US companies shift to cheaper open AI models as costs spiral

Key Points

1. Open-weight AI models are taking share from OpenAI and Anthropic inside large US companies.
2. AT&T runs 40% of AI workloads on open models; Tinder's AI bill jumped tenfold in seven months.
3. The shift threatens revenue growth at two labs valued in the trillions.

The latest:

Large US companies are routing a growing share of their AI work to cheaper open-weight models that can be downloaded, customised and hosted in-house, the Financial Times reported. Executive mentions of open-weight or open-source models on earnings calls and at investor conferences rose sixfold in August and September against the same two months a year earlier, according to research platform AlphaSense.

Details:

• The usage data: Open-weight models accounted for 56% of all tokens processed through Vercel’s AI Gateway in August, up from 7% in December, per the Financial Times. A similar pattern appears on OpenRouter, where open-weight models from Chinese labs dominate the top 10 by tokens processed.
• Who is moving: The trend is sharpest among technology companies, but PNC Financial Services, logistics group CH Robinson and industrial manufacturer Siemens have all discussed using open-weight models in recent weeks. Neither the scale nor the timing of their deployments was detailed.
• AT&T’s numbers: Andy Markus, the telecom group’s chief data and AI officer, said AT&T now runs about 40% of its AI workloads on open models and is targeting 70% within a year. He put current volume at 45 billion tokens a day, a scale at which he said cost becomes decisive if accuracy holds.
• Tinder’s cost curve: Vinay Kuruvila, chief technology officer at the Match Group-owned dating app, said AI spending ran at an annualised $1 million in January and reached $10 million by July. Tinder has begun routing some queries from non-technical users to open-weight models to contain the increase.
• The substitution argument: Kuruvila said frontier systems already carry enough capability for about 90% of the tasks Tinder is attempting, and that he may not need them at all if open-weight models close the gap. He said he does not want “another 10X increase” in spending.
• Why they are cheaper: Because their underlying parameters are published, open models can be downloaded, hosted and fine-tuned without per-token fees, the Financial Times reported, often delivering output comparable to closed systems sold by Anthropic and OpenAI.
• Beyond price: Scott Wallace, senior global director of solutions architecture at data centre operator Digital Realty, said sovereignty and security weigh as heavily as cost. He said proprietary and customer data goes into private models hosted on the company’s own infrastructure, never into a frontier system.
• The competitive field: Chinese companies dominate the open-weight market, with DeepSeek and Zhipu selling tokens directly at prices far below OpenAI and Anthropic. Comparable models on similarly permissive terms are being built by France’s Mistral and US-based Nvidia, Reflection AI and Thinking Machines Lab.
• The price war: Both OpenAI and Anthropic released cheaper versions of their flagship models last week as competition with Chinese rivals intensified. Pricing details for those releases were not disclosed.
• The stakes: Anthropic is preparing an initial public offering expected to value it at $2 trillion or more, while OpenAI is in early talks for a private round at a $1.2 trillion valuation after pushing its own listing to 2027.

Background:

Open-weight models release their parameters publicly, letting any user run and modify them on private hardware. Closed frontier systems are sold through paid access, charged by the token, with the provider controlling the model.

Between the lines:

The two frontier labs still capture the bulk of AI spending, and their demand for computing power underpins the US data centre construction boom. That makes the substitution visible in the Vercel and OpenRouter token data commercially awkward timing: the valuations now being tested in public and private markets assume revenue growth that open-weight adoption at AT&T-scale volumes would directly erode.

What’s next

Watch whether Anthropic’s IPO pricing and OpenAI’s funding round hold at the reported valuations, whether AT&T reaches its 70% open-model target within a year, and whether further price cuts follow last week’s cheaper flagship releases.