The IMF lists an AI share correction among its top stability risks

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

The IMF has put an AI share-price correction on its short list of risks to world financial stability, with dollar figures on the circular financing between hyperscalers and weaker AI firms, and a Jefferies note puts about half of the four biggest clouds' contracted backlog on two private labs. Further down: memory shares slide while phones run short of memory, agents outgrow infrastructure built for people, Google rations its newest frontier model, a bank and a wary public push back, and two studies probe the limits of safety testing.

The IMF puts numbers on AI's circular financing

The IMF's Global Financial Stability Report names an AI share-price correction as one of three "salient risks" to world financial stability, alongside rising government borrowing costs and a major cyber-attack, the Guardian reports. The fund says weaker or less durable AI investment and profitability could trigger a correction amplified by leverage, concentrated exposures and circular financing, and it puts figures on that web: US$300bn of transactions between hyperscalers and "weaker AI firms" over the past three years, supplemented by US$650bn of debt, in exchange for US$1.4trn of future financial commitments. That is the IMF's list of tail risks, not a forecast; the fund doesn't say whether or when a correction comes. A broker's note reaches the same concentration from another direction: Jefferies estimates that about half of the roughly US$2.3trn contracted backlog at Microsoft, Alphabet, Amazon and Oracle is attributable to OpenAI and Anthropic, and argues that cheap Chinese, open-weight and small models threaten the prices those two charge. It is one broker's view, and the OpenRouter token data it leans on likely skews towards developers and price-sensitive work. Jefferies also notes that "other non-operating income", which includes investment gains, accounted for all of the four hyperscalers' US$222bn rise in annualised earnings over the four quarters to 2Q26. One of those two counterparties is about to show its numbers to a select audience: Anthropic plans to disclose its third-quarter financials for the first time at a meeting with potential institutional investors on Wednesday, US time, CNBC reports, citing two people familiar with the matter; Anthropic declined to comment. CNBC says it is unclear whether Anthropic will count cloud-partner sales the same way OpenAI did last week, when it shared a US$50bn revenue figure with investors, below the widely reported US$70bn — a number that, according to a person familiar, stripped out revenue from cloud partners rather than signalling weaker demand.

Memory shares slide; phones still can't get enough

Memory and chip-equipment shares fell hard in Tuesday's session: CNBC reports SK Hynix down more than 10% and Samsung Electronics 7.4% in Seoul, with SoftBank Group off 8.5% and Tokyo Electron 7.2% in Tokyo. CNBC links the move to worries over AI financing and to China's CXMT, which said over the weekend it had made significant progress on a chip architecture designed to pack more data into less space; CXMT's own shares closed 2.8% lower in Shanghai. The analysts CNBC quoted were sceptical that CXMT changes the near-term picture: Rayliant's Phillip Wool called the reaction disproportionate, since CXMT has yet to demonstrate large-scale production and the Korean makers dominate HBM. Mark's reading is that the market simply does not grasp the scale of demand for RAM. The shortage is plain at the consumer end of the chain: Counterpoint's preliminary estimates have global smartphone shipments down 20% year on year in the third quarter, the lowest Q3 level since 2012, a decline Counterpoint attributes to memory shortages and rising DRAM and NAND prices. Samsung and Apple gained share while shipping less — Apple's 22% is its first Q3 share above 20% — while Xiaomi's shipments fell more than 30% and OPPO's and vivo's more than 20%. Counterpoint does not expect a meaningful recovery until component costs and supply improve.

Agents outgrow infrastructure built for people

Anthropic says agent-written code made source control its biggest development bottleneck, and that it replaced git for its monorepo with an in-house system, Scion, in five months. By its account, merged changes rose from about 2,500 a day in May to more than 30,000 a day by September, and Scion's infrastructure footprint is 200 times smaller in CPU and memory, with CI checkouts taking seconds rather than minutes. These are vendor claims about the vendor's own unusual workload, and Anthropic says Scion is not a general git replacement — but it expects similar pressure points to reach other heavy users of coding agents within six to 12 months. AMD is betting agents change the hardware mix as well: its new Helion rack platform is aimed at agentic AI rather than frontier training and leans heavily on CPUs, with a Hyperscale configuration of up to 128 EPYC processors designed to sit alongside a Helios GPU rack. The premise, as HotHardware explains it, is that much of an agent's work is conventional computing — calling tools, running code, querying databases — and the near one-to-one CPU-to-GPU mix it cites is an estimate, not market data. AMD expects Helion to be available from ecosystem partners next month.

Google rations the frontier and opens a private app lane

Google began the wide rollout of Gemini 4 Argon on 13 October, but only to paying users, Ars Technica reports: AI Pro and AI Ultra in the Gemini app, and Ultra only in the Antigravity coding environment and AI Studio. API access costs US$2 per million input tokens and US$10 per million output through 30 November, a 50% launch discount, rising to US$4 and US$20 after that. Our read is that this is a company rationing compute rather than racing prices down — the frontier priced for scarcity while commoditised capability keeps getting cheaper. At the other end of the market, Google is building a funnel from personal builders to registered developers: apps built in AI Studio can now go straight to a private "Personal Library" on the Play Store, unreviewed by Google and shareable with "a small group" of friends and family without a developer account; Google has not defined how small. Mark paired the news with Nate Meyvis's conjecture that AI may be replacing broadly available software with private software, the category Google is now giving distribution.

A bank and a wary public push back

JPMorgan will not let outside AI agents act on customers' behalf in their accounts, Jamie Dimon told analysts on the bank's third-quarter call: as CNBC explains it, agents can read balances through Plaid-style connectors but cannot send money or pay bills, and there is no timeline for changing that. CFO Jeremy Barnum said the bank has begun piloting its own tool, Smart Cash, to move funds to maximise yield, which CNBC reads as defensive given worries that agents could sweep cheap deposits into higher-yielding accounts. John's view is that this has to be resolved: "nope is not a policy or strategy". Resistance of a different kind is gathering around AI's physical footprint: a CNN investigation of more than 100 police reports found police have investigated dozens of threats linked to US data centres since last year, from bomb claims to online calls for violence against council members. CNN says the threats have rarely led to violence but have consumed hundreds of hours of police time, and a CNN poll conducted by SSRS in September found nearly nine in 10 Americans more concerned than excited about the rise of data centres. For investors, the exposure is less the threats than what they signal for permitting as the midterms approach.

Two studies probe the limits of safety testing

Anthropic published research arguing that small models may be "months away" from self-replicating on cheap internet-of-things devices: a 9B-parameter model got inference running on eight of nine hardware types, and in simplified evaluations GLM-5.2 completed end-to-end replication. The critics' caveats are substantial. The Register notes that Anthropic largely set aside the exploit stage, so replication ran on nodes deliberately left open inside a purpose-built harness; that only GLM-5.2 succeeded; and that the projected doubling time of days compares with Mirai's roughly hourly doubling in 2016. Mark pointed to a December 2024 arXiv paper reporting agents that created live copies of themselves, and was unimpressed by the novelty. A NeurIPS 2026 paper poses a harder problem for evaluators: its authors argue models can provably hide arbitrary capabilities, even from someone with full access to the weights, by embedding an encrypted network that runs only on a secret trigger. In their proof-of-concept, the hidden capability resisted fine-tuning, weight noising, gradient-based trigger search and latent adversarial training. On that argument, evaluations still catch capabilities that emerge by accident but cannot certify the absence of one deliberately planted — shifting weight from testing the finished model to verifying how it was produced.

Read all 13 signals →