When NVIDIA and OpenAI announced a sweeping infrastructure partnership in September 2025, most headlines treated the $100 billion figure as a completed deal. It was not. The announcement was a letter of intent — a conditional, milestone-linked plan in which NVIDIA said it "intends" to invest "up to" $100 billion in OpenAI as each gigawatt of NVIDIA systems was deployed. By March 2026, NVIDIA CEO Jensen Huang had confirmed what NVIDIA's own securities filings had warned since November: the $100 billion was "probably not in the cards." What NVIDIA finalized was a $30 billion equity stake, part of OpenAI's $110 billion funding round announced February 27, 2026. Now, as NVIDIA's Vera Rubin platform enters full production this week with OpenAI among its first customers, the story of how that original deal unraveled — and what replaced it — offers the clearest window yet into how the AI compute race actually works.
What the $100 Billion Announcement Actually Said
The September 22, 2025 announcement contained two important qualifications that most coverage omitted. First, NVIDIA used the phrase "intends to invest up to" — not "will invest" or "has committed." Second, the investment was staged: NVIDIA would release capital progressively as each of the ten gigawatts of NVIDIA-based computing infrastructure was deployed, with the first gigawatt targeted for the second half of 2026 on the Vera Rubin platform. The two elements — the equity investment and the infrastructure deployment — were linked to each other, meaning neither was a standalone unconditional commitment.
Within weeks, NVIDIA inserted explicit risk language into its Q3 2025 10-Q securities filing: "There is no assurance that we will enter into definitive agreements with respect to the OpenAI opportunity." At the UBS Global Technology and AI Conference in December 2025, NVIDIA CFO Colette Kress confirmed that no definitive agreement had been reached and that the $100 billion commitment had been excluded entirely from NVIDIA's $500 billion data-center bookings guidance. In late January 2026, negotiations went "on ice", with some inside NVIDIA expressing doubts about OpenAI's business model.
The financial relationship did not collapse. It changed form.
How the $30 Billion Deal Replaced the $100 Billion Plan
On February 27, 2026, OpenAI announced $110 billion in new investment at a $730 billion pre-money valuation. NVIDIA committed $30 billion — a direct equity stake, not tied to deployment milestones. SoftBank contributed another $30 billion and Amazon $50 billion. OpenAI also made binding hardware commitments alongside the equity: dedicated Vera Rubin inference and training capacity, representing a more concrete commercial anchor than the original letter of intent had provided.
Speaking at the Morgan Stanley Technology, Media and Telecom Conference on March 4, Huang confirmed the shift. The original $100 billion opportunity was "probably not in the cards," he said, partly because OpenAI was moving toward an initial public offering. He described the $30 billion investment as likely NVIDIA's last equity check before that IPO. OpenAI CEO Sam Altman dismissed reports of a rift as "insanity," saying the company would remain NVIDIA's "gigantic customer" for the long term.
OpenAI closed the round on March 31, 2026 with $122 billion in total committed capital at an $852 billion post-money valuation.
Vera Rubin GPU: Architecture Behind the Deal
The hardware at the center of both deals — NVIDIA's Vera Rubin platform — entered full production this week. At NVIDIA's COMPUTEX keynote on June 1, Huang confirmed that Vera Rubin has entered manufacturing, with first systems scheduled to ship to customers in the second half of 2026. OpenAI is among the named early customers.
Understanding why Vera Rubin matters for the investment story requires understanding what it does technically. Each Vera Rubin NVL72 rack integrates 72 Rubin GPUs and 36 custom Arm "Olympus" Vera CPUs into a single liquid-cooled unit, connected by NVLink 6, NVIDIA's sixth-generation GPU-to-GPU interconnect. Each rack contains nine NVLink 6 switches delivering 260 terabytes per second of total scale-up bandwidth — GPU-to-GPU communication operating at roughly 3.6 terabytes per second per GPU. For comparison, PCIe, the standard server interconnect, tops out at a fraction of that speed.
This bandwidth matters for the specific architecture of modern large language models. Mixture-of-experts models activate only a subset of their total parameters for each token, requiring massive all-to-all GPU communication during inference. At scale, that communication pattern can consume more of a system's time than the actual computation. NVLink 6's bandwidth is designed to absorb that overhead.
NVIDIA claims the Vera Rubin NVL72 can train a 10-trillion-parameter mixture-of-experts model in one month using one-fourth the number of GPUs required on its previous Blackwell generation, and can deliver inference at one-tenth the cost per million tokens. The platform achieves 8 times the inference compute per watt of Blackwell. These are manufacturer claims subject to change; NVIDIA's product pages include explicit caveats that performance projections may vary.
