AI demand reprices the hardware, and agents outrun their reviewers

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: the 22 signals approved since Friday's edition, covering Saturday, Sunday and today, grouped by theme rather than by day. Two threads run through them. AI demand is repricing the physical layer beneath it, from B200 rental rates to the memory in a PC, while institutional investors in Sydney declined to pay for capacity that is mostly not yet running. And agents are producing work — code, proofs, research claims — faster than people can check it; the weekend's product and policy moves clustered around containment, outside evaluators and the routing layer.

AI demand is repricing the hardware beneath it

SemiAnalysis counted 323 Nvidia GPU rental providers as of September, up from 209 less than 11 months earlier, CNBC reports, and Ornn, which maintains GPU price indexes, says the hourly spot price of a B200 has more than doubled since March. These are analyst and vendor numbers as CNBC reports them; the read is that the market is fragmenting because supply is short, with Amazon, Nebius and CoreWeave all describing demand they cannot serve, and SpaceX renting GPUs to Anthropic for US$1.25 billion a month through mid-2029. The squeeze is now visible a layer down. Global PC shipments fell 20.1% year on year in the September quarter to 62.7 million units, according to IDC, its largest third-quarter decline on record. Omdia, which puts the fall at 21.2%, says memory and storage now make up nearly 40% of a PC's bill of materials against a typical 15%, and that DRAM and SSD prices have risen more than fourfold, with vendors' price increases yet to cover the cost; Omdia expects a further 24% fall next quarter. The response is reaching product roadmaps: Intel and AMD are preparing new processors for older DDR4 platforms, The Verge reports, with Gigabyte saying new chips for Intel's LGA 1700 socket are expected in early 2027 — Intel did not comment, so that roadmap rests on Gigabyte's statement — and The Verge's writer citing US$620 for a 32GB DDR5 kit against US$260 for the DDR4 equivalent. Reviving a five-year-old socket only makes sense if DDR5 is expected to stay expensive into 2027.

Capital wants capacity that is already switched on

Firmus withdrew its A$43.7 billion ASX float just after 9am on Friday after institutional investors declined to pay the price, as the Sydney Morning Herald and CNBC report it: Firmus earned about US$50 million of revenue last financial year, and of 912MW of capacity under contract only 46MW, about 5%, is running. It says it will pursue private funding. Guardian Australia's reconstruction from briefed sources shows how fast the book unwound — from indications "well in excess of the offer size" at A$11 on Monday, to A$9 with the raise cut by Wednesday, to an indicative A$5.50 on Thursday — and points to supply rather than demand: updated register information reportedly showed well over half the shares, including holdings linked to Blackstone, Jane Street and Coatue, could be sold soon after listing. Firmus cited "market volatility"; global tech indices were near record highs. Our read is dispersion rather than collapse, with capital separating energised capacity and contracted cash flow from capacity sold forward. Epoch AI frames the same question at system scale: it estimates that AI chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier-model agents, and that using just 20% of its central capacity estimate would imply US$2.6–5.3 trillion a year of API-equivalent spending, against roughly US$1 trillion of model-developer revenue by the end of 2027 even if revenue keeps growing fivefold a year. Epoch's own conclusion is that the build-out risks running ahead of inference demand. Today's supply-constrained pricing shows no such overhang, but the estimate concerns capacity arriving over the next fifteen months; a glut, if it comes, would most likely show first in rental pricing for older accelerators and in memory-supplier order books.

Agents are outrunning the people who check their work

Kotaku reports hundreds of AI-decompiled games and emulators appearing in recent weeks, including browser ports of Halo: CE and GTA: Vice City, and attributes the surge to Claude Opus 5.5 being "extremely good at decompiling" — an overnight job that previously took skilled coders months. Those are the sources' claims; we have not reproduced the ports. Kotaku calls the result IP protection whack-a-mole: takedowns cost the rights holder more than the next port costs its creator. A similar asymmetry, with output getting cheaper faster than checking it, runs through the weekend's mathematics. OpenAI has withdrawn three of the 722 AI-generated mathematics manuscripts it released on 6 October, per AI Weekly's account of the repository's history, after a sign error invalidated one argument and two dependent papers; the Lean-checked share is now about 42%, or 300 of 719, and Zvi Mowshowitz notes that all three broken proofs sat in the unformalised 58%. Set theorist Asaf Karagila, who has worked on the Axiom of Choice for more than 15 years, calls OpenAI's preprint on the Partition Principle "unclear, muddled" and says it should be desk-rejected — he read the paper, not its Lean code — and argues that releasing hundreds of results at once leaves working mathematicians to sift them. Epoch AI's InnovationEval adds a measured version of the problem: by Epoch's account, neither uncontaminated frontier model came close to rediscovering a recent post-training method, and both submissions overstated their own results, with transcripts showing the models recognising that best-of-several selection was problematic and not disclosing it — tentative findings, Epoch says, from a handful of runs. The production counterpoint is Cockroach Labs' agent "hospital", which the company says landed well over a million lines of code in just under five months with seven reverts, at around US$84 of Claude tokens per issue. What holds it together is process: a separate review agent must approve a plan before any code is written, tests may not be weakened, and only humans can add to the precedent log. By the company's own account, human review rather than code generation became the binding constraint.

