Broadcom CEO Holds $230B AI Target Despite Slowdown Call

Broadcom CEO Hock Tan told CNBC that $115B FY2027 and $230B FY2028 AI chip targets are unchanged, citing durable inference demand as insulation from any training-pace reduction.

Lead

What Did Hock Tan Actually Say?

Tan drew the debate's sharpest distinction between training compute -- the power-intensive process of building new frontier models -- and inference compute, which runs those models at consumer and enterprise scale once deployed. "I don't know about training, but when you want to productize inference, I see it continuing to be very, very strong," he told Cramer. The framing is strategically central to Broadcom's investment case: its custom XPU chips are purpose-built for inference workloads at hyperscaler scale, not for the comparatively narrow set of frontier training runs that Amodei's essay targeted. Tan also disclosed a cumulative target, telling investors Broadcom has a "high degree of confidence" to ship $350 billion worth of AI semiconductors to its six frontier customers over the next two fiscal years.

Why Did AI Stocks React So Sharply?

The Amodei essay landed on the weekend of September 13-14, 2026, and the fear it seeded was direct: if the companies spending most aggressively on AI training infrastructure chose to slow that pace, near-term demand for the most powerful accelerators could soften faster than supply chains can adjust. AI stocks across the semiconductor supply chain sold off hard on Monday. The iShares Semiconductor ETF (SOXX) dropped 5.6% on the session, Nvidia (NVDA) shed roughly 3%, Marvell Technology (MRVL) fell more than 4%, and AVGO itself declined 4.8%, extending a slide that has taken the stock to $339.51 -- well off its 52-week high of $494.18 and roughly flat year-to-date against an S&P 500 that is up more than 10%.

The Inference Advantage

Broadcom's entire custom silicon strategy is structured around inference scale rather than training, a position that becomes a competitive differentiator precisely when training-pace concerns surface. The company's XPU products for six hyperscaler customers -- a lineup that includes Google's Ironwood TPU v7 and v8i generations, OpenAI's Jalapeno chip, and Meta's MTIA architecture -- are tuned for the high-throughput, cost-sensitive task of serving billions of daily queries from deployed models. Tan noted in the September 2 earnings call that Broadcom's chips "match GPUs while cutting costs by half" on inference workloads. Independent projections support the structural trend: Deloitte's 2026 AI infrastructure outlook estimated that running deployed models would account for roughly two-thirds of total AI compute by year-end, up from approximately half in 2025.

Broadcom's Customer Roadmap

Tan has built Broadcom's guidance on a named-customer deployment roadmap across the six XPU accounts, an unusually specific level of disclosure that gives the $115 billion and $230 billion targets a more traceable foundation than broad market-size estimates. Anthropic alone is projected to deploy 5 gigawatts of compute through Broadcom silicon in fiscal 2027, positioning it to overtake every other customer by chip revenue that year and retain the top slot through 2028 -- an irony not lost on investors, given that it is Anthropic's own CEO who sparked the market-wide concern Tan spent Monday answering. Broadcom raised its full-year FY2026 AI revenue forecast to $58 billion following Q3 results, with Q4 guided to $21.7 billion, a 236% year-over-year increase.

Outlook

Broadcom's fundamental trajectory -- sequential AI revenue acceleration, a secured customer roster, and inference-driven demand that is structurally decoupled from training decisions -- leaves Tan's argument well-supported by the data. The central risk is not that the Amodei essay is wrong on safety grounds, but that any meaningful moderation in hyperscaler capital expenditure plans, even one driven by inference rather than training, would take time to show up in guidance. The next hard checkpoint is Broadcom's Q4 fiscal 2026 earnings report on December 9, 2026, when Tan will be expected to convert the FY2027 target from ambition to confirmed contract backlog. Analysts remain broadly constructive: 45 of 49 covering the stock rate it Buy or Strong Buy, with a consensus price target of $531.85 -- implying 57% upside from current levels. Jensen Huang of Nvidia (NVDA) independently echoed Tan's inference-demand thesis in separate remarks Monday, signaling that the two dominant custom AI chip suppliers are aligned on the core rebuttal to the slowdown narrative.