Escape Budget

Why Nvidia may be changing customers

Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion.¹ Hugging Face generates roughly $150 million in annualized revenue and is not profitable. At that price, the acquisition is difficult to explain by looking at the business Hugging Face is today.

The price makes more sense as a bet on the customers Nvidia may need tomorrow.

The AI boom turned Nvidia into the primary hardware supplier for an unusually concentrated group of enormous buyers. Frontier model labs, hyperscalers and the cloud providers serving them bought accelerators at a scale few technology markets have ever produced. Nvidia’s second-quarter revenue reached $96.2 billion, with $89 billion coming from data center products.²

That market is still growing rapidly. There is no evidence that Nvidia’s data-center business is about to collapse.

The customers spending the most money with Nvidia are also the customers with the strongest incentive to stop depending on it.

The Escape Budget

Designing an accelerator is expensive. So is building the compiler, networking stack, serving infrastructure and engineering organization required to keep it competitive across generations. For an ordinary company, avoiding Nvidia by developing custom silicon would make no economic sense.

At sufficient scale, the calculation changes.

A company spending tens of billions of dollars on accelerators every year can spend billions developing hardware optimized for its own workloads. What is prohibitively expensive for everybody else becomes a route to lower inference cost, better power efficiency and less dependence on a single supplier.

I call that spending threshold the escape budget.

Google crossed it years ago with TPU. Amazon developed Trainium. Microsoft is deploying Maia, including the inference-focused Maia 200 introduced this year. OpenAI and Broadcom are building Jalapeño, an inference accelerator designed around OpenAI’s own workloads. Anthropic is expanding its internal silicon effort and has been evaluating partners for custom inference hardware.³

None of this means these companies will suddenly stop buying Nvidia GPUs. OpenAI explicitly expects to continue using Nvidia hardware, and hyperscalers operate mixed infrastructure because demand is too large and heterogeneous for a single architecture.

Reducing dependence on Nvidia is enough to change the structure of the market. Nvidia can continue growing while its position becomes less secure. AI demand may increase fast enough for Nvidia to sell more hardware even as custom silicon takes a larger share of inference. The transition can therefore remain almost invisible in revenue for some time.

Customers can gain bargaining power before the shift shows up in Nvidia’s revenue.

Extending the Existing Market

Nvidia has not been passive about this.

It invested $30 billion in OpenAI and $10 billion in Anthropic before Jensen Huang said in March that these were likely the company’s final private investments in either firm.⁴ Nvidia has also become involved in increasingly large infrastructure financing arrangements around the AI industry. Recent reporting puts the financing assembled with partners at roughly $500 billion, alongside substantial guarantees connected to infrastructure that OpenAI will lease.⁵

These arrangements have attracted criticism as forms of circular financing: Nvidia helps finance companies and infrastructure that in turn buy Nvidia hardware. Nvidia rejects that characterization, and the demand itself is plainly real. Its latest financial results make that difficult to dispute.

But financing can extend an existing market without changing its underlying economics.

Once a customer crosses the escape budget, reducing the margin paid to an external supplier becomes attractive regardless of who helps finance the next data center. Nvidia can make the transition slower. It cannot make vertical integration irrational.

Below the Escape Budget

Enterprises have almost the opposite economic profile.

A bank, manufacturer, pharmaceutical company or government department may eventually operate substantial AI infrastructure. Very few will ever design their own accelerator. The development cost, organizational complexity and required volume make no sense.

They sit permanently below the escape budget.

That makes enterprise-owned AI a much more attractive long-term market for a hardware supplier. Instead of a small number of customers that can eventually integrate around you, it creates a fragmented market of customers that will continue buying general-purpose hardware.

Most enterprises currently consume advanced AI as a service. They buy tokens from OpenAI, Anthropic, Google and others. The accelerator sits inside somebody else’s data center and Nvidia sells to that intermediary.

For Nvidia to turn enterprises into direct infrastructure customers, enterprises need a reason to run the models themselves.

Open weights provide one.

Nvidia is making self-hosting easier by releasing models together with training material, deployment recipes and aggressively optimized quantized variants. Nemotron 3.5 Lightning, for example, ships in an NVFP4 form designed to run efficiently across Nvidia hardware, including a validated configuration for a single DGX Spark.⁶

Optimizing open models for Nvidia hardware can help sell that hardware.

Why Hugging Face Matters

Hugging Face sits directly between those two layers.

It has become the default place where much of the open-model ecosystem is discovered, downloaded, fine-tuned, quantized and redistributed. Model developers publish there. Hardware vendors publish optimized variants there. Tooling assumes its existence.

