Justin Boitano of NVIDIA put the case for open models about as plainly as it can be put. A great deal is happening at the frontier, he argued, but if you can build open models that land near it, you unlock new use cases across whole industries. The stated goal of Nemotron is to give every company an open foundation it can adapt to its own domain and turn into new business.
For anyone holding the frontier laboratories, that framing is uncomfortable, and it deserves a straight answer rather than a reflexive one. If near-frontier capability becomes something an enterprise downloads and tunes for itself, what exactly are OpenAI and Anthropic selling? The evidence available through the middle of 2026 answers the question with reasonable clarity, and the answer is more interesting than either side of the argument usually allows.
Start with who is making the argument
NVIDIA earns its margin in silicon, not in weights. Its commercial interest is that the model layer becomes abundant, cheap and undifferentiated, because every enterprise that adapts an open checkpoint still has to buy accelerators to train and serve it. Commoditising the layer above you is a well-worn move when you own the layer below, and it does not make the argument wrong. It does mean the argument should be tested against usage data rather than accepted as neutral forecasting.
Open weights are winning volume, and losing value
The single most useful statistic in this debate comes from token-share and spend-share data on Vercel’s hosting platform for June to August 2026. On token volume, DeepSeek ran at roughly 25.2 per cent against Anthropic’s 24.5 per cent, so open weights genuinely do process comparable quantities of traffic. On money, the same dataset put Anthropic at approximately 64.6 per cent of model spend against DeepSeek’s 2.8 per cent.
That gap contains the whole argument. The open tier is absorbing high-volume, low-value workloads such as classification, extraction, routing and summarisation, where a model a few points behind the frontier is entirely adequate. The closed tier retains the work that people are willing to pay for. Anyone quoting open-weight token share as evidence of commoditisation is measuring the part of the market that was always going to be cheap.
Enterprise adoption moved the other way
Menlo Ventures found enterprise adoption of open-weight models falling from 19 per cent to 11 per cent year on year, with Chinese-origin models accounting for roughly 1 per cent of enterprise LLM API usage despite considerable developer enthusiasm outside the enterprise. Within the same survey, Anthropic held about 40 per cent of enterprise LLM spend, ahead of OpenAI at 27 per cent and Google at 21 per cent, with the remainder spread across Meta, Cohere, Mistral and others.
The reason matters more than the ranking. That position was attributed largely to coding, where Anthropic was estimated to hold roughly 54 per cent share, and where buyers proved distinctly price-insensitive when capability was on the line. Enterprise open-weight share was therefore falling over precisely the period in which the open-weights thesis was being argued most loudly.
The revenue trajectory shows no commoditisation yet
Anthropic’s annualised run-rate moved from roughly $9bn at the end of 2025 to $47bn in May 2026 and approximately $65bn by the end of July 2026, with investors reportedly expecting the year to close somewhere between $100bn and $120bn. Whatever downward pressure open weights exert on price, it has not yet shown up as a constraint on growth at the frontier. Three years of the commoditisation thesis have not appeared in the revenue line, which does not make the thesis wrong, but does raise the bar for treating any single executive remark as new information.
How near-frontier is Nemotron, in fact?
The phrase “near frontier” is doing a great deal of work. Nemotron 3 shipped as Nano at 30bn parameters, Super at 100bn and Ultra at 500bn, using a hybrid Mamba-Transformer design with mixture-of-experts routing, sparse activation and a one-million-token context window. Independent benchmarking placed the Nano variant roughly level with gpt-oss-20B and Qwen3 VL 32B, which describes a competent open model rather than anything adjacent to the frontier. Partners have marketed the family on an inference-cost advantage of up to twenty times against closed APIs, which tells you the pitch is about unit economics on commodity workloads rather than about capability parity.
The unit is commoditising; the business is not
Most bear cases collapse these two things, and the distinction is where the analysis actually lives.
The unit is unambiguously commoditising. The price of a fixed level of capability has fallen steeply and continuously, capability has converged across the leading closed vendors to the point where benchmark differences are frequently within noise, and the lag between the frontier and the best open weights has compressed to something on the order of six to twelve months. That is the textbook shape of a commoditising input.
The business has not commoditised, because two effects have more than offset the falling unit price. Cheaper inference has pulled far more work into scope rather than merely reducing the bill for existing work, and the marginal token has migrated towards long-horizon agentic tasks in which reliability compounds across many steps. A model two points behind on a static benchmark can be dramatically worse at completing a fifty-step task, and that is where the premium now sits.
Where the moat has moved
The defensible layer is no longer the weights, which are replicable, nor the benchmark score, which converges. It has migrated into the harness around the model: the agent scaffolding, tool-use reliability, the evaluation suites a customer has built against one model’s specific failure modes, procurement and data-residency arrangements, and above all the product surfaces through which the model is consumed rather than called. A dominant share of coding is not a weights advantage. It is a position in a workload where a customer who has tuned an agentic pipeline around one model’s behaviour does not re-tune on a whim.
The real risk is not the open tier
The more serious threat to model-layer economics is not NVIDIA and not open weights. It is that three or four exceptionally well-capitalised closed vendors sit at near-parity on capability with abundant compute behind them, which is the structure that produces price competition irrespective of how good the technology is. Open weights set a floor under commodity inference, and that floor is rising quickly, but a floor only binds if the frontier stops moving upward faster than the floor rises.
What would actually signal commoditisation
Four indicators are worth watching, and none of them had turned on the data available through mid-2026. The first is enterprises multi-homing at the workload level rather than the portfolio level, because routine substitution is the real test of fungibility. The second is price cuts at constant capability, as distinct from price cuts that accompany a capability upgrade. The third is gross margin per token compressing at the laboratories. The fourth is erosion of share specifically in coding and agentic workloads, which is where the premium currently lives.
Bottom line
Open weights are commoditising the price of intelligence per token while the frontier laboratories continue to capture the value, because the workloads that pay have moved to where the open tier is not yet competitive. NVIDIA’s argument is directionally sound about the input and wrong so far about the business. The thesis becomes investable the moment agentic and coding workloads become good enough at the open tier, and there is no evidence of that in the spend data yet.
Sources
Menlo Ventures, “2025: The State of Generative AI in the Enterprise”, December 2025. Token-share and spend-share data drawn from Vercel platform traffic, June to August 2026, as reported by MindStudio. TechCrunch, “Anthropic’s annualized revenue surges to $65B”, August 2026. The Register, coverage of the Nemotron 3 open-weights release, December 2025. OpenRouter, “The Open Weight Models that Matter”, June 2026.
