AVGO (Trans): AI to add 30GW over two years; Anthropic accounts for half.

Below is Dolphin Research's compiled Trans.$Broadcom(AVGO.US) for Broadcom's FY26 Q3 earnings call

Earnings take: Google wavers; Broadcom's promises don't fill the gap

I. Key takeaways

1. Capital returns and debt

Q3 cash dividends paid were $3.1bn, implying a quarterly common dividend of $0.65 per share. Over the same period, the company repaid $5.6bn of long-term debt, and a further $1.5bn of senior notes after quarter-end.

Fixed-rate debt principal stood at $59.6bn at quarter-end, with a 4% WAC and a 7.4-year weighted avg. remaining tenor.

2. Outlook: one-and-done AI revenue path through 2028

a. Q4 total: consolidated revenue of Approx. $34.8bn (+93% YoY). Semis of Approx. $26.1bn (+136% YoY), of which AI semis are Approx. $21.7bn (+236% YoY); infra software of Approx. $8.7bn (+25% YoY).

b. Q4 margins and tax: consolidated GPM of Approx. 73% (vs. 78% a year ago), OPM of Approx. 66% (flat YoY). Non-GAAP tax rate of Approx. 16%, driven by the global minimum tax and profit mix changes vs. FY25.

c. Q4 diluted shares of Approx. 4.94bn. Capex of $1.4bn, directed to semi capacity adds.

d. Cross-year AI guide: FY26 revenue of $58.0bn (+186% YoY), above the prior $56.0bn guide; FY27 Approx. $115.0bn; FY28 Approx. $230.0bn, totaling Approx. $350.0bn over FY27–FY28. EPS in FY28 is expected to top $30.

e. This multi-year view is for trajectory only. Management will not update it quarterly.

3. Quarterly financials

a. Total: revenue of $29.6bn (+86% YoY). OP of $20.1bn (+92% YoY), OPM 67.9% (+240bps YoY); non-GAAP EPS of $3.32 (+96% YoY).

b. Consolidated GPM was 75%, down 210bps QoQ as AI semis mixed higher in revenue. It still beat the prior 74% guide.

c. Semis: revenue of $20.8bn (+127% YoY), 70% of total. AI semis were 56% of total revenue, up from 49% in Q2.

d. Segment GPM was Approx. 67%, OPM 61% (+440bps YoY). Opex was $1.2bn, 6% of segment revenue, up 22% YoY.

e. Infra software revenue was $8.8bn (+29% YoY), 30% of total. GPM was 94%, with opex over $0.9bn; OPM was Approx. 84% (+650bps YoY).

f. Cash and balance sheet: FCF was $13.7bn, 46% of revenue. Capex was $532mn. Cash ended at $24.0bn, up $4.3bn QoQ from $19.6bn; inventory was $4.5bn.

4. XPV platform: moving customer financing off balance sheet

In Jun, AVGO set up the AI XPV platform with Apollo and Blackstone, targeting over 20GW of compute infra for OpenAI and Anthropic by end-2028. The first $35bn tranche closed in Jun, tied to Anthropic's 1GW buildout, which is underway.

Third-party financial partners underwrite and fund the platform; Broadcom does not lend directly. Where needed, it may provide limited residual value backstops, which management classifies as low-risk contingencies given the labs' profit trajectories and asset value retention.

II. Call details

2.1 Prepared remarks

1. AI semis: Q3 deliveries and Q4 cadence

a. Q3 AI semis revenue rose 221% YoY and 54% QoQ. XPU shipments were up over 3.5x YoY, accounting for 73% of AI revenue, while AI networking grew over 2.5x YoY.

b. The company shipped Ironwood (TPU v7) in volume to Anthropic and Google, began volume production shipments of TPU v8i to Google, and delivered OpenAI's first-gen custom accelerator, Jalapeno.

c. vs. Ironwood, TPU v8i adds memory and bandwidth, and is tuned for inference. Performance is on par with or better than the Vera Rubin GPU; Broadcom says its shipping started ahead of the earlier-initiated MediaTek version.

d. According to OpenAI last week, Jalapeno beats Grace Blackwell Ultra on latency, throughput, and power. On OpenAI workloads, it rivals the Vera Rubin GPU at roughly half the GPU cost.

e. Q4 cadence: accelerated Ironwood to Anthropic, ramping TPU v8i to Google, continued Jalapeno deliveries to OpenAI. Meta's custom MTIA for inference and large-scale recommendations enters volume shipments.

f. Both XPU and AI networking revenue are expected to be up roughly 3x YoY in Q4.

