Corporate America is turning to lower-cost “open” AI models, as spiralling IT expenses push executives to look for alternatives to Anthropic and OpenAI’s frontier systems.
So-called open-weight models, which users can customise and run on their own hardware, can be far cheaper for businesses to run than paying for access to the most advanced versions of ChatGPT or Claude.
Executive mentions of “open weight” or “open source” models in earnings calls and investor conferences surged sixfold in August and September compared with the same two-month period last year, according to data from AlphaSense, the research platform.
While the shift is most pronounced among technology companies, a cross-industry swath, including PNC Financial Services, logistics group CH Robinson and industrial giant Siemens, have discussed using open-weight models in recent weeks.
The open-weight AI market has become dominated by Chinese companies, though capable models released on similarly permissive terms are also being developed by France’s Mistral and US-based Nvidia, Reflection AI and Thinking Machines Lab.
Public market investors are closely tracking how corporate customers are allocating AI spending between frontier models and open-weight competitors.
Rapid improvements in these cheaper AI systems this year could threaten the revenue growth of Anthropic, which is preparing for an initial public offering expected to value the company at $2tn or more, and OpenAI, which is in early talks for a private funding round at a $1.2tn valuation after delaying its own IPO until 2027.
For now, the two frontier AI labs capture the bulk of AI spending, and their voracious demand for computing power has underpinned much of the data centre construction boom buoying the US economy.
But usage data from groups in the AI supply chain show the shift is more than just talk. In August, open-weight models accounted for 56 per cent of all tokens processed through Vercel’s AI Gateway, up from 7 per cent in December. A similar pattern is visible on OpenRouter, where open-weight models from Chinese labs dominate the top 10 by tokens processed.
Vinay Kuruvila, chief technology officer at Tinder, the dating app owned by Match Group, said it had begun routing some queries from non-technical users to open-weight models as the app worked to get control of rising AI costs.
“In January we were spending at the rate of $1mn per year and by July it had climbed to $10mn . . . I don’t want another 10X increase,” he said. “The frontier models like OpenAI’s Astra and Claude Fable already have enough intelligence for 90 per cent of the tasks we’re trying to do. If open-weights models catch up, I may not need to use them anymore.”
Open models can often deliver comparable outputs to the closed systems sold by Anthropic and OpenAI. The difference is that their underlying parameters, which determine how an AI model responds to queries, are released publicly.
This allows companies to download, host and fine-tune them without paying per-token fees to the AI model’s developer. Users can also buy tokens directly from the open-weight providers, typically at far lower prices than those charged by OpenAI and Anthropic, which is how companies such as DeepSeek and Zhipu, two of the leading Chinese open-weight labs, generate revenue.
Unlike Anthropic and OpenAI, whose closely guarded models are their primary commercial product, rivals building open-weight tools might distribute those for free in order to lock users into particular hardware or to sell advertising. Both US companies released cheaper versions of their flagship models last week as the price war with Chinese rivals intensifies.
The open-weight shift extends well beyond Silicon Valley. AT&T now runs about 40 per cent of its AI workloads on open models and aims to reach 70 per cent within a year, according to Andy Markus, the telecom group’s chief data and AI officer.
“We’re doing 45bn tokens a day,” Markus said. “At that scale, costs become super important. If we can move that traffic from a super expensive closed-source model to an open model that we can often serve at [a fraction] of the cost, and still maintain the accuracy, then we’ll do that.”
Markus said AT&T took open models and tuned them using the company’s proprietary data so they performed as well as, or better than, closed alternatives for specific tasks. “The open models are getting better and better,” he said. “That’s giving us more optionality.”
For some companies, cost is only part of the equation. Scott Wallace, senior global director of solutions architecture at data centre operator Digital Realty, said sovereignty and security were equally important drivers of his company’s embrace of open models hosted on private infrastructure.
Wallace said Digital Realty had built its own internal chat interface running on open-weight models and used a mix of open and proprietary systems depending on the sensitivity and complexity of the task.
“There’s proprietary data that we’re putting into this private model that I wouldn’t put into a frontier model,” Wallace said. “Any customer data, never, never, never goes into a frontier.”