Kostas Karolemeas
Writing archive

Aug 31, 2026 · AI strategy

Why NVIDIA Wants Models to Be Cheap

Joel Spolsky's old rule about commoditizing your complements explains NVIDIA's interest in Hugging Face—and exposes the harder question: can models become interchangeable without making the compute platform interchangeable too?

AI strategyNVIDIAHugging FaceOpen modelsCompetitive advantage
A low dark teal mass sits among translucent aqua and copper planes beneath a sweeping cream brushstroke

The most useful question about a technology company giving something away is often: what does that make more valuable?

An old essay by Joel Spolsky offers a useful way to read NVIDIA's interest in Hugging Face. The strategic prize may be an ecosystem in which models become easier to obtain, improve and replace, while the platform used to run them remains difficult to substitute.

That is my interpretation of the incentives, not a statement of NVIDIA's acquisition plans.

The reported NVIDIA–Hugging Face deal remains unconfirmed by either company, with conflicting accounts of whether an agreement was signed. As of August 31, I have found no official confirmation. The argument below does not require the transaction to close.

In my earlier article on Stripe and NVIDIA, I examined how the companies could organize the demand and supply sides of inference. This essay asks a different question: which parts of the AI stack would each participant like to become cheap, and where does it expect to keep earning money?

The old rule

In his June 2002 Strategy Letter V, Spolsky wrote:

Smart companies try to commoditize their products’ complements.

Complements are things consumed together. Lowering the cost of one can increase demand for the other. Spolsky used the PC industry to explain how competition among hardware suppliers could strengthen the economics of the operating-system supplier.

The practical question is: which things must customers buy or assemble before they can use your product?

You may want those things to be excellent. You may fund their development. You may also want many interchangeable suppliers competing to provide them.

This is an incentive, not a universal explanation of corporate behavior. Commercial interests and sincere support for openness can coexist. Nor does making a complement cheaper guarantee that the resulting value accrues to you.

Open models can expand the market for compute

Applied to NVIDIA, the logic is straightforward. More useful and accessible models can make more applications economical to build. Those applications need computation.

NVIDIA's investment in open models is documented independently of the acquisition reporting. Its March 2026 Nemotron Coalition announcement described a collaboration among model builders, with a shared model trained on DGX Cloud and released to the open ecosystem. NVIDIA was helping produce a complement to its own platform.

My reading is that a competitive model layer offers two potential benefits. It can expand overall demand for AI, and it can prevent a small number of model suppliers from controlling the customer's route to that demand.

A closed API customer buys a result from the model provider; the provider chooses much of the infrastructure behind it. A customer deploying open weights has more choices about who supplies the computation. Neither arrangement guarantees NVIDIA a sale. But they distribute bargaining power differently.

The objective need not be to eliminate profitable model companies. NVIDIA can benefit from successful customers. It has a strategic interest, however, in avoiding a world where a few customers can dictate the economics of the entire platform beneath them.

Why Hugging Face matters

If useful open models are the complement, Hugging Face helps make that complement usable.

Its Inference Providers documentation describes access to multiple providers through common clients and routing. The product connects model selection with execution. That reduces the distance between discovering a model and paying someone to run it.

NVIDIA and Hugging Face have already experimented with a more direct connection. In July 2024 they launched a serverless NVIDIA NIM service on the Hub, exposing models optimized for NVIDIA infrastructure. That particular service was retired on April 10, 2025; the announcement now points readers toward Inference Providers. It is historical evidence of the relationship, not a current deployment recommendation.

Ownership would potentially give NVIDIA more influence over how model discovery connects to deployment: which integrations receive attention, which optimization paths become convenient, and where developers encounter ready-to-use infrastructure.

That is a strategic possibility, not evidence that Hugging Face currently manipulates those choices or that NVIDIA has announced such a plan.

The distinction matters. A repository's code can remain open while the surrounding convenience—documentation, testing, packaging, support and default deployment paths—makes one commercial platform especially attractive. Economic dependence does not require an explicit prohibition on alternatives.

The complement can also make you replaceable

Hugging Face's value is partly that it works across the industry.

Its Optimum project supports optimization across multiple hardware platforms. Its 2023 AMD partnership explicitly aimed to improve transformer performance on AMD CPUs and GPUs and integrate ROCm into its libraries. These are not merely distribution mechanisms for NVIDIA.

