Nvidia quietly rewired its business model this week. In a July 1 blog post co-authored by CFO Colette Kress, the company unveiled what it calls the AI Compute Partnership: Nvidia guarantees a floor on how much of a cloud provider's GPU fleet stays paid-for — renting back idle capacity itself at a predetermined rate — and in exchange collects a recurring share of the cloud revenue those chips generate. The percentage? Undisclosed. Nvidia declined to elaborate beyond the post.
The first two partners are small names attached to big numbers. Sharon AI, an Australian sovereign-cloud provider that listed on Nasdaq in February, plans up to 40,000 Grace Blackwell GB300 GPUs under a six-year agreement, weeks after raising $1.6 billion privately to fund the buildout. Firmus Technologies is constructing a 360MW campus in Batam, Indonesia designed to house up to 170,000 Nvidia GPUs. Combined: roughly 210,000 chips, aimed at serving AI-native customers like Baseten, Fireworks AI, and Together AI.
This formalizes what Nvidia had been doing ad hoc. Its demand guarantees reportedly helped CoreWeave raise $6.3 billion in debt and Lambda $1.5 billion. What was a quiet supporting role is now a named program with Nvidia's balance sheet explicitly on the line — and explicitly compensated.
Why lenders wanted this
The neocloud bottleneck was never demand — it was collateral. Banks don't know what a GPU is worth in four years, so they price clusters like depreciating mystery boxes and financing drags out for 12–18 months. A buyer-of-last-resort pledge from a $4-trillion-plus vendor changes the math: idle capacity now has a floor price, which makes the whole cluster bankable. Expect buildout timelines to compress — and expect more of the record flood of infrastructure capital to flow through structures like this one.
The skeptics' version, per The Register: this is "double-dipping" — Nvidia profits on the sale, then again on the usage. The kinder analogy is GE Capital, which financed the customers who bought GE's turbines for decades. It worked brilliantly, until the financing arm became the risk.
The circularity question
Strategically, this keeps independent GPU clouds alive as a counterweight to the hyperscalers — Nvidia's hedge against its biggest customers designing their own silicon and pulling AI deployment in-house. But note who absorbs the downside: if AI demand softens, Nvidia ends up renting back capacity that its own chip sales created. The market's first read was split — Nvidia barely moved on the news, while Sharon AI shares sold off hard per market-data trackers, a hint that investors see the revenue share diluting the very upside that makes neoclouds investable.
Our take: Vendor financing is the oldest late-cycle signal in tech — Lucent wrote the cautionary tale in 1999. The difference here is that Nvidia is writing demand insurance on a product only it can restock, from the strongest balance sheet in the industry. That's not a reason to panic; it's a reason to change what you read. The AI trade's health is no longer visible in keynotes or benchmark charts — it's in commitment footnotes and utilization floors. Risk didn't leave the system this week. It moved up the chain, to the one company everyone assumed had none.
What to watch
- Nvidia's next 10-Q. The size of rent-back commitments and any receivables buildup will be the first honest disclosure of how big this program really is.
- The revenue-share percentage. Undisclosed terms on a program this structural won't stay undisclosed through earnings season.
- Who joins next. If CoreWeave or Lambda formalize under the program, this is the new default for GPU cloud financing.
- Sharon AI's stock. The only public pure-play in the program — the market's live referendum on whether backstopped neocloud economics work for anyone but Nvidia.