches aim at the same target: using advanced materials and engineering to bring the cost of the drilling down far enough that the free heat below finally justifies the trip.

What the heat is worth

None of this would feel urgent if the value of the heat energy had stayed where it was in 1996. What has changed now is the demand. Firm, always-on power has quietly become one of the most valuable things in the economy, because the build-out of artificial intelligence has turned electricity, and specifically the dependable, build-anywhere kind, into the binding constraint on the whole enterprise. A data center needs power at three in the morning in January and in every other hour of the year, for twenty years at a stretch. That steady, weatherproof profile is precisely what geothermal supplies. It’s frugal with land as well, occupying only a small fraction of the ground a wind or solar farm needs for the same output, which counts for a great deal when the sites worth powering keep getting larger and harder to place.

There’s an interesting circularity in this: the same wave of AI investment that has made always-on power so crucial and valuable is also producing the most promising tools for making the drilling cheap. So the technology that most needs the energy today may turn out to be the one that lets us reach it. The demand and the means of supply are, in a real sense, the same phenomenon.

Which brings me back to that solitary 3 MW in Cornwall. For most of a century, British geothermal has been quietly filed under geological bad luck: no volcanoes, the heat too deep, never quite worth the effort. What has actually kept it locked away is the wear life of a small piece of superhard material, and materials are something we are finally able to engineer to order. That makes the heat beneath Cornwall a great deal more reachable than a century of leaving it alone would suggest, and whether Britain now decides it’s worth the drilling is a different question, and a far more hopeful one to be left with.

ASML Business Line NXE; NXE:3800E; cleanroom building 5 Veldhoven

In 1990 America made over a third of the world’s chips. Today the most advanced ones come from a single island, patterned by machines from a single Dutch town. How did the most widely dispersed industry on earth become the most concentrated, what are the risks, and what can we learn for the future of technology resilience?

Every leading-edge chip in the world is patterned by a machine that comes from one town in the Netherlands. The town is Veldhoven, the machine is an extreme-ultraviolet lithography system, and the company that makes them, ASML, makes all of them. Each one is roughly the size of a bus, costs somewhere north of $200m, and there are only about forty built a year for a grand total of three serious customers, with no second supplier anywhere.

Those machines then feed, overwhelmingly, into fabs on a single island. That island is Taiwan, and in practice the fabs are owned by a single company, TSMC, which makes over 90% of the world’s most advanced logic. The high-bandwidth memory stacked next to that logic in every AI accelerator comes mostly from two Korean firms. A great deal of the tooling, the wafers, and roughly half of the photoresist chemistry comes from Japan.

In short, the most sophisticated object humanity manufactures at scale passes, at some point, through a supply chain you could sketch on the back of an envelope. But it wasn’t always like this, and that’s the part worth pausing on. In 1990 the United States held roughly 37% of global chip-making capacity. That rapidly became barely 12% by the early 2020s. And the slide had started even earlier: at the beginning of the 1980s, US firms held more than half of worldwide semiconductor sales, a lead Japan had taken from them before the decade was out. The geography we now treat as a law of nature is roughly forty years in the making. So the question I keep coming back to is a simple one: how did an internationally interdependent, free-trading industry end up as one of the most concentrated? And, given how much now rides on it, what can honestly be done?

A chain of rational decisions

It’s tempting to look for a villain here, a policy blunder, a stolen decade or a single bad bet, but what happened is quieter than any of those. Instead, it was the sum of a lot of individually sensible choices, made by different people in different countries, none of whom were aiming at the outcome we got.

The split that mattered most, the divorce of chip design from chip manufacturing, only became possible in the late 1980s. Through the 1970s a chip design was not a document, it was a craft: layouts were drawn by hand around the quirks of one company’s specific manufacturing process, and a design made for one fab was simply meaningless in another. What changed was abstraction. In 1979 Carver Mead and Lynn Conway published a set of simplified, process-independent design rules that let designers work without knowing the details of the factory, and through the 1980s a new industry of design software gradually grew up around that idea, turning a chip design into a portable file that could be handed to any fab running a compatible process. Once designs were portable and fabs were unaffordable, splitting the industry in two stopped being a choice and became an inevitability.

