Broadcom CEO addresses Anthropic's slowdown push, says AI revenue targets haven't changed

Broadcom CEO Hock Tan reaffirmed the company's AI semiconductor revenue forecasts of $115 billion for fiscal 2027 and $230 billion for fiscal 2028, arguing that AI compute and especially inference demand will remain durable. The reassurance followed investor concerns that Anthropic CEO Dario Amodei's call to moderate frontier-model development could reduce AI infrastructure demand; Broadcom shares fell 4.8% and the iShares Semiconductor ETF declined 5.6%. Tan said Anthropic is on track to become Broadcom's largest custom-chip customer in 2027 and 2028, highlighting both the opportunity and customer-concentration sensitivity.

Analysis

The market is conflating a possible reduction in frontier-training cadence with a reduction in total AI silicon demand. That distinction favors AVGO relative to GPU-heavy training exposure: inference deployment shifts system bottlenecks toward scale-out networking, memory bandwidth and lower-cost custom accelerators, where AVGO can capture content even if customers lengthen the interval between major model-training runs. The near-term issue is not demand direction but concentration: a customer representing an increasingly material share of custom-silicon revenue raises the probability that one customer’s capex timing can create quarterly volatility and a valuation derating.

Management’s long-dated revenue framework should be treated as an aspiration rather than an independently validated backlog measure. For the next 1-3 months, confirmation requires hyperscaler capex guidance, disclosed accelerator deployment volumes, and evidence that custom-chip programs are moving from design wins to volume ramps; absent those data, the sector-wide selloff can persist as investors haircut terminal AI capex assumptions. GOOG is a second-order beneficiary if its TPU program delivers lower inference cost per token, but it also has strategic incentive to retain that advantage internally rather than expand external silicon supply.

The contrarian read is that tighter frontier-model governance can increase inference demand rather than destroy it: fewer frontier training cycles may redirect budgets to reliable enterprise deployment, security, and model-serving infrastructure. That supports AVGO’s networking mix and potentially shifts relative demand away from pure training hardware, but only over 6-18 months; a broad capex freeze or delayed customer tape-outs would overwhelm this distinction. Falsify the AVGO thesis on a material cut to FY27 AI outlook, a disclosed delay in major custom-accelerator ramps, or hyperscaler capex guidance rolling over sequentially.

Key Decisions for Investors

• Accumulate AVGO in tranches over the next 2-6 weeks rather than chase a one-day sentiment reversal; target a 6-12 month horizon. Structure sizing around customer-concentration risk, with a thesis stop if management reduces its FY27 AI revenue framework or signals a program delay.
• Use a relative-value expression: long AVGO / short SMH for 1-3 months, only after confirming AVGO’s relative weakness has stabilized. The thesis is that custom inference and networking content should be less exposed to a moderation in frontier-training intensity than the broad semiconductor basket; exit if AVGO underperforms SMH following the next hyperscaler capex updates.
• Maintain GOOG as an indirect inference-cost beneficiary, but do not underwrite material external TPU monetization without disclosure. Monitor Google Cloud growth, capex, and commentary on TPU availability as the catalysts that could convert internal cost advantage into earnings leverage over 6-18 months.
• Watch for the next capex guidance from major AI buyers and any evidence of revised accelerator orders. A synchronized reduction in 2027 capex plans would invalidate the inference-offset narrative and argues for reducing AI-infrastructure beta rather than adding exposure.