Anthropic, Broadcom, and the Quarter That Priced the Wrong Year

On the September 2 earnings call — not in the press release — Hock Tan named Anthropic as Broadcom’s coming largest XPU customer and raised the multi-year AI revenue target to a line of sight near $230 billion. The stock fell anyway.

What Happened

This piece is the call follow-up to the already-covered press-release analysis. The numbers in that release told one story. According to the September 2 earnings call transcript, corroborated by Bloomberg’s same-day report, CEO Hock Tan told a bigger and structurally different one. He said two things that did not appear in the press release: first, that Anthropic is on track to become Broadcom’s largest XPU customer in 2027 and to sustain that position in 2028 — a frontier AI lab set to overtake Google as the single biggest buyer of Broadcom’s custom-silicon business. Second, that Broadcom now sees fiscal 2027 AI-semiconductor revenue of approximately $115 billion, a genuine raise from the “more than $100 billion” figure that had been merely reiterated in June, and — for the first time on any call — a fiscal 2028 line of sight near $230 billion, framed as “double again.” Both figures are call statements attributed here to the September 2 transcript and Bloomberg’s same-day reporting; they are management targets underwritten by orders, not delivered revenue, and should be read accordingly.

What is deliberately excluded from this analysis: the gigawatt capacity figures circulating from the transcript. The 2027 numbers match June’s call language, other reporting conflicts on them, and they cannot be cleanly sourced to tonight’s call alone. Also excluded: the roster of six custom XPU customers, which is a prior-quarter fact and not a new disclosure. What is new, and what matters, is the raise and the name.

The market’s response, per Bloomberg — whose headline read “Forecast Misses Estimates, Shares Drop on Slower AI Gains” — was to sell the stock after hours. Bloomberg attributes the decline to the near-term Q4 guide of approximately $34.8 billion landing short of roughly $35.1 billion in consensus expectations. The after-hours drop is market data, still moving at the time of writing, and should not be read as the market rejecting the multi-year AI vision. It is repricing a quarter that came in a few hundred million light of the bar.

The Structural Read

The lab overtakes the hyperscaler. Broadcom’s custom-silicon business was built on hyperscalers — Google’s TPU, first and largest among them. That is the baseline the market has priced for years. When Hock Tan says a frontier lab, not a cloud, is on track to be his biggest XPU customer, the statement reorders the competitive map of AI infrastructure at the hardware layer. Anthropic’s reported multi-tens-of-billions in compute commitments are now large enough that its custom-accelerator demand — for training, for inference, for the capacity it needs to run at the frontier — tops Google’s share of Broadcom’s business. That is the compute-commitment thesis made concrete in silicon: the frontier labs are no longer renting the frontier from clouds; they are building it into bespoke hardware at hyperscaler scale.

This maps directly to the vertical-integration layer of the Map of AI Redrawn: when a model company starts buying custom silicon at a scale that beats the hyperscalers who supply it, the boundaries between the model layer and the infrastructure layer are no longer boundaries — they are overlapping claims. Anthropic is not just an AI lab anymore; it is, by Broadcom’s measure, a compute operator of hyperscaler magnitude.

Stretch-to-floor guidance confirmed. The raise from “more than $100 billion” — a floor dressed as a ceiling in June — to a specific $115 billion for FY2027, and then a first public line of sight near $230 billion for FY2028, follows the pattern this column has tracked: Broadcom’s management sets what looks like a stretch target, watches it become anchored supply reality, then raises again. The $230 billion is a target and stated line of sight, not a fact. “Secured supply” is Broadcom’s characterization — guidance underwritten by orders, not delivered revenue. But the directional logic is now two raises deep: the number keeps moving up because the orders keep moving up, and management is now willing to give a two-year forward number for the first time. That is a confidence signal, even if the number itself carries the usual guidance caveats.

Narrative versus tape. The most clarifying data point of the evening is not any of the large numbers — it is the gap between the out-year guidance and the stock reaction. Broadcom raised its multi-year AI revenue target, named a frontier lab as its coming anchor customer, and outlined a doubling trajectory through 2028. The stock fell because the Q4 guide missed by roughly $300 million. This is the same proof-over-promise dynamic that sold MongoDB’s beat and bought Snowflake’s in the days prior: expectations have moved so far ahead of even excellent results that the marginal quarter, not the multi-year vision, sets the price. The numbers are now so large — $115 billion, $230 billion — that being told you will double to a quarter-trillion of AI silicon in two years is not sufficient to move a stock if this quarter’s forward guide is a sliver light. The market is not pricing the 2028 story. It is pricing the next four months.

Three Implications