Starting with Japan, its firms, Toshiba, NEC, Intel and Hitachi, dominated the 1980s by doing everything themselves: design, fabrication, sales, all under one roof. That integration was the source of their strength, right up until it wasn’t. When the industry split in two, with fabless design houses on one side and pure-play foundries on the other, the Japanese giants were too vertically integrated to move. They kept trying to do everything, while their competitors got very good at doing one thing. From roughly half of global production in 1989, Japan’s share fell below 10% by the late 2010s.

The opening that Japan left was widened, ironically, by American trade policy. American chipmakers accused Japanese firms of dumping memory chips below cost, and under threat of sanctions Japan signed the 1986 US-Japan Semiconductor Trade Agreement, agreeing to two things: its firms would stop selling chips below set floor prices, not just in America but in every market, and Japan would open its home market to foreign chips. (The famous tariffs came a year later, when the US judged Japan non-compliant and briefly imposed 100% duties on $300m of Japanese televisions, laptops, and power tools; a punishment aimed at Japanese electronics generally, not at chips themselves.)

With Japanese memory held artificially expensive everywhere in the world, Korean and Taiwanese entrants, who were not party to the agreement, could sell under that umbrella at prices Japan was now forbidden to match: Korea’s share of the DRAM market climbed from 5% to 40% in the years that followed, while Japan’s collapsed from 75% to 20%.

Taiwan took that room deliberately, with state money: in 1987 the government put around $100m into a new company spun out of its Industrial Technology Research Institute, run by Morris Chang, a Texas Instruments veteran. Chang’s bet was that manufacturing chips was every bit as hard, and as valuable, as designing them, and that a company that only manufactured, and kept its customers’ secrets religiously, could become indispensable to everyone at once. And he was right: TSMC made the foundry itself the product and won on manufacturing alone.

Korea did the same in memory. Samsung and what became SK Hynix built enormous fabs on government backing and flooded the DRAM market that Japan had pioneered. Today those two firms hold more than half the world’s DRAM and NAND, and dominate the high-bandwidth memory that modern AI is bottlenecked on. All of this being the product of industrial policy sustained across decades and business cycles.

And then there’s the American side of the ledger, where the story is one of walking away, and it is best told through one company. Intel was, for a generation, the most advanced manufacturer on earth, and the great exception to every rule in this piece. It stayed vertically integrated long after everyone else split apart, and under Andy Grove, whose management creed was literally “only the paranoid survive,” that integration was a weapon: through the 1990s and 2000s the x86 monopoly in PCs and then servers funded fabs that ran a full process generation ahead of the field, and Intel’s “tick-tock” cadence of alternating process and architecture upgrades effectively set the industry’s clock.

But in the mid-2000s Intel passed on supplying the processor for the original iPhone, with then-CEO Paul Otellini later admitting the forecast volumes looked too small to justify the price, and so it sat out the mobile revolution that went on to fund its rivals. Then came the stumble in manufacturing itself: Intel stumbled on its 10nm process from around 2015 and stayed stuck on 14nm for years, in part because it refused to invest early in exactly those EUV machines from Veldhoven that TSMC embraced without hesitation. While Intel wrestled with its own fabs, the fabless designers simply rented TSMC’s leading edge and pulled ahead: AMD, written off a decade earlier, took a large and growing share of the server CPU market under Lisa Su with TSMC-made chips, and Nvidia, which has never owned a fab in its life, took the AI GPU market almost in its entirety.

Line all of those dominos up and a pattern emerges: on one side of the Pacific, states treated chip-making as strategic, and paid for decades to keep it. On the other, firms treated it as a business like any other, optimized for margin and specialization, and offshored the hard, capital-heavy, unglamorous part. Both sets of decisions were defensible, and together they produced a world in which the most important physical technology of the century sits, at its most advanced tip, in just a handful of postcodes.

Why this is suddenly everyone’s problem

For most of the past thirty years, this concentration was an efficiency story, and a good one. Specialization made chips cheaper and better, faster than anyone had a right to expect, and the fragility was a footnote.

AI is what turned the footnote into the headline: of the three inputs to AI progress, data, algorithms, and compute, it’s compute that’s done the heavy lifting, and by some estimates its contribution dwarfs the rest. That means the frontier of arguably the most strategically consequential technology of our time is gated by a physical pipeline that runs Netherlands to Taiwan to Korea to Japan, and barely touches the countries doing most of the model-building.

