Voters reject data centers; Apple rations engineer AI budgets; Revolut names its 90% cut

🔥The day's one thing

🏢How organisations are changing

• Apple's internal AI budgets — Apple is cutting what its engineers can spend on AI each day, Bloomberg reports: the daily cap for engineering teams fell from $300 to $200 in October, with a drop to as little as $100 planned and a limit as low as $25 being weighed outside engineering. The clearest sign yet that per-engineer token budgets are becoming a finance-controlled line item rather than a tooling decision — and sources say the cuts are stoking staff fears of layoffs. The report landed November 4; Apple has not commented. explainx.ai

• Revolut's switch to GLM — Revolut moved roughly 2,000 engineers off Claude Code and Codex defaults onto GLM, a Chinese open-weight model it hosts itself; it says the bill is now about a tenth of what it paid Anthropic and OpenAI. Strongest public evidence that routine coding runs on cheap self-hosted weights, frontier models kept for exceptions — Revolut's own figures, not an isolated comparison. Dealroom News

  “The bottleneck for engineering is not necessarily coding,”

• Canva's restructure — Canva is cutting 1,271 people, 22% of staff, after growth slowed for the first time in a decade, and lowered its revenue growth forecast from 30% to 20% to absorb ballooning AI costs. AI hits it from both sides — a rising model bill, and chatbots that now make the posters and slides it sells — and investors had already marked it to about $35 billion. Product, growth and platform teams are being merged to reduce 'subspecialisation'. TechCrunch

🧾Vendors, pricing and risk

• Replit Pro arbitration opt-out — Replit Pro customers have until November 8 to email an opt-out of the mandatory arbitration and class-action waiver added October 9; on a team, each person opts out individually. A use-it-or-lose-it contract right — and while you're in settings, check Model Improvement: turned on by default September 9 for eligible accounts, per person, and not retroactive. Opt-outs go to [email protected] with name, mailing address, username and account email. Read the Diff

• Gemini's chat-history terms — A Hacker News post that hit 606 points flags that paid personal Google AI accounts still use saved Gemini chats for training, with no opt-out except disabling Keep Activity — which drops chats after 72 hours. Personal accounts, not workspace ones, but the details matter for policy: a subset of chats is read by human reviewers and kept for up to three years even after you delete your activity. Google's own guidance warns users not to enter confidential information. Hacker News

• GitHub's November 4 incident — Pull requests, git operations, issues, API requests and Actions all ran degraded for about two hours on November 4, from 09:49 UTC until GitHub marked the incident resolved at 11:50 UTC. With agents pushing merges around the clock, a GitHub outage now stalls machines as well as people — worth confirming your CI and agent retry behaviour rides one out. GitHub says a root-cause analysis is coming. GitHub Status

📰News the board will ask about

• Meta acquires Conductor — Meta has bought Conductor, the tool for running coding agents in parallel workspaces, and is shutting its cloud fast: read-only November 13, closed with user data deletion from November 20; the local app is supported into early 2027. A reminder that agent tooling built on other labs' models can vanish in weeks — any team using Conductor Cloud needs to export before November 20. Conductor only launched in July 2025; its agents come from Anthropic and OpenAI, not in-house. Conductor

  “We have never sold or trained on Conductor user data, and we still won’t.”

• Fal — The startup that runs image, video and audio models for developers raised $1.3 billion led by Bessemer at a $16 billion valuation — double April — after passing $800 million in annualized revenue. The layer serving models to your apps is being capitalised at extraordinary speed, though people familiar with these companies' books say margins stay thin because compute eats most of the money. Rivals Modal and Fireworks raised at $15.75 billion and $17.5 billion valuations in recent months. TechCrunch

• Technology Radar v35 — Thoughtworks' latest Radar argues agentic AI's real costs — shared context, token spend, review load, agent security — appear only once agents become the operating model, not a pilot, and calls the agent harness the new core resource for platform teams. Its advice: manage spend per finished task rather than per token, treat review as capacity with a cap on each reviewer's queue, and inventory agent identities before models. Agent sandboxing went straight to the Adopt ring in this volume. Thoughtworks

  “A cheaper model with a good harness will outperform an expensive model with a bad one.”

🌍The big AI stories today

• OpenAI status — OpenAI reported elevated error rates on paid ChatGPT plans November 4; its status metrics are aggregate, so your real availability may differ when claiming credits. OpenAI Status

• PHP2Shell — AI-assisted research turned a long-ignored PHP crash bug into remote code execution hitting sites including NASA's — your legacy 'harmless' bugs are now attack surface. IEEE Spectrum

• Tencent Hy4 — Tencent launched Hy4 (770B parameters, vision, computer-use) on OpenRouter at $0.834 per million input tokens, weights due shortly — a budget computer-use option for shortlists. explainx.ai

💡One takeaway