Broadcom Q3 FY2026 — Earnings Recap

Summary

Broadcom delivered a record quarter and then buried the headline underneath it. Revenue of $29.6bn (+86% YoY) and non-GAAP EPS of $3.32 (+96%) beat consensus, with AI semiconductor revenue of $16.7bn growing 221% year-on-year and 54% sequentially, comfortably ahead of the company's own $16.0bn guide. Non-GAAP operating margin hit a record 67.9% and free cash flow reached $13.7bn, or 46% of revenue. But the substance was in the outlook: Hock Tan guided AI semiconductor revenue to ~$115bn in FY2027 and ~$230bn in FY2028 — roughly a 4x from FY2026's $58bn — and stated that supply is already secured for both years, with demand exceeding the FY27 figure. He then went further than he ever has, putting a number on earnings: above $30 of EPS in FY2028. On our reading, the competitive picture is better than the "MediaTek share loss" narrative allows given that Broadcom remains lead TPU supplier across coming generations and is capturing the entirety of Anthropic's TPU programme (5GW in FY27 and 10GW in FY28), and the share it has conceded at Google is a pricing outcome rather than a technology one. The stock fell as much as 6% before recovering most of the losses on a 0.7% miss to the Q4 revenue guide and now trades at roughly 12x a company-guided two-year-forward EPS - a valuation that, in our view, does not survive contact with the numbers management has just put on the record.

Key Takeaways

1. The TPU "share loss" story is narrower than the narrative and Anthropic is the offsetting kicker

The bear case on Broadcom for much of 2026 has been that Google is dual-sourcing its TPU programme and that Broadcom is structurally losing content. This quarter clarified the picture, and it is considerably more favourable than the framing suggests.

First, the share loss is confined to Google's own TPU business, and even there Broadcom is winning on execution. Management disclosed that it is shipping TPU v8i ahead of MediaTek's v8t, despite the MediaTek programme having been initiated earlier in the quarter. Our channel work indicates Broadcom retains lead-supplier positioning on the TPU roadmap across several coming generations, and that the share MediaTek has taken is a pricing outcome rather than a capability one. Google pays roughly half the gross margin dollars to MediaTek that it pays to Broadcom for comparable content. That is a very different proposition from losing sockets on merit, and it caps how much further the second source can take.

Hock Tan's stated moat is specific and, unusually for a moat claim, checkable: industry-leading SerDes, chip-to-chip interconnect, leading-edge HBM and SRAM integration, differentiated advanced packaging and the fastest time to market from product definition to production without respins. The absence of respins on programmes of this complexity is the operationally meaningful part.

Second, and more importantly, Google's absolute volumes still grow. Broadcom signed a long-term agreement to develop and supply future generations of TPUs and AI networking, under which it plans to deliver "multi tens of billions of dollars of TPUs annually over the next several years," with FY28 and FY29 demand fulfilled through successive generations of increasingly complex TPU SoCs already in development. A shrinking percentage of a rapidly expanding pie can still be a much larger dollar number, and that is what is happening here.

Third and this is the real kicker: Broadcom is capturing the entirety of Anthropic's TPU purchase. The disclosed schedule is 1 gigawatt of Ironwood in 2026, 5 gigawatts of TPU v8i in 2027, and a clear line of sight to an incremental 10 gigawatts in 2028, approximately 15 gigawatts across the next two fiscal years. Management stated plainly that Anthropic becomes Broadcom's largest XPU customer in 2027 and sustains that position in 2028, ahead of Google.

That is the single most consequential disclosure in the release. It converts a story about defending a mature Google socket into a story about a second anchor customer of comparable scale arriving on top of it and it is what underwrites the visibility on revenue and earnings growth over the next eight quarters.

2. Hock Tan has followed Jensen Huang into long-dated guidance, and we think the motive is the same

One week after NVIDIA broke with a decade of practice to guide a full fiscal year forward, Broadcom went further and guided two. We do not think that is coincidence. Both companies are responding to the same thing: a market that has spent much of 2026 questioning whether AI infrastructure demand is durable or whether an air pocket is coming, and which has been unwilling to pay for near-term momentum without a view of the far side.

The numbers management put on the record:

That is roughly a 4x in AI semiconductor revenue from FY2026 to FY2028, or approximately $345bn of AI silicon shipped to six customers across two fiscal years and both figures sit well above where the Street was carrying its numbers.

Two features of the guide matter more than the headline. The supply language is stronger than NVIDIA's. NVIDIA said supply allows 70% growth against ~100% demand. Broadcom said supply is secured for a doubling, and then another doubling. That is a commitment against contracted wafers, substrates and HBM rather than an aspiration.

