Anthropic shutdown fuels case for open-source AI models
TL;DR:
- Anthropic’s sudden suspension of its top models drove home that access to closed AI can be cut off at any time.
- Companies are increasingly weighing open-source models they can download, run and control on their own infrastructure.
- The open models gaining the most adoption are largely Chinese — a strategic problem for the US and its allies.
Anthropic’s decision to disable its most advanced models last week has crystallised a hard lesson for the firms that depend on them: access can vanish overnight. The episode has sharpened interest in open-source AI — downloadable models that run on a company’s own servers, where “no political fight can switch it off”, as CNBC put it — and reframed who really holds the advantage in the AI race.
The pull towards models you own
Even Microsoft CEO Satya Nadella, a principal backer of both OpenAI and Anthropic, warned that firms need to “build agentic systems that improve over time, while still retaining control over their IP”. Investors took the cue: Chinese open-source labs MiniMax and Zhipu surged as the dispute spotlighted self-hostable models. Applied Compute’s Yash Patel said the fight “highlighted the significance of owning your own model”, with demand for a “multimodal future” free of single-vendor lock-in rising sharply. Cost is accelerating the shift too — as state-of-the-art pricing climbs, firms route routine work to cheaper models and reserve premium ones for the hardest tasks.
The catch is uncomfortable: the open models winning adoption — from DeepSeek, Tencent, Xiaomi and MiniMax — are predominantly Chinese, even as Washington and Beijing battle over AI’s future. That tension matters acutely for the UK, which has just watched US export curbs strip away access to Anthropic models and is reviewing its own dependence on foreign systems. DSIT has backed open-source AI with compute and mentoring, but Britain has no frontier open model of its own.
Looking forward
The strategic bind is plain: open-source offers the control that closed-vendor curbs have made newly valuable, yet the strongest open options are Chinese — hardly a comfortable hedge for UK firms wary of one geopolitical dependency only to acquire another. With ChatGPT barely four years old, the market is young enough that today’s valuations may mislead. For UK organisations, the practical move is to design for portability now — so that switching models, open or closed, is an engineering decision rather than an existential one.