Why Compute Is the New Oil

Models are converging into commodities. The actual fight is over compute, electricity, and the rare metals that make both possible. Whoever wins the physical stack, wins the whole bag.

The most important fact about the artificial intelligence race in 2026 is that the models themselves are becoming commodities.

A Chinese lab released a model in 2025 that mimicked the architecture and behavior of one of Anthropic’s frontier systems closely enough that researchers noticed within days.

Open-weight models from labs in France, China, and the United States now sit within months of the proprietary frontier on most benchmarks. The agent layers that sit on top of these models are converging too. The same patterns appear in every research paper, the same training tricks, the same scaffolds, the same evaluation protocols.

This is the shape of the technology itself, so no AI company has an actual ‘moat’, if we’re just talking about the models.

Software, given enough time, gets copied. The first generation of any breakthrough is rare. The fifth generation is everywhere. Anyone tracking the history of databases, web browsers, search engines, mobile operating systems, or cloud infrastructure has seen this pattern play out. The novel becomes standard. The standard becomes free. The free becomes a commodity.

The artificial intelligence race is collapsing into the same pattern at a speed nobody expected.

That sounds like bad news for whoever is currently leading the model layer, but in fact, it is just a relocation of the real prize.

If models stop being the differentiator, something else has to be. And the something else is already visible to anyone willing to look one layer down.

The race is for compute. The compute is built on chips. The chips depend on rare metals. All of it runs on electricity.

Whoever controls the physical stack controls the next twenty years of artificial intelligence, and through it, the next twenty years of national power.

Why models are becoming commodities

A frontier model in 2026 looks roughly the same as a frontier model from 2027 will look, just somewhat better. The architectures are converging. The training recipes are leaking. The talent is mobile. A senior researcher at a top lab can move to a competitor and bring most of what they know with them, regardless of non-disclosure agreements.

Ideas do not stay in safes.

What does not move easily is what built the model in the first place: tens of thousands of advanced chips, gigawatts of electricity, billions of dollars in capital, and the supply chains that fed all three.

Picture two countries trying to build the same skyscraper. Both have the blueprints. Both have the architects. One has steel mills, copper smelters, cement factories, and a working power grid. The other has none of those things and has to import them under sanctions.

The blueprints are equal, but the ability to build is not.

The blueprints are AI models. Having them is great, but the actual race is much more physical.

This is why the most consequential decisions of the next decade are no longer being made in research papers. They are being made in trade negotiations, mining rights, electricity contracts, and chip export licenses.

Compute is the new oil

For most of the twentieth century, oil determined which countries had freedom of action. A nation that controlled enough oil could feed its military, run its economy, and project power abroad. A nation that had to import oil was always one embargo away from collapse. The Cold War, the Gulf wars, the relationship between Saudi Arabia and Washington, the rise of Russia as a petrostate, the structural fragility of post-1979 Iran. Oil was not the only factor in any of those stories, but it was always present, and often decisive.

Compute is becoming the same kind of strategic resource for the twenty-first century.

A nation with abundant compute can train better models, run more agents, simulate more scenarios, predict more events, and automate more decisions than a nation without it. The advantage compounds. A government that uses artificial intelligence to allocate capital, target weapons, gather intelligence, and run logistics better than its rival will outpace that rival by orders of magnitude over a decade, not percentages over a year.

This is why the largest AI infrastructure projects on Earth are being announced at a cadence that would have looked absurd two years ago. A single campus under construction in the Persian Gulf is rated for several gigawatts of power.

The combined American buildout, often referred to under the umbrella of the Stargate program, is targeting hundreds of billions of dollars in compute infrastructure over a five year window. Chinese state planners are matching the announcements with their own.

Compute is now treated by every major government as critical national infrastructure, in the same legal and regulatory category as power generation, telecommunications, and weapons production. The category change is recent. The implications are still being worked out.

The chip layer

A modern artificial intelligence chip is the most complex object humans manufacture at scale. The advanced graphics processors that train and run frontier models contain transistors at sizes measured in single-digit nanometers, smaller than the wavelength of visible light. A single such chip requires hundreds of manufacturing steps using equipment built by perhaps a dozen specialized firms in the world.

