Broadcom Inc. "FQ2:26 (Apr) Preview: Raising PT, Adjusting Estimates [Revised]"

研报英文原文证据摘录

Broadcom Inc. "FQ2:26 (Apr) Preview: Raising PT, Adjusting Estimates [Revised]"

.05 0 2.05

now expect AVGO's Anthropic revenues per GW to track closely to the GCP

Q2E 2.38 2.38 0 2.42

TPU revenues per GW, which we estimate is closer to the bottom of the $10-20B/

Q3E 3.30 2.80 -15 3.26

GW range. AVGO suggests the revenues vary by customer and in light of GCP's Q4E 4.41 3.36 -24 3.85

use of OCS-based scale-up networking solution, we see the AI networking 10/26E 12.16 10.60 -13 11.33

opportunity for AVGO largely limited to scale-out Ethernet portfolio (Tomahawk). 10/27E 22.56 21.14 -6 18.05

10/28E 28.59 28.59 0 22.29

The latest GCP announcement of v8i and v8t chips leads us to believe that AVGO is

the supplier for v8i (our note: Thoughts On Google’s TPU v8 Generation), Timothy Arcuri

although our conversation with AVGO suggested its roadmap is capable of both Analyst

training and inferencing, which would suggest GCP is de-risking the MediaTek v8t timothy.arcuri@ubs.com

ramp with optionality for training deployment. Interestingly, TPU v8i features both +1-415-352 5676

larger HBM (288GB vs 192GB in Ironwood) as well as 384MB of on-chip SRAM (up Natalia Winkler, CFA

from 64MB on Ironwood) and compared with 500MB in the latest Groq L3 chip. Analyst

natalia.winkler@ubs.com

We are revising our AVGO TPU units to ~3.7MM in C2026 and ~6.8MM in C2027 +1-415-352 4626

(from ~4.6MM/~7.3MM prior).

Grant Joslin

META announced a multi-year partnership with AVGO and we understand it will Analyst

include multiple inference chips. We continue to model META's XPU program grant.joslin@ubs.com

+1-201-352 0442 scaling from ~$800MM in F2025 to $3B in F2026, $8B in F2027 and $13B in

F2028 - suggesting the 1st GW (at ~$20B/GW AVGO revs) deployed over ~2yrs,