AI Chip Lock-In Runs Deeper Than Hardware
The infrastructure bet in the original NVIDIA-OpenAI partnership was never purely about buying chips. It was about embedding a computing ecosystem.
NVIDIA's competitive position rests on CUDA (Compute Unified Device Architecture), a proprietary parallel computing platform introduced in 2006 that has evolved into a stack of compilers, optimized libraries — including cuDNN, NCCL, and TensorRT — domain-specific software development kits, and profiling tools. More than 4 million developers have registered for CUDA, and more than 40,000 organizations use CUDA-accelerated applications. That scale creates switching costs that are practical as much as technical: developer familiarity, debugging tooling, training materials, and thousands of small engineering decisions embedded in production code — kernel fusions, mixed-precision behavior tuned to NVIDIA's math libraries, distributed training paths optimized around NCCL assumptions, and continuous integration pipelines built around CUDA-native tooling.
AMD's ROCm platform can run many of the same workloads at roughly 70 to 90 percent of CUDA performance for standard inference kernels, but the gap widens on operations that use CUDA-specific library optimizations. Porting a production system and re-qualifying its throughput, numerical behavior, and stability under different kernels is the real switching cost — not a single benchmark comparison.
At the proposed scale of 10 gigawatts, OpenAI would not have been placing a large GPU order. It would have been coordinating future model development with several generations of NVIDIA hardware and software. That would have given NVIDIA sustained visibility into one of the industry's most demanding AI workloads while embedding its platform more deeply into OpenAI's technical roadmap.
How Circular AI Investment Draws Antitrust Scrutiny
The investment pattern that emerged from the OpenAI deal extends across NVIDIA's portfolio. CNBC reported in May 2026 that NVIDIA has committed more than $40 billion to AI equity investments in the first four months of 2026, led by the $30 billion OpenAI stake. The remainder is spread across CoreWeave, IREN, Corning, Nebius, and roughly two dozen private rounds — each position paired with compute capacity reservations, silicon roadmap alignment, or supply-chain integration.
The structure has a name: circular investment. NVIDIA invests in companies that buy NVIDIA chips; those companies grow, buy more chips, and their rising valuations support NVIDIA's balance sheet. Wedbush Securities analyst Matthew Bryson acknowledged the dynamic fits "squarely into the circular investment theme" that has driven concerns about the AI market's durability, though he argued it could also build a "competitive moat" if executed. Bernstein analyst Stacy Rasgon noted that the circular concerns "have been raised in the past, and this will fuel them further."
The pattern has attracted regulatory attention. The U.S. Department of Justice has issued subpoenas to NVIDIA and third parties, examining whether NVIDIA pressures customers to use its chips exclusively and whether it penalizes companies that buy from competitors. DOJ Antitrust Division Chief Gail Slater stated publicly that enforcement must focus on "preventing exclusionary conduct over the resources that are needed to build competitive AI systems and products." France's competition authority, the Autorité de la Concurrence, concluded after a preliminary probe that NVIDIA was likely abusing its dominant position through price-fixing, production restrictions, and unfair contractual conditions, and has opened a formal investigation. Chinese regulators separately found NVIDIA violated anti-monopoly laws related to its 2020 acquisition of Mellanox Technologies.
Rebecca Haw Allensworth, an antitrust professor at Vanderbilt Law School, identified the specific concern with the OpenAI investment structure: "They're financially interested in each other's success. That creates an incentive for Nvidia to not sell chips to, or not sell chips on the same terms to, other competitors of OpenAI." NVIDIA has said it competes on merit and adheres to all laws.
OpenAI Diversifies Compute Suppliers
OpenAI's own response to hardware concentration risk has been to reserve large volumes of computing capacity across competing suppliers.
In October 2025, OpenAI and AMD announced a multi-year agreement covering six gigawatts of AMD Instinct MI450 GPU capacity, with the first gigawatt scheduled for the second half of 2026. AMD also issued OpenAI warrants covering as many as 160 million AMD shares, tied to deployment volumes, technical milestones, and commercial conditions — a partial reversal of the NVIDIA arrangement in which the chip supplier was investing in the AI lab rather than the reverse. OpenAI has additionally committed to consuming two gigawatts of Amazon Trainium capacity through Amazon Web Services as part of the companies' 2026 partnership.
Meanwhile, OpenAI's Stargate initiative with SoftBank, Oracle, and other partners has targeted a broader $500 billion and 10-gigawatt U.S. infrastructure program.
