Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Ever since the rise of the Linux computer operating system, open source technology’s combination of defensive and disruptive properties has made it a powerful weapon in the tech world.
IBM famously used Linux as a counterweight to Microsoft, limiting the spread of Windows and preventing Microsoft from dominating the market for computer servers as it had PCs. Free software also threatened to eat into Microsoft’s main profit-earner, prompting then-CEO Steve Ballmer to call it a cancer on commercial technology.
So why would companies that are pouring tens of billions of dollars into advanced AI models, making them most at risk from open source competition, suddenly seem so interested in launching free, open versions of their own products?
Earlier this week, OpenAI released two open models — its first since 2019 — that anyone can download and run free of charge. Not to be outdone, Elon Musk, who has sued OpenAI for its earlier decision to close off its technology, promptly declared that next week xAI would also open up a model it launched a year ago.
The cause is not hard to find. A wave of free, open models from Chinese companies, starting with DeepSeek, has spread around the world this year. In some cases, these have come close to matching the performance of the most advanced — and expensive — commercial alternatives. For US companies, neutralising this push with rival open technology has become critical.
The big unknown is how powerful a force open technology will be in the AI market. US companies are betting that these models will play only a limited, supporting role alongside commercial alternatives.
Yet nearly two-thirds of the world’s servers are estimated to run on open-source Linux. That has still left a big market for Microsoft Windows, but it has greatly limited the overall size of the commercial software pie, while also preventing operating systems from being used as a chokepoint in the tech world.
The US companies hope to defend the centrality of their commercial models in the AI market by positioning open alternatives as entry-level products. Those from OpenAI only handle text, for instance, and Musk has said he will release technology that is one generation behind xAI’s latest, most cutting-edge models.
Open technology might also help by enlarging the market. Some companies that want to use their data to fine tune AI models may hesitate to send it to an outside company: Privacy is no longer a worry if they have open models running on their own servers.
As a result, most US AI companies seem to be moving towards offering a mix of open and closed technology. Meta’s open Llama models, for instance, were launched as a defensive weapon to ensure the company wasn’t shut out by more advanced AI companies. CEO Mark Zuckerberg even cited Linux as an inspiration.
The calculation has since changed and Meta’s CEO has now gone on the offensive, seeing AI as a core technology. Its huge investment in people and hardware have led to reports that it may no longer open up its most advanced models.
Building a strategy around a combination of open and closed technologies seems to match the current market. Only around one in eight commercial AI workloads are being run on open models, according to a survey of 150 tech executives by Menlo Ventures. Most customers would rather pay to use the latest and greatest commercial technology, the VC firm said. And according to OpenAI, even customers that run open models tend to mix them with commercial ones.
None of this, though, negates the deeper, disruptive risk. Companies like DeepSeek and Alibaba have shown surprising success at matching the best commercial models on some benchmarks. If US companies want to be taken seriously, they will have to match them, not hold back all their best ideas for their closed models.
AI also faces the risk of all new technologies: That as they mature, customers no longer see the need to pay up for the latest features and turn to cheaper or even free alternatives that seem “good enough”.
OpenAI’s decision to put the “open” back into its name is likely to be quickly forgotten amid this week’s release of GPT-5, the latest iteration of its most advanced, frontier model. But pushing the frontier forward is getting harder all the time.
[email protected]