The AI build-out is being tested by its lenders and its power lines

AI² — AI Investor Sydney · Friday 11 December, The Mint. One widely cited forecast puts cumulative AI and data-centre capital spending at about US$11 trillion by 2029. One day, six evidence-led briefings on where that capital is creating and destroying value — no panels, no pitches. Every ticket includes a year of NOOPS Pro. A$995 until 30 October → Book your seat

Two views aired on Monday disagree about whether the AI build-out faces a shortage or a surplus. Much of the rest of today's fourteen signals tests the build from its edges: a power delay at a Melbourne data centre, a wider spread on Oracle's credit default swaps and higher yields on CoreWeave's bonds, and Anthropic setting terms ahead of a possible listing.

Two forecasts, one build

SemiAnalysis forecasts about 50GW of new data-centre capacity in 2027, which it says is roughly equal to the entire global cumulative market up to 2024, citing more than a trillion dollars of lease value signed in 2026 so far and more than 24GW in advanced construction across 40 campuses. It argues that power is "broadly solved", through behind-the-meter generation and modular builds, and that the bottleneck moves to chips, with memory and logic possibly the number-one constraint in 2027–28. That is SemiAnalysis's forecast; its paid section, on oversupply, was not read. Mark's separate view is that a "massive compute crunch" is coming.

Gartner's chief forecaster, John Lovelock, put the competing case to CNBC the same day: there is more compute capacity than demand right now, with many tokens given away in freemium products. Gartner stands by its analysis that AI revenue will not support profitable returns on all planned AI capex between 2027 and 2030, and repeats its prediction that 80% of neoclouds could cease to function as independent companies, mostly through mergers and takeovers. The two claims are not necessarily in conflict: they cover different horizons, and Gartner's worry runs to returns and revenue as well as utilisation. A period in which capacity runs ahead of paying demand is one way both could be true, and it is the scenario highly levered neoclouds would find hardest to fund.

Memory is one place SemiAnalysis expects the constraint to land. ChangXin Memory (CXMT) said in a Friday filing that it expects a nine-month net profit of 131 billion to 132.5 billion yuan, against a loss of 5.28 billion a year earlier, on revenue of about 265 billion to 270 billion yuan, roughly eightfold the prior-year period; it attributes the turnaround to AI-driven DRAM shortages. Separately, its shares closed 7.8% lower in Monday's session, a fall CNBC does not tie to any single cause. Counterpoint puts CXMT at 10% of global DRAM revenue in the third quarter, up from 5% a year earlier.

Power arrives last

On the day SemiAnalysis called power broadly solved, NextDC, Australia's largest listed data-centre developer, sought an ASX trading halt to assess the impact of a delay in getting power to its M3 facility in West Footscray, Melbourne, which is undergoing a $750 million expansion. NextDC did not give a reason. Citi analyst Siraj Ahmed told investors the halt probably related to distributor Jemena's pause on a high-voltage pole rollout, which The Age reports stalled after residents blockaded contractor trucks. The halt runs until Wednesday or a full announcement, which should show how long the delay is and what customers' contracts allow.

Who carries the risk of building for a data centre is the question a Minnesota proposal tries to settle. Google and Xcel Energy have proposed a 780MW data centre at Pine Island with minimum demand charges: Google would pay for at least 75% of its contracted capacity even if it uses less, would owe up to eight years of those charges if it left early, and would finance 1,400MW of wind, 200MW of solar and 300MW of battery storage on Xcel's grid. The deal still needs the Minnesota Public Utilities Commission's approval. A take-or-pay floor moves stranded-asset risk onto the hyperscaler — a guarantee that is credible from Google in a way it would not be from a levered neocloud.

Lenders are repricing the borrowers

The spread on Oracle's five-year credit default swaps stands at 295 basis points, CNBC reports, up 287% in a year after briefly touching 300 on 15 October. Oracle is rated BBB-, the lowest investment-grade notch, by S&P Credit Ratings; Gavan Nolan of S&P Global Market Intelligence said that spread range is more in line with BB and B credit, which is high yield. Consensus forecasts put Oracle's free cash flow at minus US$23.7 billion this year and minus US$36 billion in 2027, and Deutsche Bank's Jim Reid wrote that a cut to junk "would risk pressuring IG-only investors into forced sales".

