Open-weight models are moving from an engineering option to a staple of corporate earnings calls. According to a Beating AI flash report on September 28 citing a Financial Times survey, data provider AlphaSense found that mentions of open-weight or open-source models in US companies' earnings calls and investor meetings rose about sixfold year over year in August and September. Open-weight models are spreading from tech companies into traditional industry.
The numbers tell the story. PNC Financial, logistics giant CH Robinson and Siemens have all recently spoken publicly about applying open-weight models in their internal operations. Tinder's AI bill is the most vivid footnote to this shift: its annualized AI spending jumped from around $1 million in January to around $10 million in July, and it has now started routing some ordinary requests to open-weight models.
AT&T plays at a bigger scale and its strategy is explicit: roughly 40% of its AI workloads already run on open models, and the company plans to push that share toward about 70%. Its internal AI systems process around 45 billion tokens a day, using a routing strategy — simple tasks go to cheap open models, complex work stays on frontier closed models from OpenAI and Anthropic.
Background: the bill rose first, not the parameter count
Why now? Because agents are moving from chat windows into coding, research and office work, pushing per-task token consumption from thousands into the hundreds of thousands or millions. Corporate AI bills are growing far faster than per-token prices are falling — Tinder's tenfold jump in six months is a snapshot of that usage explosion.
Companies are responding not by switching vendors but by building "tiered routing": ordinary requests go to open weights, complex work stays with frontier closed models. After Kimi K3 was open-sourced in July, BlockBeats noted that roughly half of Hex's customers had hooked up to Kimi and Harvey had adopted Zhipu's GLM — open models absorbing huge volumes of ordinary requests while dragging closed-model giants into a subsidy war.
Why it matters: pricing power comes under pressure before capability does
"Open-weight models" showing up on earnings calls means the CFO vocabulary has changed: AI spending is turning from a default innovation investment into a cost line to be explained and optimized. That is the real sign that open weights have arrived.
For closed-model labs, competition is sliding from "who is most capable" to "who is cheapest for the same task." The price war is no longer just about per-token prices — it is about the total cost of getting a job done, a point also made in recent Chinese media analysis of the new round of model releases from OpenAI, Anthropic, Xiaomi and SpaceXAI, all pushing intelligence to be cheaper.
For developers and small businesses, the takeaway is concrete: tiered routing is now standard practice. There is no need to wait for cheaper frontier models — moving simple tasks off expensive models cuts the bill immediately. Complex work still belongs to frontier models; the capability gap remains, but the way money is spent has gotten smarter.
Open weights on earnings calls mark the moment when the "cut costs" logic was formally adopted by management. The next round of competition will be judged not only on parameters and leaderboards, but on whose bill looks better.
