Getting your hands on enough cloud GPU capacity to actually train a model has become one of the biggest headaches in AI right now, and Nvidia just made itself the middleman trying to fix that.
The chipmaker announced a new partnership program Thursday linking AI startups directly to its network of cloud service providers, giving smaller firms a faster route into full-stack cloud computing built around Nvidia chips instead of negotiating capacity deals on their own. In exchange, the startups and cloud providers involved will share both product and cloud revenue with Nvidia, positioning the company less as a hardware vendor and more as the broker sitting in the middle of the entire cloud supply chain.
Two Australian firms are kicking things off as the initial cloud compute providers. Sharon AI plans to deploy up to 40,000 Nvidia GPUs across its cloud infrastructure, while Firmus Technologies is building a data center in Batam, Indonesia, that the company expects to scale to 360 megawatts and house as many as 170,000 GPUs. Those numbers give a sense of just how much physical cloud infrastructure sits behind every AI product people take for granted.
The move says a lot about where the real bottleneck in AI development actually is these days. It’s not always the models, it’s the cloud capacity behind them. GPU access has gotten scarce and expensive enough that people compare it to oil, with reports of futures contracts tied to availability. Startups without deep pockets or existing cloud provider relationships have struggled just to get in line for capacity.
That scarcity is pushing more AI companies into unconventional deals to keep their cloud operations funded. OpenAI, for instance, has structured arrangements involving shares and investment from partners including Amazon and AMD, according to CNBC reporting from January. Revenue and equity sharing has become a workaround for companies that need cloud compute now but don’t have the cash to buy capacity outright.
Nvidia isn’t sitting on unlimited resources either. The company said earlier this month it’s looking to raise at least $20 billion in debt, according to sources familiar with the plans, with proceeds going toward general corporate purposes, including refinancing existing obligations tied to its cloud infrastructure buildout.
The company has since corrected its original announcement to clarify which cloud services it’s offering and to which customers, and to confirm Firmus Technologies is based in Australia.
