The AI trade's numbers are getting a harder look

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 labs' revenue lines are now in dispute

OpenAI told investors it reached roughly US$50 billion in annualised revenue at the end of September, CNBC confirmed after the Financial Times first reported the figure — about US$18 billion below the US$68 billion widely reported late last month, which a person familiar with the matter said included gross revenue from OpenAI's partners. It is a question of definitions rather than growth, and it arrives as both leading labs head toward public markets: Anthropic said in July its run-rate was US$65 billion, and neither lab has published audited numbers on a common basis. The same week, independent research firm New Constructs told investors to avoid Anthropic's planned listing and put its value at roughly US$150 billion, against a reportedly sought US$2 trillion; the firm has not seen the prospectus, works from reported figures, and argues that open-weight competition leaves closed models without pricing power. Its record is mixed — WeWork's listing was pulled weeks after New Constructs' founder named it 2019's most ridiculous IPO, while his 2020 pick, DoorDash, has held up well. In the NOOPS discussion, John Allsopp put US$2 trillion at about 20 times revenue, and Mark Pesce refined that to 20 times current and about three times forward revenue — their arithmetic, resting on a New York Times projection of a US$100 billion end-2026 run-rate holding up. The open question is which input the prospectuses settle first: gross-versus-net treatment of partner revenue, the size of compute commitments, or the path of operating losses.

Investors and buyers are getting choosier

In Sydney, Firmus Technologies' float was in doubt on Thursday after its joint lead managers cut the institutional price from A$11 to around A$8 when demand fell short, the Sydney Morning Herald reported (A$8.25 by the AFR's account); people familiar with the matter said investors were concerned that 57.6% of shares would be free to trade on listing, and a separate Herald analysis puts the A$11 target at roughly 874 times last year's reported revenue of about A$50 million. Established operators NextDC and Goodman finished slightly higher, which suggests discrimination rather than an exit from the sector. The same day, Samsung estimated third-quarter operating profit of 107.4 trillion won, its first above 100 trillion, and TSMC reported September revenue up 54.6% year on year — down 0.6% on August and, per CNBC's correction, not a record month. Samsung's shares still fell 0.7% on Thursday morning, a reaction eToro's Josh Gilbert read as a measure of how demanding the AI trade has become. On the buying side, Microsoft and Meta are reportedly curbing staff use of Anthropic's Claude in favour of in-house tools, according to a summary of a paywalled report by The Information: Microsoft's monthly AI spending caps in its cloud and AI division reportedly fell from US$100,000 per employee to about US$10,000 in most cases — ceilings, not actual spend — and Meta's Claude Code users reportedly halved to about 30,000. Separately, Tomasz Tunguz reports that spending among the top 1% of AI adopters fell 9.7% in August even as token usage rose 50% since July and prices fell 41%; he reads that as models commoditising, an interpretation resting on one month of data from the heaviest adopters.

The build-out turns to credit and onshore packaging

Broadcom has been working to arrange more than US$50 billion in financing for the custom AI chip it is developing with OpenAI, with Apollo and Blackstone among the lenders approached, according to The Wall Street Journal as summarised by Stocktwits; the talks are early. Separately, Oracle is in talks with Apollo and Goldman Sachs on a structure in which investors fund a separate company that buys chips and leases them to Oracle — converting capex into lease obligations and moving the risk on the hardware's useful life onto the lenders. On packaging, GlobalFoundries signed a five-year, US$2 billion deal to make silicon interposers for TSMC's CoWoS packaging at its Malta, New York fab, but volume production is not expected before the first half of 2028, so accelerators bound for US data centres stay exposed to Taiwan for at least two more years wherever their wafers are made. Inside the memory stack, Micron will use a foundry-made base die for HBM4E and for NVHBM, the custom memory it is co-designing with Nvidia; The Elec reports the die is going to TSMC, and Micron's chief operating officer said NVHBM should lift selling prices and margins. TSMC, already making the GPU and the CoWoS packaging, is increasingly making the memory's logic layer too. On power, Wood Mackenzie finds four-hour batteries now undercut open-cycle gas turbines on levelised cost in all 43 markets where it modelled both; the reports do not discuss data centres, but if storage beats gas for peaking duty, the constraint on new load shifts from turbine order books toward battery supply chains and interconnection queues — with battery capex up about 2% this year, its first rise in three years, as a caution against assuming a smooth decline.

Value moves to the interface and open agents

Nous Research raised US$90 million from investors including Nvidia and Microsoft's M12 to build an enterprise version of its open-source Hermes Agent, which the company says, on its own internal estimates, drives about 2.5% of all global token usage; the figure is unverified, but the bet echoes Tunguz's argument that in a commoditising model market the gains go to whoever controls the interface and routing. In China, Manus's parent Butterfly Effect raised more than US$500 million in its first round since regulators forced Meta to unwind its roughly US$2 billion acquisition, led by Boyu Capital and IDG Capital; it disclosed no valuation, and CNBC notes appetite for agent start-ups has held up even as model price competition intensifies. At the application layer, ArtCraft has released open-source clones of Adobe apps, and lookalikes of Word, Excel and PowerPoint, which its developer says he built with Claude Opus 5.5; by ArtCraft's own account they are early — WordCraft is pre-alpha at an estimated 62% feature parity — and what matters is the falling cost of a credible substitute rather than any near-term threat to Adobe or Microsoft revenue. And OpenAI's Intelligent UI has ChatGPT answering with interactive components — forms, charts, maps and small generated tools — as GPT-6 rolls out to paid and free users, as MacRumors reports it: a model composing a working interface at answer time is the mechanism that, at scale, substitutes for single-purpose apps and lightweight SaaS front-ends.

Chinese labs ship fast and disclose little

SemiAnalysis's census of 857 model releases by nine leading Chinese developers finds that only 31, or 3.6%, have ever been accompanied by a published safety-evaluation result from the developer, and just nine had one at launch — with the caveat that not found means not published in the materials checked, not untested. Releases climbed from three in the first quarter of 2023 to 101 in the third quarter of 2025, more than half open-weight, while those with any safety result never exceeded seven a quarter. Chinese rules govern content, minors and applications tightly but impose no duties triggered by compute or capability, unlike the EU's 10²⁵ FLOP threshold or California's SB 53. The investment read is that if US labs slow voluntarily while Chinese labs keep up a release-and-price war, Chinese open-weight models go on setting the floor on token prices — the same open-weight pressure New Constructs leans on in its case against Anthropic's valuation.

Read all 14 signals →