Enterprises push open-weight models into production
A Fastino review of early deployments at Uber, AT&T, Amplitude and others finds companies turning to open-weight models to rein in rising agent costs and keep proprietary data under their own control.
AI startup Fastino reviewed recent open-weight model deployments and experiments at Uber, AT&T, Amplitude, Shopify, Pinterest, Harvey and Thomson Reuters. It concludes that companies are still working out, gradually, how much of their AI stack to rent and how much to own.
Cost is the main driver. Fastino cites McKinsey's May 2026 Enterprise AI FinOps survey, in which 93% of qualified respondents said their organizations had exceeded AI budgets. The post also says Uber used up its annual AI budget in the first quarter of 2026 as adoption of coding agents grew. AT&T processes about 45 billion tokens a day across more than 1,000 internal use cases, according to the Wall Street Journal. Fastino argues that lower token prices alone won't fix an uncontrolled system. In its view, the companies cutting spend most effectively also measure usage and change how work is split across models.
Data control is the second driver. Fastino notes that open weights don't guarantee it, since a managed endpoint still runs on someone else's infrastructure. What they do offer is a choice of where the model runs.
The most detailed case in the post is Amplitude. The company ran Kimi K2.7 unmodified on Fireworks' dedicated GPUs. On its 186-case internal evaluation, Kimi first scored 64, behind Anthropic's Sonnet 4.6. Amplitude then worked on the system around the model. It cached permission checks, added typed chart definitions, and put in specific fixes for problems such as tool loops and context limits. After those changes, Kimi scored 73.7 against Sonnet's 72.7, at roughly one-third the cost.
The available text stops before the details for Uber and the remaining companies. Fastino builds AI models, and these observations come from its own analysis.