Broadcom's $230bn AI target needs its quarterly chip sales to more than triple

Start from the ramp the two targets require. Broadcom's guide for the current fiscal year is $58 billion of AI semiconductor revenue ^[6]. Getting to $115 billion the following year is 98% growth on that ^[1], and $230 billion after it is exactly double again ^[2], so the two-year path is a little under four times the current year, roughly 99% compounded annually ^[3]. Per quarter the demand is plainer: $115 billion needs an average of $28.75 billion, 1.72 times the $16.7 billion Broadcom just reported ^[4], and $230 billion needs $57.5 billion, or 3.44 times it ^[5].

The figure Broadcom actually committed to moved by $2 billion, a 3.6% raise on the prior $56 billion guide ^[6]. The third quarter's $16.7 billion annualizes to $66.8 billion, which is $8.8 billion above the full-year guide ^[7]. The early quarters of that year were much smaller; the 221% growth rate implies about $5.2 billion in the year-ago quarter ^[8].

Asked whether calls from AI leaders to slow development had changed his outlook, Tan said: "No, not in the least." ^[2] He pointed to inference, the compute consumed after a model is trained each time it answers a query, as the driver ^[7].

Now the customer. Tan projects that Anthropic will be Broadcom's largest custom-chip customer by fiscal 2027 ^[8]. The letter that took 4.8% off the stock and the account that anchors the fiscal 2027 target come from the same company ^[3][4][8]. Crypto Briefing put the counterargument as concentration risk, with outsized exposure to any stumble at Anthropic, and said the $230 billion assumes multiple hyperscaler relationships keep expanding at once ^[10][11]. Crypto Briefing did not report how much of the $115 billion is already under contract.

If inference demand holds near current rates, the ramp is a manufacturing and scheduling question and the targets are reachable ^[7]. If Anthropic's spending slips, fiscal 2027 loses its largest custom-chip line, and the shortfall has to come from other hyperscalers buying custom accelerators and high-speed networking, the part of the market Broadcom serves with application-specific silicon rather than general-purpose GPUs ^[8][9]. The third case is the dull one: the curve holds but arrives a year late, and on a doubling schedule a year of slippage is the difference between $115 billion and $230 billion ^[2].

In my view the $58 billion is the only one of those three figures carrying a current-year commitment, and the other two get tested by order flow. The outside check is quarterly capital expenditure disclosure from the major cloud providers, which is where Crypto Briefing also pointed readers ^[12].