This is the part of “AI sovereignty” that tends to get lost. The phrase is usually taken to mean something about models: can a country train and run its own systems without asking permission? That matters of course, but it’s the easy half; the harder half is physical. It doesn’t help to own the model if the compute it runs on is fabricated entirely inside the potential blast radius of someone else’s geopolitics. A blockade in one strait, an incident in one Dutch town, a decision by one government about who its national champion may sell to: any of these can reach into a data center in London or Virginia and switch off the future. Sovereignty over bits is not worth much without some sovereignty over atoms.

It’s worth being concrete about this worst case scenario, because the instinct is to just picture a shortage, something the world absorbs and grinds through the way it did in 2021. When a drought in Taiwan and a fire at a single Japanese plant tightened chip supply that year, the disruption was partial and temporary, yet it still cost the global auto industry around $210bn in lost sales, because there was no spare capacity and no inventory anywhere to absorb the shock. A conflict over Taiwan would take the world’s supply of advanced chips to effectively zero, and hold it there. A leading-edge plant depends on an unbroken drip of servicing from the Netherlands, chemicals and wafers from Japan, and the tacit knowledge of tens of thousands of engineers, none of which are available during a war.

Play that out and the damage doesn’t stay in AI, or even in tech. Bloomberg Economics has modeled a war over Taiwan at roughly $10 trillion. That’s close to 10% of global GDP in the first year alone, a larger shock than the pandemic, the financial crisis, or the war in Ukraine. Their figure climbs toward 14% if the world turns out to be worse than expected at substituting the missing chips.

That’s why the physical supply of compute has quietly become one of the central vulnerabilities of the world economy, and the single central vulnerability of AI.

What can actually be done now?

The reason advanced manufacturing concentrates is that it’s punishingly hard and slow to stand up: a leading-edge fab is one of the most complex things people build, and the knowledge to run one is accumulated the expensive way, over many years of trial and error. Anything that compresses that curve (shortening the time and lowering the cost of bringing a new manufacturing capability online) buys back resilience. There is potential for AI to speed up the learning curve of new industrial facilities, so that this tacit knowledge is generated faster and hard physical capability becomes feasible to build in more than one place.

But that will only ever be part of the story, because the history above argues against a purely technological fix. What created this concentration was policy, capital and time: with governments who were willing to fund a strategic industry through decades of losses. Reversing our current situation would need the same things in place.

Thankfully, that seems to be beginning to happen. The CHIPS Act of 2022 put $52bn on the table, enabling TSMC to start making advanced chips in Arizona. A few years ago that looked impossible, so I’m pleased to see genuine progress. But it’s also years behind schedule and far over budget, precisely because a fab is not a software problem and cannot be treated as one. AI can help you climb the learning curve, but it cannot vote the subsidies, resolve the labor shortages, or conjure thirty years of accumulated process knowledge that, in some cases, was never even written down.

So the honest version is this: AI is likely a necessary ingredient of a more resilient supply of compute. It shortens the odds; it does not change the game on its own.

And that brings me back to where we started. What strikes me most about the geography of compute is how little of it was lost in a fair fight. For the most part the West had the technology and chose, rationally, incrementally, one sensible quarter at a time, to let it go somewhere else. In some ways that is good news: if the constraint was a set of choices, it can be chosen differently, though the fix will look a lot like the thing that built East Asia’s lead in the first place (patient capital, industrial policy, and a long time horizon). The wider point is one that hardly needs arguing anymore. In a more volatile world, ownership of core technology has stopped being a preference and become a non-negotiable, for AI and for everything downstream of it. The compute has to come from somewhere, and the countries that want a say in their own technological future will have to make some of it themselves

.

When people invest in these companies they are making a speculative technology bet: either the frontier labs lower their prices for their best models, or customers will be happy with something worse but cheaper.

What happens if neither of these things come to pass?

These businesses will have to raise prices in order to become profitable. I think they’ll be able to do this. But the margins will be thin - I don’t know anyone who is willing to pay double for Cursor over Claude Code.