And the guide embeds a deliberate haircut. Asked by Bernstein to reconcile the gigawatt schedule with the revenue guide, Hock confirmed the ~10GW (2027) and ~20GW (2028) arithmetic and then volunteered that the guides do not assume all 30 gigawatts enters production in the window, because data centre shells must be ready before chips ship: "We think we judge it conservatively to be somewhat less." At the $15–25bn of Broadcom content per gigawatt he later specified, the $115bn and $230bn guides imply only roughly 45–50% of committed gigawatts actually deploying. Management is guiding to what it believes it can physically ship, not to what customers have asked for. That is the opposite of a promotional guide, and it is why we treat the numbers as a floor.

3. $30 of EPS in FY2028 and the multiple that implies

The most important sentence on the call was delivered almost in passing: "We are very much on target to exceed $30 in earnings per share in fiscal 2028."

Context matters here. Hock Tan is not a chief executive who throws numbers around for effect. Across a two-decade record of acquisition integration and margin delivery he has been consistently, sometimes frustratingly conservative in what he commits to publicly, and he has a long history of meeting or exceeding the targets he does set. When he puts a number on the record, our working assumption is not only that it is achievable but that it is more likely to be beaten than missed.

The claim is also internally consistent with everything else guided. FY28 AI revenue of $230bn, plus roughly $18bn of non-AI semiconductors and roughly $40bn of infrastructure software, gives approximately $288bn of revenue. At the ~66% operating margin management says it can sustain through the mix shift, that is around $190bn of operating income; less interest on $59.6bn of debt at a 4.0% coupon, taxed at the guided 16%, and divided across ~4.94bn shares, it produces approximately $32 of EPS. The arithmetic reconciles.

What that means for the stock is straightforward:

A 25x multiple is not a heroic assumption for a business compounding AI revenue at a 4x over two years, running 68% operating margins, converting 46% of revenue to free cash flow on capex of under 2% of sales, and carrying a 94%-gross-margin software franchise alongside it. It is a discount to where the shares have traded at points in the last eighteen months. On that basis we get to roughly $750 and a two-year CAGR comfortably above 40% from current levels.

The market is not there. Broadcom is up around 6% year to date against a semiconductor sector up roughly 60%, and remains around 19% below where it traded before the December print. Sell-side price targets cluster between $490 and $525. The gap between those targets and the share price is not a disagreement about whether Broadcom can build the chips it is a refusal to underwrite the customers.

4. GPUs and XPUs will co-exist and this quarter, alongside Marvell's, settles it

We have made this point before and this earnings season reinforces it. The custom-silicon-versus-merchant-GPU debate is repeatedly framed as zero-sum. It is not. The scale of business now being disclosed by both Broadcom and Marvell is arriving alongside NVIDIA guiding to roughly 70% revenue growth in its own next fiscal year, not instead of it.

Marvell, reporting a week earlier, raised its FY2027 revenue outlook to ~$12bn from $11.5bn (+~45%) and its FY2028 outlook to ~$18bn from $16.5bn (+~50%), the second consecutive quarter it has lifted both with data centre revenue guided to grow ~60% in FY27 and above 60% in FY28, and custom silicon expected to more than double in FY28. It also disclosed an expanded commercial agreement with a large hyperscaler carrying warrant coverage tied to revenue milestones. Three suppliers, all guiding up, all at once.

The division of labour is becoming clear, and it is a functional one rather than a competitive one:

• GPUs remain favoured for training and for general-purpose, fungible workloads — where model architectures are still moving, where the workload is unknown in advance, and where a single platform that runs everything is worth paying for.

• XPUs are being deployed by established players with the engineering depth to design their own silicon and the workload scale to justify it - overwhelmingly for inference, where the workload is known, stable, repetitive and enormous, and where a chip tailored to one company's models delivers materially better performance per watt and per dollar.

The evidence from this call is direct. Broadcom's TPU v8i is described as inference-optimised, with more memory and bandwidth than Ironwood, and comparable to or better than Vera Rubin in performance. OpenAI's "Jalapeno" is stated to outperform Grace Blackwell Ultra on performance per watt, latency, throughput and power, and to be comparable to Vera Rubin on OpenAI's own workloads. Meta's MTIA is explicitly built for inference and recommendation at scale.

None of that is a threat to a training franchise. It is the inference layer of the same buildout being served by a second architecture, and both are growing at triple digits simultaneously.

5. Land, power and shell is the binding constraint and that points somewhere else entirely

For the second time in a week, the chief executive of a company at the epicentre of AI compute told investors that the cap on his growth is not silicon.

Hock Tan was unambiguous and repeated it across three separate answers. When Broadcom sets its outlook it does not simply take what a frontier-model customer asks for; it is "very engaged with each of them on how much LPS — land, power, and shell they have before we actually believe it will happen," because that element carries the longest lead time and is fundamentally a construction project. Asked to rank the constraints, he described a multi-dimensional problem in which leading-edge silicon, substrates, HBM memory and even system memory each become the bottleneck at different moments but LPS is the one that "dictates specific timing of when this capacity gets deployed." It is why the revenue guides assume fewer than half the committed gigawatts actually deploy.

