The Product Desk

AT&T told The Information it will hold OpenAI and Anthropic spending flat while running 40% of its 45 billion daily tokens on open weights, with a stated target of 60-70%. After adopting the LiteLLM router, its AI coding costs fell 56% while measured quality fell 2%. That is the first at-scale exchange rate an enterprise buyer can hold your product to, and the same buyer caps what each developer may spend on premium AI tooling.

Nvidia agreed to pay $6 billion for a non-exclusive license to Poolside's model technology plus $1 billion in equity at a $12 billion pre-money valuation, per a leaked investor letter obtained by Newcomer. Non-exclusive means AMD, Google or AWS can license the same technology next quarter. Fortune's Term Sheet reports investors marking 10-20% of portfolios as vulnerable to foundation-model encroachment. Any feature whose edge is model quality is a two-quarter asset, not a strategy.

A research team wrapped database ACID semantics — validate before commit, isolate failed attempts, keep durable state — around a multi-step agent and beat Anthropic's Claude Code by up to 10.6% on a data-agent benchmark, with no better model. Two separate studies found memory-based self-improving agents look materially worse once task order and evaluation variance are controlled. Reliability is an architecture decision your backend team can ship today, and 'our agent learns from every session' will not survive a technical buyer's diligence.

1Password, Cloudflare and Google each published a competing agent-identity architecture in the same week, and Uber open-sourced its Agent Detection & Response system with a public benchmark covering 300+ tasks, 17 attack techniques and 133 MCP servers. Independent vendors converging on one problem in one week is a spec forming, not thought leadership. Specs that form in August arrive as procurement questions in Q1, so your agent PRD needs an authorization section before a security reviewer writes one for you.

SolarWinds found 84% of IT service management teams say AI met or exceeded ROI expectations, while 52% report their total workload increased anyway. An SSRN working paper on Chinese secondary students found generative AI raised homework scores 18% while closed-book exam scores fell 20%. Census pulse data shows 55% of US workers used AI on the job, and 13% of those users report no time savings or more work. 'Saves X minutes per task' is a weak renewal story when the saved time gets reabsorbed by validation.