So we may end up with something that looks like a standard retail business: buy some widgets from a manufacturer and sell them on, repackaged or with a better shopping UX.

I don’t mean to trivialise these businesses. Retailers have moats and accrue value. Some have very differentiated user experience - just try buying an iPhone from an Apple store vs your local department store. It’s just that these moats don’t allow you to make 80% margins like the best hardware or software businesses.

So, you’ll end up with quite large and profitable wrappers. But they’ll look more like retailers or distributors, and be valued as such when the noise dies down.

And we’ll end up with the foundation model companies, with 50%+ margins, being the final winners.

Does this look like a bubble? I don’t think so. The amount of intelligence being consumed in this scenario doesn’t change much from the bull case - we’ll still need the chips, compute, and frontier models - it’s just that the majority of the profit will accrue to a few companies. Startup investors in these “retail” platforms will lose lots of money, but VC is still a pretty small industry and these companies are private - hardly a recipe for a bubble.

They winners probably already exist and were founded before ChatGPT with the exception of xAI. I can’t see many others accessing the capital required to build a competitive model at this point.

Full disclosure: I run a company making AI-designed hardware & advanced materials for data centers.

Six years is a long time in technology, especially AI. But a think piece by Benedict Evans in 2018 contains a lot of wisdom that still rings true. Even the tagline feels apposite in 2024, if you replace “AI” with “Chatbots”:

Everyone has heard of machine learning now, and every big company is working on projects around ‘AI’. We know this is a Next Big Thing. But we don’t yet have a settled sense of quite what machine learning means - what it will mean for tech companies or for companies in the broader economy, how to think structurally about what new things it could enable, and what important problems it might actually be able to solve.

There was one particular insight that stuck in my memory and is worth revisiting today. Ben suggests a surprising, but very helpful way to think about AI — relational databases:

Why relational databases? They were a new fundamental enabling layer that changed what computing could do. Before relational databases appeared in the late 1970s, if you wanted your database to show you, say, 'all customers who bought this product and live in this city', that would generally need a custom engineering project.

This analogy is more true than ever today. Language models are an enabling layer that change what, and how, we use computers. Like databases, most if not all software products to be using a language model within the next 5 years.

Is there anything we can learn about language models the way the database market has evolved and matured?

One very active debate at the moment is whether most companies will use a free and transparent model (a Mixtral or Llama), or a closed source model (a ChatGPT or Gemini).

Here parallels with databases seem striking. Databases, like LLMs, deal with a companies’ most valuable data. This privileged level of access requires additional security and compliance hurdles, potentially leading to an advantage for free and transparent models where weights and serving code can be externally analysed and verified. A quote from Mistral’s Chief Business Officer seems to support this hypothesis:

Open-source models were particularly attractive to state-owned or highly regulated entities, such as defence companies or banks, who wanted to experiment with generative AI but could not do it with proprietary software because of compliance reasons, Bressand said.

Given these parallels you might expect to see open source databases have a larger market share than closed source alternatives. It’s striking, however, that the total market cap of 3 of the largest open source database providers (MongoDB, Elastic) is around $50bn - but Snowflake, a company of a similar vintage, has a market cap of $70bn. This leaves out the major closed source database providers - Oracle, Google, AWS, Microsoft.

How much this analysis transfers to LLMs is debatable, but I think the conclusion is directionally correct — there is a place for the unique selling points of open source, but ultimately quality and cost are primary determinants of market share. It may be that the company who can offer the best service is ultimately also an open source company - but to suggest a causal link is, I think, wrong.

At the moment, it is only the closed sourced providers who have the financial firepower to make billions of dollars in losses while training and serving the best LLMs. This observation leads me to a slightly strange conclusion: no matter how crazy some of the venture rounds in LLM foundation startups seem to be, perhaps, it’s still too small.

If you enjoyed this post, please consider following me on twitter.

.

My personal guiding belief is that the most powerful technologies in the world should be applied to the most important problems we face. At the time AlphaFold was released I had just started work on applying large scale machine learning to what I believe is the foundational physical science problem of the 21st century, known as “Materials Science”.

