AT&T moves 40% of AI workloads to open-weight models, and Chinese labs are winning the cost race
机器之心 · wechat · 2026-10-01
Per the Financial Times, US enterprises are shifting AI workloads from OpenAI and Anthropic frontier models to open-weight alternatives, with Chinese labs capturing much of the cost-saving wave.
Key numbers:
- AT&T: handles 45B tokens daily, runs 40% of workloads on open models, targeting 70% within a year; fine-tunes them on proprietary data, sometimes matching or beating closed models
- Tinder: annual AI spend jumped from $1M estimated in January to $10M/year by July — a 10x rise in six months; its CTO says frontier models are overkill for 90% of tasks
- Market signals: mentions of open-weight models in earnings calls rose 6x YoY (AlphaSense); 56% of tokens on Vercel's AI Gateway now come from open-weight models, up from 7% in December
Beyond price, data sovereignty and private deployment drive the shift. DeepSeek and Zhipu are among the most active open-weight players, alongside Mistral and Nvidia. For OpenAI and Anthropic, the real competitor may be a CFO's spreadsheet, not a bigger rival.
More from Venture
- VC Warning: Board Observer Seats Are a Liability — remilouf · 2026-10-01
- Indie dev's AI video site hits $2,013 in 30 days with zero paid traffic, crediting GEO — gefei55 · 2026-10-01
- Cheaper AI will drive adoption more than boost hyperscaler profits, argues investor — TansuYegen · 2026-10-01
- tuttihq's September numbers: 14,269 creators onboarded, $81,480 paid out in one month — nikola_mr64990 · 2026-10-01
- 90-Minute Zero-to-One SEO Guide for Going Global: Keywords, Pages, Links, Technical — yihui_indie · 2026-10-01
- ChatGPT now recommends plugins mid-conversation, echoing the 2008 App Store gold rush — sherwinwu · 2026-10-01