On July 1, 2026, Nvidia published a blog post co-authored by its CFO Colette Kress announcing what it called a revenue-sharing and credit-support model for AI cloud operators. The language was anodyne. The structure was not. Under the new program, AI companies receive Nvidia's latest Grace Blackwell GB300 GPUs without paying the full cost upfront. In return, Nvidia collects its standard hardware revenue plus a recurring percentage of whatever cloud revenue those GPUs generate. If the GPUs sit idle because the operator cannot fill them with paying customers, Nvidia guarantees to buy back the unused capacity at a predetermined price.
The first two partners are Sharon AI, an Australian company that listed on Nasdaq in February 2026 through a $125 million IPO before raising an additional $1.6 billion in private placement in June, and Firmus Technologies, which is building a 360-megawatt AI factory campus on the Indonesian island of Batam. Between them they have committed to deploying 210,000 Grace Blackwell GB300 GPUs under six-year agreements. Sharon AI will operate 40,000 for sovereign AI workloads. Firmus will house up to 170,000 in a facility built to Nvidia's DGX SuperPOD reference architecture.
What Nvidia announced is three financial instruments bundled into one program. The first is deferred payment: customers access hardware now and pay from future earnings. The second is a revenue share: Nvidia earns a royalty on every dollar of cloud income generated by its chips, converting a one-time sale into a recurring stream. The third is a put option: the buyback guarantee sets a floor price on GPU capacity, meaning Nvidia has written an option obligating itself to purchase its own product back if demand fails to materialize. The chipmaker is simultaneously the manufacturer, the creditor, and the insurer.
Nvidia reported $81.6 billion in revenue for the quarter ending April 2026, up 85 percent from a year earlier. Data center revenue alone reached $75.2 billion. Gross margins held at 75 percent. The company's market capitalization as of July 1 was $4.7 trillion, making it the most valuable company on Earth. These are not the financials of a company that needs to extend credit. These are the financials of a company whose customers need to receive it.
The historical parallel is exact. In 1999 and 2000, Cisco Systems extended $2.4 billion in vendor financing to telecom companies building out internet infrastructure, roughly 10 percent of its annual revenue at the time. Lucent Technologies committed $8.1 billion. Nortel extended $3.1 billion. The logic was identical: demand for networking equipment was insatiable, customers were capital-constrained, and the manufacturers were so confident in the end market that financing the purchase seemed like free money. Between 2000 and 2003, forty-seven competitive local exchange carriers declared bankruptcy. Cisco wrote off nearly $900 million in bad loans. Its stock fell 89 percent from its peak.
Nvidia's program differs from Cisco's in one structural respect that makes it either more defensible or more dangerous depending on your assumptions. Nvidia controls the depreciation schedule of its own hardware. When Nvidia announces the next generation of GPUs, current-generation hardware loses value on a timeline that Nvidia alone determines. This means Nvidia is the only entity in the market that can accurately price the residual value of a Grace Blackwell GPU three years from now, because Nvidia is the entity that decides when Grace Blackwell becomes obsolete. The buyback guarantee is not a bet on the market. It is a bet on Nvidia's own product roadmap.
In the Cisco analogy, the equipment had value independent of the manufacturer. A Cisco router depreciated based on bandwidth demand, which Cisco did not control. Nvidia's GPUs depreciate based on the next chip Nvidia ships. If Nvidia launches a successor that delivers twice the performance per watt, every GB300 under a buyback guarantee is worth half what Nvidia promised to pay for it. The guarantee becomes a liability precisely when Nvidia succeeds. The faster Nvidia innovates, the larger the gap between the guaranteed price and the market value of the hardware it must buy back.
The program also creates a secondary incentive that no one is discussing. Under a pure sales model, Nvidia benefits from selling as many GPUs as possible regardless of whether customers use them productively. Under a revenue-share model, Nvidia benefits only when customers generate cloud revenue from those GPUs. This aligns Nvidia's incentive with customer utilization for the first time. But it also means Nvidia now has a financial interest in the success of every AI cloud startup running its hardware. If the AI workload market grows slower than the GPU supply, Nvidia is exposed on both sides: the revenue share underperforms, and the buyback guarantee activates.
Two hundred ten thousand GPUs is a pilot. Nvidia shipped the equivalent of millions of GPUs worth of data center hardware last quarter. But pilots become templates. The revenue-share structure solves a real problem: AI infrastructure requires capital that most companies do not have, and the gap between GPU supply and customer purchasing power is widening as chip performance improvements outpace the growth in AI workload revenue. If the program scales to a meaningful fraction of Nvidia's shipments, the company's balance sheet transforms from a hardware manufacturer into something closer to a specialty finance company with a chip design arm.
The market has not priced this transformation. Nvidia trades at approximately 58 times trailing earnings, a multiple appropriate for a high-growth semiconductor company with 75 percent gross margins. If a material portion of future revenue becomes tied to customer cloud income rather than hardware sales, the earnings profile changes. Revenue becomes more recurring but also more contingent. Margins depend not on Nvidia's manufacturing cost but on whether an Indonesian data center in Batam fills its racks with paying AI workloads six years from now.
Cisco's stock in March 2000 reflected a company selling equipment into unlimited demand. By March 2001, it reflected a company that had financed its own customers' inability to generate revenue. Nvidia in July 2026 has taken the first step on the same path, with the same confidence, the same structural logic, and one additional complication: it also controls the clock that determines when its guarantees become expensive. The world's most valuable company just became a lender. The collateral is its own product. The borrowers are companies that do not yet have the customers to pay for it.