AI work got cheaper as the money behind it got dearer

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 week's argument is that the price of a unit of AI work fell from every side while the price of the money behind it rose. Anthropic cut list prices on its small model by 90 per cent and two Western labs previewed large open-weight models a day apart. In the same days SpaceX was reported in talks on a US$40 billion GPU-backed loan, Broadcom and Oracle turned to private lenders, the IMF's managing director warned about hyperscaler leverage, and institutional investors baulked at buying Firmus at A$11 a share. The two movements meet at the question the Haiku release leaves open: whether volume grows fast enough to offset the price cuts. Lenders are now, in effect, underwriting the answer.

The unit of work got cheaper

Claude Haiku 5.5 lists at US$0.10 per million input tokens and US$0.50 per million output for prompts up to 100,000 tokens, a 90 per cent cut, and Anthropic says it costs on average about 75 per cent less to run than Haiku 4.5. The same release halved cache reads on Sonnet 5.5. Travis Smith's Telegraph Test found another lever: asking models to write machine-read records in a compressed telegraph register cut billed output tokens by 40 to 49 per cent on several small open-weight models, with downstream answers as good as or better than from plain text. It is one author's benchmark, and the technique did not pay on the one model that could not switch off reasoning.

Spending data already shows the pattern. Tomasz Tunguz reports that spending among the top 1 per cent of AI adopters fell 9.7 per cent in August while token usage rose 50 per cent since July and prices fell 41 per cent; he reads that as models reaching commodity status, an interpretation of one month from the heaviest users. SemiAnalysis estimates that subscriptions can take more than 40 per cent of Anthropic's inference compute while supplying about 10 per cent of its revenue. Cheap tokens are a margin question before they are a volume win.

Open weights came back from the West

Reflection put Beam, a 501-billion-parameter mixture-of-experts with 23 billion active, into early access with Apache 2.0 weights promised this month. It concedes that Kimi K3 is ahead on raw capability and pitches Beam on efficiency, by its own approximation. A day later Mistral opened a preview of Mistral Large 4, a trillion-parameter model with 49 billion active that Artificial Analysis scores at 38 on its Intelligence Index, against a median of 26 for comparable models; its pitch is cyber defence and sovereignty. Aleph Alpha's Kolibri, 78 billion parameters with 3 billion active, makes the sovereignty case from Germany, sized for serving cost rather than benchmark rank.

None of these claims the open frontier. What they offer is a Western, licensable alternative at a moment when a SemiAnalysis census finds that only 31 of 857 Chinese model releases since 2021 carried a published safety-evaluation result. That gap becomes a lever only if enterprise buyers start requiring the documentation. Money is also following the layer above the model: Nous Research raised US$90 million for an enterprise version of Hermes Agent, which it says, on its own internal estimate, drives about 2.5 per cent of global token usage.

Who pays for the volume

Cheaper tokens help the labs only if someone buys many more of them, and the consumer evidence is thin. Andreessen Horowitz found that in August 4.5 per cent of US consumers had an active paid personal subscription to ChatGPT, Gemini or Claude, with the top 1 per cent of payers accounting for 19.5 per cent of observed spend. The labs are converging on advertising, as Meta has with Muse. OpenAI said on Monday it would test visual ads during ChatGPT image generation, with free users able to opt out only by losing the feature. On Wednesday it began rolling out GPT-6 with Intelligent UI, answers composed as interactive tools, with free users getting GPT-6 Luna from the following day; a larger free audience is also a larger audience for those ads.

Meta, Walmart, Stripe and Sierra published a personal agent protocol that lets a business authenticate an agent acting for a real person. Amazon, which has blocked Meta's agents, is not among the members CNBC names, and neither OpenAI nor Anthropic has joined yet.

