Nvidia's Openness Is How It Closes Every Exit

When Jensen Huang announced Nvidia's $12.9 billion purchase of Hugging Face, the company that hosts most of the world's open-source AI models, he said the most generous thing a dominant company can say ^[1].

NVIDIA compute will not be required to build on or deploy through Hugging Face. — Jensen Huang

It reads like surrender — the chipmaker promising not to favor its own silicon on the platform where models are built and shared. It is the opposite. Owning Hugging Face means Nvidia owns the distribution layer regardless of whose chip runs the model underneath. Whoever wins the silicon war, the models still flow through a platform Nvidia bought. The pledge is what makes that tolerable to everyone else. This is the shape of Nvidia's strategy now, and it only becomes visible when you line up a dozen separate moves from the past year. Each one takes the form of openness: a platform, a partnership, a pledge. Each one closes a different exit. Start with the exit that got bought outright. OpenAI had been testing Groq, a startup whose chips run inference far faster than Nvidia's GPUs, as an alternative for about a tenth of its inference work ^[2]. Nvidia paid $20 billion to license Groq's technology and hire its chief executive and senior leadership. The challenger is now integrating Nvidia systems into its own data centers — the company that was supposed to be the escape route became a customer ^[3]. Then the wiring. Arm's Neoverse CPUs are the architecture Google, Meta, Microsoft, and Amazon use in their custom data-center chips — the silicon they build to reduce their dependence on Nvidia. Arm joined Nvidia's NVLink Fusion program so those chips connect natively to Nvidia GPUs through a coherent fabric, and Nvidia's stated aim is explicit ^[4].

Arm is integrating NVLink IP so that their customers can build their CPU SoCs to connect Nvidia GPUs. — Nvidia

Qualcomm separately adopted NVLink Fusion to connect its own Oryon CPUs to Nvidia GPUs ^[5]. Even the rival silicon routes through Nvidia's wiring. Then the physical supply: $279 billion committed through fiscal 2029 to lock up memory chips, up from $119 billion three months earlier ^[6]. Then the money itself: up to $500 billion in data-center financing arranged, backstopping $105 billion for a single Ohio project ^[7]. Nvidia is not just the supplier of the chips; it is the bank that funds the buildings they go in. The walls narrow with each beat. A customer who wants faster inference finds the alternative bought. A customer who wants its own silicon finds its chips wired into Nvidia's fabric. A customer who wants memory finds it prepaid. A customer who wants to build at all finds Nvidia holding the loan. But the exits are not all closed, and it would be wrong to pretend otherwise. OpenAI has deployed a model on Cerebras hardware and is expanding to AMD and Broadcom — but it frames the whole thing as diversification around a Nvidia core ^[8].

That’s why we are anchoring on Nvidia as the core of our training and inference stack, while deliberately expanding the ecosystem around it through partnerships with Cerebras, AMD and Broadcom — OpenAI

Google is building its Ironwood TPU and copying Nvidia's own financing playbook — financial guarantees, billions in equity — but it still buys Nvidia's Blackwell Ultra and frames its own chips as additive rather than a substitute ^[9].

I would love to hear them demonstrate the cost advantage of TPUs. It makes no sense in my mind. — Jensen Huang

And in China, DeepSeek has gone somewhere Nvidia cannot follow: it excluded Nvidia and AMD from optimizing its V4 model, giving Huawei a head start, and is building a 160,000-chip Huawei cluster ^[10]. Huang himself conceded the logic.

if they can’t buy from us, they’ll build their own — and then we’ll have a competitor we wouldn’t have otherwise created. — Jensen Huang

These are real exits. But measure their boundaries. OpenAI's diversification is at the margins — a tenth of inference, the core still Nvidia. Google's parallel stack is additive, not a substitute. China's is a bifurcation, a rival ecosystem Nvidia cannot reach and cannot serve. The exits that remain open are marginal, or on the other side of a geopolitical wall, or still wired through Nvidia's interconnect. Which leaves the question Huang answers with a shrug. Asked about the TPU threat, he was dismissive.

Our market reach is far greater than any TPU or ASIC can possibly have. — Jensen Huang

AllianceBernstein has begun warning that AI infrastructure stocks now drive nearly half of S&P 500 earnings growth, and that the Russell 1000 Growth Index has 57% exposure to its top ten holdings ^[11]. A system with one layer every path runs through is efficient until it isn't. Nvidia has spent a year making sure every road leads through it. Either that is the confidence of a man who has closed every exit — or it is the blind spot of the chokepoint he is building.