Nvidia didn't grow into a $5 trillion company earlier this year by selling tchotchkes. Its graphics processing units (GPUs) -- the computer chips that are powering the AI revolution by handling and calculating multiple streams of data simultaneously -- are in high demand. OpenAI, for instance, is estimated to need 4.8 billion hours of access to cutting-edge GPUs next year just to serve its customers' demand as they use ChatGPT, up from an estimated 2.3 billion hours this year, according to Barclays.
Yet Nvidia, which remains largely the only game in town when it comes to chip supply, isn't able to meet demand at the scale the market wants. "The clouds are sold out," explained Nvidia chief financial officer Colette Kress on the company's recent earnings call, with GPU output "fully utilized". Those who can get their hands on GPUs are having to pay high prices for them: the AI computer chip market is set to hit $286 billion next year, according to analysts Omdia, up from $207 billion in 2025, and $123 billion in 2024.
Startups are looking to get their hands on GPU time any way possible -- including by hiring access hours to harness their capabilities. "The demand is crazy," says Greg Osuri, founder of Akash Network, a peer-to-peer marketplace for underused GPU supply to be rented out to those who want it, often startups seeking the ability to capitalize on AI inference (or live queries) or training, but who can't afford, or can't wait, for GPUs of their own.
It's a burgeoning industry with plenty of players. Vast.ai aggregates large fleets of third-party GPUs and markets them directly for AI workloads, letting customers pick specific architectures for model training or multimodal inference. Render Network, a company previously specialized in cloud infrastructure for 3D rendering, is now exploring offering enterprise grade AI access via the community's RNP-021 proposal.
One of the biggest is CoreWeave, which saw three-year growth of 5,896 percent to the end of 2024, according to prior Inc. reporting. That's despite the company losing $863 million in 2024, as well as being $8 billion in debt
But that's just the tip of the iceberg: io.net positions its pooled capacity explicitly for training, fine-tuning, and model serving, while Aethir is aiming to serve AI inference at the edge as well as in data centre deployments.
There are other companies that capitalize on the demand for access to AI chips, even if they don't own them themselves. They include Featherless AI, which gives users access to more than 12,000 AI models as an inference provider. "Inference providers, generally speaking, they're a kind of cloud service provider," says Wes George, chief operations officer at Featherless AI. Those inference providers sit between GPUs and the applications -- and businesses that run them. "We help applications run AI models, and we take a different approach than most inference providers, in that we can run models at incredibly small scales yet efficiently."
George also admits that there are "GPU constraints" in the market, with the largest companies unable to fill their demand -- but he stresses that the squeeze is uneven, and that "in the last year and a half, there was always inventory for small scale experimentation," even if smaller users have to tolerate quirkier community clouds, older hardware and the occasional bout of flaky infrastructure.
The GPU rental market itself is still growing. Market researchers estimate that "GPU as a service" will expand from low-single-digit billions of dollars in annual revenue today to tens of billions by 2030, as AI workloads move off-premise and more firms choose to rent rather than own capacity. Osuri says Akash's "utilization right now was 70 percent which is very healthy," with room for new suppliers to plug in idle cards, while George argues there will "always be some rental inventory" -- but cautions that entrepreneurs entering the space need to be ready for patchy availability, high capital costs, and customers who want them to take away the issue of juggling last-generation and cutting-edge GPUs scattered across multiple providers.
Given those costs and issues, renting tends to make most sense for startups testing ideas, and larger companies that need to go beyond their existing capacity for a specific project. For the former, the issue is simple sticker shock. "This stuff is still very expensive," says George, pointing to the costs of renting an Nvidia A100 GPU at $150 an hour. If you're doing a training run, you're probably going to do about a node of eight," he adds. Run over 24 hours, the cost quickly gets into the thousands.
Entrepreneurs eyeing up the GPU rental space need to recognize they're selling raw GPU time and predictability, with customers not necessarily whether their workload is running on H100s in a tier-one cloud or last-generation tech in someone's basement.
Osuri's bet is that, over time, that "basement" might literally be your neighbor's house. He's been pitching lawmakers on a more distributed future for AI, where training jobs can spill out of giant data centres and tap into idle capacity at the edge, including homes with excess solar. A pilot running across dozens of nodes across the country will train a model, fully distributed, "so it can demonstrate this thing works," he says.
If that vision comes to pass, the GPU rental market could look less like a handful of hyperscale landlords buying into the sector and more like a group of mini-utilities, with everyone from miners to hobbyists plugging their cards into global pools. "Owning a GPU becomes a commodity. Owning a GPU becomes valuable because now it's as valuable as keeping water on your property," says Osuri. He believes that GPU access could become a "substrate of civilization" in the same way AI will itself. "It has to be a utility, like the air we breathe and the water we drink," he says.
The final deadline for the 2026 Inc. Regionals Awards is Friday, December 12, at 11:59 p.m. PT. Apply now.
