In March 2026, NVIDIA released Nemotron 3 Super at GTC and scored 60.47% on SWE-Bench Verified, beating every open-weight model ever tested on code generation. The model is free. Anyone can download it, run it, fine-tune it, deploy it commercially.

Two weeks earlier, Alibaba released Qwen 3.5 9B, a model with 9 billion parameters that scored 81.7 on GPQA Diamond, a benchmark designed to test graduate-level reasoning. That score matched models 13 times larger. Also free.

Meta continues shipping Llama variants. Mistral ships open models from Paris. DeepSeek ships from China. The cadence is now measured in weeks, not quarters. And the performance gap between the best open-weight models and the best closed models is narrowing with every release.

This is not generosity. It is strategy. And it is rewriting the economics of every company that charges for AI inference.

Why NVIDIA Gives Away Models

NVIDIA made $130 billion in revenue in fiscal 2025, almost entirely from selling GPUs. The company does not need to charge for models. It needs people to run models. Every developer who downloads Nemotron and deploys it on a GPU cluster is an NVIDIA customer. Every startup that fine-tunes an open model for production creates demand for H100s, B200s, and whatever comes next.

Giving away a frontier model is the cheapest customer acquisition strategy in the AI industry. The training cost of Nemotron 3 Super is a rounding error against NVIDIA’s R&D budget. The downstream GPU purchases it generates are not.

The same logic applies to NVIDIA’s open datasets, training recipes, and inference optimization tools. Every piece of open infrastructure that makes self-hosted AI more practical increases the total demand for NVIDIA hardware. Jensen Huang does not sell software. He sells the shovels. The more gold rushers there are, the more shovels he moves.

Alibaba’s motivation is different but equally strategic. Qwen models position Alibaba Cloud as the inference platform of choice for Chinese and Southeast Asian developers. Meta releases Llama to prevent any single company (read: Google) from controlling the foundation model layer. Mistral ships open models to establish European AI sovereignty and sell enterprise support contracts around them.

Every major open-weight release has a business model behind it. None of them depend on charging for the model itself.

What This Does to Closed Model Pricing

OpenAI charges $15 per million output tokens for GPT-4o. Anthropic charges $15 per million output tokens for Claude Opus 4.6. Google charges $12 per million output tokens for Gemini 3.1 Pro.

Running Nemotron 3 Super on your own B200 cluster costs the electricity, the hardware amortization, and the engineering time. For high-volume workloads (10 million+ tokens per day), self-hosting can be 5 to 10 times cheaper than API pricing. The break-even point depends on your scale, but for any company running agents at enterprise volume, the math favors self-hosting on open models.

This does not mean closed models are dead. GPT-5.4 and Claude Opus 4.6 still outperform open models on the hardest tasks: complex multi-step reasoning, novel mathematical problems, long-horizon planning. The gap is real. But it is narrowing. And for the majority of production AI workloads, which are not the hardest tasks but rather routine classification, extraction, summarization, and code completion, open models are already good enough.

The pricing pressure is visible in the data. GPT-4-class inference dropped roughly 90% in price between 2023 and 2025. That drop was driven partly by efficiency improvements and partly by competitive pressure from open alternatives. When a free model can do 80% of what a $15-per-million-token model does, the $15 model has a pricing problem.

The Winners

Application companies. Harvey, Perplexity, Glean, and every other company that sits between the model layer and the end customer benefits from model commoditization. Their costs go down. Their margin goes up. A model-agnostic application company can route routine tasks to a free open model and reserve expensive API calls for the 20% of tasks that require frontier capability. That blended cost is far lower than paying frontier prices for everything.

Enterprises running high-volume agent workloads. A company processing 100,000 documents per day through an AI pipeline saves millions annually by switching from API-based inference to self-hosted open models. The operational complexity is higher (you need ML engineers to manage the deployment), but the unit economics are dramatically better.

NVIDIA. Every shift from API-based to self-hosted inference means more GPU purchases. NVIDIA wins regardless of which model wins.

Developers and researchers. Open weights mean open experimentation. Fine-tuning, distillation, architecture modifications, and novel applications that are impossible with closed API access become standard practice with open models. The pace of AI research outside the major labs has accelerated precisely because open models provide a foundation to build on.

The Losers

Pure-play model providers without a platform. Companies that sell AI exclusively through per-token API pricing face margin compression from two directions: open models from below and platform subsidies (Google, Microsoft) from above. OpenAI and Anthropic are both racing to build platform businesses (enterprise features, agent frameworks, vertical applications) precisely because they recognize that selling tokens alone is not a durable business.

AI wrappers with no domain depth. A startup that built a product by calling GPT-4’s API and adding a nice interface has no moat when the same API call gets 90% cheaper and open alternatives emerge. The “thin wrapper” critique has been valid since 2023. Open models make it lethal. If your entire product can be replicated in a weekend with Nemotron and a React frontend, you are not a company. You are a demo.

The Speed of Commoditization

Consider the timeline. GPT-4 launched in March 2023 as the most capable model in the world. Within 18 months, open models matched its performance on most benchmarks. GPT-4o launched as the efficient frontier model in May 2024. Within 12 months, open models matched it. The cycle is compressing.

Each generation of closed model buys maybe 6 to 12 months of performance advantage before open alternatives catch up. That window is the closed model provider’s entire pricing premium. Everything the model can do after the open models catch up is priced at commodity rates.

The only sustainable advantage for closed model providers is to keep the frontier moving fast enough that the 6-to-12-month window always contains capabilities open models cannot match. OpenAI is doing this with GPT-5.4’s million-token context and autonomous workflow execution. Anthropic is doing it with Claude’s extended thinking and tool use. The question is whether they can maintain that pace indefinitely while also building enough platform value to survive when any given model generation gets commoditized.

What This Means for the Agent Economy

The agent economy described in our analysis of AI agent economics depends on inference costs for its margin structure. As open models improve and inference costs drop, the economics shift in predictable ways.

Agents built on narrow, well-defined tasks become dramatically cheaper to run. Document processing, code review, customer routing, data extraction: these are the workloads where open models already match closed models. The per-task cost drops from dollars to cents. This expands the total addressable market because tasks that were not worth automating at $5 per run become worth automating at $0.10 per run.

Agents built on complex, multi-step reasoning remain expensive because they still require frontier models. Legal analysis, financial modeling, medical diagnosis: these are the workloads where GPT-5 and Claude Opus still outperform. The application companies serving these markets (Harvey, for example) maintain pricing power because the alternative to their AI agent is not a cheaper AI agent but a $600-per-hour human.

The market splits. Commodity agents get cheap fast and scale wide. Premium agents stay expensive and serve high-value verticals. The companies that thrive are the ones that know which category their product belongs to and price accordingly.

NVIDIA, as always, sells to both sides.