Open-source AI is imperfect hedge against US clout

MiniMax

29/07/2026 19:41

The author is a Reuters Breakingviews columnist.  The opinions expressed are her own.
By Karen Kwok
NEW YORK, July 30 (Reuters Breakingviews) - Modifiable artificial intelligence has a new sales pitch. Large language models that users can download, change and run on their own infrastructure already promise lower per-token costs than renting proprietary technology from leading labs like OpenAI or Anthropic. The Trump administration's temporary restrictions on Anthropic's Fable give companies and governments another reason to seek alternatives. The so-called open-weight models offer an imperfect hedge.
The best-known and most capable algorithms, such as Anthropic's Claude, Alphabet's GOOGL.O Gemini and OpenAI's ChatGPT are proprietary. Users cannot freely inspect, modify or run these systems but typically pay to access them through an interface. That is expensive but convenient. Closed models are a one-stop shop with performance, chips, cloud storage, and service guarantees. Brookings reckons proprietary models are likely to keep dominating enterprise functions, national security and other sensitive uses.
Open-weight alternatives are catching up, however. Epoch AI estimates such models are only about four months behind the most capable closed systems. Kimi K3, released by Chinese startup Moonshot AI earlier this month, closely challenges Anthropic's Fable 5 in some tasks.
This makes it easier for users to route tasks according to cost, risk and performance, reducing their reliance on Anthropic or OpenAI. For narrow business-specific problems that draw on proprietary data, a smaller tailor-made system may equal or beat a general frontier model, especially after fine-tuning.
For companies, the main attractions of open-weight models are economic flexibility and privacy, as well as lower token costs. Chinese models like Z.ai's GLM 5.2 can be 60%-90% cheaper than proprietary systems like Anthropic's Opus 4.8 and OpenAI's GPT 5.5, OpenRouter's Justin Summerville reckons. But cheaper tokens do not always mean lower total costs. Large open-weight models can end up consuming more tokens, and running them privately brings extra bills for technology infrastructure and cybersecurity. For governments, the bigger issue is security and control. They want to know where sensitive data is stored, who can inspect the model, and whether access can be cut off by a foreign government.
Corporate customers are already voting with their wallets. As of early June, 30% of AI calls made by business-travel service Navan ran on systems built by fine-tuning open-source small language models. Coinbase Chief Executive Brian Armstrong says frontier models may be best for planning, but overkill for execution. He expects roughly 80% of AI workloads to shift to models that are about 99% cheaper in the next 18 months. That undermines the pricing power of OpenAI and Anthropic and pressures the revenue growth they need to justify sky-high valuations.
The beneficiaries of this shift are likely to include model routers, cloud providers, fine-tuning specialists and sovereign-AI vendors that help customers assemble cheaper, more customized systems. Nvidia-backed Baseten and Fireworks AI, last valued at $13 billion and $17.5 billion, respectively, help users build customized open-source systems and find the computing power to run them. Canada's Cohere said it would acquire Germany's Aleph Alpha with a pitch to offer governments and businesses an alternative to U.S. tech. Cohere expects its annualized revenue to grow three times faster next year and has worked with companies including Japan's Fujitsu to build custom models for government use.
The catch is that many of the strongest open models, including DeepSeek, Z.ai, Xiaomi 1810.HK, MiniMax 0100.HK, Tencent 0700.HK and Alibaba 9988.HK, are Chinese. The share of tokens used by U.S. users on Chinese open-source models via OpenRouter — a platform that enables developers to access a range of AI models — hit 64% in the week of July 13.
It is a difficult balancing act. Users switching to models developed by Chinese labs cannot be sure that future updates or restrictions will remain politically neutral. Hidden security risks may surface only after deployment, especially if models generate or execute code. Washington is already debating whether to restrict, or effectively ban, some Chinese open-weight models in the United States, Reuters reported, citing sources. Beijing is similarly discussing limits on overseas access to its own advanced models, Reuters has reported. Open-source AI, therefore, risks becoming another contested strategic technology, much like advanced chips made by Nvidia NVDA.O.
Western firms are racing to offer alternatives. Giant hedge fund Bridgewater Associates used Tinker, a new open-source model released by U.S.-based Thinking Machines, to build a custom version of Alibaba's Qwen, which outperformed top proprietary models at a lower cost. Meta's Llama, Europe's Mistral and Nvidia's model offerings are part of the same shift. Consultant McKinsey reckons the market for sovereign models, data and tooling could reach $100 billion to $140 billion by 2030.
Governments face the hardest dilemma. Most states lack the money, computing power and expertise to train competitive models independently. Countries that lack capital may trade high-quality domestic data for access, as Iceland did in its partnership with OpenAI for language preservation.
Richer countries can go further. India's information technology ministry chose local startup Sarvam AI to develop the country's first indigenous foundation model. South Korea provides funding, processing power and data-centre support under a national "Sovereign AI Foundation Model" project. Japan has bought thousands of Nvidia's latest chips to build homegrown foundation models for robotics and other uses.
Even so, sovereignty rarely means full independence. Countries still rely heavily on adapting American-made open-weight models. Data from the Center for a New American Security shows that Meta’s Llama family appears in 36% of projects, ahead of France’s Mistral and Alibaba’s Qwen.
Countries outside China and the United States may therefore need to cooperate. Project Tapestry, backed by the AI Alliance and advised by former Meta chief scientist Yann LeCun, aims to let countries and institutions co-train an open base model while keeping sensitive data local and building sovereign derivatives. These efforts may increase control, but they will take time.
Open-weight AI will therefore not kill proprietary models, just as the open-source Linux operating system did not spell the end for Microsoft or Oracle. But the shift may reduce the routine enterprise work customers send to OpenAI and Anthropic, crimping their pricing power. What began as a choice between cheap open models and expensive closed ones is becoming a contest between competing forms of dependence. Companies and governments will discover that a perfect AI hedge remains expensive and elusive.
Follow Karen Kwok on LinkedIn and X.

(Editing by Peter Thal Larsen; Production by Pranav Kiran)
((For previous columns by the author, Reuters customers can click on KWOK/karen.kwok@thomsonreuters.com))