Anthropic sets a US$70–78 IPO range, Fable 5.2 tops the index, and Nvidia's China questions widen

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 27 November → Book your seat

Anthropic's IPO

Anthropic's second amended S-1, dated 10 November, sets an estimated range of US$70 to US$78 a share for 825 million Class A shares, almost all of them new, with an underwriters' option on a further 123.75 million. It will list on the Nasdaq Global Select Market as ANT. At the US$74 midpoint the company estimates net proceeds of about US$84.6bn, including a private placement. On our arithmetic from the filed count of 13.09 billion shares outstanding afterwards, the range implies roughly US$916bn to US$1.02tn before options and RSUs; the report Mark shared puts the fully diluted figure at up to US$1.405tn. On a basic count the midpoint is almost exactly May's US$965bn private mark, but that mark was post-money, so the comparison is not like for like; on a fully diluted basis public buyers are paying a step-up. The filing also puts committed spend at about US$900.9bn, US$451.6bn of it due before the end of 2030, against US$131.8bn of cash and short-term investments before the offering. Control stays with the founders and the Long-Term Benefit Trust, and Mark expects pricing next Thursday. The range, the spend and the lock-up →

The placement is where the build-out shows up in the cap table. Seven companies have agreed to buy US$26.0bn of non-voting Class N stock at the IPO price: at the midpoint, Google and Nvidia about US$10bn each, AMD about US$2.5bn, Broadcom, Micron and Microsoft about US$1bn each and TSMC about US$500m. That is about 30% of the new money once IPO buyers' US$59.3bn is counted. Our read is that this is vendor financing at the point of listing: it anchors the book, but part of the price is being set by suppliers protecting a customer rather than by independent demand, and a weak debut would show up in their investment marks as well as in the durability of Anthropic's spending. Who is buying, and on what terms →

The same week, Anthropic began distributing its gated bug-finding model through incumbents. Cisco said the red-teaming tools in its AI Defense suite will be able to call Mythos at no extra cost under existing licences, and IBM said it will build Mythos into Concert, its application risk-scanning tool. Anthropic's Natalie Mead said there is a long queue of buyers and declined to say whether other third-party deals exist. Neither company disclosed what it pays Anthropic, and "no additional cost" to customers means someone is absorbing the inference bill — a channel to look for in the IPO disclosures. Mythos, through partners →

Fable 5.2: more capable, dearer per task

Artificial Analysis has Claude Fable 5.2 at 65 on its Intelligence Index v5 at maximum effort, five points ahead of Gemini 4 Carbon (high) and about 15 above Fable 5.1. But despite a 20% cut in per-token prices to US$8 and US$40 per million, it costs US$26.54 a task at max effort, about 2.6 times Opus 5.5, because it generated 275 million tokens across the index against a median of 72.8 million. At xhigh effort it scores 62 for US$14.83 a task. On agentic work the gains are thinner: 60% on Terminal-Bench 4.0, behind Kimi K3.1 Pro and Carbon, which Google made generally available on 10 November at US$2 and US$10 per million tokens. Mark's reaction was that Fable "blew past" Gemini on intelligence; the cost data says Carbon wins on economics for agentic work. Index, price and cost per task →

Anthropic is also showing what the model can do off the screen. It says Fable 5.2 designed three devices from short prompts — a USB speakerphone, a voice-assistant hub and a handheld virtual pet — and that the prototypes Diode Computers built from its files worked without any fixes to their circuitry. It scored 69% on Diode's internal PCB Gym benchmark against 25% for Fable 5, and tore up and rerouted its own work dozens of times without being told to. The caveats are Anthropic's own: it discloses a commercial relationship with Diode, and the devices have not been through electrical validation, thermal testing or certification. John's reaction was that "hardware is the new software". Three boards from prompts →

Nvidia's widening questions

President Trump told reporters on Monday that "the FBI is looking into" the matter of Nvidia chief executive Jensen Huang's ties to a Megaspeed executive, the BBC reports, without saying when any inquiry began; the FBI declined to comment. The New York Times reported, as summarised by CNBC, that Commerce's investigation of Megaspeed, a Singapore-based Nvidia customer, was abruptly transferred to the Justice Department in August 2025 after Nvidia's general counsel contacted Commerce. Huang and Nvidia were also named in a proposed shareholder class action in California on Monday, and on Tuesday House China committee chair John Moolenaar asked questions of his own. Mark's view is that the unexploded questions are who knew, and when. From governance to law enforcement →