This is exactly what NVIDIA said a week earlier, and the read-through is the same: the marginal dollar of AI capex is migrating from silicon toward everything that surrounds it. Where we think that leads:

• Behind-the-meter and dedicated power generation — the ability to bring firm, off-grid capacity to a site is now the scarcest input in the entire chain.

• Electrical infrastructure and specialist contracting — switchgear, transformers, medium-voltage distribution, and the engineering and construction capacity to install it. Broadcom's own customers are being gated by exactly this.

• Industrials supplying the data centre shell — structural build, thermal management and liquid cooling, and the mechanical equipment required to stand a gigawatt-class facility up.

Notably, Broadcom is itself responding to the adjacent bottlenecks by vertically integrating: capex steps from $532m to a guided $1.4bn in Q4, funding owned substrate capacity in Singapore from FY2027 and more than tripling indium phosphide capacity year-on-year for EML, CW and VCSEL lasers across the US and Singapore, where Hock says demand "far surpasses" industry supply. When a fabless business starts building factories, that is a statement about where the constraints have moved.

6. XPUs pull Broadcom's networking business and the wider ecosystem with them

The under-discussed half of the AI story here is networking, and it benefits from custom silicon in a way that is structurally different from the accelerator business.

The numbers first. AI networking revenue grew over 2.5x year-on-year in Q3 and is guided to triple year-on-year in Q4 alongside XPUs, and management stated explicitly that AI networking revenue is expected to grow just as fast as XPUs over the next few years. With XPUs at 73% of AI revenue, that means roughly a quarter of the AI business compounds at the same extraordinary rate — off a product portfolio that is best-in-class:

• Tomahawk 6 — first to market at 100 terabits, in both 100G and 200G SerDes versions, and per management deployed at essentially every AI hyperscaler building XPUs with Broadcom, and also at those that are not.

• Tomahawk 7 — just taped out; the industry's first 200 terabit-per-second Ethernet switch, with 400G SerDes.

• Tomahawk Ultra — scale-up over low-latency Ethernet, i.e. open-standard competition for proprietary scale-up fabric. Adoption "surprised us", deployment starts this quarter and scales through FY27, and it is going into both XPU and GPU clusters.

The strategic point is the one management made almost in passing, and it is the most important thing in this section: because these are Ethernet-based and therefore open, anyone can connect to them. An XPU deployment does not lock the operator into a single vendor's networking architecture. That is a material part of why hyperscalers pursue custom silicon in the first place, it preserves the freedom to assemble the rest of the rack to their own specification.

And that freedom is what makes this an ecosystem trade rather than a single-name one. Operators pairing Broadcom switch ASICs with system and platform vendors — the merchant-silicon-based switching and networking OEMs, and the ODMs building the racks — capture a share of the same buildout. The switch silicon, the optics and DSPs, the co-packaged optics roadmap, the systems integration and the rack assembly are all being pulled by the same XPU volumes. The custom accelerator is the entry point; the networking, optical and systems content around it is where a meaningful part of the value accrues, and it accrues to a wider set of names than the accelerator vendor alone.

Appendix — Figures

A. Q3 FY2026 delivered versus expectations

B. Guidance issued versus expectations

Management has stated it does not intend to update the multi-year AI revenue guidance on a quarterly basis.

C. Headline P&L progression

D. Segment detail — Q3 FY2026

Semiconductor segment gross margin of ~67% is derived from the disclosed 61% segment operating margin plus segment opex at 6% of segment revenue, and cross-checks against consolidated and software gross margins. Note that within AI revenue, XPUs represented 73% (~$12.2bn) and AI networking 27% (~$4.5bn).

E. Disclosed XPU customer roadmap

Broadcom guides to shipping ~$345bn of AI semiconductors across FY27 and FY28, which at management's stated $20–30bn of content per gigawatt implies roughly 45–50% of committed gigawatts actually deploying in the window — the deliberate haircut for land, power and shell readiness.

F. Valuation

Context: Broadcom is up roughly 6% year to date against a semiconductor sector up roughly 60%, and sits around 19% below its pre-December level. Sell-side consensus is Buy / Strong Buy across 25–27 analysts, with average price targets of approximately $490–500 and a median around $525.

G. Cash flow, balance sheet and capital allocation

H. Custom silicon peer context — Marvell Q2 FY2027 (reported 27 August 2026)

Second consecutive quarter in which Marvell has raised both its FY27 and FY28 outlooks. Three custom-silicon suppliers guiding up simultaneously, alongside NVIDIA guiding to ~70% revenue growth in its own next fiscal year, is the clearest available evidence that the GPU and XPU markets are expanding together rather than at each other's expense.