NVIDIA: AI GPUs, Infrastructure Deals and Investments

From graphics cards to the engine of AI

Jensen Huang co-founded NVIDIA in 1993 and remains its president and chief executive. The company built its early business selling graphics chips for PC gaming. Researchers later found that the same chips, built to do millions of parallel matrix calculations for 3D rendering, were also good at the matrix math behind training neural networks. That second market now dwarfs the first. [2][3]

For the quarter ended July 26, 2026, NVIDIA reported $96.2 billion in revenue, up 106 percent from a year earlier, with its data center segment alone bringing in $89.0 billion, about 93 percent of the total. Gaming, the business the company started in, is now a rounding error next to chips for AI training and inference. [4]

The current product line runs from the Blackwell architecture through the Vera Rubin platform NVIDIA launched in March 2026, which pairs 72 Rubin GPUs with 36 Vera CPUs per rack and which Huang said had a trillion dollars of orders behind it through 2027. NVIDIA's own estimate puts its share of GPUs used to train and run AI models above 80 percent, a dominance that draws both customers and antitrust attention, and that AMD, Amazon's Trainium, and Google's TPUs are each built to erode. [3][1]

The money behind the chips

NVIDIA's role in AI history is not just the chip design. It is the position that sits at the center of every large buildout, supplying the hardware and increasingly financing the customers who buy it. In September 2025 NVIDIA committed up to $100 billion to OpenAI, paid out as at least 10 gigawatts of NVIDIA systems come online, meaning the chip vendor helps fund the purchase of its own chips. [5]

That pattern repeated in August 2026, when NVIDIA joined Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion of third-party capital for AI data centers, underwritten against the GPUs themselves, and again a week later when NVIDIA guaranteed up to $105 billion in financing for an Ohio data center built to run OpenAI's workloads, alongside a $1.5 billion stake in SB Energy itself. [6][7]

Atlas interpretation: Larry Fink, one of the signatories to the compute-financing platforms, compared the arrangement to the invention of mortgage-backed securities. That comparison has two possible endings: it either builds a durable market for compute the way securitization built one for mortgages, or it recreates the part where the collateral's value depended on the same demand curve continuing. NVIDIA is now both the supplier and, through these deals, a lender whose collateral is its own product. [6]

Buying into every layer of the stack

Beyond financing its customers, NVIDIA has spent 2025 and 2026 acquiring pieces of the AI software stack it does not want to depend on others for. In November 2025 it joined Microsoft in putting up to $15 billion into Anthropic, buying a stake in the one frontier lab it does not already fund through OpenAI. In December 2025 it licensed Groq's inference technology and hired much of Groq's leadership team, a deal reported at $20 billion that left Groq running as a separate company under a new chief executive. In August 2026 it did something similar with a $6 billion license for Poolside's model-building software, plus a $1 billion equity stake and offers to more than 100 of Poolside's engineers. [8][9][10]

The largest of these moves came on September 3, 2026, when NVIDIA agreed to acquire Hugging Face for roughly $12.9 billion, buying the model and dataset hosting platform much of the open-model ecosystem runs on. NVIDIA said the Hub would keep supporting hardware other than its own, a commitment the company itself will have to be judged against once the deal closes. [11]

On the cloud side, AWS committed to 2 million more NVIDIA GPUs for 2027 and 2028 in August 2026, on top of a million already pledged, taking Amazon's total commitment past 3 million chips and setting aside 100,000 for secure US government workloads. Every major cloud provider now runs NVIDIA hardware at that kind of scale, which is a large part of why the company's valuation moved the way it did. [12]

The parts that could go wrong

The market's confidence in NVIDIA has already broken once and recovered. In January 2025, NVIDIA lost roughly $600 billion in market value in a single session on the theory that DeepSeek had shown frontier models could be trained cheaply, meaning fewer GPUs would be needed. Nine months later, in October 2025, NVIDIA became the first company to reach a $5 trillion valuation. The market had settled on the opposite read: cheaper inference meant more of it would be run, not less. [13][14]

Export controls are the other constraint on NVIDIA's growth, and they now reach individuals as well as the company. In August 2026, Taiwanese prosecutors indicted nine people, including NVIDIA and Super Micro staff, over a scheme that allegedly routed 130 Blackwell-based servers toward buyers in China through Indonesia, Japan and Hong Kong. The charges, breach of trust and forgery, are what export-control enforcement looks like when it reaches employees rather than corporate policy. [15]