Earnings Call Highlights: Broadcom (AVGO) | Q3 FY2026 — $16.7B of AI in One Quarter, FY2028 Called at $230B, and the 30 GW Math Ends Up Hinging on Land and Power

2 days ago

11 min read

This quarter Broadcom made its case that "custom accelerators (XPUs) are not a GPU substitute but a main line of their own" with $16.7B of AI semiconductor revenue in a single quarter. Three signals matter most: first, AI revenue grew 221% YoY and 54% QoQ, and its share of total revenue jumped from 49% to 56% within one quarter; second, management broke precedent with a two-year outlook (FY2027 AI revenue of about $115B and FY2028 of about $230B), saying in both cases that "supply is already secured"; third, Hock Tan personally laid out roughly 30 GW of deployments across six customers over the next two years, then immediately added that this doesn't mean all 30 GW will go into production within those two fiscal years.

The third point is the real information in this call. Broadcom is solving chip-side bottlenecks itself (its own substrate plant in Singapore, laser capacity up 3x YoY), but land, power and shell (LPS) are not in its hands. That is why the $230B figure must be read together with "we judge it will be somewhat less than 30 GW".

This article is for technology and industry trend analysis only and does not constitute investment advice.

1. Executive takeaway: the three most important signals

Signal one: AI semiconductors are rewriting Broadcom's P&L into a different company. AI semiconductor revenue was $16.70B for the quarter, up 221% YoY and 54% QoQ, 56% of total revenue (49% last quarter). XPU shipments grew more than 3.5x YoY and made up 73% of AI revenue; the remaining 27% (about $4.5B) was AI networking, up more than 2.5x YoY. This isn't "AI business growth"; it's a change in who the company fundamentally is.

Signal two: two-year revenue guidance, explicitly not updated quarterly. FY2026 AI revenue was raised to $58B (from $56B), FY2027 to about $115B and FY2028 to about $230B: doubling two years in a row. CFO Amie Thuener said plainly that this is "to give you the growth trajectory; we don't intend to update it quarterly". Management also set the FY2028 EPS bar at above $30. Guidance like this is rare in semiconductors, and it also means the market loses a calibration point for the next few quarters.

Signal three: the customer list was laid out at GW scale for the first time. Google (long-term agreement, up to tens of billions of dollars of TPUs per year), Anthropic (1 GW in 2026 → 5 GW in 2027 → an incremental 10 GW in 2028, becoming the largest XPU customer from 2027), OpenAI (1.3 GW in 2027 → more than 5 GW in 2028, the second largest), Meta (three generations of MTIA, 3 GW cumulative by 2028). That adds up to about 30 GW, and Hock tempered expectations himself: it will be less than 30 GW, because the data centers may not all get built.

2. Revenue and financials: gross margin gives up 3.4 points in exchange for another 240 bps of operating margin

All figures below are on a non-GAAP basis and fully consistent with this quarter's earnings data card. The fiscal quarter ended 2026-08-02.

Core P&L (Q3 FY2026)

Revenue $29.59B: up 85.5% YoY and 33.4% QoQ;

Gross margin 75.0%: down 3.4 pp YoY and 210 bps QoQ;

Operating income $20.10B, operating margin 67.9%: up 92.2% YoY, with operating margin up 240 bps YoY;

Net income $16.37B: up 94.8% YoY; diluted EPS $3.32: up 96.4% YoY.

Segments (mutually exclusive, summing to total revenue)

AI semiconductors $16.70B (up 221% YoY, 54% QoQ); infrastructure software $8.75B (up 29% YoY, ARR up 15% YoY); non-AI semiconductors $4.14B (up 5% YoY, flat QoQ, with broadband and server storage growth offset by wireless).

The two segments have very different profit structures

The semiconductor solutions segment posted revenue of $20.84B (up 127% YoY) with a 61% operating margin (up 440 bps YoY); operating expenses were just $1.2B, 6% of segment revenue. Revenue grew 127% YoY while expenses grew only 22%, which is the entire reason operating margin can rise while gross margin falls. The infrastructure software segment has a 94% gross margin and about an 84% operating margin (up 650 bps YoY).

Cash flow and balance sheet

Free cash flow was $13.67B, 46% of revenue (operating cash flow $14.20B, capex $532M); ending cash was $23.98B (vs. $19.6B last quarter); inventory $4.5B. Broadcom repaid $5.6B of long-term debt during the quarter and another $1.5B of maturing senior notes after quarter-end; gross fixed-rate debt is $59.6B, with a weighted average coupon of 4% and weighted average maturity of 7.4 years. The dividend was $0.65 per share, totaling $3.1B for the quarter.

The gross margin line needs to be read correctly: the decline isn't pricing weakness; it's product mix. XPUs carry ever more memory content (HBM), and memory is a pass-through cost that naturally dilutes gross margin. In the Q&A Hock told analysts outright to "stop focusing on gross margin", because revenue is growing far faster than operating expenses and all the leverage shows up in operating margin. That argument is acceptable, but its flip side is: future profit improvement now has only one path, operating leverage; gross margin is no longer a variable.