2. Roadmaps for the four major XPU customers

a. Google: a new long-term agreement covers future TPU and AI networking generations, with annual multi-ten-bn deliveries over the next few years. Google becomes the largest XPU customer in 2027 and remains so in 2028, with '28–'29 demand carried by subsequent TPU generations in development.

b. Anthropic: deploying 1GW of Ironwood in 2026, then 5GW of TPU v8i in 2027. There is clear visibility for another 10GW in 2028.

c. OpenAI: Jalapeno plans to deploy 1.3GW in 2027. In 2028, Jalapeno and successors exceed 5GW, making OpenAI the No. 2 XPU customer; the next-gen XPU is near tape-out, and a third gen is in development.

d. Meta: three MTIA generations will ship from now to end-2027. Combined, they have 3GW of deployment visibility by 2028.

e. Moat: SerDes, die-to-die interconnect, leading-edge HBM and SRAM integration, differentiated advanced packaging, and the fastest path from product definition to mass production without respins.

3. AI networking and optical interconnect

a. Tomahawk 6 (100Tbps) launched first for both scale-up and scale-out Ethernet switching. Tomahawk 7 has taped out, the industry's first 200Tbps Ethernet switch chip.

b. The company maintains generation-to-generation leadership in PCIe switching on the scale-up side. Optical DSPs are cutting-edge, and EML, VCSEL, and CW laser capacity is ramping quickly.

c. Management expects AI networking revenue growth to remain as fast in coming years, and will invest heavily to sustain the broadest, leading AI portfolio.

4. Non-AI semis and infra software

a. Q3 non-AI semis were $4.2bn, +5% YoY and flat QoQ. Strength in broadband and server storage was partly offset by weakness in wireless; Q4 is guided to Approx. $4.3bn, a modest QoQ increase.

b. Software ARR grew 15% YoY. VMware Private AI Cloud launched, bundling AI infra, security/compliance, and tools to build/operate trusted AI agents so enterprises can run AI alongside existing apps at low cost while protecting their data.

c. VCF is helping customers repatriate workloads from public cloud to private cloud, for greater control and meaningfully better infra economics. Management views enterprise AI consumption as incremental to software.

2.2 Q&A

Q: Where are the supply bottlenecks for next year's doubling outlook? What could push the number higher?

A: $115.0bn is a conservative view on locked-in supply; demand could support more shipments. The company believes it could ship materially more on the demand side but is deliberately conservative, focusing on whether customers can deploy on time after receiving chips.

$115.0bn corresponds to supply already secured, and the number could go up if the supply chain expands. The FY28 $230.0bn follows the same logic, based on the six customers' confirmed 2028 DC site capacity and Broadcom's secured leading-node wafers, substrates, and HBM.

Q: How are substrate and other bottlenecks addressed? Will Singapore capacity be used for XPUs?

A: In-house substrate capacity in Singapore will ramp from FY27, removing a key bottleneck. Management said the Singapore fab will start producing substrates in FY27, covering a critical choke point. No further details were provided on other links.

Q: Does the 10GW mentioned refer to Google plus Anthropic, or Anthropic only?

A: 10GW refers to Anthropic only, excluding Google. Management confirmed directly, without further elaboration.

Q: How many of the six XPU customers use Tomahawk for scale-out? What is Tomahawk Ultra adoption?

A: Tomahawk 6 is deployed at almost all hyperscalers co-developing XPUs with Broadcom. It comes in 100G and 200G SerDes versions, ramping first in scale-out; both versions have landed. Non-Broadcom XPU customers also use it, and management did not provide sell-through updates.

Tomahawk Ultra uses low-latency Ethernet for scale-up. Adoption is ahead of the company's expectations, with deployments starting this quarter and a scale-up ramp in FY27.

Management added it is the first time scale-up within rack-level GPU/XPU clusters runs on Ethernet, with packet loss and latency no worse than prior non-Ethernet approaches.

Q: Adding ~10GW in '27 and ~20GW in '28 to get 30GW, is that the right lens?

A: 30GW is valid as a sum, but actual production over two years will be below that. Management confirmed the sum across the six customers.

However, deployment is more than shipping chips; DC shells must be in place and production-ready. On a fiscal-year basis, management conservatively expects realized production below 30GW.

Thus, $115.0bn in FY27 and $230.0bn in FY28, Approx. $350.0bn combined, are discounted numbers. What the company has high confidence in is the $350.0bn shipment value, not that all 30GW lands in-period.

Q: Back-solving implies $11–12bn revenue per GW, while competitors cite closer to $40bn. Is that reasonable?

A: No direct comment on the gap; XPU cost is less than half that of GPUs. Management said custom XPUs are optimized for each customer's LLM workloads, matching or exceeding GPU performance at 'less than half' the cost. It did not provide an official per-GW revenue metric.