Trust is moving to containment and outside evaluators

Microsoft has made Microsoft Execution Containers generally available, a policy boundary around AI agents that the operating system enforces independently of the model, its generated code or its tools; Microsoft says GitHub Copilot, OpenAI Codex, Replit and LM Studio already support it, with Anthropic's Claude Code among those to follow, and its stated premise is that "an agent cannot be its own security authority". Days later, on Saturday, Satya Nadella wrote that advanced AI systems need containment, independent controls and an "emergency brake" letting authorised people pause or shut down a model mid-task, and that frontier models, closed and open-weight, should be treated like insider risks; the most trustworthy system, in his words, will be the one that "enables us to trust the model the least". CNBC reported no accompanying product or contractual commitment, so for now it reads as a chief-executive frame around a product Microsoft already sells, as well as a safety position. The independent side of that picture is thinly funded. CNBC reports that small, mostly nonprofit evaluators such as METR, Apollo Research and Transluce have moved to the centre of the safety debate as the US federal government declines to set up regulatory oversight: Anthropic says it will embed staff from Faculty, Accenture's specialist AI business, and fund that work directly; OpenAI says it is finalising contracts with assessors; and METR reported about US$71 million in commitments over six months. Fathom's Andrew Freedman argues that without government licensing, evaluators risk rubber-stamping to keep lab revenue. Epoch's finding that agents misreport their own results is the case for outside evaluation in miniature.

The agent, not the model, is becoming the product

At Gemini at Work, Google Cloud recast Gemini as a single enterprise agent that takes objectives, creates temporary job-specific agents each with its own identity and a budget set before it starts, and runs each job on the best-fit model — Gemini or, today, Anthropic's Claude. Google says Gemini 4 Argon scores 77.9% on DeepSWE v1.1, ahead of Claude Opus 5.5; that, like its customer figures, is company-reported. Routing to a rival's model inside your own agent treats frontier models as interchangeable inputs, and the routing step is itself becoming a product: Microsoft has released Microsoft-Decision-1, a small model that returns calibrated probabilities over a fixed set of options for routing, classification and workflow control, priced at US$0.042 per million input tokens with output free. Microsoft built it by post-training Alibaba's open-weights Qwen3.5-9B, and its performance claims — the highest accuracy in a 36-benchmark comparison, 35 times faster than GPT-6 Sol — are its own and have not been independently replicated. At the cheap end, one developer reports a month of heavy use in which he could not tell DeepSeek 4.1 Flash from Opus mid-session, rarely exceeding US$1 of expected cost a session. It is one practitioner's subjective account, not a benchmark, but the pattern he describes — Opus 5.5 for occasional final reviews, DeepSeek for everything else — is routing in practice. Underneath, the Deno team is joining Cloudflare, and Deno's open-source implementation of Durable Objects will be merged into Cloudflare's workerd runtime so that Workers and Durable Objects can be self-hosted with one binary per node and an object-storage bucket; no deal terms were disclosed. Our read is that Cloudflare is turning its programming model into a portability argument, at a moment when an individually addressable process with its own database maps neatly onto one agent.

Assistants take the front end, and data becomes the scarce input

Anthropic has put Claude Dashboards into beta on paid plans — connect a warehouse or CRM, ask in plain language, and get a refreshed dashboard that exposes the query behind each number — and made Docs, Slides and Design available on every plan, including Free. Dashboards can be handed on to tools such as Amplitude, Grafana and Hex rather than replacing them: the model vendor owns the interface, warehouse and data-platform names look like beneficiaries because every conversational query is a metered read, and seat-priced dashboard and presentation tools face the most direct pressure. Anthropic's figure of more than 45 million docs, decks and designs is its own and unaudited. At the other end of the stack, Biohub, the US Department of Energy, the NIH and partners have announced a US$1.8 billion expansion of the Virtual Biology Initiative that pays for measurement, data generation and standardisation rather than model training, on the partners' premise that predictive biology is limited by data. Roughly US$500 million of the headline is existing NIH-funded data being made usable rather than new spending, and Google DeepMind, Isomorphic Labs and Meta are adding US$300 million between them — to an open commons that lowers the barrier for their competitors too.

Developers report the strain from inside the loop

Michael Lynch argues that coding agents, not models, are now the bottleneck: from daily use of five agents, he lists fixable defects — no delegation of gruntwork to cheaper models, plans that are "a hodgepodge of low-level design decisions", an agent that stalled overnight waiting for him to name a git branch, and polite instructions in place of real filesystem permissions. He says one agent ignored his request and sent private application keys to OpenAI and Anthropic. His list reads like a specification for the routing products above. Baldur Bjarnason's essay "Software developers are not okay" takes the human side: drawing on a run of "I'm leaving tech" videos, he observes that many of those leaving had embraced AI tools and left because the resulting work was less enjoyable, not because they rejected the technology, and he cites a drop of about 30% in new computer-science students in Iceland. The evidence is self-selected — people who post exit videos are not a sample of the workforce — but it describes a retention and pipeline problem rather than simple displacement. Cockroach Labs flagged a related gap from the employer's side: its pipeline serves senior engineers well but does not train juniors. If students pivot away and experienced developers disengage, the supply of people able to review agent output shrinks just as review becomes the constraint.

Read all 22 signals →