The quarter AI became the business

Consolidated

• Revenue $29.6B, up 86% year over year — a company record.
• Gross margin 75% of revenue, down 210 bps sequentially on richer AI semiconductor mix, but ahead of the 74% guide.
• Operating income $20.1B, up 92% Y/Y. Operating margin 67.9% — a record and up 240 bps Y/Y despite the lower gross margin, purely on operating leverage.
• Non-GAAP EPS $3.32, up 96% Y/Y.

Semiconductor Solutions segment

• Revenue $20.8B (record), up 127% Y/Y, 70% of total company revenue.
• AI semiconductor revenue $16.7B = 56% of total revenue, up from 49% in Q2.
• Segment gross margin approximately 67% — Amy Tiner corrected this on the call after initially citing 76%.
• Operating expenses $1.2B, just 6% of segment revenue; opex grew 22% Y/Y against 127% revenue growth.
• Segment operating margin 61%, up 440 bps Y/Y.

Infrastructure Software segment

• Revenue $8.8B, up 29% Y/Y, 30% of total revenue.
• Gross margin 94%; opex just over $900M; operating margin ~84%, up 650 bps Y/Y.
• Annual recurring revenue sustained at +15% Y/Y.

Non-AI semiconductors

• Revenue $4.2B, up 5% Y/Y and flat sequentially.
• Broadband and server storage were both up; wireless declined, partially offsetting.

• AI semiconductor revenue more than tripled Y/Y to $16.7B; up 54% sequentially. Hock Tan: "Q3 demand was simply hot and we're just getting started."
• XPU shipments up over 3.5x Y/Y, representing 73% of AI revenue in the quarter.
• AI networking revenue up over 2.5x Y/Y, with momentum continuing into Q4.

Silicon shipped in Q3

• Ironwood (TPU v7) delivered in high volume to both Anthropic and Google.
• TPU v8i entered production shipment for Google — more memory and bandwidth than Ironwood, inference-optimized, and per management comparable to or surpassing the Vera Rubin GPU in performance.
• Competitive marker: Broadcom is shipping TPU v8i ahead of the MediaTek version, which was initiated earlier.
• Jalapeño — OpenAI's first-generation custom accelerator — shipped, outperforming Grace Blackwell on inference workloads.

Q4 ramp already lined up

• Accelerating Ironwood shipments to Anthropic.
• Ramping high-volume TPU v8i to Google.
• Continued Jalapeño shipments to OpenAI.
• Production shipments of Meta's MTIA accelerator, optimized for inference and recommendation at scale.
• Both XPUs and AI networking expected to triple Y/Y in Q4.

Why the labs are choosing custom silicon

• OpenAI disclosed that Jalapeño outperforms Grace Blackwell Ultra on performance-per-watt, latency, throughput and power — and is comparable to Vera Rubin GPUs on OpenAI's own workloads.
• Jalapeño also demonstrated it can run other frontier models, not just OpenAI's — at roughly half the cost of a GPU.
• Tan's moat argument for Google: strongest semiconductor design IP portfolio — SerDes, chip-to-chip interconnect, leading-edge HBM and SRAM integration, differentiated advanced packaging — plus consistently fastest time to market with no respins.
• On why labs won't defer to their cloud/funding partners: these customers want first-party compute and are becoming hyperscalers in their own right. Near term they rent third-party capacity; long term Tan expects them to run their own data centers and serve generative AI APIs directly. "It's not speculation, it is actually happening."
• Only two of the six XPU customers need financing help — the other four are financially able to fund deployments themselves.

• Tomahawk 6 — first to market at 100 terabits. Ships in both 100G and 200G SerDes versions; deployed at essentially all AI hyperscalers building XPUs with Broadcom and at those that aren't. Described as the fastest-ramping switch family in Broadcom's history.
• Tomahawk 7 taped out — the industry's first 200 Tb/s Ethernet switch.
• Tomahawk Ultra — scale-up networking over low-latency Ethernet, enabling rack-level GPU/XPU clustering on open Ethernet for the first time, matching prior proprietary options on loss and latency. Adoption "surprised us"; first deployments start this quarter and expand in FY27.
• Attach rate: XPU customers are deploying Tomahawk 5 → 6 → Ultra, and scale-up solutions using them appear in both XPU and some GPU clusters — the open-Ethernet standard is the stated reason.
• Leader in every generation of PCI Express switching.
• Leading-edge optical DSPs, plus a leadership position in EMLs, VCSELs and CW lasers for optical interconnect — capacity being more than tripled year over year across US and Singapore indium phosphide fabs.