Beyond its $150 million in annualized revenue, Hugging Face influences how an open model travels from its creator to the machine that eventually runs it.

Nvidia already has the accelerator. It has CUDA and the surrounding software ecosystem. It has systems such as DGX. It increasingly has its own open models. Hugging Face gives it a position at the distribution layer where those models meet the people trying to deploy them.

Nvidia could influence deployment without overt exclusion. It could not turn Hugging Face into an Nvidia-only repository without destroying much of what makes the platform valuable. AMD, Intel, Google and almost every important model developer participate in the same ecosystem.

Deployment defaults give Nvidia another way to influence where workloads land: checkpoints that are easy to deploy, quantizations available at release, tested serving recipes, and hardware configurations that work without days of integration.

The Pressure Moves Upward

This is also where the acquisition becomes uncomfortable for the frontier model labs.

There is no broad enterprise migration toward self-hosted open models today. Menlo Ventures estimated the open-model share of enterprise LLM usage at 11 percent in late 2025, down from 19 percent the year before.⁷ Whatever the long-term trajectory, enterprises have so far shown a strong preference for managed frontier services.

Frontier models have generally been better, hosted APIs are dramatically easier to operate, and deploying a model yourself introduces utilization, hardware and engineering problems that disappear when somebody else charges you per token.

Price alone has not been sufficient to overcome those disadvantages.

For an enterprise, owning the inference environment also provides control over data location, model versions, availability and network boundaries. Some workloads cannot leave a jurisdiction. Others cannot leave a physical network. Some organizations need models that can operate without an external service at all.

Those requirements change the threshold at which an open model becomes competitive.

The open model has to cross the perception threshold for the workload being performed. Once the remaining capability difference becomes smaller than the operational advantage of running locally, the superior hosted model can still be technically better while becoming commercially less attractive.

There are already hints of how unevenly this transition may happen. Vercel reported that open-weight models accounted for 29 percent of the tokens passing through its AI Gateway in June while representing less than 4 percent of spend.⁸ That is not evidence of an enterprise-wide replacement cycle, but it does show the shape such a transition can take. Cheap, bounded workloads move first. Expensive frontier work stays behind.

Volume can leave before revenue does here as well.

For companies whose economics depend heavily on enterprise token consumption, that distinction is not comforting indefinitely. Anthropic in particular has an unusually enterprise-heavy revenue mix; Reuters reported last year that business and enterprise customers represented roughly 80 percent of its revenue.⁹

Nvidia’s largest model customers are building hardware that reduces their dependence on Nvidia. Nvidia can respond by helping their customers reduce dependence on the models.

The Limit of the Play

None of this guarantees that enterprise-owned AI becomes an Nvidia market.

Open weights weaken the model provider’s control over deployment precisely because they are portable. The same model that runs on an Nvidia system can potentially be optimized for AMD, Intel or entirely different architectures.

China makes that constraint particularly visible.

DeepSeek has already adapted models for Huawei hardware, while Xiaomi recently demonstrated an AI Cube prototype built from its own XRING processors and designed for local deployment of large models.¹⁰ The broader direction is toward tighter integration between locally deployable models and increasingly specialized local hardware.

That produces the same escape mechanism one layer lower.

If sufficiently capable open models become common, Nvidia benefits only if the infrastructure beneath them continues to belong to Nvidia.

Hugging Face helps with that fight, but cannot settle it.

There is another risk. Hugging Face derives much of its value from being perceived as neutral infrastructure for the open-model ecosystem. Nvidia would be buying a coordination point used by its own competitors. If ownership begins to distort that neutrality, the ecosystem can move. Model weights are considerably easier to relocate than semiconductor fabs.

Nvidia’s ability to preserve Hugging Face’s neutrality while directing more workloads toward its hardware will help determine whether $12.9 billion eventually looks cheap or absurd.

A Change of Customer

Nvidia’s next market may have a different structure. Frontier labs and hyperscalers increasingly build custom infrastructure because their scale makes specialization economical. Enterprises sit on the other side of that threshold. They can own AI infrastructure, but most cannot justify designing it.

Open-weight models can turn those enterprises from consumers of remote tokens into owners of compute. Nvidia has spent the past year making those models easier to run on Nvidia hardware. Buying Hugging Face would place it inside the distribution system through which much of that market already moves.

The acquisition could extend Nvidia’s hardware business to a much larger number of customers.

The frontier labs will continue buying Nvidia accelerators for a long time.

Nvidia also has a direct economic interest in making enterprise dependence on frontier APIs weaker.