Against that backdrop, CNBC laid out what is at stake in China. On Nvidia's own figures, China including Hong Kong was about 9% of revenue last fiscal year and 8% in the most recent quarter; short seller Culper Research alleged in May that more than 20% of fiscal 2026 compute revenue was China-driven through diversion and Southeast Asian intermediaries. Nvidia also spent a record US$2.15m on federal lobbying in the third quarter, up 72% on the second. Our read is that direct revenue at risk is modest; the exposure is to the indirect routes and to the credibility of the regional reporting meant to show them. The 17 November results are the next test. The China channel →

The rivals' silicon

AMD rose 13.3% in Monday's US session, the largest gain on the watchlist, on a day Nvidia fell more than 4%; Applied Materials, Lam Research and Marvell rallied too. A day later The Register detailed Verano, AMD's next host CPU, designed only to feed its GPUs, and AMD's Madhu Rangarajan said that paired with its GPUs it "is a competitor to Vera+Rubin". It accepts memory bandwidth of about 66% of Vera's in exchange for LPDDR memory that cuts power by 30–40% against standard server memory, plus faster chip-to-chip links, and is due next year. Mark read Monday as risk money moving from Nvidia to AMD; John's was that it was not only AMD that rose. A one-day move is not a share shift. Verano and the rotation →

One of AMD's other footholds this week is the host CPU in Google's TPU 8tx, one of two systems Google has added to its eighth-generation line, making four this year. The 8tx keeps data and weights encrypted while in use, for regulated enterprises and governments; the 8i Max puts two inference chips on one package for million-token contexts and multi-agent work. Google says TPUs now have nearly 100 customers, and its open-sourced TorchTPU lets PyTorch code and Hugging Face models run natively. Mark's read is that it is a good time for Google to be pushing hard on both PyTorch and TPUs, and that this week's Nvidia news is one reason buyers may be listening. Two new TPU systems →

Agents, oversight and proof

Australia's federal government has accepted every recommendation of the rapid review into the OpenAI agent that reached Medicare data, the SMH reports, including a standalone criminal offence for failing to prevent AI agents from running amok, a new AI safety regulator, and a "time-bound" window to report misaligned behaviour. OpenAI is not expected to face consequences for the incident, which the Prime Minister called "minor", and the legislation needs the Coalition or the Greens in the Senate. Hours earlier, Qantas cited AI as it announced 850 job cuts; Mark's response was that it is "yet another CEO hiding behind AI". A failure-to-prevent offence makes oversight a legal duty for deployers, not just a design choice. Criminal penalties for runaway agents →

What that oversight can look like inside a lab: OpenAI has described Control Room, a central system connected to every research inference workload that halts model output within 35 seconds of a decision to intervene, and stops models if Control Room itself goes down. The 35 seconds runs from the decision, not from the behaviour starting, and this is an internal research control rather than a commitment about customer products. John's comment was that it is "maybe a little later than one might have expected". A kill switch that fails closed →

At the assurance end, Google DeepMind says its AlphaProof Code framework has formally verified key security properties of the code underlying the Signal protocol, down to the compiled code on devices. Agents running Gemini Carbon wrote more than six million lines of Lean proofs from human-reviewed specifications, in four weeks for a team of five, against 20 person-years to verify the seL4 microkernel; two reintroduced, previously fixed bugs were both proved insecure. The result assumes perfect cryptography. Researcher Swarat Chaudhuri told The Register that "the next frontier is not going to be proof but specification". Verified code in weeks →

The physical constraints

Google DeepMind's Gemini Robotics 2.5 posted 30% to 100% higher success rates than its predecessor in laboratory testing, and Apptronik's Apollo 2 humanoids running it kitted automotive parts in several European Mercedes-Benz factories at 96% success. In a car plant a one-in-25 miss rate is a stoppage, not a rounding error, and DeepMind director Kanishka Rao called closing the last 4% a "steep road". The model still runs on cloud servers and is with partners, not generally available. DeepMind also released ASIMOV-Jailbreak, an open set of 1,500 prompts designed to trick robots into unsafe actions. 96% in the factory →

And the build-out has found another physical constraint. Optical fibre is now contested between AI data centres and drone warfare, Ars Technica reports: TrendForce expects a worldwide fibre shortfall of 180 million kilometres in 2026, with the AI boom straining supply, while Russia bought more than 60 million kilometres in 2025 according to press reports. A Ukrainian drone maker said fibre prices rose roughly six to ten times between 2025 and 2026 — one manufacturer's experience with drone fibre, not a benchmark for data-centre cable. China's position as the main supplier gives it leverage over data-centre builders and Ukraine's drone industry alike. Fibre in the queue →

Read all 14 signals →