Materials Science encompasses the design of new batteries, materials that capture CO2, the chemistries that can produce sustainable, carbon neutral fuels and the processes by which we convert biomass into natural chemicals. Versions of these products power the world economy but require dramatic innovation - by itself, the global market for aviation fuel is projected to be ~$650bn in 2030, and we need a sustainable and, crucially, affordable alternative. 

The progress over the past few years in AI for materials has been breathtaking, as we and the wider community started taking the principles from models like AlphaFold, ChatGPT & Stable Diffusion and adapting them for our uses. It’s often felt to me that decades of progress was happening in a matter of months. And it seemed almost impossible to me that incumbent companies were going to leverage AI with the speed they needed to - we’ve heard stories of large chemistry companies with a single computational materials scientist.

It’s this trajectory, and the urgency of the problems we face, that led me and the rest of the founding team to start Orbital Materials - an AI-first chemistry & materials company, building upon the incredible work in generative AI over the past 24 months.

I’m incredibly excited that other people share this vision, including the extraordinary individuals who have joined us since starting late last year - people who have done foundational work on AI for language, generative models, and AI for chemistry, but also people who have spent decades at the coal face of bringing these products to market. I think this is a testament to the ambition of what we're building. And, last week, we had our first public coverage in Bloomberg.

You’ll find us in the lab, discovering industry-defining technologies in areas like Clean Air (e.g. CO2 capture), Clean Energy & Natural Chemicals (e.g. Sustainable Aviation Fuel) and Clean Water (e.g. removing forever chemicals from our drinking water). Please reach out if you'd like to learn more.

As someone who uses AI to explore new materials (think semiconductors, carbon capture, and the like), I’m especially excited about the possibilities here. It's worth noting that while computer-aided design has revolutionized many fields, materials science has remained largely resistant to these kinds of advances. We don’t, for example, design new batteries on a computer the way we design an aeroplane wing. But what if a foundation model could change that? Perhaps it could help us design new computer chips, batteries, and climate technologies at an unprecedented pace.

Of course, this is easier said than done. Language models like ChatGPT are designed to model languages, not the physical world. They can't predict the results of physics simulations, for instance - try asking GPT-4 to simulate three particles according to Newtonian physics, and you'll get an incorrect answer (and one that takes much longer than it should to produce).

So what's the solution? We should take the idea of "foundation modeling" - using large generative models trained on lots of data - and adapt it to scientific computing. For materials science, this could mean a foundation model that could be used to simulate a wide range of materials with high fidelity.

As a generative model, it may also be capable of ‘inverse design’ - as ChatGPT can generate a poem with a particular rhyming scheme, generating new materials based on desired properties, such as conductivity, or hardness, as well as optimising existing materials to improve their performance.

This approach has the potential to revolutionize materials science. Rather than relying on trial-and-error experimentation, scientists would use a foundation model to create materials that meet their specific needs. With the ability to simulate a wide range of materials, a foundation model for materials science could open up new possibilities for discovery and innovation.

The major breakthroughs in foundation modelling over the past year have broadly been the result of pairing a scaleable, reliable model architecture with the correct generative modelling task. Once these have been identified, they must then be engineered to an exacting standard. Building a foundation model for materials science requires finding the right architecture for the job, as well as the right generative modeling task. 

For large language models, this was pairing the transformer architecture with the task of predicting the next word in a sequence. A huge variety of knowledge about the world can be encoded in this way  - and transformers have a very well documented ability to continue to absorb as much data as you can provide.

For images, it was pairing convolutional neural networks (CNNs) with diffusion modelling. For reasons that are yet to be fully understood, CNNs perform extraordinarily well when trained with a denoising diffusion loss.

A combination that would result in a breakthrough is starting to emerge for materials science. Graph neural networks (GNNs) seem to be the correct architecture for materials science, but we have not had that ‘break out’ moment where, as we add more data, performance keeps getting better. Diffusion models have some nice properties, but will surely need adapting to be more physically informed if they're going to be used for generative modelling in materials science. And then, the engineering is different for graphs and will need innovation and dedication.

But even once we have the right architecture and task, there's still a lot of work to be done. The bar for transformative impact in materials science is much higher than it is in other fields. A foundation model for materials science will need to be able to design molecules and materials that actually work in the lab, not just perform well on computational benchmarks. Surely this is a more worthwhile thing to work on for startups than thin layers on top of OpenAI APIs.