At CoreWeave the numbers are starker. Bonds it issued in June at a yield of 9.675% rose to yield about 15% last week, trading below 80% of face value, according to a Sydney Morning Herald analysis, which also reports it costs almost US$1.2 million to insure US$10 million of CoreWeave debt — by the paper's reading, a more than 60% chance of default within five years. The SMH's core argument is a duration mismatch: long commitments to landlords, short subleases to customers, and GPUs that may be superseded before the centres are fully earning. Chief executive Mike Intrator has said markets are acting "irrationally".

Anthropic prepares for the public market

Of 23 companies on a Kalshi market asking whether each will officially announce an IPO in 2026, Anthropic is the only one traders put above 50%, at 75% as of Monday morning, CNBC reports; OpenAI sits at 3%. Those are traders' odds, and CNBC discloses a commercial relationship with Kalshi. Anthropic confidentially filed a prospectus in early June, so the price is less a forecast of demand than a bet on process. 22V Research's Dauvin Peterson wrote that a revised valuation range starting at US$1.5 trillion, against a prior expectation of at least US$2 trillion, "was viewed as disappointing".

The company is setting terms before any listing. Anthropic said on Sunday that it plans to commit 5% of its shares to a new Public Benefit Fund, subject to stockholder approval, with the board able to add 2 percentage points for every 50% rise in value above a baseline it has not disclosed; the fund's equity will not be used for lobbying or political donations. Class V Group's Lise Buyer told CNBC the timing is "mandatory to prepare for an initial public offering". A Wall Street Journal piece, which we could not retrieve, says in the excerpt we have that Anthropic's existing governance trust has no explicit vote in many shareholder matters and exercises its power through appointments instead.

If it proceeds and, after at least a year of trading, meets the index's eligibility tests, a company that size could eventually add to that concentration. An ABC News analysis finds the top 10 stocks make up about 39% of the S&P 500, against about 27% at the dot-com peak in 2000, quoting RMIT economist Angel Zhong. On the ABC's calculations, six names — Meta, Microsoft, Apple, Micron, Nvidia and AMD — account for 57% of the index's gains since April; without them, it would have risen 4.35% over six months rather than about 10%.

Cheap models keep arriving

MiniMax released M3.1 Flash at US$0.15 per million input tokens and US$0.60 per million output, and Artificial Analysis independently scores it 42 on its Intelligence Index, while noting it generated 180 million tokens during evaluation against a median of 99 million. A model that verbose gives back part of its per-token discount, so cost per task is the comparison that matters. Its other benchmark scores are self-reported, and the weights "will be released later".

Mistral's news is a domain partnership. CERN and Mistral have signed a statement of intent to build a Large Physics Model, trained on billions of simulated proton-proton collisions from the Large Hadron Collider to reconstruct collisions, identify particles and flag unexpected signals. Simulated collisions have known answers, which makes evaluation unusually rigorous. There is no disclosed funding, compute commitment or timeline, and claims about discovery remain prospective.

Agents change the economics of data and intrusion

A Business Insider investigation describes a fast-growing market in companies' Slack messages, email and CRM records, bought by brokers and resold, anonymised, to AI developers building workplace agents. Training startup Micro1 runs a programme paying midsize companies US$100,000 to US$2 million for access, and Gartner estimates annual spending on data for AI will rise from less than US$1 billion last year to almost US$6.5 billion in 2027. Emory law professor Ifeoma Ajunwa told BI that, absent federal privacy law, there are few legal barriers.

Agents are also doing the attacking. According to AI security firm Norforge, as reported by The Verge, a hacker broke into more than 100 websites using AI agents that mostly operated on their own, through exposed instances of the open-source Cairn agent platform. Norforge does not name the targets; Reuters, as cited by The Verge, reports victims included John Deere, SharkNinja and Cummins, and says the attacker set "preferred intrusion tactics" in fewer than 100 Chinese characters, with successful breaches taking about 23 hours. The companies' responses suggest limited damage; the significance is that the labour was the agents'.

Read all 14 signals →