Compute moved onto credit, and the market pushed back

The buyers of compute are now borrowing for it. Apollo and several banks are in preliminary talks with SpaceX on a US$40 billion GPU purchase financed mainly with investment-grade debt, likely secured on the GPUs themselves; a credit-market source told CNBC that lenders still expect GPUs to hold their value for about seven years, longer than most hyperscalers' depreciation schedules. Broadcom has been working to arrange more than US$50 billion for the custom chip it is developing with OpenAI, and Oracle is negotiating a vehicle that buys chips and leases them to Oracle, moving the risk on the hardware's useful life onto the lenders. On Wednesday the IMF's Kristalina Georgieva warned that hyperscaler leverage could turn an earnings shortfall into a far-reaching shock. The seven-year collateral assumption rests on compute staying scarce, and Epoch's 2 October report models the opposite risk: memory shipped by 2027 could support far more agent hours than demand looks able to fund. Those are modelled ceilings, not forecasts.

Public markets were less accommodating. Firmus's ASX float was in doubt on Thursday after institutions baulked at A$11 a share and the lead managers cut the institutional price to around A$8; an SMH analysis put the target valuation at roughly 874 times last year's reported revenue. The AFR had already reported that Firmus missed its first rent payment, a fact that may have mundane explanations. NextDC and Goodman finished slightly higher, which suggests discrimination rather than flight. OpenAI told investors of roughly US$50 billion in annualised revenue, about US$18 billion below the figure widely reported last month, which a person familiar said had included partners' gross revenue. New Constructs, working from reported figures rather than the prospectus, told investors to avoid Anthropic's listing, valuing the company at about US$150 billion against a reported target near US$2 trillion. Even Samsung's record quarter was met with a 0.7 per cent share-price fall on Thursday morning.

Memory, packaging and permits set the pace

The physical inputs are not getting cheaper either. Micron said demand will exceed supply through 2027 and 2028, with a large share of 2027 high-bandwidth memory sold at prices "much higher" than this year's, and it is moving the HBM4E base die to a foundry, reportedly TSMC, including for custom memory co-designed with Nvidia that it says will lift prices and margins. GlobalFoundries signed a US$2 billion deal to make TSMC's CoWoS interposers in New York, but volume is not expected before the first half of 2028.

Permission is the other clock. Finland ordered Google to halt work at two data-centre sites until environmental assessments are complete, a month after Google pledged US$15 billion to the country; the order is procedural, but such assessments typically run for months. In Canberra, freedom-of-information documents show the government gave Anthropic a draft of its data-centre expectations while negotiating the memorandum in which Anthropic agreed to follow them; they establish early access, not influence, which Anthropic disputes. One input moved the other way: Wood Mackenzie finds four-hour batteries now undercut gas peakers on levelised cost in all 43 markets where it modelled both, which would shift the constraint on firming new load from turbine order books toward battery supply.

Containment became a product

AWS open-sourced Strands Box, a sandbox whose policy engine remembers what an agent has already done and enforces rules the agent cannot talk its way around. Microsoft announced Execution Containers for coding agents the same day, and Cohere is offering a lockdown mode, so three large vendors reached the same layer in one week. The week's threat reports show the surface it has to cover. Google and four other organisations have acknowledged protocol-pivoting flaws in which one compromised agent passes harmful instructions to others. Lumen's Black Lotus Labs found a campaign that has infected more than 3,000 servers, mostly exposed LiteLLM and Ollama deployments, and set their GPUs mining cryptocurrency. South Korea's president said signs had emerged of AI being used in hacks on banks, though no technical evidence has been published. Cheap self-hosted inference is also cheap attack surface.

What we'd hold

First, the Firmus book: whether the float proceeds at a lower price and how much stock existing holders sell into it, the clearest read on appetite for AI infrastructure without an operating history. Second, the pricing of the SpaceX GPU loan, its coupon, spread and advance rate against GPU collateral, which will show how far lenders believe the seven-year assumption. Third, whether OpenAI and Google reprice their small tiers within weeks of Haiku 5.5, and whether the labs' next revenue disclosures, on a basis that separates gross from net, show volume growing faster than price is falling. The week's question is whether the work gets cheaper faster than the money gets dearer; those three will say.

All 72 of this week's signals are on the feed →