Open-weight AI models have become a defining force in the artificial intelligence landscape. In July, Nvidia and more than 200 other companies and organizations signed an open letter titled “Open Weights and American AI Leadership,” underscoring the growing importance of making powerful AI models freely available. The letter argues that open weights can spur innovation, foster competition, and keep the United States at the forefront of AI development. But behind the enthusiastic endorsements lies a practical question that the industry cannot ignore: who will pay for these models?
Former Red Hat CEO Jim Whitehurst has argued that open weights can play the same catalytic role for AI that open source played for software. Open source did not eliminate proprietary software, but it created a broader competitive landscape where innovation happened faster and the benefits were widely shared. Similarly, open-weight models like Kimi, Nemotron, or Llama are unlikely to topple leading proprietary systems from OpenAI or Anthropic in the near term. Yet they can expand the arena, lower barriers to entry, and give a wide range of organizations the ability to build and customize AI applications.
The economics of self-interest
Open source has always thrived on corporate self-interest. Companies contribute to projects like Linux not out of altruism but because they gain strategic advantages. At any given moment, one firm may decide it no longer benefits from contributing, while another discovers new reasons to get involved. This constant churn of incentives has kept open source alive for decades. The same logic applies to open weights.
In 2016, the prevailing wisdom was that “there’s no money in open source.” That remains true in 2026 for both open source and open weights. The smart bet is not that any particular company will keep pouring money into training models that it then gives away. Rather, the bet is on the perpetuity of the overall supply of open-weight contributions. Different players have different reasons to keep the ecosystem alive. Meta wants to avoid dependency on rival platforms. Alibaba wants to drive cloud consumption. DeepSeek and Moonshot want global attention. Nvidia wants to sell more chips. These asymmetric incentives make the system resilient, even when individual contributors face high costs.
Take Meta as an example. CEO Mark Zuckerberg has been a vocal champion of open source and open weights, releasing models like Llama to the world. In 2024, he offered an unusually candid explanation for this strategy. First, Meta did not want to depend on another company’s AI platform the way it depends on Apple’s mobile platform. Second, a large Llama ecosystem would produce silicon support, inference optimizations, tools, and integrations that Meta could not build on its own. Finally, Meta could pursue these benefits because giving away Llama was unlikely to cannibalize its primary business: advertising.
Will Meta always see enough self-interest to release open weights? Probably, but not necessarily. The company now says it expects to train a combination of open and closed models. Even if Meta becomes more selective, there are others ready to fill the gap. Alibaba’s Qwen releases create demand for its cloud services. Z.AI, formerly Zhipu, offers GLM-5.2 under an MIT license while charging for API access. The incentives differ, which is precisely the point.
As much as we may praise Meta today, the industry does not need any single company to keep contributing forever. It only needs at least one ambitious model builder in each generation to decide that distribution is worth more than exclusivity. Market leaders generally have the strongest reason to protect scarcity. Challengers, by contrast, tend to embrace openness to catch up and shift the playing field. If today’s underdog becomes tomorrow’s leader and pulls back, someone else inherits that incentive. Competition drives contribution.
Good enough is enough
Open models do not have to outperform every closed model to be valuable. As of March, Stanford researchers found that the best closed model led the best open model by just 3.3%. Epoch AI estimates that open models have trailed the closed frontier by roughly four months during 2026. Four months may feel like an eternity in a fast-moving field, but it is largely irrelevant to an enterprise trying to summarize documents, classify support requests, extract data, or run routine agents. The model does not need to be the best in the world. It needs to clear the company’s evaluations at the right price, latency, and level of control.
Inference companies are increasingly fine-tuning open-weight models for enterprise use cases, making “good enough” open models arguably “better than” closed frontier models for specific tasks. Vercel’s AI Gateway provides a window into this evolving market. In July, open-weight models processed 36% of the gateway’s tokens while accounting for only 8.6% of spending. DeepSeek became the second-largest lab by token volume, while Anthropic captured 65% of spending on 30% of tokens. One provider’s traffic does not represent the entire market, but the pattern is telling: open models absorb a growing share of high-volume work, while closed frontier models retain premium workloads. This mirrors trends seen in other markets, such as databases, where open source options like PostgreSQL coexist with commercial offerings.
‘Open core’ comes to open weights
Weights alone are not a product. Someone must make them fast and reliable, serve them efficiently, help customers evaluate and adapt them, and make a newly released model available in production on day one. This ecosystem is precisely what Whitehurst envisions. But it is also where tensions will grow. A model creator wants inference providers to broaden adoption, optimize performance, and create demand. It may become less enthusiastic once those providers start capturing significant revenue. Open source history offers a warning: when downstream companies begin making serious money from someone else’s work, the original creator often starts to question whether it is getting a fair return. This dynamic played out in the database world, where vendors changed licenses to defend against cloud competitors. Something similar is now happening with model weights, ushering in a new version of “open core.”
Moonshot’s Kimi K3 license, for example, states that a model-as-a-service provider with more than $20 million in annual revenue must negotiate a separate agreement. MiniMax M3 similarly requires prior authorization once products or services built on the model exceed $20 million in annual revenue. By contrast, DeepSeek V4 and GLM 5.2 use permissive MIT licenses. Restrictive licensing is not inevitable, but the market is splitting between models designed to commoditize the entire model layer and models designed to win adoption while preserving tollbooths around the most valuable commercial uses. The weights may remain downloadable even as the right to build a large business with them becomes less open.
These restrictions are not necessarily a bad thing. They reflect the reality that creating cutting-edge models is expensive. If no one can monetize open weights, the supply will dry up. The challenge is balancing openness with sustainability. Some companies choose a permissive license because their business model depends on widespread adoption. Others choose a more restrictive approach because they want a share of the value created downstream. Both strategies are legitimate, but they create a complex landscape for enterprises to navigate.
Bet on supply, not the supplier
What should an enterprise do in this environment? The safest approach is not to bet the company on any single model, whether open or closed. Organizations should build evaluations that reflect their actual work, preserve the ability to move data and tuning, keep application logic from becoming unnecessarily dependent on one provider, and read the license carefully before assuming that “downloadable” means “unrestricted.” Technical portability without legal portability is not real portability. It is lock-in.
The industry can reasonably bet on more and better open-weight models in the years ahead. The forces that drive open source—asymmetric incentives, competitive pressure, and the constant emergence of challengers—are just as strong in the AI market. But no one should bet on any specific company remaining permanently committed to openness. Nor is such a bet necessary. If open source history is a guide, open weights will endure because somebody will always profit by breaking rank with closed-model incumbents. Competition is openness’s best friend.
As the AI landscape evolves, the definition of “open” will likely face pressure. The term “open source” has been stretched and diluted over the years, and the same could happen to “open weights.” Some vendors may label their models open while attaching conditions that undermine practical freedom. That is why enterprises must remain vigilant. They should look beyond marketing language and examine the actual license terms. They should also consider the broader ecosystem: Is there a community of developers? Are there multiple providers offering the model? Is the model portable across clouds? These factors matter more than the reputation of the original creator.
Ultimately, the future of open-weight AI depends on a delicate balance of incentives. No single company will carry the torch forever, but as long as there are challengers looking to disrupt incumbents, openness will retain its strategic value. For enterprises, the opportunity is clear: open weights offer flexibility, control, and cost efficiency. The risks are manageable if approached with careful planning. By staying agnostic and prioritizing portability, organizations can reap the benefits of open-weight innovation without becoming hostages to any vendor’s changing strategy.
Source: InfoWorld News