That creates a tension an acquisition cannot erase. The same ecosystem that makes models more accessible can help developers run them on competing hardware.

Spolsky identified a related tension in Sun's promotion of Java: software portability could weaken differentiation in the hardware Sun sold. The lesson is not that portability always destroys the supplier. It is that a strategy for cheapening adjacent layers can reach your own layer too.

For NVIDIA, the favorable outcome would be abundant model choice combined with a compelling reason to keep using its platform: performance, reliability, tools, support, or the cost of changing a production system.

The unfavorable outcome would be model abundance accompanied by sufficiently cheap hardware substitution that customers can redirect the new demand elsewhere. Google's TPU documentation provides a concrete example of an accelerator platform designed for machine-learning workloads outside NVIDIA's GPU stack. The competitive alternative is not hypothetical, even though substituting it for a particular deployment may be difficult.

Hugging Face would therefore be both a potential source of demand and a continuing test of neutrality. Restricting alternatives could damage the trust that made the platform valuable. Supporting them preserves that trust while keeping competitive routes open.

More AI does not automatically mean more NVIDIA compute

There is another gap in the easy version of this argument.

Cheaper models may increase usage. Better models and inference software may reduce the computation needed for each useful result. Other platforms may capture a larger share. These effects can happen together.

For a fixed workload category, consider this deliberately simplified accounting identity:

NVIDIA compute consumed = completed tasks × average compute per task × share served on NVIDIA.

Here, share means the fraction of tasks served on NVIDIA, and compute per task means the average for those tasks. Use a consistent measure of compute across the comparison. This describes consumption, not revenue, GPU shipments or profit. Training, workload changes, utilization and hardware generations would require separate treatment.

Suppose task volume doubles, compute per task halves, and NVIDIA's share stays constant. Compute consumption is unchanged. If volume rises fivefold under the same efficiency improvement, it grows 2.5 times. If NVIDIA's share also halves, that growth falls to 1.25 times.

These are illustrative scenarios, not market forecasts. They show why “cheaper models are good for NVIDIA” needs conditions. The commercial outcome also depends on prices, margins and how much existing capacity can serve the work.

The strategic bet is that expanding use, platform competitiveness and access to demand will outweigh efficiency gains and substitution. A model distribution platform could help with that bet. It would not settle it.

For builders, turn the theory into a substitution test

This framework is useful well below the scale of NVIDIA.

An AI application company should usually welcome better, cheaper models. But if its entire distinction is an interface around one model, the same improvement may make its product easier to reproduce.

Consider an exam-preparation product. Parsing lecture slides and generating questions are inputs. A stronger product might diagnose a student's misconceptions, choose the next exercise and measure whether the intervention improved understanding. That is a product-design example, not a claim that such a system has already demonstrated superior outcomes.

The question becomes which part of the system improves through use and which parts can be purchased from competing suppliers. I developed the learning side of that argument in The AI Product Is Not the Moat. The Operating Loop Is..

For an architecture review, I would make the complements argument concrete with four checks:

  1. Name the outcome customers pay for. A model call, retrieved document or generated answer may be an input rather than the purchased result.
  2. List the dependencies that could become interchangeable. Models, inference providers, parsers and orchestration components deserve separate entries. Sharing an API does not make their behavior equivalent.
  3. Run one real substitution. Hold an evaluation set, quality threshold, latency budget and security requirements fixed. Replace a dependency and record engineering time, regressions and total operating cost.
  4. Ask whether the same substitution could remove you. If the customer's platform can reproduce the whole outcome just as easily, cheaper inputs alone will not create a defensible business.

This is why a portability claim needs evidence. The interface is only the start; behavior, operational constraints and accumulated learning determine the actual switching cost.

What remains valuable when the surroundings get cheaper?

I read NVIDIA's interest in Hugging Face as an attempt to strengthen the connection between an abundant model ecosystem and a differentiated compute platform. That interpretation fits its existing open-model investments, but it remains an interpretation.

The harder question is whether NVIDIA can keep that platform attractive as the same ecosystem makes alternatives easier to use.

For everyone building above it, the test is similar: if every model and infrastructure dependency became ten times cheaper tomorrow, would your business become stronger—or easier to replace?

Sources