Many of the grand challenges in AI have fallen in recent years, faster than certainly I expected. The challenges laid out above are certainly very difficult, but I believe such a system could be achieved with the right team and dedication. If we can do that, we could meaningfully accelerate scientific progress in atoms, not just bits.

In three short years OpenAI has released GPT3, Dalle-2 and ChatGPT — a stunning set of products that have reframed what many believe is possible with machine learning. ChatGPT has persuaded a large fraction of the world that Artificial General Intelligence (AGI) is a matter of how long, not a matter of feasibility - an extraordinary accomplishment.

You’d be forgiven for having deja vu here - in early 2020 you may have read a similar introduction, but describing DeepMind and the extraordinary successes of AlphaGo, AlphaZero, AlphaStar and AlphaFold. What happened? It’s not so obvious. DeepMind and OpenAI’s similarities are stronger than you might think - both are funded by Big Tech motherships, are of a more comparable size than many appreciate, and have charismatic, formidable, techno-utopian leaders.

Trying to answer the question “Why didn’t DeepMind initiate and deliver GPT3?”[0] is one way of shedding light on this puzzle. I say specifically GPT3 because that was the significant innovation — we’ve been following a playbook since then, and most of the perceived advantage of OpenAI stems primarily from how fast they ship, and their appetite for it, not from the pace of discovery.

As someone professionally interested in how you build extraordinary scientific teams, there are three things that strike me quite profoundly about GPT3.

The first is that there is no real evaluation metric or target for GPT3. Nothing was “solved” when GPT3 was released, in the way that Go or Protein Folding was “solved”. Nobody knew in advance how long you’d have to train GPT3 before it would start to count, and the eerie experience of interacting with GPT3 is not in any way captured by question answering benchmarks. This lack of easily quantifiable measurement is striking in its departure from previous grand challenges in AI.

The second is that there are comparatively few people with traditional elite academic machine learning backgrounds in the GPT3 author list (PHDs in machine learning, people with many thousands first or last author papers) - an organisational departure from the prevailing wisdom on how to build teams in AGI.

The third is the scale of organisational-level risk taking involved in building GPT3. It seems obvious now, but it was in no way clear in 2019 that reducing the language modelling loss on the whole of the internet would lead to the amazing properties we see in large language models. There was significant risk it wouldn’t work out, and the costs - opportunity and financial - to OpenAI would have been significant.

These points are related. They stem from strong organisational, almost philosophical, differences. OpenAI is an exceptionally engineering focused research company, concerned first and foremost on how to build systems that appear to have intelligence when interacted with. This stands in stark contrast to most academic machine learning that is focused more on algorithmic understanding than system performance. Engineering focused papers often have a hard time getting into conferences, with reviewers saying “clear reject” because of “lack of novelty” — it’s “just engineering” after all.

This focus on building systems, not discovering algorithms, leads to technology that looks more like distributed systems software than traditional machine learning. And, consequently, built & led by a team that reflects the different skills and attitude that requires. It’s very unlikely that an organisation whose core cultural metric is elite academic venue papers would invest very significant resources into a project where the route to publication in Nature/Science etc. may be unclear.

OpenAI has realised that the race for perceived leadership in AGI is not run in the most highly cited academic journals in the world. It’s run in the subjective experiences of the users of AI — the use of AI as a product. Nothing has done more for persuading the world AGI is round the corner than ChatGPT, and it is no coincidence that OpenAI is majority led by people whose professional careers have been focused on shipping products. But, the strategy of shipping fast and frequently has its downsides - it’s a little chilling to see the extreme financial pressure to release powerful technology on the world prematurely, as in the case of ChatGPT (and similar)’s rushed integration into Bing.

Finally, in 2019 OpenAI had something to prove. They were commonly viewed as a company without clear focus. Now the shoe is on the other foot, DeepMind (and Google) have to respond.

EDIT: An earlier version of this article had an incorrect statement about GPT3 getting into major ml conferences.

[0] In attempting to answer this question I’m not primarily interested in whether DeepMind had the technical ability or resources to build and serve large language models - clearly they did and still do. My former DeepMind colleagues are extraordinarily talented and not to be underestimated. The race isn’t won yet.

Subscribe now