In 1901, prospectors struck oil at Spindletop in East Texas so violently it took nine days to cap the well. Within a year, more than 200 oil companies had set up shop nearby. Wildcatters, chancers, people who had no business being in the oil industry were suddenly in the oil industry. That is what a genuinely open resource market looks like at its birth: messy, crowded and full of outsiders.
Compute, the resource that will shape the rest of this century the way oil shaped the last one, is not getting that kind of beginning.
Last week Nvidia CEO Jensen Huang called compute chips an "investable asset class," the first time the world's most valuable chipmaker has spoken about silicon in the language usually reserved for real estate, commodities or gold. The occasion was a reported $500 billion infrastructure financing arrangement, the latest in a series of mega-deals that have turned GPU capacity into something banks now price, structure and trade.
The comparison to oil is tempting, and people at the top of the industry are leaning into it. But there is a key difference: oil built an open economy. Compute is building a closed one.
The 20th-century American economy flourished because oil was accessible. Hundreds of producers competed, while refiners, distributors, toolmakers and service companies grew up around that competition. The resource was valuable, but the market around it was open enough that value spread outward rather than concentrating in a handful of firms. Early Silicon Valley ran on the same principle: the same chips and components were available to two guys in a garage as they were to IBM, and that openness is what let an entire industry grow from nothing.
A market that only rents to landlords
Compute is going the other direction. Microsoft, Google, Meta and Amazon have locked up the overwhelming share of next-generation GPU allocation through 2027, secured through multi-year supply agreements, direct stakes in chip and data centres, and financing arrangements that put entire data centre buildouts on their balance sheets before a single chip ships. They are pre-purchasing capacity that does not yet exist, reserving tomorrow's compute today.
AI labs then lock in their own access through exclusive cloud partnerships with those same hyperscalers. By the time a growth-stage AI company or an ambitious research team goes looking for GPU time, the good capacity is spoken for, the remaining inventory is overpriced, and the leverage sits entirely on the side of Big Tech.
The result is a curious market, with no real shortage of silicon and a meaningful share of global GPU capacity sitting idle across data centres that have no efficient way to pool or resell it. Supply is not the problem. What is missing is access. A startup with funding and a team still cannot get the GPUs it needs because it does not have a relationship with the right cloud provider.
Crypto learned this lesson the hard way
Anyone who has been in crypto long enough will recognise this pattern. The original promise of blockchain was open networks where participation did not require approval from an incumbent. Then liquidity concentrated on a few exchanges, development consolidated around a few chains, and the ecosystem started looking a lot like the financial system it was designed to replace.
In the rush to decentralise everything, people forgot that decentralisation is a tool, not the destination. The most successful projects were the ones that kept asking whether they were actually useful, actually working, and actually giving more people access than before. Stablecoins did not win because they were maximally decentralised. They won because they gave more people access to dollar-denominated transactions than the banking system was willing to.
Compute is at the same fork in the road. Decentralised compute networks have proven that GPUs distributed around the world can be organised into usable AI infrastructure. But to solve the access problem that is holding back the global supply of AI, the move has to go further. Just like web3 gamers do not necessarily care what is going on underneath, a founder trying to train a model or a company running inference at scale just wants to know: can I get reliable access to the compute I need, at a price I can sustain, without signing a multi-year contract with one of four companies?
The missing layer
Locked-up markets do not stay locked up forever. Every time a resource gets concentrated in too few hands, an intermediary eventually shows up that the incumbents did not bother to build.
Cloud computing was exactly this kind of correction. It took servers out of corporate basements and turned them into something anyone could rent by the hour. But cloud has since drifted back towards concentration, with a handful of firms controlling most global data centre capacity and GPU allocation inside it following relationships rather than open bidding.
What is emerging now is a new layer of companies sitting between the hyperscalers and the people who actually need to build. Their job is straightforward: find the compute that already exists but sits underused, whether idle in a data centre running below capacity or committed under a contract nobody is fully drawing down, and route it to the broader market without demanding a multi-year lease to get in the door.
The smartest of these are learning from crypto. Instead of treating decentralisation as an identity, they treat it as one supply source among several, and mix it with centralised infrastructure where reliability demands it. The point is not architecture. The point is whether more people can access compute than could before.
What gets built when access opens up
The real cost of a locked-up compute market is paid by the companies that never get started. When a founder has to plan a roadmap around whatever compute they can secure rather than the strongest technical bet, caution wins. And caution does not build the next breakthrough in protein folding or autonomous systems.
Research teams design experiments around available GPU hours rather than what the science calls for. Entire categories of ambitious, compute-hungry ideas simply never get funded, because everyone in the room already knows the compute costs will strangle the company before the market gets the chance to.
Every prior computing era is remembered for who it let in. Mainframes gave way to minicomputers, minicomputers to PCs, PCs to cloud. Each transition expanded who got to build. If the compute layer underneath AI hardens into a permanent hierarchy of four owners and everyone else as tenants, the next generation of consequential AI companies will be limited to whoever was lucky enough to be there at the start.