Anthropic just asked the world to slow AI. Its biggest chip supplier says the order book hasn't moved.

Amodei's three-step plan calls for embedded evaluators, frontier-lab coordination, and international cooperation on AI risks — explicitly framed as a \"pacing\" not a \"stopping\" of progress. Sam Altman endorsed the spirit on X. But the financial translation of \"slowing the pace\" is what's roiling chip stocks: if the largest AI labs throttle capability gains, the marginal gigawatt of compute gets harder to justify. The market reacted in textbook fashion — CoreWeave dropped roughly 7%, Vertiv about 8%, and AI-exposed REITs Equinix and Digital Realty each gave back more than 3%.

Tan walked straight into that tape with a counterweight. On the same CNBC interview he confirmed Broadcom's existing AI semiconductor revenue trajectory — $58B for fiscal 2026, $115B for fiscal 2027, and a fresh disclosure of $230B for fiscal 2028. He also named the customers: Anthropic (largest in 2027 and 2028), followed by Alphabet and OpenAI (Jalapeno, second-largest by 2028).

Broadcom's $58B → $115B → $230B staircase is now the most concentrated AI revenue projection on a US tape. The FY2027 number alone equals roughly 38% of NVIDIA's trailing-twelve-month revenue of $302.9B (per NVIDIA's Q2 FY2027 10-Q, filed Aug 26, 2026) — and it depends almost entirely on two customers: Anthropic and OpenAI's Jalapeno. That concentration is the bear case for the trade, but it's also exactly the kind of customer concentration that explains why Tan went on CNBC the day after Amodei's essay.

Read the table as a single statement: the cohort most exposed to a compute slowdown got hit hardest, and the cohort most exposed to AI-misuse risk rallied hardest. Marvell's 4% drop is the cleanest tell — it shares Broadcom's customer mix and Q1 FY2027 revenue grew to $2.418B (per Marvell's Q1 FY2027 10-Q, filed May 28, 2026) on custom-AI ASIC traction. The trade that's been working — long the AI infrastructure cohort — got a haircut, and the cross-trade (cybersecurity as AI-misuse hedge) got paid.

The interesting economic fact is not the essay or the interview — it's the gap. Anthropic's public posture (slow down) and its private procurement posture (5+ GW of TPU capacity in 2027, with options to 10 GW) are simultaneously true. That gap is the trade. If the public posture wins — Anthropic's compute orders slow — Broadcom's $115B 2027 guide is impossible. If the procurement posture wins, the $115B number is conservative and the cybersecurity trade unwinds. Neither side has to be lying; the essay is about model-training pace, the deal is about inference infrastructure (the part Tan explicitly highlighted as \"very, very strong\").

What Tan's intervention does is move the question from \"will AI labs slow?\" to \"can they?\" — and the procurement contracts say no, not in 2027. That's the structural driver behind the $230B FY2028 number: the deal pipeline is sized in gigawatts, and the foundry capacity to fulfill it is the bottleneck, not the demand.

Broadcom's chips are made by TSMC and packaged with HBM from SK Hynix and Micron. Broadcom told investors in March 2026 that it had locked in HBM and TSMC advanced-node capacity through 2028 (per Chosun Biz, March 6, 2026). TSMC reported Q2 2026 revenue of TWD 1.27 trillion (~$39.7B at current FX, per TSMC's Q2 2026 financial statements filed Aug 14, 2026) and called 2026 capacity \"choked\" on its March 24 earnings call. The supply chain is sold out, not slowing.

TSMC's Q2 2026 revenue implies a year-on-year step-up of roughly 30% in TWD terms — the largest in three years, driven by AI accelerator demand from both NVIDIA and Broadcom's hyperscaler-custom programs. The bottleneck is the trade: whoever controls advanced packaging (TSMC's CoWoS-S / CoWoS-L) and HBM allocation (SK Hynix lead, Micron catching) gets the margin, regardless of whether the AI labs agree with each other on safety.

Palo Alto Networks and CrowdStrike each jumped more than 13% on Sept 14, with Zscaler, SentinelOne, and others up double digits. The move is the cleanest expression of the PR risk: if Amodei's essay is read as \"AI misuse is a real, headline-grade risk,\" then AI-era security spend has a fresh tailwind. CrowdStrike's TTM revenue of $4.76B (per its most recent 10-Q) and Palo Alto Networks's larger installed base both look like a way to play the same narrative from the defensive side — without taking silicon-cycle risk.

But the cybersecurity trade is the mirror image of the chip trade, and the two can't both be right. If Tan is correct that AI revenue targets are intact, the cybersecurity trade fades as fast as it ran. If Amodei is correct that frontier-capability progress will be paced, the chip trade bleeds further and cybersecurity extends. The investor question is which side of the spread to fade.

Short-term (days–quarters): Tuesday's session is the live test. If Broadcom reclaims Monday's drawdown, the cross-check held and the AI infrastructure cohort gets a relief bid — Marvell, NVIDIA, TSMC, and Micron likely follow. If it doesn't, the cybersecurity trade extends and the chip trade bleeds further. The Q3 FY2026 earnings call already locked in $21.7B of Q4 AI revenue, so any gap-down Tuesday is a tape story, not a fundamentals story.

Long-term (1–3 years): the $115B FY2027 / $230B FY2028 Broadcom number is the load-bearing assumption across the whole AI infrastructure cohort. If even half of it lands, TSMC advanced-node revenue, SK Hynix HBM volumes, and Applied Materials packaging-tools demand all stay on the upward slope. The two real risks to the thesis: (1) Anthropic/OpenAI/Goggle/Tencent/Meta actually slow orders — not essays, orders — which would be visible first in TSMC monthly revenue prints and Broadcom's Q4 FY2026 call in early December 2026; (2) a shift from training to inference changes the silicon mix in ways that favor NVIDIA inference GPUs over custom ASIC, which would compress Broadcom's customer-concentration premium.

Bottom line: Broadcom's CEO didn't argue with Amodei's safety premise — he argued with the market's conclusion that the essay changes demand. The order book says it doesn't, at least through 2027. That makes Tan's CNBC appearance the cleanest data point the AI infrastructure trade has had in weeks.