These figures cannot simply be added. Some represent long-term targets, others conditional deployments, and some describe infrastructure that has not yet been built, connected to the electrical grid, or placed into full operation. Announced capacity is not the same as commissioned, energized, or utilized capacity.
AI Compute Race Becomes Power-and-Land Race
The scale of the original proposal becomes clearest when measured in electricity rather than processors.
One gigawatt is one billion watts. The International Energy Agency estimates worldwide data-center electricity consumption could grow from approximately 485 terawatt-hours in 2025 to around 950 terawatt-hours by 2030, with AI-focused facilities growing faster than the overall average. The U.S. Department of Energy's national assessment estimated that data centers consumed about 4.4 percent of U.S. electricity in 2023 and could reach between 6.7 and 12 percent by 2028.
Deploying infrastructure at multi-gigawatt scale requires more than hardware procurement. Developers need land, transmission access, substations, transformers, cooling equipment, backup generation, and in many regions new power plants or long-term electricity contracts. Grid connection can take years. The binding constraint on AI capability development is consequently shifting from chip supply alone toward a combination of semiconductor access, power infrastructure, and land acquisition. The compute race has become a power-and-land race.
NVIDIA's $30B Commitment Signals Discipline, Not Retreat
The retreat from the original $100 billion plan shows that even the world's dominant AI chip supplier remains cautious about underwriting infrastructure whose long-term economics have not been validated. NVIDIA's own filings reflected that caution months before the restructuring became public: the November 2025 10-Q risk language and the exclusion of the commitment from bookings guidance were early signals that the letter of intent structure was less certain than the original announcement implied.
What replaced it — a $30 billion direct equity stake alongside binding hardware commitments — is in some respects a more durable arrangement. The equity investment is not contingent on gigawatt deployment milestones. The hardware commitments are binding commercial contracts rather than intentions.
NVIDIA's broader investment pattern underscores the same point. More than $40 billion deployed into AI equity in the first four months of 2026, across OpenAI, Anthropic, CoreWeave, IREN, Corning, and two dozen private rounds, reflects a company that is not retreating from AI infrastructure but repositioning how it participates in it — moving from a single massive milestone-linked commitment toward a diversified portfolio of equity stakes paired with hardware supply agreements.
The winning AI laboratory may not be the one with the most elegant model architecture. It may be the company that coordinates capital, chips, networking, data centers, and electricity quickly enough to keep its researchers supplied with compute. As Vera Rubin enters production and the first systems reach customers this fall, that competition is no longer hypothetical. It is active.
Frequently Asked Questions
What happened to NVIDIA's $100 billion OpenAI investment?
The September 2025 announcement was a letter of intent, not a completed transaction. NVIDIA pledged to invest "up to" $100 billion progressively as each gigawatt of computing infrastructure was deployed, but the companies never signed a definitive agreement. By March 2026, NVIDIA CEO Jensen Huang confirmed the full $100 billion was "probably not in the cards," and NVIDIA finalized a $30 billion direct equity stake in OpenAI's February 2026 funding round instead.
Why did NVIDIA only invest $30 billion in OpenAI instead of $100 billion?
Jensen Huang cited OpenAI's planned IPO as the primary reason — once a company is preparing to go public, external pre-IPO equity investments of that magnitude become structurally complicated. The $30 billion stake, committed at OpenAI's $730 billion pre-money valuation, was described by Huang as likely NVIDIA's last investment in OpenAI before the offering. The original deal's milestone-linked structure — releasing capital as each gigawatt of NVIDIA systems was deployed — also created contingencies that neither company ultimately chose to execute.
What is the NVIDIA Vera Rubin platform, and why does it matter for OpenAI?
Vera Rubin is NVIDIA's current-generation AI data center architecture, now in full production. Each Vera Rubin NVL72 rack packs 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled unit, connected by NVLink 6 at 260 terabytes per second of internal bandwidth. NVIDIA claims the platform trains large mixture-of-experts models with one-fourth the GPUs required on its previous Blackwell generation and delivers inference at one-tenth the cost per token. OpenAI is a named early customer, with first systems scheduled to ship in the second half of 2026.
What is circular investment in AI, and why does it concern regulators?
Circular investment describes a structure in which a chip supplier invests in an AI company that in turn spends most of its capital on that same supplier's products, reinforcing both parties' valuations. The concern is that such arrangements create preferential access — the chip supplier may favor its equity partners with better pricing or faster delivery at the expense of competitors. The U.S. Department of Justice has issued subpoenas to NVIDIA examining whether it pressures customers to buy exclusively from NVIDIA, and France's competition regulator has opened a formal investigation after a preliminary finding of likely abuse of dominant position.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
