One document filed on Monday names the thing the rest of the day’s news was circling. In its S-1, published on the SEC’s EDGAR system on 26 October, Anthropic discloses about $265.6 billion of fixed-or-minimum compute commitments “with terms ranging from one to 20 years” — including long-term agreements carrying “take-or-pay provisions… that obligate us to pay for resources regardless of actual utilization.” Read that clause beside the day’s other three stories — Microsoft inviting a quarter of its workforce to exit, GXO ordering humanoids its own CEO says do not yet return their cost, and a student map of the annotation workforce underneath it all — and the pattern is hard to miss: this economy no longer waits for the proof before signing.
🔍 THE BOTTOM LINE
Monday’s coverage is four versions of one move: the AI economy locks in decade-scale obligations — $265.6 billion of take-or-pay compute, 60,000 severance packages, multi-site humanoid agreements, an offshore labelling workforce with real people reading the worst content on the internet — years before the returns arrive. When the bets clear, whoever signed early owns the market; when they miss, the obligations are still due. For New Zealand, the practical question is not whether to join a race but to time entry to the curve the big spenders are already committed to underwriting.
The S-1 puts the number nobody had to say out loud
Anthropic’s own announcement of ticker ANT on Nasdaq is the headline, but the filed document underneath it is the substance, and it is unusually literal about what a frontier lab’s runway is made of. End-2025 non-cancellable commitments stood at $54.6 billion; by 30 June 2026 the fixed-or-minimum total was roughly $265.6 billion across terms of one to 20 years — as CNBC worked through in its filing coverage, including $110 billion to Amazon’s Trainium through 2036 and $114.5 billion to Google TPUs through 2033. First-half revenue of $16.3 billion sits alongside an $80.7 billion half-year net loss, with an 85 percent gross margin implying the unit economics turn once inference efficiency holds.
The filing’s own words are the sharpest available description of what this means: compute is “procured and committed in advance and the expense is incurred as it is brought online,” there is “a lag between when investment is committed and” when the workloads arrive, and the risk factors concede plainly that the company may pay for capacity it cannot use. Anthropic is not hiding the structure — it is asking public shareholders to price it. Our S-1 coverage this morning walked the whole filing; the point for this week is the verb tense: the money is committed, the proof is scheduled.
Microsoft is paying to shorten its own commitment to people
Employment contracts are obligations too, and Microsoft’s Voluntary Exit Offer — roughly 60,000 employees across the US, UK, Canada and India eligible, per CNBC’s reporting on the program, with exits on 31 December — is the AI-capex era converting people-obligations into cash obligations on the company’s own timetable. Notably, teams in AI, Copilot and CoreAI are among those eligible: the most AI-exposed staff are being offered the door. Our morning piece, following April’s 8,750-worker voluntary retirement, carries the rehiring-corrected caution: about a third of AI-cut roles later get rebuilt, so the sign-now-verify-later reflex runs on both sides of the ledger — Microsoft is pre-clearing its headcount bet before the AI revenue arrives to justify it.
GXO is buying a learning curve its own CEO says has no ROI yet
The GXO–Agility expansion of 26 October adds North Carolina and Tennessee from early 2027 — and on GXO’s Q2 call in August, CEO Patrick Kelleher said the company had run 45 humanoid pilots without achieving return, estimating “a couple of years away.” Signing the expansion anyway is not irrational; it is purchasing optionality on someone else’s capital — the learning-curve position costs GXO a deployment contract today against a machine economy Agility’s investors are underwriting for the rest of the decade. As covered this morning, the same company also tested Apollo and Reflex — a portfolio of bets, each small, each placed before the economics work.
The annotation map shows who the commitment economy rests on
Ninety pages of risk factors never itemise the humans in the stack, but the fourth story on Monday’s site did: the Brown student map linking 13 labelling firms, 141 customers and 31 countries — with annotation already the largest new job category this site has tracked, 282,000 US annotators among 750,000 AI-linked roles. The map catches the obligation economy from below: the take-or-pay compute, the buyout cheques and the humanoid pilots are all funded by a bet that model quality keeps improving — and model quality is still, per advisor Suresh Venkatasubramanian, “all humans underneath.”
What it means for New Zealand
CJ Murden has written this site’s AI-economy position with unusual directness: public AI money should target export earners, not institutional experiments — distributed, revenue-returning bets rather than one-off cheques. Monday’s pattern is the private-sector version of the same test, at a thousand times the scale: obligations signed first, verification scheduled later. The S-1’s own risk factor is the whole argument in one line — take-or-pay means paying regardless of utilisation.
And the private bills anchor the public ones: as this site put it when the S&P 500 rewrote its admission rules for giant AI listings, KiwiSaver global index funds already hold the compute bet whether they opted in or not. There are also two practical entries for a small economy positioned on the curve’s edge. First, the 2026–2027 window is when committed capacity comes online — inference prices are the one input a small economy’s AI adoption plan can actually forecast, and buyers clubbing together on take-or-pay compute (the S-1’s own fungible-pool model, at village scale) is the version CJ’s solar-organisation logic points to. Second, the verification ladder does not need the capital: the trust-migration thesis, the competence-gap data and now the annotation map all say the jobs with pricing power are the judgement roles — the ones who check what the committed machines produce. A decade of take-or-pay is somebody’s guarantee of demand; the honest play is to be paid on the outcome, not the obligation.
❓ FAQ
What did Anthropic actually commit to in the S-1? About $265.6 billion of fixed-or-minimum compute and infrastructure commitments as of 30 June 2026, with terms from one to 20 years, including take-or-pay provisions that require paying regardless of utilisation — per the filing’s own disclosure.
Does GXO regret its humanoid pilots? No — its CEO said in August that 45 pilots had not yet achieved ROI and put payback “a couple of years” out, then signed an expansion to two more sites in October. The bet is on the learning curve, not today’s unit economics.
Why is Microsoft offering buyouts instead of layoffs? The Voluntary Exit Offer spreads the choice across roughly a quarter of the workforce while the company absorbs AI infrastructure spending; terms match standard severance. The memo does not mention AI — the efficiency framing comes from the reporting around it.
What can New Zealand actually do with this? Time adoption to the committed capacity coming online in 2026–2027 (forecastable inference prices), avoid signing long-term obligations on unproven internal AI benefits, and invest in the judgement roles — verification, evaluation, expert annotation — that the committed machine economy still runs on and pays premiums for.
🔍 THE BOTTOM LINE
The defining habit of this phase of AI is signing before proving: $265.6 billion of one-to-20-year take-or-pay in one filed document, 60,000 invited exits at one software company, robots ordered by a customer who said on the record that they do not pay for themselves yet — all resting on a workforce of labellers and annotators that the filing never names but the map does. Obligation has become the industry’s strategy. For everyone not holding the pen, the moves that pay in 12–24 months are the ones timed to the committed build-out, not the ones copying the commitment.
📰 Sources
- Anthropic PBC S-1 (SEC EDGAR, filed 26 October 2026)
- Anthropic — S-1 and Nasdaq listing announcement
- CNBC — Anthropic IPO live filing coverage
- Microsoft — Chief People Officer memo, Voluntary Exit Offer
- CNBC — Microsoft buyout reporting
- Agility Robotics — GXO partnership expansion announcement
- The Motley Fool — GXO Q2 2026 earnings call transcript
- Brown University — AI data workers project
