Nvidia-OpenAI partnership
Last reviewed
May 31, 2026
Sources
16 citations
Review status
Source-backed
Revision
v3 ยท 2,607 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
16 citations
Review status
Source-backed
Revision
v3 ยท 2,607 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Nvidia-OpenAI partnership is a strategic agreement announced on September 22, 2025, under which NVIDIA said it intends to invest up to $100 billion in OpenAI as OpenAI builds out a large amount of new computing capacity. The two companies described the deal as a letter of intent to deploy at least 10 gigawatts of NVIDIA systems, a quantity they said would represent millions of graphics processing units. NVIDIA said it would invest progressively, with each portion of the money tied to OpenAI bringing more capacity online, and that the first gigawatt would run on NVIDIA's next-generation Vera Rubin platform in the second half of 2026.[1][2]
The arrangement quickly became one of the most discussed examples of what critics call "circular financing" in the AI sector, the pattern in which a chip supplier invests in the customers that turn around and buy its chips. Supporters describe the structure as a normal way to fund capital-heavy infrastructure and to align a supplier with the growth of a key customer. The debate sits inside a larger argument about whether the AI industry's spending on data centers and accelerators can be sustained by the revenue those systems eventually produce.[3][4]
NVIDIA and OpenAI said NVIDIA intends to invest up to $100 billion in OpenAI over time. The money was meant to be deployed in stages rather than all at once, with each piece linked to OpenAI deploying another gigawatt of NVIDIA-based computing. In total the companies pointed to at least 10 gigawatts of NVIDIA systems for OpenAI's next generation of AI infrastructure, which they said would amount to millions of GPUs.[1]
In its release, NVIDIA said it would become a "preferred strategic compute and networking partner" for OpenAI's AI factory growth, and that the first gigawatt of capacity would use the Vera Rubin platform, NVIDIA's planned successor architecture to its current Blackwell generation. The first gigawatt was targeted for the second half of 2026.[1] The chips that underpin the relationship today are part of NVIDIA's Blackwell line, the data-center accelerators that power much of the current AI training and inference market.
NVIDIA founder and chief executive Jensen Huang said in the announcement that the company's work with OpenAI had run for about a decade, from early DGX supercomputers through the launch of ChatGPT, and that the new investment and infrastructure partnership would help power "the next era of intelligence." OpenAI co-founder and chief executive Sam Altman said that "everything starts with compute" and that compute infrastructure would be "the basis for the economy of the future."[1]
A central feature of the announcement is that it was a letter of intent rather than a finished, binding contract. NVIDIA said an initial $10 billion would be tied to the execution of a definitive purchase agreement for the first gigawatt of systems, with later money following as each additional gigawatt was deployed.[1][2] Reuters reported that the investment would take the form of non-voting shares in OpenAI, while CNBC reported that the funding would be in cash, most of which OpenAI would then use to lease NVIDIA chips. The companies said definitive terms were still to be worked out.[3][5]
That milestone-based design is the source of much of the discussion about the deal. Because the funding was tied to deployment, the headline figure of $100 billion was a ceiling that depended on OpenAI actually building and powering the systems over a multi-year period, not a sum that changed hands at signing. NVIDIA shares rose on the day of the announcement.[3]
The table below summarizes the main figures the companies and reporters cited, with the caveat that several were described as targets or as subject to definitive agreements.
| Element | Detail as announced or reported |
|---|---|
| Investor | NVIDIA |
| Recipient | OpenAI |
| Maximum investment | Up to $100 billion, deployed progressively |
| First tranche | About $10 billion, tied to a definitive purchase agreement for the first gigawatt |
| Compute commitment | At least 10 gigawatts of NVIDIA systems |
| Scale | Described as millions of GPUs |
| First-phase platform | Vera Rubin |
| First gigawatt timing | Second half of 2026 |
| Legal nature | Letter of intent, described as non-binding |
| Announcement date | September 22, 2025 |
Sources for the table: NVIDIA and OpenAI announcements and contemporaneous reporting.[1][2][3]
The NVIDIA deal was one part of a much larger set of computing commitments OpenAI made across 2025. The anchor is Stargate, a data-center program OpenAI announced in January 2025 alongside SoftBank, Oracle, and the investment firm MGX. OpenAI said the project intended to invest $500 billion over four years in AI infrastructure in the United States, with $100 billion deploying in an initial phase, and it named Arm, Microsoft, NVIDIA, and Oracle as technology partners.[6] The first flagship site is the Abilene data center in Texas.
OpenAI also signed or expanded compute deals with several other suppliers in the same period. The pattern in these agreements is that OpenAI commits to large amounts of capacity, and in some cases the supplier takes on equity exposure to OpenAI's growth.
| Counterparty | Reported scope | Notable terms |
|---|---|---|
| NVIDIA | At least 10 GW of systems | Up to $100B investment; Vera Rubin first GW in H2 2026 |
| AMD | 6 GW of GPUs | Warrant for up to 160M AMD shares; first GW of MI450 in H2 2026 |
| Broadcom | 10 GW of custom accelerators | OpenAI-designed chips; deploy H2 2026 through end of 2029 |
| Oracle | About $300B of cloud compute | Roughly 4.5 GW over about five years; linked to Stargate |
Sources: company announcements and press reporting for AMD, Broadcom, and Oracle.[7][8][9]
In October 2025, Advanced Micro Devices and OpenAI announced a partnership for OpenAI to deploy 6 gigawatts of AMD GPUs over multiple years, starting with 1 gigawatt of AMD Instinct MI450 chips in the second half of 2026. As part of that deal AMD issued OpenAI a warrant to buy up to 160 million AMD shares, vesting as deployment and share-price milestones were met, a stake that could reach about 10 percent of AMD if fully exercised.[7] Days later, OpenAI and Broadcom said they would co-develop and deploy 10 gigawatts of custom AI accelerators that OpenAI designs, with racks scaled using Broadcom Ethernet networking, deploying from the second half of 2026 through the end of 2029.[8] Separately, news outlets including The Wall Street Journal reported that OpenAI agreed to buy roughly $300 billion of computing power from Oracle over about five years starting in 2027, a contract tied to the Stargate buildout.[9]
Taken together, these commitments implied tens of gigawatts of planned capacity and, by various estimates, more than a trillion dollars of future spending obligations. The gap between OpenAI's current revenue and its forward commitments became one of the reasons the NVIDIA deal drew scrutiny.[10]
The label "circular financing" describes a loop in which money flows from a supplier to a customer and then back to the supplier as a purchase. In this case, NVIDIA would invest in OpenAI, and OpenAI would direct a large share of its resources toward NVIDIA systems. Critics argue that this kind of structure can flatter demand, because some of the buyer's purchasing power originates with the seller, and can make it harder for outside observers to judge how much underlying, independent demand exists.[3][4]
Several analysts said as much in the days after the announcement. Stacy Rasgon of Bernstein wrote that the deal would "clearly fuel 'circular' concerns." Jay Goldberg of Seaport Global Securities called the structure "bubble-like" and compared it to "having your parents co-sign on your first mortgage," and elsewhere described NVIDIA's investing in its own customers as "very murky." Gil Luria of D.A. Davidson argued that OpenAI would need hundreds of billions of dollars to meet its obligations and questioned whether the company was in a position to make commitments of that size.[4] The investor Michael Burry, known for betting against the 2000s housing market, later said he was shorting NVIDIA and likened it to Cisco at the peak of the dot-com boom, and the short seller Jim Chanos compared the AI deals to the vendor financing that hurt the telecom equipment maker Lucent around 2000.[11][12]
Supporters and the companies themselves pushed back on the framing. NVIDIA characterized its OpenAI investment as optional and not a precondition for OpenAI to buy its products, and Huang said OpenAI could use the money "for whatever they want." He called the circular-financing characterization, in his words, "nonsense," and a memo NVIDIA circulated to analysts argued that its strategic investments were a small fraction of its revenue and that AI developers raised most of their capital from outside investors.[11][12] Rasgon, despite flagging the circular concern, also said he could argue there was "no better use of NVIDIA's cash right now."[11] Some investors framed the web of deals as a normal way to line up suppliers and customers for a fast-growing market rather than as a warning sign.[4]