Q: What is the off-balance-sheet risk cap under backstop agreements? The last 10-Q showed a max exposure of Approx. $29bn for the first tranche; can that be a per-GW proxy?

A: No total cap disclosed; prior figures remain valid with nothing new today. Management reiterated there is nothing new to announce on residual value guarantees or backstops, and previously disclosed numbers still hold.

Future strategic financings will be evaluated site by site and tranche by tranche, with terms tailored to each lab and investor. There will be no single exposure cap framework, and disclosures will come at appropriate times.

Q: How will XPU mix and rising memory costs impact GPM in FY27–FY28?

A: Only quarterly margin guidance; Q3 semi GPM corrected to Approx. 67%. Management corrected a prior misspeak and did not guide FY27–FY28 margins.

Margins depend on the semi vs. software revenue mix, product mix within semis, and rising memory content. A higher XPU mix weighs on GPM, but the company guides only one quarter at a time.

Management added investors should focus less on GPM and more on OPM. Revenue is outgrowing the opex needed to support it, providing operating leverage to keep OPM stable even as mix dilutes GPM.

Q: Two frontier labs are set to become the largest AI customers while Google still cites tight supply; these labs rely on CSPs for land, power, and funding. Do they really control chip choices?

A: Broadcom provides financing vehicles only for Anthropic and OpenAI; the other four self-fund. Management said the other four are financially capable of self-funding, which Broadcom welcomes.

The unit economics are that each 1GW can drive Approx. $3.0bn in annualized ARR for these labs. Helping them scale is a good investment for Broadcom and aligns with its own economics.

Management acknowledged that near term these two deploy models on third-party cloud, but expects them to look like hyperscalers over time. They will build and operate their own DCs and offer GenAI APIs and model services as first parties, and this shift 'is not hypothetical; it is underway.'

Q: How should we think about Tomahawk Ultra's attach rate to Broadcom ASICs?

A: No attach-rate disclosed; customers split between Tomahawk 6 and Ultra. Among co-developed XPU customers on the scale-up side, some upgraded from Tomahawk 5 to 6, while others moved to Tomahawk Ultra.

Because it is Ethernet-based, open, and standards-driven, anyone can connect. Deployments are visible across XPU clusters and some GPU clusters, but the company did not provide attach-rate figures.

Q: How much sensitivity is incorporated for land, power, and DC shell constraints?

A: The outlook already discounts for customers' actual land/power/site progress, not just chip demand. Management said it works with each customer to verify land, power, sites, and system timelines, not just chip or rack demand.

These are long-cycle construction projects that directly determine when capacity is usable. The analysis is embedded in the outlook provided.

Q: Will XPV cover most of Anthropic's 10GW and OpenAI's 5GW financing? What are the timing and hurdles?

A: No coverage ratio commitment; evaluated site by site. Management noted the two customers' credit profiles will evolve over the next two years — Anthropic is progressing toward IPO with improving investment-grade prospects, while OpenAI's path differs with lower visibility.

The CFO added customers tap multiple funding sources. Broadcom may participate only when it fits its framework and economics, and it may choose not to; there is no blanket commitment.

Q: Beyond land and power, what other supply constraints exist, and can they be eased?

A: Substrates and HBM are two more bottlenecks, and different constraints rotate over time. Management described a multi-dimensional issue: land, power, and DC shells not only raise concerns but set the timing of usable capacity.

The second layer is leading-node silicon output and delivery timing. Next are substrates — the reason for building scaled in-house substrate capacity in Singapore.

There is also HBM, and system memory beyond HBM for AI servers; Broadcom does not supply the latter, but customers must lock it in. Constraints will each become the bottleneck at different times, and management is 'glad to have only six customers to handle.'

Q: How will per-GW dollar content evolve for future XPU generations?

A: Per-GW content stays at $20–30bn; growth comes from more GW. Each XPU/GPU generation improves performance and moves to more advanced nodes, lifting per-die ASPs.

But per-die power also rises, meaning fewer next-gen XPUs fit into 1GW. Net, per-GW dollars remain relatively stable at $20–30bn and are expected to hold, which is above the $11–12bn implied by $350bn over 30GW.

Incremental growth comes from accelerating total GW to meet exponential compute demand.

Q: Rising capex — is it to expand EML, CW, and indium phosphide capacity?

A: US and Singapore capacity for InP, EML, and CW has more than tripled YoY. Management said capex continues to fund in-house plants: substrate lines are about to start up, and US/Singapore EML, CW, and InP fabs have expanded to over 3x YoY. The expansion was completed this year, with significant adds over the next two years, driving higher capex.

Industry demand for EML and CW lasers is clearly outpacing supply. Broadcom has substantial share and is doubling down on capacity to support ecosystem growth.

<End>

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosure