Riccardo Milani/gettyimages.com

Broadcom's fiscal third quarter produced $16.7 billion in AI semiconductor revenue — more than triple what the company generated in the same period a year ago — and then delivered something the $16.7 billion figure alone does not convey: Hock Tan told investors Wednesday that actual customer demand already exceeds even Broadcom's own $115 billion fiscal 2027 target, meaning the company's remarkable forecast is intentionally conservative because fabrication capacity and advanced packaging supply, not orders, are the binding variable, per the official Q3 earnings press release.

The disclosure inverts the narrative that has dominated Broadcom coverage through the first half of 2026. The risk was supposed to be whether demand would hold. It is holding — and then some.

Anthropic Is Now Broadcom's Biggest Customer in Waiting

The single most structurally significant statement from Wednesday's earnings call was not about the $230 billion fiscal 2028 target, striking as that figure is. It was that Anthropic is on track to become Broadcom's largest XPU customer in 2027, surpassing Google — and to hold that position through 2028, according to the Benzinga Q3 2026 earnings call transcript. Hock Tan said Anthropic will deploy an additional 5 gigawatts of next-generation TPU v8i capacity in 2027 and that Broadcom has clear line of sight to deliver another 10 gigawatts in 2028. That progression — 1 gigawatt of Ironwood in 2026, 5 gigawatts of v8i in 2027, 10 more gigawatts in 2028 — represents one of the most aggressive compute capacity ramps any AI laboratory has publicly committed to.

The shift matters editorially because Broadcom's AI story has been constructed, quarter by quarter, around the Google relationship: Google's TPU programs were the revenue anchor, and every analyst model asking whether the $56 billion fiscal 2026 forecast was achievable ran primarily through Google. Anthropic's emergence as the new anchor customer in 2027 means the company's forward revenue is less exposed to the Google concentration risk that prompted Macquarie's June downgrade — Macquarie cut Broadcom to Neutral from Outperform over that concern — a move that sent the stock sharply lower and has weighed on it relative to peers ever since.

OpenAI, whose first-generation Jalapeño accelerator Broadcom shipped during the third quarter, is on track to rank as the second-largest XPU customer in 2027, per the Benzinga call transcript. Tan said Jalapeño outperforms Nvidia's Grace Blackwell GPU for inference workloads, placing the custom chip at performance parity with or above Nvidia's latest merchant silicon for the specific workloads OpenAI needs. Broadcom and OpenAI are already in development on a second-generation chip and outlining a third.

For Meta, Broadcom expects to begin production shipments of the latest MTIA custom accelerator in the fourth quarter, optimized for inference and recommendation workloads at scale. "Four of our six customers will grow to enormous scale," Tan said, per the Q3 2026 Investing.com transcript.

Q3 by the Numbers

Total consolidated revenue for the quarter ended August 2 came in at $29.59 billion, an 86% year-over-year increase that beat the Wall Street consensus of approximately $29.36 billion, per the official earnings press release. The quarter extended Broadcom's streak of adjusted EPS beats to nine consecutive quarters, with non-GAAP diluted EPS of $3.32 clearing the consensus of $3.22 to $3.24.

Profitability metrics showed the operating leverage that custom silicon generates at scale. Non-GAAP operating income grew 92% year over year to $20.1 billion, representing an operating margin of 68% — a rate that reflects how little incremental overhead the AI chip business requires relative to the revenue it produces, per the earnings press release. Free cash flow reached a record $13.7 billion, equal to 46% of quarterly revenue. The company ended the period with $24.0 billion in cash after paying down $5.6 billion of debt and returning $3.1 billion in dividends.

AI semiconductor revenue of $16.7 billion now accounts for 56% of Broadcom's total revenue — a share that has grown from negligible two years ago, per Benzinga's Q3 call transcript. The Semiconductor Solutions segment as a whole rose 127% year over year to $20.8 billion. Infrastructure Software — anchored by the VMware portfolio — grew 29% to $8.75 billion, marginally short of the $8.82 billion consensus estimate, though it did not generate the kind of concentrated analyst focus that a software miss did in the second quarter.

For the fourth quarter of fiscal 2026, Broadcom guided total revenue to approximately $34.8 billion, 93% above the same period a year ago. AI semiconductor revenue is expected to reach $21.7 billion — a 236% year-over-year increase — as XPU and networking revenue together triple year on year, per the earnings press release. The guidance landed approximately $230 million below the analyst consensus of $35.03 billion, enough to push AVGO down as much as 6% in after-hours trading before partially recovering. The stock had closed the regular session Wednesday at $367.47, up about 6% year to date — meaningfully behind the S&P 500's 12% gain over the same period, a divergence that has puzzled observers given the company's underlying AI chip growth rate.

Cody Acree, an equity research analyst at StoneX with a Buy rating on the stock, captured the reaction precisely: "The magnitude is not quite enough from a top and bottom line standpoint on the beat and raise when you have a company that is this levered to AI," he told Yahoo Finance.

The board declared a quarterly cash dividend of $0.65 per share, payable September 30 to shareholders of record as of September 21.

The Roadmap: $58B This Year, $115B Next, $230B in Two

The financial figures for fiscal 2027 and 2028 that Tan disclosed Wednesday were prepared remarks, not responses to analyst questions — a signal that Broadcom views them as sufficiently reliable to state publicly rather than let them emerge from Q&A hedging.

Full-year fiscal 2026 AI semiconductor revenue is now projected at $58 billion, up from the prior $56 billion guidance — a 186% year-over-year increase, per Benzinga's earnings call transcript. For fiscal 2027, Broadcom has secured the supply to double AI revenue to approximately $115 billion. For fiscal 2028, the company has line of sight to double again to $230 billion — with supply secured for that target as well.

The crucial qualifier, stated plainly by Tan: demand for fiscal 2027 already exceeds the $115 billion outlook, per the Investing.com Q3 earnings transcript. Broadcom is not holding back a bullish forecast to be conservative. It is holding back a larger number because it has not yet secured all the supply chain capacity needed to deliver it. The forward demand roadmap across Broadcom's six XPU customers extends to 30 gigawatts of aggregate compute capacity — a figure that maps to revenue well above the disclosed targets, per TheStreet's live earnings call updates.

Whether Broadcom can convert that order pipeline into delivered silicon — and whether TSMC can expand fabrication capacity quickly enough — will define the AI infrastructure story for the next two years.

What Actually Limits the $115B Target: CoWoS, HBM, and Power

Tan named the supply chain inputs that are pacing Broadcom's AI revenue growth: leading-edge wafers at the 3-nanometer node, CoWoS advanced packaging substrates, high-bandwidth memory, power, land, and system components needed for large-scale data center deployment, per Yahoo Finance's Q3 earnings call highlights.

Understanding why packaging rather than silicon is the binding constraint requires a brief look at how modern AI chips are built. A custom accelerator like Google's Ironwood TPU is not simply a logic die — it is a system assembled through CoWoS (Chip-on-Wafer-on-Substrate), the advanced 2.5D packaging process TSMC developed over 12 years and now controls at roughly 90% of global capacity at AI-chip scale, according to Epoch AI's chip supply research. CoWoS sits the logic die on a silicon interposer alongside stacks of high-bandwidth memory, connecting them with thousands of short, dense electrical paths that enable the terabytes-per-second memory bandwidth modern AI model weights require. Without this packaging step, a fabricated 3nm wafer cannot become a functional, shippable accelerator.

TSMC has confirmed that CoWoS capacity has been sold out through 2026 and into 2027, with packaging lead times of 52-78 weeks. Scaling from approximately 35,000 CoWoS units per month in late 2024 to 130,000 by end-2026 is itself an extraordinary manufacturing achievement — and it is still short of demand. HBM memory from SK Hynix, Samsung, and Micron represents a parallel constraint: AI chip designers consumed roughly 90% of global HBM supply in 2025, leaving virtually no margin for demand growth without new capacity coming online, per the Epoch AI analysis.

Broadcom announced it will increase capital expenditure to $1.4 billion in the fourth quarter, directing spending toward substrate production capacity in Singapore and indium phosphide capacity for co-packaged optics — the optical interconnect technology that Broadcom's AI networking portfolio depends on to link XPU clusters across data centers, per the Benzinga Q3 transcript.

How Ironwood v8i Changes the Technical Benchmark

During the third quarter, Broadcom delivered Google's Ironwood TPU v7 in high volume to both Google and Anthropic — simultaneously. At the same time, the company began production shipments of the next-generation TPU v8i, a chip Tan described as having substantially more memory and bandwidth than Ironwood and as designed specifically for inference workloads, per the Alphastreet Q3 2026 transcript.