The most consequential of those firms sits in the Netherlands and makes the lithography machines that print the smallest features. The most consequential foundry sits in Taiwan and runs the bulk of the world’s advanced chip production. The most consequential designer of artificial intelligence chips sits in California and depends entirely on the foundry in Taiwan to actually manufacture its products.

Three countries hold the keys to the entire frontier. Pull any one of them out of the chain and the global frontier of artificial intelligence stops moving.

This is the geometry that makes Taiwan the single most important geopolitical pressure point on Earth in 2026.

It is why the United States has spent the past five years subsidizing the construction of foundry capacity in Arizona and pressuring its allies to host parallel facilities. It is why the Chinese leadership treats reunification with Taiwan not as a sentimental ambition but as a strategic necessity.

The country that controls the foundry controls the chip.

The country that controls the chip controls the model.

The country that controls the model… controls the rest.

The Americans currently hold the design layer and the lithography layer through allies. The Chinese are racing to build a parallel domestic supply chain that does not depend on either. The race is years behind on the most advanced nodes and closing fast on the older ones.

Whoever closes the gap first, changes the entire balance.

The rare earth layer

Underneath the chips is a layer most readers outside the industry have never heard of. It is the layer that makes the chips, the magnets, the lasers, the optical fibers, the radar systems and the electric motors physically possible.

The processed feedstock for that layer is dominated by China to a degree that has no parallel in any other strategic resource.

Roughly four out of every five tons of refined rare earth elements consumed globally pass through Chinese processing facilities. For some specific elements, the Chinese share approaches the totality of world supply. Neodymium and dysprosium, which together make the high-strength permanent magnets inside electric vehicles, wind turbines, and precision-guided weapons, are largely processed in two provinces in southern China. Gallium and germanium, which are essential to advanced chip packaging and certain laser systems, were placed under Chinese export licensing in 2023. The licensing regime has expanded since.

This is not an accident. The Chinese government identified rare earth processing as a strategic priority in the late 1980s and spent thirty years building the infrastructure, while Western governments shut down their own processing facilities under environmental pressure. The resulting asymmetry is the kind of vulnerability that gets noticed only when it is too late to fix in a single political cycle.

The Americans are building processing capacity in Texas. Australia is expanding mining and partial processing. The European Union is funding strategic stockpiles. The Saudis and Emiratis are bidding for ownership stakes in African mineral plays. None of these projects will produce significant volume before 2028. Most will not produce significant volume before 2030.

In the meantime, the country that processes the metals can also withhold them. Every advanced chip, every electric vehicle and every guided missile sits downstream of that decision.

The energy layer

The final layer is the one that makes everything else possible: electricity.

A modern data center campus capable of training frontier models draws somewhere between several hundred megawatts and several gigawatts of continuous power. A single gigawatt is roughly the output of a large nuclear reactor, or the electrical demand of a city of a million people. The largest campuses currently announced will draw the equivalent of medium-sized nations.

The electricity has to come from somewhere, and it has to be cheap, abundant, and reliable. Intermittent sources like solar and wind, on their own, do not work for training runs that have to operate continuously for months. Batteries help but cannot bridge the gap at the required scale. The realistic options are nuclear, natural gas, hydroelectric, and in some places geothermal.

This is why the most aggressive infrastructure buildouts are happening in three categories of countries. The first is the United States, which is reopening retired nuclear plants, fast-tracking new natural gas turbines, and approving small modular reactor designs at a pace not seen in decades. The second is the Persian Gulf, where natural gas is abundant, sovereign capital is patient, and regulatory approvals can happen in months rather than years. The third is China, which is building nuclear capacity faster than the rest of the world combined and treats grid reliability as a national security matter.

The countries that cannot bring power online quickly will rent compute from the countries that can. The rent is not just financial. It is informational, strategic, and political.

What an AI-enhanced government actually looks like

The phrase “artificial intelligence” still suggests, to most people, a chatbot. The actual capabilities being deployed at scale by serious governments are something else.

Military targeting systems are using machine learning to identify, track, and prioritize objectives in real time across satellite, radio, and signals intelligence feeds. The window between detection and engagement is collapsing from minutes to seconds.

Economic forecasting and sanctions modeling departments inside major treasuries are running large language models against entire archives of trade data, shipping records, and financial flows to find evasion patterns that human analysts would never catch.

