Broadcom reports Wednesday after the close against roughly $29.25 billion of revenue and $3.21 of adjusted earnings. The test is not whether AI demand is strong. It is whether hyperscalers are building their own machines, and whether a $100 billion-plus fiscal 2027 AI book is conversion or already priced.
Broadcom reports fiscal third-quarter results Wednesday after the close. Street numbers sit near $29.25 billion of revenue and $3.21 of adjusted earnings — more than 80 percent growth if the print lands there. Management has already pointed to about $16 billion of AI semiconductor sales in the quarter, on a path toward roughly $56 billion for fiscal 2026. That is not the information the market is missing.
The missing information is architectural. Is the AI buildout still a Nvidia cluster with accessories, or a hyperscaler race to design the machine? Broadcom sits in the second chain: custom architecture, then Broadcom or Marvell, then foundry, then packaging, memory, and Ethernet. The custom-silicon book we mapped in March made the company look like the forge. Wednesday tests whether the forge’s fiscal 2027 line — more than $100 billion of AI revenue — is converting into shipments or remaining a slide.
June already sketched the trap. Broadcom grew 48 percent, posted $10.8 billion of AI semiconductor revenue, beat on adjusted earnings, lifted the fiscal 2026 AI outlook to $56 billion, and repeated the $100 billion-plus 2027 figure. Revenue came in a tick light versus consensus. Gross margin was guided toward 74 percent as lower-margin accelerators mix up. The stock still fell about 13 percent. Buyers wanted a higher 2027 number and cleaner conversion. They got confirmation. At this multiple, confirmation is a sale.
Nvidia can remain the dominant general-purpose accelerator while the stack around it gets larger, more specialized, and less synonymous with one logo.
Hyperscalers are commissioning dies, not only buying SKUs
Google’s TPUs are the oldest public case. Amazon, Meta, Anthropic, and OpenAI are now in the same business of specifying silicon rather than waiting in a GPU line. OpenAI’s Jalapeño tape-out with Broadcom is one named instance of that impulse. A narrower Nvidia franchise does not require Nvidia to lose. It requires the rack around Nvidia — custom accelerators, Ethernet fabrics, HBM, CoWoS — to become a market of its own. CUDA still binds a generation of software. Custom XPUs bind a generation of workload.
That is Broadcom’s collection point. ASICs are only half the AI line. Networking — switches, SerDes, Ethernet displacing InfiniBand lock-in — is the other. If third-quarter AI revenue clears $16 billion with fabric still a material share, the second layer is no longer a Google anecdote. If the mix stays concentrated in one customer’s architecture, the 2027 figure is still a single design wearing a sector label.
Marvell already paid the timing tax
Last week Marvell beat, raised longer-term forecasts, and was sold anyway. An expanded Google custom-silicon agreement can be described as on the order of $120 billion of cumulative revenue into the early 2030s. Chief executive Matt Murphy said the meaningful contribution arrives in fiscal 2029. Near-term growth, the company said, still lives in connectivity and optics. Custom silicon is the franchise. It is not yet the quarter.
That is the comparison Broadcom cannot dodge. A $100 billion fiscal 2027 AI target is only as good as the conversion calendar. Three figures do the work after the close: AI semiconductor revenue, the fourth-quarter guide, and any change to 2027. An in-line print with a restated slogan is Marvell with a larger market cap.
The firms writing the checks may not keep the returns
Amazon, Alphabet, and Microsoft are already spending like utilities — capex that can swallow a year of cloud revenue before the racks are fully sold. Broadcom sells pieces of that plant. It does not have to win the model war. It has to ship the die and the switch that make someone else’s cluster run. Nvidia’s own record quarters have already taught the same lesson: spectacular growth can be priced as a question about payback rather than as a victory lap.
The inversion is the economics of the boom, not a sector color story. The companies spending the most on AI may post the weakest returns on it. The companies selling the scarce intermediate goods — custom silicon, memory, packaging, power — collect first. Broadcom is the cleanest public test of that inversion because its AI growth is the custom layer made visible.
Wednesday will not settle whether Nvidia still “wins AI.” It will show whether winning now means a larger, more fragmented infrastructure market: GPUs plus XPUs plus fabric, with profit accruing to whoever sits between the hyperscaler’s architecture and the foundry’s tape-out. If those three numbers widen the gap between booked ambition and shipped silicon, the boom is still Nvidia’s. If they close it, the boom has already moved.
Continue reading
Sources
Broadcom Q2 FY2026 results (revenue $22.19B, AI semiconductor $10.8B, non-GAAP EPS $2.44) and Q3 guide (~$29.4B total, ~$16B AI semiconductors, ~74% gross margin); management FY2026 AI semiconductor outlook of ~$56B and FY2027 AI revenue of more than $100B; Street Q3 consensus near $29.25B and ~$3.21 adjusted EPS; post-Q2 share-price coverage of the ~13% decline; Marvell Q2 FY2027 results and Reuters/CNBC coverage of Google custom-silicon timing (meaningful contribution in FY2029); prior Culled analysis of custom silicon, hyperscaler capex, and Nvidia's prove-it prints.