The v8i's performance claim carries competitive significance: Tan said it is "comparable to, if not surpasses" Nvidia's Vera Rubin GPU — the generation of Nvidia hardware that will not reach volume production until 2027 at the earliest. Whether that benchmark holds at scale across diverse inference workloads remains to be demonstrated in production, but the design achievement is notable: Broadcom is now simultaneously ramping two generations of Google TPU, with v8i shipping ahead of MediaTek's competing v8T design that was initiated before v8i's program began, per Benzinga's full call transcript.

For context on the infrastructure that connects these chips at data center scale, Broadcom's Tomahawk 6 and Tomahawk Ultra Ethernet switching products continued ramping across AI customers during the quarter, and AI networking revenue grew more than 2.5 times year over year. The networking portfolio — Ethernet switches, PCIe switching, optical DSPs, lasers, and co-packaged optics — is what allows thousands of XPUs to operate as a coherent AI training or inference cluster rather than as disconnected processors, per Benzinga's Q3 earnings call.

What Does Broadcom's AI Chip Business Actually Do?

Unlike Nvidia, which sells graphics processing units to a broad market — including gaming, scientific computing, and enterprise inference — Broadcom's AI semiconductor business is built on co-design partnerships with individual hyperscalers. The company's engineers collaborate with a customer's architecture team over an 18-to-24-month design cycle to translate the customer's specific model architecture — the structure of their neural network, the memory access patterns, the precision requirements — into a physical chip layout that TSMC can fabricate.

The resulting chip is an application-specific integrated circuit: it can do one thing extraordinarily well and cannot be repurposed for a different customer or a different class of workload, as IEEE Technology Navigator describes for this class of semiconductor. This specificity is precisely what makes the economics work. A chip optimized for running Google's specific transformer architecture can devote die area to matrix multiply-accumulate units calibrated for that workload rather than the general-purpose floating-point logic a GPU needs to serve thousands of different customers. Independent operators have confirmed the economics: Midjourney, the AI image platform, reported cutting monthly inference compute costs from approximately $2.1 million to $700,000 — a 65% reduction — after moving inference workloads from Nvidia GPUs to Google's seventh-generation TPUs, per TechTimes' AI chip cost analysis.

The tradeoff is lock-in and lead time. Once a hyperscaler has co-designed a chip with Broadcom, switching design partners costs roughly one to two chip generations — three to five years of roadmap continuity — because the design knowledge, the verified IP blocks, and the TSMC process relationships do not transfer. That switching cost is Broadcom's moat, and it deepens with every generation of TPU, MTIA, and Jalapeño that ships.

Hyperscaler Capex as the Upstream Driver

The five largest hyperscalers — Alphabet, Microsoft, Amazon, Meta, and Apple — collectively increased capital expenditures approximately 80% in 2026 to more than $700 billion, with AI infrastructure representing the primary growth driver, per Bloomberg's Big Tech capex tracking. That spending flows directly into the order book of the companies that supply the chips, packaging, and networking equipment required to build AI data centers at scale.

Broadcom's AI chip order book entered Q3 carrying more than $30 billion in bookings from the prior quarter alone — a figure that, annualized, exceeds the company's entire AI semiconductor revenue for fiscal 2025, per Reuters reporting via KELO.

VMware's AI Push and the Enterprise Layer

Broadcom released its Q3 earnings immediately following VMware Explore 2026, the four-day conference held this week in Las Vegas that opened September 1. At that conference, Broadcom unveiled VMware Private AI Cloud — a platform designed to bring production AI inference inside enterprise data centers — alongside AgentMinder, a governance control plane for autonomous AI agents that Broadcom said handled 43 million API calls per day in internal production. The enterprise AI product line represents Broadcom's bet that the AI infrastructure cycle is not limited to hyperscalers: as enterprises move AI inference onto private infrastructure to manage cost and data governance, VMware Cloud Foundation becomes the software layer those workloads run on, per the VMware Private AI Cloud announcement.

The VMware business has faced significant customer retention headwinds since the $61 billion acquisition in 2023 — a Gartner survey from April 2026 found 76% of VMware customers hold negative views of Broadcom's ownership — but the AI inference narrative gives the platform a retention argument that licensing disputes cannot: migrating away from VMware Cloud Foundation now means rebuilding the AI infrastructure stack from scratch on a competitor platform.

