Buried on page four of Broadcom's Sept. 2 earnings call, past the record revenue and the record margin and the record free cash flow, Hock Tan dropped the number that actually explains the quarter. "Overall, our XPU shipments for the third quarter were up over 3.5x year-on-year and represented 73% of AI revenue during the quarter," he said. "Our AI networking revenue was up over 2.5x year-on-year."
Seventy-three percent. Of the $16.7 billion in AI semiconductor revenue Broadcom booked in its fiscal third quarter, roughly $12 billion came from accelerators that Nvidia does not design, does not sell, and does not price — chips co-developed with six customers who decided the cheapest way to buy compute was to build it.
The consolidated numbers, reported for the quarter ended Aug. 2, are large enough on their own. Revenue hit $29.59 billion, up 86% from $15.95 billion a year earlier. GAAP net income more than tripled to $13.09 billion from $4.14 billion, a 216% increase. Non-GAAP operating income rose 92% to $20.1 billion, putting operating margin at a record 67.9% of revenue. Free cash flow reached $13.67 billion, or 46% of revenue. Adjusted earnings came in at $3.32 a share against an LSEG consensus of $3.24, and revenue beat the $29.36 billion estimate.
The mix has flipped
Semiconductor Solutions revenue grew 127% to $20.84 billion and now accounts for 70% of the company. Infrastructure software — the VMware franchise Broadcom bought for the recurring cash — grew 29% to $8.75 billion with 15% ARR growth, respectable and suddenly beside the point. Non-AI semiconductors did $4.2 billion, up 5%.
AI semiconductor revenue of $16.7 billion grew 221% year over year and 54% sequentially, and CFO Amie Thuener noted it represented 56% of total company revenue, up from 49% in the prior quarter. Broadcom is now, by revenue mix, an AI company with a large software business attached.
Tan walked through the customer roster in unusual detail. Broadcom shipped Google's Ironwood TPU v7 in high volume to both Google and Anthropic, began production shipments of the next-generation TPU v8i for Google, and shipped Jalapeño, OpenAI's first custom accelerator. Meta's MTIA part enters production in Q4. The named pipeline runs to gigawatts: 5GW of TPU v8i for Anthropic in 2027 with line of sight to an incremental 10GW in 2028; 1.3GW of Jalapeño for OpenAI in 2027 and over 5GW in 2028; 3GW across three MTIA generations for Meta through 2028.
Guidance is where the story got loud. Broadcom forecast Q4 revenue of about $34.8 billion, up 93%, with AI semiconductor revenue of $21.7 billion — up 236%. Full-year fiscal 2026 AI revenue lands at $58 billion, up 186%. For fiscal 2027, Tan raised the AI target to roughly $115 billion, from a prior forecast of over $100 billion, and put fiscal 2028 at $230 billion. "In 2027, we have secured the supply to again double AI revenue to approximately $115 billion," he said. "Our demand actually exceeds this outlook."
Reuters reported that bookings for Broadcom's AI chips topped $30 billion in the quarter alone.
Custom silicon versus the merchant GPU
The strategic argument Tan made is the one Nvidia has spent three years insisting does not scale. "When you co-develop a chip that is optimized for your particular LLM workloads, you will outperform any GPU," he told analysts, adding that customers "can do all this at half the cost of a GPU." He claimed TPU v8i is comparable to or better than Vera Rubin on inference, and that Jalapeño outperforms Grace Blackwell on OpenAI's own workloads.
Discount the vendor framing and a real structural point survives. Frontier labs now have workloads stable enough and volumes large enough to amortize a custom tape-out, and Broadcom is selling them the SerDes, the chip-to-chip interconnect, the HBM and SRAM integration, and the advanced packaging required to do it. That is not a GPU substitute across the market. It is a substitute for exactly the handful of buyers who make up most of Nvidia's data center revenue.
The constraint is physical, not commercial. Tan repeatedly named substrates, HBM memory, leading-edge wafers and factory capacity as the binding variables, and said the fiscal 2028 outlook was built against a supply chain, not a demand forecast. Broadcom will begin using its Singapore fab for substrates in fiscal 2027 and is adding indium phosphide capacity; capex rises to $1.4 billion in Q4 from $532 million in Q3.
The margin math is the tell. Q3 gross margin was 75%, down 210 basis points sequentially. Broadcom guided Q4 gross margin to roughly 73% against 78% a year ago. Thuener was explicit about the cause: the shift "reflects the increasing mix of XPUs with their increasing memory content, which is diluting our consolidated gross margin." Custom accelerators carry more HBM per unit and pass more of the bill of materials through to the customer. Broadcom is trading gross margin for volume — and holding operating margin at 66% anyway on sheer operating leverage.
For Nvidia, that is the uncomfortable read. Broadcom is proving a merchant GPU competitor can be built at structurally lower gross margin because the customer absorbs design risk. Analyst Patrick Moorhead of Moor Insights & Strategy told Reuters of the committed gigawatt pipeline: "That is committed capacity, not aspiration, and it closes most of the gap to what the market wanted."
What to watch
The market shrugged. Shares fell over 1% after hours because $34.8 billion in Q4 guidance sits below the $35.03 billion consensus, and the stock is up only about 6% this year — badly trailing the semiconductor index amid Marvell's recent custom-chip win at Google.
Three things next quarter. Whether AI revenue lands at $21.7 billion, which requires the Meta MTIA ramp and Google TPU v8i volume to arrive on schedule. Whether gross margin holds near 73% or keeps sliding as HBM content rises. And whether the AI XPV financing vehicle Broadcom built with Apollo and Blackstone — first $35 billion tranche closed in June for Anthropic's 1GW deployment, targeting over 20GW by end-2028 — stays off Broadcom's balance sheet. Tan says the residual value guarantees are low risk. That claim is untested.
“That is committed capacity, not aspiration, and it closes most of the gap to what the market wanted.”— Patrick Moorhead, CEO, Moor Insights & Strategy