Firmus targets a $5B IPO, Australia's biggest listing since Telstra
The pacing debate moved markets all morning. Now it has a counterweight: the biggest public-market test of the AI buildout, roadshowing this week in the least likely place.
Firmus Technologies is seeking to raise up to $5 billion (A$7 billion) in an initial public offering, targeting an October launch with a debut on the Australian Securities Exchange as early as late October. The Australian AI-infrastructure company began meeting investors in Asia on Monday, with Europe and the United States before the Australian leg of a two-week international roadshow, Bloomberg first reported and Reuters corroborated; the size and timing can still change, and Firmus declined to comment on all of it. The numbers already sit against history: at $5 billion, Dealogic says it would be the fourth-largest IPO in the world this year, behind SpaceX, CXMT Corp and Cerebras Systems, and the second-largest Australia-listed flotation on record, behind only Telstra's $10 billion in 1997. Firmus raised $2 billion of equity in August at a post-money valuation above $10.5 billion, with Nvidia, Coatue Management, Blackstone funds and Jane Street participating, to scale its "AI factories" across Australia and Asia Pacific. The timing is the story: Firmus is selling itself the same week the pace-the-frontier call took 10% off SoftBank and dragged Nvidia lower in U.S. premarket — the first live test of whether public investors will fund AI data centers in the week three CEOs told them the pace is the risk. Today's deep dive has the other half of that ledger — Tokens halved in price. The bill to make them didn't.
The pacing camp has its first heavyweight published opponent. Ben Thompson's Monday Stratechery piece, "Pacing the Frontier, AI's Digital Limits, AI Commissars," rejects Dario Amodei's proposal in its own summary: "an unrealistic proposal that seems mostly geared to political control of AI." The headline says how far he takes it — a call for embedded evaluators and lab-run standards bodies reads, to Thompson, as an industry reaching for political control dressed as safety. That is not a doomer warning Washington ignores or a researcher's resignation letter; it is a centrist, subscriber-priced analyst disputing the legitimacy of the mechanism itself. We flagged the same fracture yesterday — Nadella backs AI pacing, and picks a fight over who referees — and today it produced its first formal brief for the prosecution.
A Hacker News front-page repo is making frontier architectures readable rather than runnable. OpenArch implements modern open-source LLMs — Llama, Qwen, DeepSeek, Gemma, GPT-OSS, Kimi among them — in plain PyTorch written from scratch, with a target list mirroring the 72 architectures in Sebastian Raschka's LLM Architecture Gallery. Some implementations are marked usable for forward passes, others under construction, and the README is candid that the point is learning, not performance. The value is dull and real: reading a working mixture-of-experts or multi-token-prediction variant side by side, in code you can step through, closes the gap between calling an API and knowing what is inside it — the literacy a hiring wave keeps pricing.
What to watch: whether Firmus prices before or after the hyperscalers' autumn earnings — the roadshow's first reaction will be the market's verdict on the pacing debate, not the labs'.
Would you buy stock in an AI data-center builder the same week its three loudest customers said the pace is too fast? Tell us in the comments.
Sources: OpenArch (GitHub)
