Hock Tan defends Broadcom's $230 billion fiscal 2028 AI forecast on inference demand

Tan gave Jim Cramer two answers about the same forecast. "No, not in the least," he said when asked whether anything in the AI slowdown debate had caused him to reconsider Broadcom's fiscal 2027 and 2028 AI semiconductor forecasts ^[2]. On what underwrites those numbers he was narrower: "I don't know about training, but when you want to productize inference, I see it continuing to be very, very strong," Tan said ^[4]. Amodei's essay argues for moderating the pace of model development ^[5], and CNBC describes inference as the day-to-day usage of models after they have been trained ^[12].

Tan vouched for demand that shows up after training is finished.

The targets are $115 billion of AI semiconductor revenue in fiscal 2027 and $230 billion in fiscal 2028, both set out on the September 2 earnings call ^[7]. The second figure is twice the first, so it requires another $115 billion of AI revenue inside a single fiscal year ^[1]. Broadcom counts custom accelerators and networking chips in that line ^[8]. Tan said Anthropic is on track to be the largest custom chip customer in 2027 and to hold that position in 2028 ^[9], ahead of Google, whose tensor processing units Broadcom co-designs ^[10]. CNBC did not report Anthropic's share of the $115 billion.

Monday's selling was less discriminating than the story suggests. Broadcom fell 4.8% while the iShares Semiconductor ETF fell 5.6% ^[11], which means the company whose incoming largest custom customer wrote the essay dropped 0.8 percentage points less than the basket ^[2].

Tan said he agrees with Amodei about the need for some restrictions on AI ^[13]. "Like any tool, it's important to put governances, safeguards on how we use the tool," he told Cramer ^[14]. He also said, "It's not a live animal that will run wild by itself" ^[15], and likened generative AI to the Industrial Revolution that began in England in the 18th century ^[16].

If inference volume grows the way Tan describes, fiscal 2027 survives a slower frontier training cycle and the doubling into 2028 turns into a supply question. Should Anthropic act on its own three-step proposal and trim accelerator orders ^[6], the customer concentration Tan himself described is how that reaches the forecast. The third case is that the labs change nothing and Monday's 4.8% was noise ^[11]. I would weight the first, partly because Tan restated the target twelve days after issuing it ^[3]. A narrowed or withdrawn $115 billion at the next earnings call settles it the other way.