Broadcom CEO: AI Revenue Target Unchanged, Anthropic Set to Become Largest Customer
nashnova research
Broadcom CEO Hock Tan said Anthropic's call to slow frontier AI development will not change the company's AI revenue forecast — $115 billion for FY2027 and $230 billion for FY2028 — and revealed Anthropic will overtake Google as Broadcom's largest custom chip client by 2027.
Anthropic says "slow down" — why does Broadcom say "no impact at all"?
Anthropic CEO Dario Amodei recently called for slowing frontier AI model development. Broadcom shares fell 4.8% on the day; the iShares Semiconductor ETF dropped 5.6%.
Tan's response was blunt: "Not at all." This means → Broadcom has decoupled its growth thesis from the question of whether bigger models keep getting trained.
In plain terms = even if frontier training truly slows, Broadcom believes its revenue won't shrink — because its bet is on AI's everyday-use layer, not the training layer.
How did Anthropic become Broadcom's biggest customer?
Tan disclosed that Anthropic is expected to become Broadcom's largest custom chip client in 2027 and hold that position through 2028.
That seat previously belonged to Google, which co-designs its TPUs — tensor processing units, Google's in-house AI chips — with Broadcom.
This reflects a reshuffling of the AI chip client landscape: Anthropic's compute procurement has grown large enough to surpass Google's custom chip orders.
Betting on inference, not training — does the logic hold?
Tan said he is especially bullish on inference demand — the compute consumed when a trained AI model serves real users day-to-day. He called it "very, very robust."
This means → the $115 B → $230 B revenue path rests not on "who is training a bigger model" but on "how many people use already-trained models every day."
In plain terms = training is like building cars; inference is like driving them. Broadcom is betting that more cars will be on the road — not that factories will keep building pricier ones.
$230 billion target — where are the verification points?
Broadcom's FY2028 AI revenue target of $230 billion beat market expectations and was a headline number from its last earnings call. It represents a straight doubling from FY2027's $115 billion.
Broadcom's AI revenue spans two business lines: custom AI accelerator chips (designed bespoke for major clients) and AI networking chips (connecting thousands of chips inside data centers).
Two verification points matter most: ① whether Anthropic's compute purchases land on schedule; ② whether inference demand can independently sustain the growth trajectory if training spend slows.
How does Tan view the AI safety debate?
On Amodei's safety arguments, Tan expressed "partial agreement," saying governance and safeguards on any tool are important.
But his concern level is visibly lower than Amodei's — he compared AI to the Industrial Revolution, called it "ultimately a tool," and stressed that AI "is not a living thing that will run off on its own."
This reflects a stark gap in how chip suppliers and AI model companies price risk: model companies worry about technology running out of control; chip companies worry about demand slowing down.
市场有风险，内容仅供研究参考，不构成投资建议。