# Morning digest: Microsoft's Essex exit, agent lockouts, and Copilot

> Microsoft ended its Nscale Essex lease over grid delays. Major sites are locking out AI agents, and six GitHub Copilot models retire today.

Published 2026-10-19T07:26:12-07:00 · Daily Digest · Canonical: https://news.littlewolflabs.com/posts/2026-10-19-morning-digest

Morning digest for Monday, Oct 19. It was a quiet night for the big labs, so this one leans on power, agent access and a couple of developer stories.

## Microsoft walks away from Nscale's Essex AI data centre over grid delays.

Microsoft ended its deal to lease capacity at Nscale's planned Loughton site, which was pitched as the UK's largest sovereign AI data centre: about £2bn, aiming for roughly 23,000 Nvidia chips in 2027. UK Power Networks says enough power may not arrive until the early-to-mid 2030s. Nscale says it's still building, is looking at other power options and UK sites, and still works with Microsoft elsewhere. *What to watch:* it's the same story as San Francisco's pause and the Halifax protest. The bottleneck is the grid, not the chips. ([BBC](https://www.bbc.com/news/articles/c9m26kl5y7epo), [Guardian](https://www.theguardian.com/uk-news/2026/oct/18/microsoft-drops-huge-ai-datacentre-project-in-essex-over-power-supply-concerns))

## OpenAI's Dean Ball says distillation may explain about half of how close Chinese models look on benchmarks.

Ball, who works on strategic futures at OpenAI and says these are his personal views, wrote Sunday that since joining OpenAI he'd raise his guess from about 30% to about 50%. He also said a lot of "distillation" in practice means stealing a model's chains of thought, not just copying its normal outputs. That's his estimate, not a published study, and he works for a competitor. It lands right after DeepSeek's Flash release and OpenCode's distillation-attack breach. *What to weigh:* if you're choosing a cheap open-weight model on leaderboard scores alone, test it on your own work first. ([Ball's post](https://x.com/deanwball/status/2111802397148471496), [his follow-up on the 50%](https://x.com/deanwball/status/2111807534940185051))

## Major sites are locking out third-party AI agents.

The Verge reported Oct 17 that Muse, Grok, OpenAI Dots and Instinct users are hitting blocks. Confirmed so far: Yelp (paid licensing only), Delta (no third-party flight booking), and JPMorgan and Charles Schwab (no agent access to accounts). Schwab's CEO said they won't let someone hand Muse a password and act on a client's behalf. Amazon already blocks Muse and Perplexity-style shopping, while Meta is partnering with Walmart and others instead. *What this means for you:* agent shopping and booking work where there's a deal in place. Everywhere else, expect hard stops. ([The Verge](https://www.theverge.com/ai-artificial-intelligence/1013195/ai-agents-blocked-web), [Schwab](https://www.theverge.com/ai-artificial-intelligence/1013303/charles-schwab-is-blocking-ai-agents-too))

## Six GitHub Copilot models retire today.

Gemini 3.7 Flash is replaced by 3.8 Flash. GPT-5.5 and GPT-5.4 are replaced by GPT-5.6 Sol. GPT-5.4 mini and GPT-5 mini are replaced by GPT-5.6 Luna. Grok 4.5 is replaced by Grok 4.6. Business and Enterprise plans usually get the replacements switched on automatically, unless an admin has locked the model policy. *Try it:* update any workflow that's pinned to an old model name. ([GitHub changelog](https://github.blog/changelog/2026-09-18-upcoming-deprecation-of-selected-github-copilot-models-in-mid-october/))

## Gemini 4 Argon's 97.5% on ARC-AGI-3 depends on the test setup.

ARC Prize reports that with Google's own test setup, which keeps the model's reasoning between steps, Argon scores 97.5% at low reasoning. With the standard setup, it scores about 49.8%. The intro API price is still $2 per million input tokens and $10 per million output tokens through Nov 30. *What to weigh:* when a vendor quotes the big number, ask which setup produced it. It's the same caveat as the Minecraft agents paper. ([ARC Prize](https://arcprize.org/results/google-gemini-4-argon))