For the magnitude of change versus last quarter, see our earlier piece Earnings Call Highlights: Broadcom (AVGO) | Q2 FY26 — $30B of AI orders in one quarter pushes optical communications visibility out to 2028: back then, visibility ended at 2028; this quarter turned 2028 into a concrete revenue number.

3. Technology and business highlights: TPU v8i ships early, and a chip called Jalapeño

Ironwood (TPU v7) shipped in high volume to Anthropic and Google this quarter. In the same quarter, the next-generation TPU v8i began volume shipments to Google. It has more memory and bandwidth than Ironwood and is likewise optimized for inference; in Hock's words, its performance is "comparable to, or even better than, the Vera Rubin GPU".

There's an industry signal worth noting here: Hock specifically pointed out that Broadcom's v8i shipped earlier than MediaTek's v8t, even though v8t actually started first. Google split TPUs between two design partners and Broadcom won on schedule; it was the only time in the call he named a competitor directly. For the architectural background on the TPU line, see our full breakdown in One architecture, five generations, 3,600x growth: Google dissects its TPU, and the hidden hero called "optical switching".

Jalapeño is OpenAI's first-generation in-house accelerator, and it shipped this quarter. OpenAI itself announced last week that Jalapeño beats Grace Blackwell Ultra on latency, throughput and power, and matches the Vera Rubin GPU when running OpenAI's own workloads. Hock distilled the conclusion into one sentence:

When you co-design a chip for your own LLM workload, it will beat any GPU, at half the cost.

Meta's MTIA will enter volume shipments in Q4, optimized for inference and large-scale recommendation systems.

Networking was just as significant this quarter:

• Tomahawk 6 (100 Tb/s) has become the fastest-ramping switch product line in recent years since launch, with 100G and 200G SerDes versions in parallel, used even by hyperscalers that don't use Broadcom XPUs;

• Tomahawk 7 (200 Tb/s) was announced, the industry's first;

• Tomahawk Ultra uses low-latency Ethernet for scale-up; deployment began this quarter with a ramp in FY2027, and it appears in both XPU clusters and some GPU clusters.

We already covered the full version of this "Ethernet pushing into the rack" story in 2026 OCP APAC Summit | Ethernet starts eating scale-up: Broadcom's keynote takes aim at NVLink; this call is the first time it shows up in the financials as "already shipping".

On optical components, it's a capacity story: optical DSPs remain in the lead, and EML, VCSEL and CW laser capacity is up more than 3x YoY (with indium phosphide fabs in the US and Singapore expanding in parallel); Hock said laser demand "far exceeds industry supply". This is the same bottleneck we tracked in The shortage map for "the engine of light": the wafer even NVIDIA is fighting for, and now even Broadcom is building its own.

4. Management outlook: Q4 guidance, a two-year trajectory, and "$20B–$30B per GW"

Q4 FY2026 guidance (official company guidance)

Total revenue of about $34.8B, up 93% YoY; semiconductors about $26.1B, up 136% YoY, of which AI semiconductors $21.7B, up 236% YoY (XPUs and AI networking both up 3x YoY); non-AI semiconductors about $4.3B, up 5% QoQ; infrastructure software about $8.7B, up 25% YoY; gross margin about 73% (78% a year ago); operating margin about 66% (flat YoY); tax rate about 16%; diluted share count about 4.94 billion; capex $1.4B.

Two-year trajectory: FY2026 AI revenue $58B (raised from $56B) → FY2027 about $115B → FY2028 about $230B. That's about $345B of AI semiconductor shipments over two years. FY2028 EPS target: above $30.

Management's wording needs to be read piece by piece:

• "We have secured the supply": used for both the FY2027 and FY2028 numbers. This is a confirming tone, referring to things Broadcom can contract for, such as wafers, substrates and HBM;

• "Demand actually exceeds this outlook, and we will try to increase supply": this is upside, but it isn't built into the numbers;

• "The question is, even if we ship the chips, will they be deployed on time": this is downside risk Hock raised himself, and the most honest line of the entire call.

The $/GW debate deserves its own look. Analyst Stacy Rasgon summed the GW figures and worked backward to Broadcom content of about $11–12B per GW, comparing it with the roughly $40B a competitor cites. Hock's answer came in two parts: first, "an XPU costs less than half of a GPU, you just proved my point"; then he gave a range: Broadcom's content is about $20–30B per GW, and it will stay at that level. The reasoning: each chip generation gains performance but also draws more power per chip, so fewer chips fit in 1 GW, and the two effects offset to keep dollar content per GW relatively stable.

These two figures ($11–12B vs. $20–30B) coexisted in the same call, and the difference likely lies in "whether AI networking and optical content are included". This is the most worthwhile follow-up question for next quarter. For changes in the competitor's $/GW pricing, compare Earnings Call Highlights: NVIDIA (NVDA) | Q2 FY2027 — the price of a GW jumps from $25B to $40B: the gap between how the two companies value "a GW" is itself the best quantification of the difference between the XPU and GPU business models.

