The GPU Access Problem NVIDIA Can’t Solve Alone

_{Guest contributor: Gaurav Sharma is CEO of io.net, focused on AI and decentralized infrastructure. He previously led infrastructure teams at Amazon, Binance, and Agoda.}

In July, NVIDIA launched a revenue-sharing program that lets AI startups trade a cut of their future earnings for access to GPU compute they cannot otherwise afford. The move was widely praised as a step towards democratizing AI infrastructure. It was also an admission from the biggest chipmaker that the companies trying to build on its technology cannot get to it.

Within weeks, NVIDIA reportedly paused parts of the program over concerns it could invite antitrust scrutiny - a sign that even the company behind the initiative recognizes how much control it would concentrate in one place.

The reason why? NVIDIA disclosed $1 trillion in locked orders for its Blackwell and Vera Rubin chips at GTC 2026, with capacity booked through 2027. Hyperscalers and frontier AI labs have secured the bulk of that pipeline years in advance. By the time the rest of the world shows up, the interesting startups, the cutting-edge researchers, teams with strong ideas and real funding, the good capacity is gone and what remains is overpriced.

While all that is happening, a meaningful share of GPU capacity sits idle across data centers that have no efficient way to pool or resell it.

The word fix has two meanings, and NVIDIA is caught between both of them. It helped fix the market in the match-fixing sense, through the bulk supply deals that gave hyperscalers priority access to its chips and left everyone else competing for scraps. Now it is trying to fix the market in the repair sense, while collecting from the arrangement that broke it. There is a reason this program feels like a pharmaceutical company selling you a partial cure to an illness it helped cause.

Here is how it works. NVIDIA helps cloud partners finance hardware in exchange for a share of the cloud revenue that hardware generates. Those cloud partners then sell compute to startups and smaller companies. So NVIDIA earns its standard product revenue on the chips, takes a recurring share of the income those chips produce, and the startup at the end of the chain is still going through a middleman. The only difference is that now NVIDIA is also clipping the ticket.

The deeper issue is that none of this changes who decides where compute goes. Access still flows through a small number of operators, still depends on relationships, and still leaves most of the market negotiating over leftovers.

This is not the first time a critical resource has been locked up by a handful of incumbents, and it is not the first time a correction has followed.

Before deregulation in the 1990s, a handful of telecom carriers controlled all access to communications infrastructure. The rise of virtual network operators, companies that did not own towers but resold capacity across multiple carriers, broke that open. Cloud computing did something similar a decade later: it took servers out of basements and turned them into something anyone could rent by the hour. Airbnb, Slack, Spotify, and Instagram could not have existed if they had each needed to build their own data centers.

But the cloud, like every "democratized" industry before it, has drifted back towards concentration. AWS, Azure, and Google Cloud now control roughly two thirds of global cloud infrastructure, and GPU allocation within it follows relationships and contract size rather than open bidding.

So what does the correction look like this time? What the industry needs most is something like a clearing house for compute. A layer of companies sitting between the large cloud providers and the rest of the world, whose function is simple: find GPUs that exist but sit underused, aggregate them, and make them available without requiring a multi-year commitment or a hyperscaler relationship. Some of this supply sits in centralized data centers. Some of it is distributed across decentralized networks. The best approaches mix both.

Stripe made banking infrastructure accessible to anyone who could write a few lines of code. Virtual network operators made telecom capacity available to customers the carriers were not interested in serving. A compute access layer would work the same way.

The real cost of a locked-up compute market is paid for by companies that never get built. When a founder has to plan a roadmap around whatever compute they can afford rather than the strongest technical bet, caution wins out.

Research teams design experiments around available compute hours rather than what the science calls for. Ambitious ideas never get funded because everyone in the room already knows the infrastructure costs will strangle the company before the market gets the chance to.

NVIDIA recognizing the access problem is a signal. But the real answer is not one company opening a door and taking a cut every time someone walks through it. It is a market where there are enough doors that nobody has to wait for permission to build.