AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Anthropic is asking Wall Street to believe in AGI

Collected Oct 29, 2026

Anthropic has filed its S-1 registration statement to go public on the Nasdaq, and shares could begin trading as soon as next month at a valuation reported to fall between $1.5 trillion and $2 trillion. That range would place the Claude maker among the most valuable companies in the world on its first day of trading. The filing runs to hundreds of pages and offers the first detailed, audited look at the economics of a frontier AI lab. Joey Brookhart, an analyst at the chip and AI research firm SemiAnalysis and co-lead of its Tokenomics model, joined The Verge's Decoder podcast to walk through the numbers.

Brookhart's overall read was strongly positive. He described a company that has scaled efficiently, citing a valuation above $1 trillion and more than $70 billion in annualized recurring revenue (ARR) reached in a short period. He said gross margins, compute margins and free cash flow all looked like what an investor would want to see, and linked the filing to a rally across semiconductor, neocloud and compute stocks in the two days after it landed. He also noted a rough few weeks before the filing, after some investors were disappointed by September ARR figures discussed at an investor day.

One headline figure is the swing in inference economics. Anthropic had negative 94 percent gross margins on inference two years ago, according to Brookhart, and now runs a positive 85 percent margin. A big part of that comes from revenue per megawatt of inference capacity. He put Anthropic at roughly $65 million of revenue per inference megawatt today, up from the mid-teens a year ago, and said SemiAnalysis research suggests a future model such as Opus 5.5 running on Nvidia Blackwell hardware could reach $100 million to $120 million per megawatt. Against that, Anthropic pays a little over $10 million per megawatt in compute costs.

The efficiency gains come from a deliberately mixed chip fleet spanning AWS Trainium, Google TPUs and Nvidia GPUs, plus software that wrings more tokens per second out of the same silicon. Brookhart said token throughput on Trainium chips is up 5x to 10x year over year. Because Anthropic rents capacity from SpaceXAI, Google and Amazon rather than making its own chips, the scale of those cloud commitments is a central part of the S-1. This is a different position from Google, which designs its own accelerators; SemiAnalysis's view is that Anthropic has compensated through throughput and model-hardware co-design rather than owning the stack.

Perhaps the most important detail for anyone modeling the business: only about 30 percent of Anthropic's compute goes to inference that generates revenue. The rest funds experimentation, research, pretraining and post-training of new models. That is where Brookhart sees the returns and defensibility accumulating. He said the company would actually like to raise the training share, and would accept non-GAAP EBITA margins closer to zero, or even below it, if it could put the freed compute into training. Anthropic has reportedly made a similar point to investors on its roadshow. The company currently sits at about 11 percent non-GAAP EBITA margin, higher than it expected, and its message is effectively that the recent margin leverage should not be extrapolated even if long-term margins end up considerably higher.

That framing has a familiar ancestor in Amazon, whose founder told investors profits would come later and large. A public Anthropic will face pressure to harvest its best current model across more industries rather than keep spending on the next frontier system. Brookhart argued the company is structured to resist that, pointing to voting stock arrangements, and drew a parallel to Mark Zuckerberg's willingness to fund expensive long-horizon bets at Meta.

The revenue base is narrow. Brookhart estimated the two leading labs together at $120 billion of ARR, but said the top 1 percent of customers, spending $50,000 to $100,000 per employee, account for most of it, while the broader Fortune 500 and Global 2000 spend comparatively little. Anthropic's demand is heavily coding-specific, making it close to a single-vertical company. Getting to the $1 trillion revenue figure Dario Amodei has described would require moving into white-collar work, including cybersecurity, where Brookhart sees eight-figure enterprise contracts scaling quickly amid rising incident volume, and into areas such as biology. It would also require shifting from augmenting workers to saving or replacing labor in some roles.

Competition is intensifying. Brookhart said Anthropic captured about 90 percent of the industry's net new ARR in the first half of the year, while in the third quarter OpenAI was probably the majority. He attributed part of OpenAI's enterprise slowdown to its Fable release and the absence of zero data retention, a feature enterprises care about and one he expects could be addressed with Fable 5.5. Open-weight models are also growing fast: open-source inference revenue went from roughly $1 billion in the first quarter to more than $10 billion by the third, by his estimate, with enterprises and AI-native firms doing custom training and post-training on open models and running inference on them.

Coding alone will not triple the market, in his view. Switching costs in parts of coding are low, so customers move between providers readily. New verticalized products are where growth has to come from. Anthropic is already building them, with beta programs such as Claude for financial services and Claude for legal, and a cyber program that gives participants 25,000 credits per month. Brookhart expects that cyber offering to be a significant product line by 2027. He also argued some sectors will always demand frontier capability, cybersecurity being the clearest case, since defenders need to match adversaries who use frontier models. Routine tasks such as email summarization need not run at the frontier.

A striking data point in the filing: 80 percent of revenue in the last three months came from models released in the last six months. That is a reminder of how short the commercial life of a frontier model can be, and how relentless the release cadence has to be. It contrasts with hardware cycles like the iPhone, where a platform is reused across several generations to extract more value before the next leap.

Why it matters: The filing gives developers and product teams audited unit economics for a frontier lab for the first time, showing that inference is now highly profitable but that most compute still goes into building the next model. That suggests (my inference) buyers should expect rapid model turnover and vertical-specific offerings rather than stable long-lived platforms, and that pricing and capability will keep shifting as open-weight alternatives and rivals compete for the same enterprise budgets. For teams planning multi-year AI roadmaps, the filing is also a warning that the current generation of models may account for most revenue only briefly.

Read at The Verge · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Today on Decoder, I’m joined by Joey Brookhart, an analyst at the chip and AI research firm SemiAnalysis, to help me dig into the story inside the numbers of the Anthropic IPO filing — a document known as an S-1. Anthropic just filed its S-1 to go public on the Nasdaq, and it could start […]