Nvidia just told the next wave of AI startups they don't need the cash up front. They just need to give up a piece of what they make later.
On July 2, Bloomberg reported that Nvidia is rolling out a revenue-sharing structure aimed at researchers and early AI companies that don't have the capital to build or rent large-scale compute. Instead of billing a customer once for a rack of GPUs, Nvidia now lets AI cloud providers procure its infrastructure, resell that capacity to startups, and hand Nvidia a cut of the resulting cloud revenue on top of the original hardware sale. Colette Kress, Nvidia's chief financial officer, co-authored the announcement on Nvidia's own blog, pointing to a gap the company says it's built to fill: emerging AI companies have historically had limited access to capital-intensive infrastructure.
That's the polite version. The blunt version is that Nvidia would rather own a slice of your revenue for years than lose you to a cheaper chip once you can afford to shop around.
The mechanics run through cloud partners, not directly through a founder's bank account. Sharon AI is deploying up to 40,000 Nvidia Grace Blackwell GB300 GPUs under the new structure. Firmus Technologies is building a DSX AI factory campus in Batam, Indonesia, that Nvidia says will scale to 360 megawatts and 170,000 GPUs. On the customer side, Nvidia named Baseten, Fireworks AI and Together AI as the kind of company the program is built for: AI-native businesses that need training, fine-tuning and inference capacity but have no interest in building a data center to get it. They rent the compute. Nvidia takes a cut of what the cloud provider bills them, for as long as they keep using it.
That's a fundamentally different bet than a one-time GPU sale. It ties Nvidia's revenue to how much a startup actually uses, not to what it could afford to buy on day one.
Part of a much bigger pattern
This isn't Nvidia's first move into owning a piece of its customers' future. The company has already committed more than $40 billion to equity stakes in AI companies so far in 2026, according to CNBC and TechCrunch, led by a roughly $30 billion investment in OpenAI along with multi-billion-dollar positions in Corning and data center operator IREN. Wedbush analyst Matthew Bryson has acknowledged the obvious circularity in all this, capital flowing from Nvidia into the same companies that turn around and buy Nvidia chips, but argues it could build the company a durable moat if it works. Tom's Hardware has been blunter, describing Nvidia as turning GPUs into capital and warning that AI companies are increasingly financing hardware the way they would finance debt.
The revenue-share model announced this week applies the same instinct one rung down the ladder, to companies too small or too early to be worth an equity check.
There's also a defensive logic here, and Nvidia isn't hiding it. Custom silicon is getting real. Google's TPUs already run a meaningful share of internal AI workloads, and chip startups are no longer theoretical threats. Etched, the transformer-specific chip maker, hit a $5 billion valuation and $1 billion in system orders this week, according to TechCrunch. Nvidia's own $20 billion deal to license Groq's technology and bring over most of its team late last year was as much about neutralizing a rival as it was about acquiring anything. A startup that can defer its GPU bill by handing Nvidia a cut of future revenue has a lot less reason to spend a year evaluating a TPU cluster or a Groq LPU instead.
Frankly, the pitch to founders is simple. Skip the funding round earmarked for infrastructure, skip the multi-year colocation contract, and start training today. The cost shows up later, as a percentage, quietly baked into every dollar the product eventually earns. For a founder staring down a compute bill they can't pay, that's a real offer. It's also a bet Nvidia is comfortable making across thousands of startups at once: some will fail and cost it nothing, and the ones that scale will keep handing over a share of their revenue for years, without Nvidia ever needing to buy them out first.
Also read: OpenAI Is Discussing Giving Washington a Stake Instead of Selling One • SoftBank pushes for a full $10 billion loan against its OpenAI stake • Robinhood Stops Renting Blockchain Rails and Builds Its Own