Foreign ministries are running scenario simulations against summit positions, treaty texts, and rival statements to anticipate counterparty moves before they happen.

Domestic security services are running facial recognition, gait recognition, and language analysis against population-scale data to track individuals of interest.

Information operations groups are generating synthetic content, translating it into dozens of languages, and distributing it through automated networks at a volume that drowns out organic communication on contested topics.

These capabilities scale with compute. A government with twice the compute can run twice the simulations, train twice the models, surveil twice the population, and generate twice the content. The advantages compound across decades.

The countries that fall behind on compute will not lose dramatically in any single moment. They will lose slowly, in every negotiation, every market move, every military exchange, every information cycle. The accumulated disadvantage is what matters. Once it sets in, it is almost impossible to reverse inside a single political generation, because the compute gap that produced it has only widened in the time it would take to respond.

The map

The current race breaks roughly along the following lines.

The United States holds the deepest model labs, the most advanced chip designs, the largest pool of patient capital, and a network of allies who control the most critical foundry capacity in Taiwan, the most critical lithography supplier in the Netherlands, and significant portions of the rare earth mining base in Australia and Canada. It does not have processing capacity for most rare earths. It does not have abundant cheap electricity in every region where compute is being built. It depends on Taiwan in ways that no responsible national security planner would accept if there were any alternative.

China holds the processing capacity for nearly every critical metal, an unmatched manufacturing base for the components that surround the chips, the sovereign ability to coordinate buildout across decades, and the fastest electricity generation expansion of any major economy. It does not yet have the most advanced chip designs or the lithography to produce them domestically. Its model labs have closed most of the gap on the open-weight frontier and are still chasing on the proprietary frontier, but the gap is measured in months now, not years.

The Persian Gulf states, especially the United Arab Emirates and Saudi Arabia, hold a unique position. They have abundant cheap energy, sovereign capital, neutral political alignment between the two giants, and a willingness to host infrastructure both sides need. Their bet is that hosting the physical layer makes them indispensable to whoever wins the software layer. So far the bet is paying off.

Europe has the lithography, partial chip design talent, regulatory weight, and almost none of the energy or processing capacity. Its current trajectory points toward becoming a customer rather than a player.

Russia has energy, raw materials, and a small but real research base. It has been excluded from the most advanced chip supply chains and is unlikely to close that gap inside the relevant window.

Every other country either rents compute from one of the above or accepts that it will not be a serious actor in this race.

The conclusion

The race the public sees is the race for better models. Each new release is announced with fanfare, scored against benchmarks, and celebrated or criticized depending on the audience. That race is real, and it matters, but it is also closing on itself. The models are becoming commodities. The performance gaps are narrowing. The architectures are converging. Five years from now the leaderboard at the top of the open-weight rankings will be crowded, and the difference between the best proprietary model and the best open one will be small enough to argue about.

The race that decides the next twenty years is happening underneath. It is being fought over chips, the metals that build the chips, the machines that print the chips, the foundries that manufacture them, and the gigawatts of continuous electricity that the chips require to do anything useful.

The countries that solve the physical stack will set the pace of artificial intelligence development, deploy the most capable systems across their militaries and economies, and exert the strongest influence on every other country that needs to rent what they have built. The countries that fail to solve the physical stack will become customers, dependents, and eventually clients of the ones that did.

This is not a software race. It never was. It is a chip race, a metals race, an electricity race, and a logistics race. The visible competition over models is the surface phenomenon. The structure of the next world order is being decided one substation, one foundry, and one processing plant at a time.

Whoever moves first, builds fastest, and locks in the supply chains wins. Whoever waits loses, regardless of how good their software is.

The window to choose which side of that line a country, a company, or a portfolio ends up on is open right now.

It will not stay open for long.

A note for institutional readers.

If you’re reading this professionally, at a sovereign fund, a family office, an investment desk, a policy shop, or a defense or intelligence team, what you’re getting here is the public-tier version of work that can be produced privately, sharper and tailored to a specific mandate.

Private engagements are taken on a selective basis. If your team would benefit from custom synthesis on the geopolitics-energy-AI stack, leave your details here: https://forms.gle/1asdUqUfp5fZzus8A

No follow-up unless you request one in the form itself.