NVIDIA's new AI cloud business model lands with Sharon AI and Firmus

NVIDIA is opening the door to its own capital. A new revenue-sharing and credit-support model lets AI cloud partners stand up DSX AI factories on NVIDIA hardware without bearing all of the capex, with Sharon AI and Firmus already in flight. The bet is that the next wave of AI clouds is no longer bottlenecked on silicon, but on the multi-year build cycle. The primary source is the NVIDIA blog post on the new model.

The deal NVIDIA is selling

The structure lets AI clouds procure NVIDIA accelerated computing for their end customers without bearing all of the capex, because NVIDIA takes a share of the cloud revenue on top of standard product revenue. NVIDIA CFO Colette Kress and Raj Mirpuri framed the move as a way to put full-stack accelerated compute in reach of model builders, inference providers, agent platforms and enterprises that have been locked out by the long build cycle. For the AI cloud, the trade is clear: give up some economics on the customer-facing layer in exchange for being able to sell capacity before the data center exists.

Sharon AI is the first named partner with a concrete deployment, up to 40,000 NVIDIA Grace Blackwell GB300 GPUs. The deal positions Sharon AI as a sovereign-grade infrastructure provider in a market that has been closed to new entrants by the long lead time on Blackwell. James Manning, cofounder and CEO of Sharon AI, called it a "pivotal moment" in his company's mission to deliver "sovereign, large-scale AI compute infrastructure," language that is also a quiet signal that the new model is built for regions and customers that need to keep data and silicon inside national boundaries. The 40,000 GPU figure is large enough to be a regional anchor but small enough that the site selection, power provisioning, and bring-up can be measured in quarters rather than years, which is the throughput constraint the new model is trying to compress.

Where DSX AI factories are landing

The second named partner is Firmus, which is building a DSX AI factory campus in Batam, Indonesia, sized at 360 megawatts and up to 170,000 NVIDIA GPUs. Tim Rosenfield, co-CEO of Firmus, said the company needs scalable, energy- and cost-efficient compute to compete globally, and that the DSX-aligned reference design is the lever. The Indonesia site is meaningful: Batam is the closest Indonesian free-trade zone to Singapore, sits on a subsea cable corridor that already serves Singapore, Malaysia, and Hong Kong, and has a power profile that is easier to expand than the Singapore grid that has capped hyperscale buildouts. A 360 MW campus is in the same band as a mid-sized hyperscaler region.

NVIDIA's customer-facing language points to AI-native inference providers as the immediate buyer. Baseten, Fireworks AI, and Together AI are named in the blog post as companies "where compute demand is headed," all of them running high-volume agentic inference, post-training, and fine-tuning workloads. The framing implies the DSX factories are being built around inference scale, not training scale, which is a notable shift for an industry that has spent the last two years chasing H100 and Blackwell training capacity. AIntelligenceHub's AI infrastructure resource page walks through the chips, cloud, and capacity tradeoffs that the new model is trying to compress.

It is worth pausing on the language NVIDIA is using to describe the new program. Colette Kress and Raj Mirpuri did not call it a financing product, a leasing product, or a vendor-financing arrangement, all of which would have implied a debt relationship on NVIDIA's balance sheet. They called it a "revenue-sharing and credit-support model" that "accelerates adoption" of NVIDIA platforms. The phrasing is the same one NVIDIA has used for its hardware-plus-software bundles since the DGX Cloud launch in 2023, and it matters because it signals that the AI cloud's customer revenue is the asset NVIDIA is underwriting, not the AI cloud's enterprise value. That is a different risk profile than a vendor-financing deal, and a different exposure than an NVIDIA Ventures equity investment.

The second-order effect on AI cloud economics

The most important sentence in the announcement is buried in the second paragraph: "with NVIDIA earning both standard product revenue and a share of the cloud revenue on the supported capacity." That structure turns NVIDIA into a partial principal on the customer-facing compute business, not just a hardware vendor. It also means NVIDIA now has a financial reason to favor partners who route revenue back into the NVIDIA ecosystem over partners who sell to the same customers on competing stacks. AI clouds that are heavy on AMD Instinct or custom silicon have a clear incentive to stay outside this new model, because accepting the model would mean letting NVIDIA see the unit economics of a competing stack in the same customer account.

The credit-support piece is the second lever. AI-native companies that have access to long-term compute commitments can now use those commitments as collateral to obtain financing, which has been the missing ingredient for the second tier of AI clouds. The pattern is similar to the way aircraft engine manufacturers sell engines to airlines on power-by-the-hour, except the meter is tokens rather than flight hours, and NVIDIA keeps equity-like exposure to the customer's commercial success. AIntelligenceHub's recent coverage of Claude on Microsoft Foundry running on NVIDIA Blackwell Ultra silicon shows the parallel demand story: hyperscaler channels are full, so model labs and inference providers are looking for new routes to NVIDIA silicon at production scale. Sharon AI and Firmus are now two of those routes.

The open question is whether the model extends beyond the first two named partners. The blog post's language is open enough that any AI cloud with a sovereign, regional, or capacity-as-a-service story could plausibly join the program, and the customer-facing framing around Baseten, Fireworks AI, and Together AI reads as an invitation to inference providers that are already running on NVIDIA hardware to consolidate their spend through a DSX-aligned partner rather than a hyperscaler channel. If even a handful of mid-sized AI clouds sign on, the new model will have shifted the center of gravity for AI infrastructure procurement away from the hyperscalers and toward a federated set of regional operators backed by NVIDIA's balance sheet. The board-direct implication is that AI factory capex is no longer a question of who can write the largest check, but of who can build the most reference-aligned, customer-funded footprint the fastest. NVIDIA is now selling that second capability as a service.