AI brief, Sunday 25 October 2026
Meta's quarter shows what superintelligence costs a shareholder, Britain buys into AI biosecurity mostly with other people's credits, Moonshot's new flagship arrives with its weights unpromised, Agent Skills moves to a neutral foundation, and two utility filings measure the gap between the data-centre pipeline and the megawatts actually switched on.
Meta's quarter puts a number on what "personal superintelligence" costs
Meta reported its third quarter on 21 October. Revenue grew 26% to $64.6bn. Costs and expenses grew 68%, to $51.6bn. Operating income fell 37% to $13.0bn and the operating margin halved, from 40% a year ago to 20%. A $9.9bn charge for legal proceedings is part of that, but not most of it.
The capital side is where the AI build-out shows. Purchases of property and equipment were $33.9bn in the quarter and $83.1bn over nine months, against $48.3bn in the same nine months of 2025. Meta narrowed its 2026 capital expenditure guidance to $135–145bn including finance-lease principal, from $130–145bn, and told investors it is "still developing our budget and capacity plans for next year" while expecting "to continue investing aggressively". Headcount fell 15% year on year to 66,775. Reality Labs losses are now guided to $18–19bn this year and about $15bn next, as investment is reallocated "toward other priority areas, primarily AI".
Alongside the results, Meta's board approved a new class of publicly listed, non-voting Class C common stock, effective 19 January 2027, with a one-time one-for-one stock dividend to Class A and B holders that will have the effect of a two-for-one split. The board's stated reasons begin with capital raising, and name "the scale of investment required to compete, sustain growth, expand infrastructure".
Mark Zuckerberg's quote in the release: "We hit major milestones this quarter toward delivering personal superintelligence for everyone — including launching Muse and Muse Wonder, our first frontier model."
Why it matters. This is the first quarter in which the cost of the frontier is unmistakable in a hyperscaler's margin rather than only in its guidance, and the company's answer is not to slow down but to build a new share class partly to help fund it. Watch whether the rest of the group — reporting over the coming week — show the same shape: revenue growing well, expenses growing faster, capex guidance going up rather than down.
Source: Meta Platforms, Q3 2026 results, Exhibit 99.1 to Form 8-K, filed 21 October 2026 (SEC EDGAR).
Britain's £300m AI biosecurity lab is £100m of cash and £200m of credits
Two UK government pages went up on 24 October. The Cabinet Office introduced an AI Biosecurity Lab, an independent non-profit it calls the world's first dedicated AI biosecurity organisation, with two initial jobs: earlier detection of emerging biological threats, and faster development of vaccines and countermeasures. HM Treasury set out the wider AI package the Chancellor is taking into the Budget, of which the lab is one line.
The funding split is worth reading closely. The government is putting in £100m. OpenAI is contributing £100m in compute and model credits plus access to its models through its Rosalind Biodefense programme; Anthropic up to £100m in model credits, model access and technical support. Google DeepMind will collaborate on research and "will look to provide" trusted access to Gemini and specialised science models. Basecamp Research is contributing AI-biology models. UKHSA, APHA, Dstl and the MHRA are the public bodies attached. No site, director or start date is named, and a scientific advisory committee is still to be set out.
The Treasury page puts the lab inside a £165m National AI Biosecurity Initiative (the other £65m goes to biosurveillance and countermeasures for priority pathogens), with a further £120m for the AI Security Institute, £70m for a National Exposure and Real-Time Vulnerability centre, £60m for a National Cyber Shield, £60m more for the Sovereign AI Fund, £60m for sovereign AI R&D procurement, £18m for a UK Advanced Chip Design Academy, £30m for local AI partnerships and £23m over three years for an AI Economics Institute. Named sites: chip design in Edinburgh, an AI Centre of Excellence and an NCSC AI cyber centre in Manchester.
Why it matters. The headline is "up to £300m", and a third of it is public money. The rest is credit on the donors' own platforms — which means a national biosecurity capability whose compute budget is denominated in two American vendors' inference, and spent back with them. That may still be the fastest way to stand the thing up; it is a different proposition from £300m of public funding, and the distinction is in the primary source rather than in most of the coverage.
Source: Introducing the AI Biosecurity Lab and Chancellor unveils Britain's ambitious plan for AI, both GOV.UK, 24 October 2026.
