Open vs Closed AI Models: How Each Business Model Makes Money

Closed AI models make money directly, by charging for access through APIs, subscriptions and enterprise contracts. Open-weight models make almost no money directly: the weights are a free sample, and the releaser earns elsewhere, from advertising, cloud rental, chips, enterprise licences or talent. Closed labs keep the margin at the frontier; open models put a ceiling on what anyone can charge for capability the open frontier has reached, which pushes value toward users and chip makers.

The clash became a market event on Monday 27 January 2025. A week earlier DeepSeek, a Hangzhou lab of about 160–200 people (reported), had released R1, a reasoning model that roughly matched OpenAI's o1 on mathematics and coding benchmarks, under the permissive MIT licence, with an API priced at $0.55 per million input tokens and $2.19 per million output against $15 and $60 for o1. Nvidia's shares fell about 17% that day, a loss of about $589 billion of market value. The fear passed within weeks. The lesson about what open models do to prices did not.

What is the difference between open and closed AI models?

A closed model is available only through its owner's API or apps. The owner sets prices, rate limits and terms, and can retire or change the model. OpenAI's GPT models and Anthropic's Claude are the leading examples.

An "open" model usually means open weights: the trained parameters are published for anyone to download, run and fine-tune. The training data and much of the code usually stay private, so a user can run and modify the model but cannot rebuild it. Openness also comes in grades, and the licence is a commercial term, not a formality.

How do closed AI models make money?

Closed labs charge for capability while it is scarce. OpenAI earned about $13.1 billion in 2025 (reported), mostly from ChatGPT subscriptions, and added advertising to free tiers in February 2026. Anthropic earned about $16.3 billion in the first half of 2026, about 80% of it from API usage (from S-1 coverage). Their pricing power depends on staying far enough ahead of the open frontier, on enough valuable tasks, to justify their prices.

The window is short. A frontier model typically leads its nearest closed rival for a few months and the best open model for perhaps six to twelve. The book's stylized model in Chapter 7 shows that a $3 billion training programme with a six-month commercial life barely covers its cost, while a nine-month life leaves about $2.5 billion to fund the next generation. Every month the open frontier closes the gap is a month of lost margin.

How do open-source AI models make money?

Nobody sells the weights. The research groups releasers by what they sell instead.

Around the weights sits an economy of hubs, hosts and tools. Hugging Face, where nearly every lab publishes, had more than 3 million public models and 18 million users by September 2026, yet its revenue was reported at about $150 million annualized in mid-2026 (The Information). Nvidia agreed to buy it for about $12.93 billion in total on 2 September 2026. Inference hosts such as Together AI and Fireworks serve identical weights, so their prices converge on cost. The research's summary is that the most valuable activity, training the model, earns least directly.

Mistral shows what an open-first lab can build. Founded in Paris in April 2023, it made its name with Apache-licensed models, then built a business on commercial models, an assistant and sovereign deployments. It claimed a run-rate above $400 million in February 2026 and raised €3 billion at more than €21 billion in September 2026.

Open vs closed: which is cheaper to use?

For a buyer, open models change the price of commodity work. The book compares three options in a stylized example: a closed frontier API at a blended $3.00 per million tokens, the same open model from a hosting firm at $0.60, or self-hosting it on rented GPUs at $2.20 an hour with a minimum of eight GPUs and a $50,000-a-month engineering team.

Self-hosting beats the closed API above about 19 billion tokens a month, but beats a hosted open model only above about 190 billion, because the host's price already sits close to cost. For a company with volume, a capable open model cuts the price of commodity work by about 80% before any engineering. That ceiling, not any loss of customers, is how open models discipline closed labs.

The closed model still wins where it is meaningfully better: at the leading edge, in the newest reasoning models, coding, agents and reliability. Commoditization happens at the trailing edge.

Open vs closed: side-by-side comparison

Why do AI labs switch between open and closed?

The research identifies a reliable pattern: labs open while they trail and close when they lead. The flagship is rarely open.

Meta is the clearest case. It released LLaMA to researchers in February 2023, and the weights leaked within about a week. Having no chatbot, cloud or model API of its own, it embraced openness: Llama 2 in July 2023 with a commercial licence, and Llama 3.1 405B in July 2024 with Mark Zuckerberg's essay "Open Source AI Is the Path Forward". Llama downloads passed 1 billion by March 2025 (company claim). Then Llama 4, released in April 2025, had a disappointing reception and became shorthand for benchmark gaming. Within about ten weeks Meta paid about $14.3 billion for 49% of Scale AI and created Meta Superintelligence Labs. On 8 April 2026 it launched Muse Spark, its new flagship, with closed weights.

DeepSeek moved the other way, from a self-funded research lab toward a business. Funded for almost three years by the quantitative hedge fund High-Flyer, it took its first outside money in 2026 at a reported $52–59 billion valuation, raised its API prices for the first time in August 2026 by up to about four times on some output tokens (reported), and reported an annualized run-rate of about $1 billion (September 2026, unaudited). The lowest-priced frontier lab became a company that has to earn a return.

By 2026 the center of gravity of open models had shifted to China. Qwen passed Llama in Hugging Face downloads in late 2025 (reported), and the largest open models of 2025–26, including Moonshot's Kimi K2, DeepSeek V3 and V4, Qwen3 and Zhipu's GLM, were Chinese.

What should builders and buyers check?

The research offers six checks for any open-model strategy:

1. Ask what the releaser sells instead. Open weights pay only where they drive ads, cloud, chips, enterprise contracts or talent.
2. Locate the capability gap. Margin lives at the leading edge; commoditization at the trailing edge.
3. Read licences, not slogans. User thresholds, naming rules and acceptable-use terms differ from MIT or Apache.
4. Treat downloads with care. They are the best market-share data available, but easily inflated.
5. Price support without control. You keep the weights, but have no contract for the roadmap.
6. Watch for the turn. DeepSeek's price rise and Meta's closed flagship were both turns.

For applications, an open model is also insurance against platform risk: a lab can cut off API access with days' notice, as Anthropic did to Windsurf in June 2025. Our guide to AI wrapper startup moats covers that risk.

Which factor of the intelligence equation does openness affect?

In The Intelligence Equation, open models are a weapon aimed at the third factor of everyone else: a position the next layer cannot route around. By making last year's frontier free, Meta, DeepSeek, Alibaba and Mistral removed the position of any lab or host whose product the open frontier had reached. For the releasers themselves, openness creates durable value only when they own something the free model makes more valuable. The open-source models and open ecosystem study traces each release, and our guide to how AI companies make money shows where both models sit in the wider stack.

The takeaway

Closed models earn money directly while they lead; open models earn almost nothing directly and are funded by whatever the releaser sells instead. Open weights cap prices for any capability they reach, cutting the cost of commodity work by about 80% for heavy users, while closed labs keep the margin only at the frontier. Labs open when they trail and close when they lead, as Meta and DeepSeek both showed in 2026. Chapter 8 of the book, Open Models, explains who makes money around open weights and why.