AI's build-out is becoming a credit story

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

AI capex becomes a credit story

Apollo and several banks are in preliminary talks with SpaceX to help finance a US$40 billion purchase of Nvidia GPUs, CNBC reported, mainly through the investment-grade debt market and likely with the chips themselves as collateral. A credit-market source quoted by CNBC says lenders still expect GPUs to hold their value for about seven years — longer than the depreciation schedules most hyperscalers use — and CNBC also reports that AI-related bonds have sold off since June, so the deal tests how much appetite is left for GPU-backed paper. The same day, in Singapore, the IMF's Kristalina Georgieva said "the AI building boom is inflationary" and warned that, should earnings fall short, hyperscaler leverage and large global holdings of US equities "could turn a disappointment into a far-reaching shock"; she added that long-dated bond issuance by AI-related borrowers now competes with governments for capital. On the consumer side, a16z's seventh Top 100 consumer-apps report, drawing on YipitData card spending, finds that only 4.5% of US consumers held a paid personal subscription to ChatGPT, Gemini or Claude in August, up from 2.1% a year earlier, and that the top 1% of payers account for 19.5% of observed spend — more than the bottom half combined, at 16.6%.

Sites, permits and power set the pace

Finland's Licensing and Supervision Agency has ordered Google to suspend construction at two data-centre sites, in Muhos and Kajaani, until environmental impact assessments are complete, CNBC reported — a month after Google committed US$15 billion to the country. The order is procedural rather than a rejection, and Google says it will follow the regulator's guidance, but such assessments typically run for months. In Canberra, freedom-of-information documents reported by the Sydney Morning Herald show the federal government gave Anthropic a draft of its data-centre expectations while negotiating the memorandum in which Anthropic agreed to follow them. Anthropic says it had no influence on the final version, and the documents establish early access, not influence; what makes it matter is that the expectations are set to become mandatory, and Treasury's foreign-investment policy already says aligned proposals will be prioritised — Anthropic's proposed A$32 billion Western Downs campus still needs FIRB approval. Where time to power is the constraint, Fervo's Cape Station in Utah began selling enhanced-geothermal electricity to the grid on 30 September: its first block, about a third of a planned 100 megawatts, took 23 months from groundbreaking, Google is among the committed buyers, and Fervo aims to bring future blocks online in as little as 18 months, though it has not disclosed per-megawatt cost.

Output keeps getting cheaper

Anthropic released Claude Haiku 5.5, which it says costs on average around 75% less to run than Haiku 4.5, with list prices cut 90% for prompts up to 100,000 tokens; it also halved cache-read prices on Sonnet 5.5. The benchmark gains it reports over Haiku 4.5 are vendor numbers. At the other end of the stack, Microsoft opened preorders for the Surface Laptop Ultra, from US$2,599 with Nvidia's Blackwell RTX Spark GPU and up to 128GB of unified memory, and demonstrated apps that route some coding requests to the local GPU and others to the cloud — a hedge on local inference rather than a switch away from the cloud. And one author's own test suggests part of the token bill is style rather than substance: Travis Smith reports that instructing models to write machine-read records in telegraph-style "cablese" cut billed output tokens by 40–49% on several mostly small open-weight models, with downstream models answering from those records as well as or better than from plain text. It is not an independent replication, GPT-5-mini saved only 17.7%, and accuracy fell from 91.5% to 74.4% when models answered people directly in that register.

Checking the work is the scarce part

A preprint by Alexander Bastounis, Fabian Circelli and Anders C. Hansen argues that having AI translate a prose proof into Lean and then checking it mechanically may offer no confidence in the original argument, and states that the Lean formalisation attached to OpenAI's announced Navier-Stokes blow-up proof does not correspond to the prose proof. That is one paper's claim, and OpenAI has not, as far as we have seen, responded. Separately, OpenAI has published 722 AI-written maths manuscripts in 372 families, produced by an internal model that was given about 4,000 open problems and spent on average about three hours of ChatGPT Pro thinking compute per result. Lean formalisations cover many but not all of them, and OpenAI itself says some of the unformalised results could have issues. A catalogue that size moves the bottleneck from producing candidate results to checking them — and on the preprint's argument, a formal check certifies only the statement it was handed.

AI widens the attack surface

Lumen's Black Lotus Labs says a campaign it calls Canto Incognito has infected more than 3,000 servers since April, mostly internet-facing LiteLLM and Ollama deployments, mining cryptocurrency on their GPUs and locating its command server through an IP address encoded in a poem on GitHub. The researchers' own explanation is that AI makes these tools easy to deploy but does not check that they are patched. In Seoul, President Lee Jae Myung told his cabinet that "signs have emerged of AI being used" in recent hacking incidents against banks, Reuters reported; the Financial Services Commission said Shinhan, KB Kookmin and others reported cyberattacks. Authorities have not said what tools were involved, and no technical evidence has been published. On the defensive side, AWS has open-sourced Strands Box, an agent sandbox whose policy engine remembers what an agent has already done — allowing, say, no more than three Slack posts in ten minutes — and enforces rules deterministically rather than relying on the model to comply. The same day, Microsoft announced Execution Containers to confine coding agents on Windows.

Read all 14 signals →