• Q3 revenue $8.8B, up 29% Y/Y; Q4 guided to stabilize at ~$8.7B.
• VMware Private AI Cloud announced — a secure, cost-effective platform for enterprises to build and run AI alongside existing applications, bundling AI infrastructure, security, compliance and tooling for trusted AI agents while keeping enterprise data protected.
• VMware Cloud Foundation is making it easier to repatriate workloads from public cloud to private cloud for greater control and materially better infrastructure economics.
• Management frames enterprise AI consumption as a new growth vector for the software business, not just a defensive play.

• Full-year FY2026 AI revenue: $58B, up 186% Y/Y — raised from the prior $56B guidance.
• On gross margin, Tan was blunt: "Stop focusing on gross margin… look at operating margin at the end of the day," because revenue growth far outpaces the opex needed to support it.

• FY2027 AI semiconductor revenue ~$115B — supply already secured. Demand exceeds this outlook; Broadcom will work to improve supply.
• FY2028 AI semiconductor revenue ~$230B — doubling again, supply also secured.
• Together, roughly $350B of AI semiconductors shipped to these six customers over two years.
• Explicitly framed as a trajectory, not quarterly guidance — Broadcom does not intend to update these figures each quarter.
• Content per gigawatt: $20–30B, and expected to stay in that range. Reason: each XPU generation is more expensive, but also draws more power — so fewer chips fit in a gigawatt, holding dollar content roughly stable while the number of gigawatts accelerates.
• Compute demand from this concentrated group of frontier-model developers is expected to inflect further in 2027 and 2028.

• Free cash flow $13.7B — a record, and 46% of revenue.
• Cash ended at $24B, up $4.3B sequentially from $19.6B.
• Inventory $4.5B, built to support semiconductor demand.
• Capex $532M in Q3, stepping up to $1.4B guided for Q4.
• Dividends paid $3.1B at $0.65 per share quarterly.
• Paid down $5.6B of long-term debt in Q3, plus another $1.5B of senior notes at maturity after quarter end.
• Gross principal fixed-rate debt of $59.6B carries a 4% weighted-average coupon and 7.4 years to maturity.

• Established in June with Apollo and Blackstone to enable more than 20 GW of compute infrastructure for OpenAI and Anthropic by end of 2028.
• First tranche of $35B closed in June, funding Anthropic's 1 GW deployment, already underway.
• Structure: third-party financial partners independently underwrite and capitalize the assets — Broadcom is not the direct lender. Where necessary it provides modest residual value guarantees, treated as contingent liabilities it views as low risk given lab profitability trajectories and asset value.
• Three stated rationales: bridge the labs' cash flow versus upfront investment; concentrate capital where demand is already locked in; and review every financing through both a commercial and balance-sheet lens.
• Tan's investment case: each gigawatt of deployed compute can generate roughly $30B of ARR for the customer. "That's a hell of a business model."
• No new backstops or residual value guarantees announced. Asked whether the ~$29B maximum exposure disclosed in the last 10-Q scales per gigawatt, Tiner declined to generalize — future tranches will be tailored deal by deal to each lab and investor.
• Credit context: Anthropic is described as well on its way to an IPO, which would change its investment credit; OpenAI's path is less clear.

• Land, power and shell (LPS) — the dominant timing constraint. Long-lead construction dictates when capacity actually comes online, and Broadcom validates customer LPS plans before putting demand into its forecast.
• Substrates — Broadcom is standing up its own substrate capacity at scale in Singapore, in production starting fiscal 2027, with a partner.
• Memory — HBM plus the system memory in AI servers, which Broadcom does not supply but its customers must secure.
• Leading-edge wafers.
• Lasers — EML and CW laser demand far exceeds industry supply; Broadcom is more than tripling indium phosphide capacity Y/Y in the US and Singapore.
• Tan on the guidance philosophy: "We try to be conservative… we can ship significantly more." The open question is whether shipped chips get deployed on a timely basis. If supply improves, the outlook goes up.
• A wry note on customer concentration: "Which is why in some ways I'm so glad we have only six customers to deal with."