A thought piece this week, in reaction to Anthropic's filing to go public: companies are starting to buy AI wholesale, in bulk, through spending commitments they already have with a big provider, rather than one tool at a time.
Companies are starting to buy AI wholesale - About 48% of Anthropic's revenue in the first half of 2026 came through the marketplaces run by Amazon, Google and Microsoft, up from about 11% in 2023. AI bought there can count toward the multi-year spending commitments large companies sign with their cloud provider in exchange for discounts, so moving AI spending onto the cloud bill can help a company meet its commitment or earn a deeper discount at renewal, even though the AI itself isn't automatically cheaper. It works like a warehouse club: better terms on what's on the shelf, but only for what the club lists, and the club takes a cut from suppliers. Anthropic and OpenAI are now opening marketplaces of their own. The speculative next step is bringing your own tokens: using AI a company has already bought inside other vendors' software. If that spreads, vendors that mark up AI usage lose that margin and have to charge for the workflow, the data and the controls around it. If you have a cloud commitment, ask finance whether your AI spending counts toward it today.
Fake Claude Code add-ons in Cursor's store - For about 11 hours on October 29, Cursor's add-on store carried fake versions (2.1.312 and 2.1.313) of Anthropic's Claude Code add-on that tried to give an attacker a way to log in to developers' computers, and uninstalling may not remove what they left behind. Ask IT whether anyone installed those versions, and whether developers can install add-ons from any publisher or only approved ones.
A near-match for an eighth of the cost - In Artificial Analysis's benchmark, OpenAI's GPT-6.1 Sol scored within three points of Anthropic's Claude Opus 5.5, both at maximum effort, for about an eighth of the cost per task. Ask the teams using AI for coding or analysis which model and effort setting they use by default, and whether anyone has tested a cheaper option on your own work.
Also in this issue:
- News Bits - Morningstar's CEO says he dropped some internal candidates partly for barely using AI; an IBM survey finds most AI incidents reach the board but few companies change their AI governance afterward; Britain's financial regulator prepares guidance on AI and quantum threats to encryption.
- The AI Bill - GitHub's Copilot Fusion claims near-top coding scores at about half the cost, and Claude Enterprise adds shared monthly team budgets with spending alerts.
- Overheard on Reddit - Claude tests an app by acting like a very realistic customer.
- Meanwhile... - Colossal Biosciences uses AI to narrow down the genetic changes behind the dodo's beak.
- What I'm Consuming - Lenny's Podcast on how AI models actually think, Alex Stamos on the PHP flaws, BCG on pivot-ready SaaS pricing, The AI Daily Brief on internal AI champions, IBM on context engineering for agents, McKinsey on quantum, and a walkthrough of P-Stack for running teams of coding agents.
- After Hours - John le Carré's The Spy Who Came in from the Cold, and a Halloween rewatch of The Thing.