NVIDIA is rewriting the deal it offers the cloud operators who buy its chips. Instead of booking revenue when a box leaves the loading dock, the chipmaker is now financing the AI cloud itself, accepting credit risk on the buildout, and taking a stated share of the cloud revenue those GPUs generate over time. For investors who have spent two years pricing NVIDIA like a semiconductor company, that is a meaningful reframe: a slice of the business is starting to behave like a utility lender with a royalty attached.
The first public anchor of the new arrangement sits in SEC filings. Sharon AI disclosed the deal as a material event on a Form 8-K and followed it with a prospectus supplement under Rule 424(b)(3), putting a smaller, less-covered AI cloud operator on the record about exactly the kind of capital partnership NVIDIA is now selling to the broader market. Third-party coverage from HotHardware pegs the eventual Sharon AI deployment at as many as 40,000 GB300 GPUs and describes Firmus Technologies' campus as a 360-megawatt buildout targeting up to 170,000 GPUs. A GB300 is NVIDIA's latest-generation Grace Blackwell server module. A 360-megawatt campus draws roughly the power of a midsized regional data-center hub.
The strategic case NVIDIA is making in its own announcement of the program is that always-on AI factories handling training, fine-tuning, post-training, and inference at scale have run into a financing wall. AI-native startups and specialized clouds cannot easily raise the upfront capital to buy and operate the newest GPU generations against future, uncertain customer commitments, so NVIDIA is packaging that financing with the chip sale itself.
The economic terms of the package, however, are still being filled in. CNBC reported that the arrangement includes revenue-sharing components and credit backing for the GPU deployments, and Benzinga framed the launch as opening access for AI startups to compute capacity they could not finance alone. What neither outlet quotes yet are the specific numbers: what equity percentage NVIDIA is taking, the revenue-share rate, the term length, the collateral structure, the recourse profile if utilization falls short, and whether the credit facility sits on NVIDIA's balance sheet, in a partner vehicle, or is partially syndicated.
Those details matter because they change the economics. A revenue share with no credit exposure is a royalty, smoother revenue with limited downside. A revenue share paired with full-recourse lending is closer to project finance, with utilization risk flowing back to NVIDIA. In the first case the market reads it as a recurring-stream upgrade. In the second, it is that plus an embedded credit line item that moves with GPU utilization and customer cash flow. Investors used to model NVIDIA around hardware shipment cycles will need to start layering in a longer-duration, utilization-linked revenue line with credit sensitivity on top.
The launch roster already hints at the workloads NVIDIA is underwriting. Alongside Sharon AI and Firmus, the program runs through inference-focused platforms including Baseten, Fireworks AI, and Together AI, all built to serve high-volume model serving rather than one-off training runs. Inference is the workload that runs continuously on a deployed GPU cluster, as opposed to training, which tends to spike and fall. That mix is consistent with the recurring-revenue framing, but it also raises the question of how exposed NVIDIA becomes if inference pricing compresses.
The most useful next moves are concrete. Watch for the next Sharon AI 8-K or 10-Q that itemizes the equity stake, revenue-share rate, and term length. Watch for any disclosure from NVIDIA itself, in the next earnings call or 10-Q, on whether the credit exposure sits on the balance sheet and how it is provisioned. Watch for additional named partners or a syndicate structure that would spread the credit risk. Until those details surface, the structural shift is real and the scale is real, but the income-statement translation is still a forecast rather than a settled number.