Six customers underwrite Broadcom's $230bn AI chip target for 2028

The ramp Hock Tan laid out on the September 2 call carries an arithmetic requirement that this quarter's guide does not yet meet ^[16]. His $115 billion of AI semiconductor revenue in fiscal 2027 averages $28.75 billion a quarter, about 32% above the $21.7 billion guided for the period ending November 1 ^{[17][7][9][8]}, and the $230 billion he put on fiscal 2028 averages $57.5 billion a quarter, nearly twice the $29.6 billion the entire company booked in the three months to August 2 ^[16][3][9]. Getting from roughly $58 billion this fiscal year to $230 billion is a factor of about four, which is 99% compounded two years running ^[16][10]. (Cryptopolitan dates the year-ago comparison to 2025 while Crypto Briefing labels the same quarter fiscal Q3 2026, so read the fiscal labels loosely ^[1][16].)

Composition tells you more than the multiple does. Fourth-quarter guidance takes total revenue up about $5.2 billion sequentially and AI silicon up $5.0 billion of that, so 96 cents of every guided incremental dollar arrives on one line while everything else adds roughly $200 million on a $12.9 billion base ^{[8][3][7][1][3][4][5][6]}. Inside that line, custom accelerators were about 73% of the $16.7 billion, call it $12.2 billion, leaving some $4.5 billion of networking ^[19][1][7]. A year earlier the whole AI line was about $5.2 billion ^[1][2]. Infrastructure software grew a respectable 29% to $8.75 billion and is now under 30% of revenue, against AI's 56% ^{[5][13][1][14]}.

The contracts underneath are long, which is the part that reads like an annuity even though the accelerators do not. Google's custom chip supply agreement, signed in April, runs through 2031 on top of a TPU relationship of roughly ten years ^[10], and the Jalapeño inference chip OpenAI unveiled in June with Broadcom is aimed eventually at about 10 gigawatts ^[12]. The trade is the ASIC trade: less flexible than a general-purpose GPU, cheaper, tuned to one job, and finding its market in inference rather than training ^[15].

Or rather, the more interesting version of the bear case treats the demand as borrowed rather than imaginary. Every name on the customer list still buys from Nvidia; Sam Altman still says Nvidia makes the best AI chips in the world ^[13], and Anthropic runs Claude on AWS Trainium and Google TPUs as well as Nvidia GPUs ^[14] while committing to 3.5 gigawatts of Google-powered capacity from 2027 that could approach 5 ^[11]. The stock dipped on a quarter whose total revenue slightly missed ^[22], so the price already contains some version of the ramp.

This is probably wrong, but the figure I would underwrite here is the customer count rather than the growth rate. Six ^[20]. With three quarters of AI revenue coming from customer-specific parts ^[19], Broadcom's fiscal 2028 is a function of architecture decisions made inside four or five buildings it does not own. What would prove it wrong is a seventh and eighth name on that list plus a January AI print near the $28.75 billion fiscal 2027 needs ^[8], at which point the concentration argument becomes an unearned discount.