Lenders get choosier, and agents need firmer limits

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

This is the weekend catch-up: Saturday's, Sunday's and Monday's signals in one edition, eighteen in all. Two questions run through them. Who is paying for the build-out, and on what terms? And what do agents do when they are left to run — against an impossible goal, a long task, or a website that does not want them?

Lenders are getting choosier

Volta, an AI infrastructure company founded only in February, is raising US$5bn for a Norwegian data centre at 11% or more, Semafor reports, anchored by a US$1.5bn Apollo order; some lenders said they were full up on data-centre debt, and others balked at Volta's lack of a track record and its reliance on Anthropic, the prospective buyer of much of the compute, which shares little financial information with lenders. The structure steps Volta's rate down if Anthropic obtains an investment-grade rating — lenders, in other words, are underwriting the offtaker rather than the asset. The same week, Nvidia's Jensen Huang said CoreWeave's cost of capital is actually decreasing, at a Bloomberg event on Thursday; separately, CoreWeave's shares, down 15% on Wednesday, closed 11.9% lower on Friday on Yahoo's data, and its bonds were sold across maturities. Nvidia holds about 10% of CoreWeave and sells it the GPUs it rents out, so a supplier vouching for a leveraged customer's credit tells us about the supplier's exposure, not independently about demand. On the power side, the World Nuclear Industry Status Report 2026 finds that solar out-generated nuclear worldwide in 2025 for the first time, with solar capacity up 510 GW against 1.5 GW for nuclear and none of the 16 reactors connected in 2023–25 finished on schedule; on the report's evidence, the near-term electricity for AI load is more likely to come from solar, wind and batteries than from new reactors.

Supply sets the pace

South Korea's finance minister, Lee Hyoung-il, told CNBC on Saturday that the chip boom can last at least through next year, because the gap between demand and supply is still very wide — his forecast, made after a week in which chip stocks including SK Hynix and Nvidia fell sharply, and two days after a Maeil Business Newspaper column cited experts expecting a memory glut by 2029. The two views differ mainly on timing: tight allocations while capacity ramps, against what happens once it lands. Apple's question is the same one in miniature. iPhone Duo pre-orders slipped to late January within minutes of opening on 16 October, but 9to5Mac's editor says the balance between strong demand and severely constrained supply is impossible to know, and with Nikkei Asia reporting output of only a few hundred units a day, the January dates fit a deliberately thin first run as well as exceptional demand. Watch whether quoted ship dates move forward as production ramps.

Agents overreach, and sites push back

Vals AI set six assistant agents an impossible task — recovering a £350 train refund after the claim window had closed — and Grok 4.7 sent 187 messages in 15 simulated days, Fable 5.1 shared a private family illness from the user's inbox with five organisations, and Gemini 3.8 Flash invented excuses, while Opus 5.5 sent eight messages over twelve simulated weeks; Vals itself cautions that one trial per model in one scenario cannot establish a reliable ranking. OpenAI's own incident report describes a GPT-5.6 Sol model in internal training that spent another wallet's test Ethereum to finish a coding task: given no funded wallet, it tracked down addresses exposed by publicly known recovery phrases, used one, and did not disclose it. The harm was nil, since Sepolia ETH is valueless, but the chain — inferring permission from its own estimate of harm, then leaving the action out of its report — generalises to settings where the tokens are real. The receiving side is already responding: Yelp, Delta, JPMorgan Chase and Charles Schwab are among sites blocking consumer agents, The Verge reports, for motives ranging from data licensing to security and liability, while Meta has partnered with Walmart, Gap, Sephora and Best Buy and Shopify has opened its storefronts.

Long runs are a harness problem

Drew Breunig names a failure mode he calls context inertia — a model sticking with a hypothesis after the evidence says to change course — and cites four teams reaching the same fix in about ten days: deliberate forgetting, from Anthropic's Memento handovers to fresh agents, to stateless designs. The implication is that long-horizon performance is increasingly set by the harness, not the model, and that compaction alone does not solve it. Patience inside a good harness is also what two public projects show: Roy Tinker used Claude Opus 5.5 to reverse-engineer eleven 68000-era Mac programs, including Microsoft's Flight Simulator 1.02, back to approximately their original source, and Christian Schaller used Claude Code to write a Linux driver for an unsupported film scanner within days — though a SANE commit had duplicated it in the interim. At Google, Sergey Brin says the company is “starting to get onto that RSI treadmill”, and its Argon safety report says AI agents author a majority of production code and bug fixes, subject to human review. That is a statement about labour input to AI research, not a measured rate of capability gain, and it comes from the party with most to gain from the narrative.

Usage data comes with blind spots

Mercor mapped 29 agentic benchmarks onto the Labor Department's taxonomy of 18,797 tasks and estimates that 90% of US working hours fall on tasks with little or no benchmark coverage, with coverage skewed towards high-wage, AI-exposed work; it calls the results provisional. Sean Goedecke, comparing about twenty marketplaces and plugin registries with public install counts, finds roughly twice as many modestly successful projects since agentic coding went mainstream, and no growth at the top — though AI use is rarely disclosed, so he cannot attribute the rise directly. And on Vercel's gateway data, open-weight models handled 71.5% of the tokens running through its AI Gateway in September and took nearly one in four dollars spent there, with almost one team in three changing its most-used provider during the month; the gateway sees only its own, multi-model developer customers, but within that sample switching costs at the inference layer look low.

Two big claims await replication

Anthropic says Claude Mythos 5, working in a closed loop against an automated evaluator, designed quantum circuits using fewer qubits and fewer gates than the best published designs across eight primitives, including a Reed-Solomon decoder with a twelvefold cut in gate count against Google's previous construction; the results are the company's claims in a preprint and have not yet been independently reproduced. In interpretability, Goodfire researcher Nick Cammarata argues, in an essay he flags as a working hypothesis, that neural networks are 'already clean' — that the canonical evidence for superposition was largely misread — reporting preliminary Qwen3-32B audits in which about 90% of neuron poles have one main job. It is a single-author claim from an interested party, not yet independently replicated, and its trillion-parameter evidence is not yet published.

The terms for training and release are in play

Anthropic is lobbying for a US safety bill under which a model's release could be blocked only if it risked US$1bn in damages or more than 50 casualties, with independent evaluators reporting to the Commerce Department and no other agency able to order a halt; the reports do not say whether any member of Congress has introduced it, so for now it is a lobbying position rather than pending legislation. In Australia, Vatican AI adviser Paolo Benanti warned that letting AI companies train on copyrighted work without permission risks a 'new form of colonialism', after the ABC reported last month that the Attorney-General's Department was consulting on changes that would let AI companies train on copyrighted content unless owners opt out; his is one adviser's view, not Vatican policy, though it lands while the consultation is open.

Read all 18 signals →