5. Supply chain and customer clues: a Singapore substrate plant, the XPV platform, and LPS as a non-technical bottleneck

Broadcom is starting to build its own supply chain. Its own substrate capacity in Singapore will start production in FY2027, in a joint venture with a partner. This is a rare move for a fabless company: it means Broadcom judges the substrate shortage to be structural, not cyclical. At the same time, capacity at its indium phosphide fabs (EML/CW/VCSEL) in the US and Singapore is up more than 3x YoY.

The XPV platform is the item in this quarter's financial structure that most deserves attention. In June, Broadcom formed the AI XPV Platform with Apollo and Blackstone, aiming to facilitate more than 20 GW of compute infrastructure for OpenAI and Anthropic by the end of 2028; the first $35B tranche closed in June, for Anthropic's 1 GW deployment. The CFO made three things clear:

1. Broadcom does not fund it itself; third-party financial partners independently underwrite and fund it;

2. When necessary, Broadcom provides a "modest residual value guarantee": this is a contingent liability, which management views as low risk;

3. The terms of each future tranche will differ and be evaluated case by case; no numbers were updated today.

An analyst asked whether the maximum exposure on the first tranche disclosed in last quarter's 10-Q (about $29B) could serve as a per-GW reference for future tranches; the CFO explicitly declined to answer. Hock added some context: of the six customers, only Anthropic and OpenAI need this structure, while the other four can fund themselves; and Anthropic is "already on the road to an IPO, at which point its credit rating will change".

Hock listed the real bottlenecks himself. In order: land, power and shell (LPS, with the longest lead time, directly determining the schedule); leading-edge wafers; substrates; HBM; and AI server system memory, which Broadcom doesn't supply but customers must secure themselves. His remark was self-deprecating: "that's also why I'm somewhat glad I only have to deal with six customers".

The six-customer model is a double-edged sword. Concentration brings depth of collaboration (doing LPS analysis together, scheduling tape-outs together), but it is also the single biggest risk. The most direct peer comparison is Marvell, which this quarter set its FY2028 target at $18B, with the biggest upgrade coming from scale-up optics: a different path, but likewise betting on a handful of custom customers. For details see Earnings Call Highlights: Marvell (MRVL) | Q2 FY2027 — FY2028 target jumps to $18B in one go.

The software side is the quiet but steady other half: VMware Private AI Cloud was announced, and VCF lets enterprises move workloads from public cloud back to private cloud. $8.75B of quarterly revenue, 94% gross margin, 84% operating margin: this business is paying for the capex and gross margin dilution of AI semiconductors, though no one asked about it on the call.

6. Risks and counterarguments: what has to go right for this guidance to hold

• 30 GW ≠ 30 GW of revenue. Hock said explicitly it will be less than 30 GW, but not by how much. $115B/$230B are numbers they have "judged", and that judgment assumes LPS arrives on time.

• Gross margin is a one-way street. Q4 guidance is 73% vs. 78% a year ago. Memory content will only rise, so this line won't reverse in the near term; profit improvement rests entirely on operating leverage.

• The two-year guidance won't be updated quarterly. That means for the next three to four quarters the market has no official calibration point, and any deviation can only be inferred by dissecting quarterly numbers.

• The residual value guarantee is a contingent liability. The amount is undisclosed and varies case by case. It creates an indirect link between Broadcom's balance sheet and two AI labs that are not yet consistently profitable, even if management assesses it as low risk.

• Non-AI semiconductors are only up 5%. Broadband and server storage growth was offset by wireless. This business can no longer provide a meaningful buffer when the AI cycle swings.

• Two versions of $/GW coexist. $11–12B and $20–30B both appeared in the same call, and the basis was never clarified.

7. Conclusion

What Broadcom said this quarter is actually very simple: when you co-design a chip for a specific LLM workload, performance wins and the cost is half, and now six companies want to do that at the same time. The financials turned that statement into $16.7B of quarterly AI revenue, a 56% revenue share, and a trajectory that doubles twice in two years.

But the most valuable line of the call wasn't any of those numbers; it was the one Hock volunteered: "Even if we ship the chips, will they be deployed on time?" Broadcom has already controlled everything it can control: its own substrates, 3x laser capacity expansion, and supply locked in through FY2028. The remaining variables are concrete, substations and grid permits, and those lie outside any semiconductor company's circle of competence.

So over the next three quarters, these three indicators are worth tracking more than revenue:

1. Whether AI networking can hold its 27% share of AI revenue. Management says networking will grow as fast as XPUs, but Q3's 73/27 split already reflects a big XPU ramp; if this ratio falls, it means networking attach isn't keeping up.

2. Where gross margin lands after Q4. 73% is guidance; the key is whether memory dilution in FY2027 Q1 exceeds what operating leverage can absorb.

3. Whether the Singapore substrate plant starts production on time in FY2027, and the terms of subsequent XPV tranches. The former determines whether supply can rise further; the latter determines what shape Broadcom's contingent liabilities will take.

In one sentence: Broadcom has drawn its FY2028 revenue very clearly, but has honestly left the biggest uncertainty outside of semiconductors.