N. Baratte - TechStock01

N. Baratte - TechStock01

Share this post

N. Baratte - TechStock01
N. Baratte - TechStock01
Broadcom results good. 55-60% AI growth in 2025, same in 2026. Consensus forecasts for 2026 are ~10% too low, but the stock is pricey

Broadcom results good. 55-60% AI growth in 2025, same in 2026. Consensus forecasts for 2026 are ~10% too low, but the stock is pricey

AVGO is more expensive than NVDA

Nicolas Baratte - TechStock01's avatar
Nicolas Baratte - TechStock01
Jun 06, 2025
∙ Paid
1

Share this post

N. Baratte - TechStock01
N. Baratte - TechStock01
Broadcom results good. 55-60% AI growth in 2025, same in 2026. Consensus forecasts for 2026 are ~10% too low, but the stock is pricey
1
Share
  • 2025 AI revenue ~US$19bn, up 55-60% YoY. Similar growth in 2026 or $~30bn. Mngt declined to update 2027 addressable market (SAM), previously stated at $60-90bn. That looks too high.

  • Maybe a risk of disappointment for 2027? No, because Consensus didn’t factor that in and is too low. My estimate for 2026 revenue is $79bn versus consensus at $73bn, $95bn in ’27 vs Consensus $83.

  • In general, Consensus for AI stocks is too low but AVGO stock is expensive, trading at 33x FY26 Consensus, most likely lower ~30x, and ~25x 2027 if I’m correct. That’s more expensive than NVDA. Why bother given the semi-atrocious transparency

May-25 (2Q25) results and Aug-25 (3Q25) guidance in line

Nothing much to say here.

AI semiconductors

Broadcom bundles together the AI Accelerators XPU and the Networking products (Tomahawk, Jericho, NIC) that goes into hyperscalers data centers.

  • Revenues 2Q25: US$4.4bn up 40% YoY 60% Accelerators, 40% Network

  • Revenues 3Q25: US$5.1bn up 60% YoY

Share

CEO Mr. Tan mentioned:

  • 3 customers and 4 prospects for XPU Accelerators.

  • Customers are:

    • Google TPU should be very dominant in Broadcom revenues

    • Meta MTIA chip production should be just starting now

    • I do not know if design is completed for Apple and Open AI

  • Prospects: or Apple or Open AI are still in the “prospect” category.

  • Clear statement that XPU revenues are now extending to Inference

  • Mr Tan: “we expect at least 3 customers to deploy 1m clusters in 2027 for model training, of which a significant portion will be XPU”. This line remains unclear to me: how much of that will be Accelerators?

  • Customers are “doubling down on inference, could accelerate demand in 2H26 for inference” and as a result “FY25 growth in AI Semi revenue will sustain into 2026”. This means ~55-60% YoY growth.

  • From other comments, it sounds like Broadcom is still expecting a ~30% long-term Cagr after 2026.

  • In previous calls, Mr. Tan was affirmative about his 2027 Serviceable Addressable Market: total market opportunity for XPUs and AI networking components at US$60 billion to US$90 billion by 2027. Today, Mr. Tan refused to provide and update “stop talking about the SAM now”, “we don’t intend” to provide any further update “probably until 2026”. But “intend to give more visibility in 2026”.

Switch: the just announced Tomahawk 6

N. Baratte - TechStock01 is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Broadcom launched its new Tomahawk chip a few days ago, and it’s a major breakthru that Data Centers need with a bandwidth of 102.4Tbps, twice the bandwidth of current Tomahawk 5. The price also doubles to ~$20k per chip. This is specifically required by AI training clusters. Broadcom mentioned that Tomahawk 6 will accelerate ethernet from 800G to 1.6T.

This places Broadcom 1 generation beyond Cisco Silicon One, Nvidia Spectrum-X, Marvell Teralynx, and Xsight Labs X3 and X4.

Taiwan beneficiaries:

  • Accton 2345.TW (switch ODM)

  • PCL Tech 4977.TW (high-speed SerDes interface)

  • ShunSin 6451.TW (optical transceiver module)

  • Elite Materials 2383.TW (high-frequency, high-speed CCL materials)

Interesting discussions regarding ethernet switch versus the NVLink (Nvidia) and UALink (Intel, Microsoft). Is Broadcom an ecosystem of one? Mr. Tan’s answer is factually correct: Ethernet-based scale up is the existing ethernet protocol, its open standard and open source ethernet, no need to invent something new. The new protocols are interesting but they are proprietary.

Non AI semiconductors

  • revenue close to bottom, slow recovery

  • 3Q: revenue around 4bn, ie similar to 2Q25.

  • No indication for 2026

Infrastructure Software

  • 2Q: 25% YoY growth organically

  • 3Q: guidance implies 16% YoY growth

Consensus forecasts and Valuations

Here are my estimates. They imply that Consensus is seriously too low.

2025:

  • AI 55-60% growth (we already know 1-3Q25)

  • Non-AI -6% decline (we already know 1-3Q25)

  • Software 26% growth, but with acquisition base effect, 3Q at 16%

2026:

  • AI 55-60% growth as guided by mngt

  • Non-AI let’s assume a recovery at +10% YoY, after a very severe -30% decline over the previous 2 years

  • Software let’s assume that 3Q growth rate at 16% continues to normalize to 12%

Comparing revenues

This post is for paid subscribers

Already a paid subscriber? Sign in
© 2025 Nicolas Baratte
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share