The most valuable AI companies may discover that the model itself is not where the durable margins are.
For the past few years, much of the AI industry has behaved as though the model were the product: build the smartest one, put it behind an API and charge by the token. It is a seductive proposition: intelligence as a metered utility, supplied by the small group of laboratories with enough capital and talent to manufacture it.
Open-weight models make that story messier.
An open-weight model makes its trained parameters available for others to download and run. That does not necessarily make it “open source” in the conventional sense. The training data and training code may remain private, and the license may restrict certain uses. Researchers writing in Nature have made this distinction forcefully: releasing weights is only one dimension of openness and does not, by itself, make a system transparent or reproducible.
Still, access to the weights changes the buyer’s position. A company can run the model in its own environment, adapt it to a narrow task, move between hosting providers or keep a working version if the original developer changes its product. That is why I think open weights are, on balance, good for business. They turn AI from something a company can only rent into something it can operate and shape.
The awkward corollary is that they also make it harder to see how frontier labs such as OpenAI and Anthropic sustain software-like margins while relying on infrastructure built at industrial scale.
The business case for open weights
The first advantage is control. If a model handles customer records, source code or internal research, sending every prompt to an outside API creates another dependency to govern. Self-hosting is not a magic privacy shield; it usually creates additional security and maintenance work, but it does let a company decide where its data travels, how long it is retained and exactly which model version is running. In a peer-reviewed study of low-resource deployments in government, research and health care, researchers found that smaller open-weight models could be competitive with GPT-4 Turbo on selected tasks while offering practical benefits in privacy and adaptability. That does not settle the question for every workload, but it shows that local deployment is more than a theoretical option.
The second advantage is customization. A general model must be competent at thousands of tasks; most businesses need it to be excellent at a handful. With access to the weights, a company can fine-tune, compress or otherwise adapt a model to its own language and workflow. OpenAI’s own gpt-oss models are a useful example. The company released them under the Apache 2.0 license and describes them as customizable and suitable for local deployment. Even the lab most closely associated with closed frontier models now treats openness as a product feature, not merely an ideology.
The third advantage is bargaining power. A company built entirely on one proprietary API inherits that provider’s prices, limits, policies and roadmap. Open weights create a credible exit option. The broader economics literature shows that switching costs can weaken competition and leave customers worse off; compatibility and portability can work in the opposite direction. In AI, a portable workload does not eliminate every switching cost, but it makes a multi-model architecture far more plausible.
Then there is cost. An OECD analysis comparing API use with private hosting found that self-hosting did not pay for itself at small volumes. Under the report’s assumptions, a medium workload of one billion tokens a month took about 30 months to break even. At 10 billion tokens a month, the estimate fell to roughly two months; at 50 billion, it was about one month. Those exact numbers will move with hardware prices, utilization and model choice. The important point is the shape of the curve. Open weights do not make compute free; they allow a sufficiently large buyer to choose who captures the operating margin.
“Free” models are not free operations
The strongest argument against open weights is operational reality.
Downloading a model is easy. Running it reliably is a different business. Someone has to provision accelerators, optimize inference, monitor quality, patch vulnerabilities, evaluate new releases and build safeguards around the system. A managed API turns much of that fixed burden into a variable expense. For a startup or a modest internal workload, paying a frontier lab may be cheaper than building an inference team, even when the per-token price looks high. The OECD’s cost analysis points to exactly this conclusion at lower volumes.
Open models can also lag the frontier. Stanford’s 2026 AI Index reports that, after narrowing sharply, the measured gap widened again during 2025. As of March 2026, the top closed model led the top open model by 3.3 percentage points on the report’s aggregate measure, up from 0.5 points in August 2024. A few points may sound trivial until they determine whether an agent completes a workflow or hands it back to a person. And benchmark scores are only part of the purchasing decision: reliability, tool use, latency and support can matter more than the license.
There is a real governance tradeoff as well. An API provider can investigate abuse, update safeguards and revoke access. Released weights cannot be recalled. The US National Telecommunications and Information Administration’s review of widely available model weights found both meaningful benefits in competition, innovation and research, and risks that become harder to mitigate after release. OpenAI makes the same point in its gpt-oss model card: a determined actor can fine-tune a downloadable model to bypass refusals, beyond the developer’s ability to revoke access or add server-side safeguards.
None of this means every company should self-host. The more defensible claim is that businesses benefit from an ecosystem in which self-hosting is possible. Open weights create an outside option; managed APIs remain attractive when convenience, frontier performance or low initial cost matters more.
The capex problem is really a margin problem
This brings us to OpenAI and Anthropic. Their growth is extraordinary, although the figures available to the public are company-reported rather than audited financial statements. OpenAI says its annual recurring revenue rose from $2 billion in 2023 to more than $20 billion in 2025. Anthropic said in May 2026 that its run-rate revenue had crossed $47 billion.
The compute bill is extraordinary too. Stanford’s AI Index, drawing on Epoch AI estimates, puts OpenAI’s 2025 compute spending at $16.3 billion and Anthropic’s at $6.8 billion. The report stresses that these figures largely represent rented cloud capacity, not data centers owned by the labs. The distinction matters: the labs do not carry all the capital expenditure themselves. But their businesses sit on top of it, and the whole stack, from chips and power to cloud capacity, eventually needs a return.
If model quality converges and capable weights can be hosted by many suppliers, raw inference starts to look like a commodity. Competition pushes token prices toward the cost of compute. That is good for customers and uncomfortable for a company valued on the assumption that intelligence itself will remain scarce.
Commoditization does not mean OpenAI and Anthropic cannot build enormous businesses. It means they probably cannot rely on the weights alone.
The more durable business may sit one layer higher: consumer subscriptions, enterprise software, coding tools, agents, identity, security, compliance and integrations into everyday work. Companies still pay cloud providers despite the existence of Linux, open databases and other widely available infrastructure. What they buy is convenience, reliability, support and an integrated product. Frontier labs can sell the same things, provided they become indispensable products rather than expensive model endpoints.
They may also retain a frontier premium. A closed model that is meaningfully better at a high-value task can command a high price even when cheaper systems handle routine work. The problem is that the premium may be temporary. Today’s frontier capability becomes tomorrow’s distilled, optimized or open alternative. Each new generation must create enough valuable demand, quickly enough, to justify the next round of spending before competitors catch up.
That is a demanding treadmill. It favors companies with distribution, enterprise relationships, distinctive data and the discipline to route each task to the cheapest model that can complete it. It does not necessarily favor whichever laboratory trains the largest model.
Intelligence may be abundant; good products are not
My working thesis is that open weights will be good for the AI economy while making the pure frontier-model business less attractive. They spread capability, reduce lock-in and let more of AI’s gains accrue to the companies applying it. A broad academic review led by Stanford researchers reaches a similar, if more qualified, conclusion: open foundation models can distribute decision-making power, support innovation and research, and enable scrutiny, though weights alone do not guarantee any of those outcomes.
Open weights also expose a tension at the center of the boom. The industry is financing intelligence as though it were a scarce natural resource just as the technology is making useful intelligence easier to reproduce and substitute.
OpenAI and Anthropic can still become enormous businesses. But their most defensible assets may not be their model weights. They may be their brands, distribution, interfaces, enterprise trust and ability to turn unreliable intelligence into dependable work.
That is the paradox of open weights: they make the technology more useful by making the technology itself less special. For businesses adopting AI, that is mostly a feature. For the companies spending billions to stay at the frontier, it is the question their business models now have to answer.