It is worth distinguishing two things that the "circular" label can blur. A supplier investing equity in a customer is common across industries and is not by itself improper. The sharper concern raised by critics is round-tripping, in which revenue is effectively manufactured by funding one's own customers, which can mislead investors about true demand. Reporting on the NVIDIA deal noted the circular structure but did not establish that the arrangement crossed into improper accounting, and the deal was disclosed publicly rather than hidden.[3][12]
The NVIDIA-OpenAI deal landed in the middle of a wider argument about whether AI capital spending can be sustained. Through 2025, the largest cloud companies raised their capital-expenditure plans sharply to build AI data centers, with combined spending by the biggest firms roughly doubling over two years. Questions grew about how quickly that spending would turn into profit. This is the heart of the AI capex discussion and a major thread in talk of a possible AI bubble.[10][13]
Skeptics pointed to the size of OpenAI's commitments relative to its revenue. By late 2025, Altman said OpenAI had crossed roughly $20 billion in annualized revenue and had about $1.4 trillion in data-center commitments over the following years, a gap critics found hard to reconcile. When pressed on a podcast about how a company with that revenue could make such commitments, Altman said the company was doing well more revenue than the figure quoted and, to one questioner, offered to find a buyer for his shares.[10] Altman had earlier acknowledged that parts of the AI market looked overheated, saying that "when bubbles happen, smart people get overexcited about a kernel of truth" and that "someone is going to lose a phenomenal amount of money," while maintaining that the technology would deliver large long-run value.[14]
Optimists countered that compute demand had consistently outrun forecasts, that frontier models kept finding new commercial uses, and that staged, milestone-based deals like the NVIDIA one were a reasonable way to fund growth while limiting downside. They noted that the NVIDIA investment was capped and conditional, so it scaled with actual deployment rather than committing the full sum up front.[1][12] Both camps tended to agree that the scale of spending made the sector's economics unusually sensitive to whether AI products generate durable revenue.
In the months after the announcement, the partnership stayed at the letter-of-intent stage and the companies did not complete a definitive agreement. Huang later described the original $100 billion figure as "never a commitment," saying OpenAI had invited NVIDIA to invest up to that amount and that NVIDIA would proceed "one step at a time."[15] In early 2026, The Wall Street Journal reported that the plan to invest the full $100 billion had stalled amid doubts inside NVIDIA, and NVIDIA's stock moved on the news.[16] NVIDIA's actual investment in OpenAI's subsequent funding round was reported at about $30 billion, well below the original ceiling, and Huang said that investment "might be the last" before a potential OpenAI public offering.[16] The companies continued to describe their commercial relationship as intact even as the headline investment figure changed.[15][16]
The NVIDIA-OpenAI partnership matters less for any single dollar figure than for what it signaled about how AI infrastructure is being financed. It tied the most valuable chipmaker to the most prominent AI developer through a multi-year, milestone-linked investment, and it became a reference point for similar deals that followed with AMD, Broadcom, and cloud providers, as well as for a related Nvidia-Microsoft partnership.[7][8]
For NVIDIA, the deal helped secure a flagship customer for its next-generation Vera Rubin systems and extended a relationship that runs back to OpenAI's early research. For OpenAI, it added a large potential funding source and supply commitment to a buildout that already spanned Stargate and deals with multiple vendors. For the broader market, it crystallized a question that remained open at the time of the announcement, which is whether the web of investments and purchase commitments among AI suppliers and developers reflects genuine, independent demand or a more fragile, self-reinforcing cycle. The way the original $100 billion pledge later shrank and slipped into uncertainty became part of that debate rather than a resolution of it.[3][10][16]