Why Demand Still Outstripping Capacity Is the Story

Broadcom's disclosure that customer demand exceeds its own $115 billion fiscal 2027 guidance is the fact that changes how to read the rest of the numbers. A $16.7 billion quarter and a $21.7 billion Q4 guide are extraordinary. But they are smaller than they could be if CoWoS capacity, HBM allocation, and data center power were unconstrained. The $230 billion fiscal 2028 target is not Broadcom's ceiling — it is the number that Tan believes the supply chain can deliver, with the acknowledged caveat that actual demand exceeds even that figure.

The practical implication for the technology industry is that AI infrastructure build-out is being paced not by customer appetite — which appears essentially unlimited — but by the speed at which TSMC can expand packaging lines, SK Hynix and Micron can produce HBM stacks, and utility operators can permit and commission the power substations needed to run data centers consuming gigawatts of electricity.

RBC has projected the total AI semiconductor market could exceed $550 billion by 2028, per RBC's AI semiconductor forecast. If Broadcom's $230 billion target holds, it would represent roughly 40% of that market — through a business model built not on selling chips to everyone, but on designing chips for six specific customers who are each individually building AI infrastructure at a scale the world has not previously attempted.

Frequently Asked Questions

Why did Broadcom stock fall after what appears to be a record quarter?

The quarter itself was not the source of disappointment. Broadcom beat analyst estimates on both revenue ($29.59 billion vs. the $29.36 billion consensus) and adjusted EPS ($3.32 vs. $3.22 expected). The sell-off — which initially reached approximately 6% in after-hours trading before partially recovering — was driven by Q4 guidance of $34.8 billion landing roughly $230 million below the analyst consensus of $35.03 billion. At the elevated valuations that AI infrastructure stocks carry in 2026 — Broadcom trades at roughly 25 times forward earnings — even a minor guidance miss against high expectations produces outsized price reactions. As StoneX analyst Cody Acree put it, for a company this exposed to AI, the market expects the beat-and-raise to be larger than what Broadcom delivered, per Yahoo Finance's analyst commentary.

How does an XPU differ from an Nvidia GPU, and why does the difference matter for AI economics?

An XPU (Broadcom's term for a custom ASIC it co-designs with a specific hyperscaler) is built to run one customer's specific model architecture and nothing else. An Nvidia GPU is a broadly programmable processor that can run any workload — the same chip trains a language model one week and runs a video rendering job the next. The GPU's programmability is its commercial advantage for Nvidia (it can be sold to millions of customers) but its efficiency disadvantage for any single buyer (die area is devoted to general-purpose compute rather than the specific matrix operations the buyer needs). An XPU eliminates that overhead: every transistor is allocated to the operations the customer actually runs. The result, as demonstrated by production deployments, is a 30% to 65% reduction in inference compute cost compared to GPU-based alternatives for the specific workload the chip was designed for, per TechTimes' XPU cost analysis.

What is limiting Broadcom's AI chip growth if customer demand exceeds even the $115 billion forecast?

The binding constraint is not customer orders — it is the supply chain required to manufacture and package the chips those orders represent. CoWoS (Chip-on-Wafer-on-Substrate), the advanced packaging process that bonds AI logic dies to high-bandwidth memory on a silicon interposer, is sold out at TSMC through 2026 and into 2027, with lead times of 52 to 78 weeks. HBM memory from SK Hynix, Samsung, and Micron is a parallel constraint — AI chip manufacturers consumed roughly 90% of global supply in 2025. Beyond chips, the data centers that will house these accelerators require power at gigawatt scale, and permitting and construction timelines for new substations are measured in years. Broadcom's $115 billion fiscal 2027 target is the number Hock Tan believes the supply chain can support, with the explicit acknowledgment that orders from six XPU customers already exceed even that figure, per the Investing.com earnings call transcript.

Which companies are Broadcom's most important AI chip customers, and is Google still the dominant one?

Google has historically been Broadcom's largest custom AI chip customer, through the long-running TPU program whose latest generations (Ironwood v7 and v8i) are shipping now. But Hock Tan disclosed Wednesday that Anthropic is on track to surpass Google in 2027 and hold that position into 2028 — with a deployment trajectory of 1 gigawatt of Ironwood in 2026, 5 gigawatts of TPU v8i in 2027, and another 10 gigawatts in 2028. OpenAI, whose first-generation Jalapeño accelerator began shipping in the third quarter, is expected to rank second. Meta's latest MTIA accelerator is entering production in the fourth quarter. Google remains a major customer under a long-term agreement that covers "multi-tens of billions of dollars" of TPUs annually for several years — but Anthropic is emerging as the new anchor relationship in Broadcom's forward revenue model, per the Benzinga earnings call transcript.