Open-weight AI models are a big deal, as Nvidia and more than 200 other companies and organizations attested in their “Open Weights and American AI Leadership” letter in July. The question is who will pay for them.
Why are they a big deal? Well, former Red Hat CEO Jim Whitehurst will tell you that “open weights can play the same catalytic role [as open source]” to drive innovation in AI. This isn’t to suggest that open source, or open weights, will “win” in some absolute sense. Linux, for example, has become essential to the fabric of enterprise computing, but there are still plenty of Windows servers running. So open-weight models like Kimi or Nemotron aren’t going to topple Anthropic or OpenAI anytime soon. However, according to Whitehurst, they can “create a broader competitive landscape where innovation happens faster, and the power of AI is broadly shared.”
Again, only if we can figure out how to pay for them.
That was the concern I expressed to Whitehurst, much more politely than when he became Red Hat CEO back in 2007, and it’s something the industry must confront. Or maybe not. Open source, for example, has always depended on widespread, ever-evolving corporate self-interest to thrive. At any given time, one company might decide it no longer had anything to gain from contributing to, say, Linux, but at the same time, another company would discover reasons it should start contributing.
Self-interest, it turns out, is one heckuva drug.
I wrote in 2016 that “there’s no money in open source.” That’s still true in 2026, both for open source and open weights. In open weights, the smart money isn’t that any particular company will keep funneling copious quantities of cash into training new, soon-to-be-given-away models. It’s never wise to bet on any particular vendor’s goodwill. Instead, the bet is on the perpetuity of the overall supply of open-weight contributions. Why? Because Meta, Alibaba, DeepSeek, Nvidia, and inference companies all have different businesses that become more valuable when the model layer becomes cheaper and more interchangeable. If Meta becomes more closed, Alibaba still wants cloud consumption. If Alibaba holds back, DeepSeek or Moonshot may want global attention. Nvidia wants chip demand.
These asymmetric incentives make the system resilient, even when the costs to individual contributors may be high.
Take Meta. It’s great to see Meta CEO Mark Zuckerberg championing open source and open weights in AI manifestos, following the release of open-weight models like Llama. In 2024, Zuckerberg offered an unusually candid explanation for opening Llama. First, Meta didn’t want to depend on another company’s AI platform the way it depends on Apple’s mobile platform. Second, he reasoned that a large Llama ecosystem would produce silicon support, inference optimizations, tools, and integrations that Meta couldn’t build alone. Finally, Meta could pursue these other benefits knowing that opening Llama was unlikely to cannibalize its primary business, advertising.
Will Meta always have sufficient self-interest to release open weights to the world? Probably? The company now says it expects to train a combination of open and closed models. Fine. Even if Meta becomes more selective, there’s Alibaba, whose open Qwen releases create demand for proprietary models and cloud compute, and Z.AI (formerly Zhipu), which can release GLM-5.2 under an MIT license while charging for API access.
The incentives differ, which is, of course, the point.
Hence, as much as we may laud Meta for its open-weight contributions today, the reality is we don’t need it to keep contributing forever. We just need at least one ambitious model builder in each model generation to decide that distribution is worth more than exclusivity.
The market leader generally has the strongest reason to protect scarcity. A challenger, by contrast, tends to look to open source (or open weights) to catch up and shift the playing field in their favor. Should today’s underdog become tomorrow’s leader and pull back from openness, someone else inherits the incentive. Competition drives contribution, as it were.
Nor do open models have to beat every closed model to matter. As of March, Stanford found the best closed model was ahead of the best open model by 3.3%, while Epoch AI estimates that open models have trailed the closed frontier by roughly four months during 2026.
Four months may feel like an eternity to the labs racing at the frontier, but it’s largely irrelevant to an enterprise trying to summarize documents, classify support requests, extract data, or run a routine agent. The model doesn’t 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. As I wrote recently, we just need the Fireworks-esque companies of the world to more finely tune those open-weight models to enterprise use cases, making “good enough” open weights arguably “better than” closed frontier models, at least for particular use cases.
Vercel’s AI Gateway offers a glimpse of how that market is developing. In July, open-weight models ran 36% of its tokens while capturing 8.6% of spending. DeepSeek became the gateway’s second-largest lab by token volume, while Anthropic collected 65% of spending on 30% of tokens. One provider’s traffic isn’t the whole market, of course, but the pattern makes sense: Open models absorb an increasing share of high-volume work, while closed frontier models retain the premium workloads, which is similar to what we’ve seen happen in other markets, like databases.
Of course, weights alone aren’t a product. Someone still has to 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. That’s the ecosystem Whitehurst is calling for.
It’s also where the tension 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 meaningful revenue. Open source history suggests that once downstream companies start making serious money from someone else’s work, the original creator tends to doubt it’s reaping a fair return on its investment. We saw this with databases and cloud providers, as vendors changed licenses to defend against cloud competitors. We’re starting to see the same dynamic with model weights, perhaps ushering in the latest instantiation of so-called open core models.
Moonshot’s Kimi K3 license, for example, says 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 isn’t inevitable. 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.
So, what’s an enterprise to do? As I argued recently, enterprises shouldn’t bet the company on any one model, open or closed. They should build evaluations that reflect their actual work, preserve the ability to move their data and tuning, keep application logic from becoming needlessly dependent on one provider, and read the license before confusing downloadable with unrestricted.
After all, technical portability without legal portability isn’t really portability. It’s lock-in.
We can reasonably bet on more and better open-weight models. We just shouldn’t bet on Meta, Alibaba, DeepSeek, or anyone else remaining permanently committed to openness. Nor do we need to. If open source history is any guide, open weights will endure because somebody will always profit by breaking rank with their closed-model cousins. Even so, it’s worth paying attention to ensure that “open weights” doesn’t come to mean quasi-open or pseudo-open, which happened in open source. But there again, there’s always reason for someone to go fully open as it seeks to outflank competitors. Competition is openness’ best friend.