Moonshot's new flagship is listed, and its weights are not promised
Moonshot AI's developer platform now carries Kimi K3.1 Pro as its flagship, described in the company's own model list as "the first model of the new K3.1 series and Kimi's most capable flagship to date: a 3T-class model with native visual understanding and a 1M-token context window", with gains in coding, 3D creation and financial modelling. A kimi-k3.1-pro-highspeed variant is listed as the same intelligence at about two to three times the speed. The documentation also sets a retirement path for the older K2 family: closed to newly registered users from 31 October 2026 and fully retired by the end of December.
One thing the documentation does not say is anything about open weights for the new model. The same page describes the previous generation, K3, as "the world's first open-source 3T-class model (2.8 trillion parameters)". Press reports say full K3.1 Pro weights are due in early November; that date is not on any Moonshot page we could open, so we are not stating it as fact.
Why it matters. The open-weight frontier has been set by a Chinese lab for several generations now, and the interesting question about this release is not the benchmark table but whether the weights follow the way K3's did. If they do, the best openly available model stays within reach of the best closed one, and the economics of everyone building on top of models change with it. If they quietly do not, the frontier closed a little this month.
Source: Kimi model list, Kimi API Platform documentation (Moonshot AI), retrieved 25 October 2026.
Anthropic hands the Agent Skills standard to a Linux Foundation body
On 22 October Anthropic donated the Agent Skills standard — along with its name, logo, GitHub organisation and the agentskills.io domain — to the Agentic AI Foundation, a body under the Linux Foundation that Anthropic co-founded with Block and OpenAI. Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals, with two named lead maintainers, and the foundation plans versioned releases, a formal proposal process and security guidance. Anthropic says it is "committed to keeping Agent Skills open, community-driven, and vendor-neutral", and that the standard can be used or contributed to without joining the foundation. Anthropic puts adoption at more than 70 agent platforms, naming ChatGPT, Cursor, GitHub Copilot and VS Code among them.
Why it matters. This is the second time in two years that the labs have taken a piece of agent plumbing they could have owned and parked it in a neutral foundation. The commercial logic is not charity: a standard that every platform implements makes the capability layer a commodity and pushes competition back onto the models underneath it, which is where the firms doing the donating believe they win. For anyone building on agents, the useful signal is that skill definitions are becoming portable — the switching cost is moving away from the integration and towards the model.
Source: Donating Agent Skills to the Agentic AI Foundation, Anthropic, 22 October 2026.
Two utility filings measure the gap between the pipeline and the megawatts
The data-centre pipeline is quoted in gigawatts. Two quarterly filings this week put numbers on what has actually been energised.
PG&E, reporting on 22 October, says it has "energized 174 megawatts of new data center projects since 2024" — and argues that "with appropriate pricing, every one gigawatt of new data center load could help customers save 1% or more on their monthly electric bill". NextEra, reporting on 21 October, describes "a significant pipeline of large-load projects that it expects to own and operate": 30 potential hubs under discussion with the market, expected to reach 40 by year-end, plus a coalition developing a data-centre campus at the Department of Energy's Paducah site in western Kentucky where NextEra expects to build and own up to 2 GW of gas generation and up to 2.6 GW of battery storage. It added 3.7 GW to a renewables and storage backlog now at about 36.7 GW, and its pending combination with Dominion Energy is through shareholder votes with state hearings in November and December and final orders expected in January 2027.
Why it matters. Both filings are true at once: the pipeline is enormous and the delivered load is small. 174 MW energised over roughly two years at a large western utility is the rate at which this actually happens when interconnection, permitting and generation have to arrive together — and capex forecasts of the kind Meta published this week are built on power showing up. Note also the argument PG&E is making: data-centre load as a reduction in everyone else's bill. That is a political case as much as an economic one, and it will be tested in rate cases long before it is settled in spreadsheets.
The same gap sits underneath single-site projects too, including Ark7's Project Soperton in Georgia, whose own published project status lists grid power as a written supply proposal still awaiting a signed service agreement. Ark7 is an advertiser on this site and we are paid if you open that link — see About — and the status is quoted from Ark7's own page because it is the relevant part: an unsigned power agreement is the normal condition of a project at this stage, not an exception.
Source: PG&E Corporation Q3 2026 results, Exhibit 99.1 to Form 8-K, 22 October 2026 and NextEra Energy Q3 2026 results, Exhibit 99 to Form 8-K, 21 October 2026 (SEC EDGAR).