Broadcom is the AI silicon company with the deepest dominance after Nvidia. Four pillars hold up the empire. Three pressure vectors are denting it at the same time.
↓pillar 1 — custom ASIC
four pillars · three pressure vectors
partner · v7 + v8
1.2 T/port
worldwide
non-Nvidia
Google. Broadcom is still the design partner on TPU v7 (training) and the TPU v8 family (training + inference). Broadcom owns the physical design and SerDes; TSMC owns the fab; Google keeps the architecture in-house.
blueprint into real silicon.
OpenAI. Exclusive design partner on OpenAI’s in-house training ASIC. Program size: ~2,000 racks, $3 M per rack, total $6–8 B. NPI today; volume ramps 2H 2026.
the chip before
volume ships.
→switch silicon
Tomahawk and Trident are inside almost every brand-name switch — Cisco, Arista, Huawei, H3C. Current generation: Tomahawk 6 at 1.2 T per port.
inside every
data-centre box.
Of the four pillars this is the one that looks most secure on a 2–3 year horizon. Marvell is the credible challenger but the share gap is wide.
⚡pillar 3 — optical chips
EML (electro-absorption laser). The light source inside 800G / 1.6T modules. Only two real suppliers worldwide: Lumentum and Broadcom. 200 G EML lead times stretch a year. Eoptolink alone has booked >50% of Broadcom’s capacity.
blinks 200 G of
data into fiber.
DSP (digital signal processor). The brain of an optical module. Outside Nvidia, Broadcom holds ~90%. The cap is TSMC’s 3 nm allocation — Broadcom and Marvell compete for the same overbooked wafers.
reads the blinks
and turns them
back into 1s and 0s.
⌖pillar 4 — CPO, the catch-22
CPO (co-packaged optics) collapses the optical engine onto the switch ASIC. Broadcom’s bet is a Mach–Zehnder route inside an open ecosystem — the opposite of Nvidia’s closed stack.
moves inside
the chip package.
Volume window: 2H 2027 to 2028 — about 6–12 months behind Nvidia. Today: prototype and pilot only, no shipments.
The deeper problem is structural. Broadcom is the EML and DSP supplier whose products CPO is built to replace. Push CPO hard, cannibalise the legacy book. Push it slow, cede the next generation to Nvidia.
kills a Broadcom
EML + DSP sale.
※three pressure vectors
MediaTek peels off Google inference. Google has shifted the V8E inference TPU work to MediaTek. Training (v7, v8) still Broadcom; inference partly MediaTek. The Broadcom premium is what Google wants to stop paying.
TPU monopoly.
Marvell takes Trainium and a Google FFN socket. Marvell now owns AWS’s Trainium 2/3 design-service contract — direct loss to Broadcom. Google also brought Marvell in as a strategic backup for FFN (feed-forward network) workloads.
is now two firms,
not one.
OpenAI conservative on 2026. The Broadcom program is $6–8 B total. OpenAI’s Nvidia spend runs ~$30 B/yr. Broadcom is one line in a portfolio — pricing power is bounded.
— Nvidia, AMD, Arm,
Broadcom.
↓one-line read
Still the most-dominant non-Nvidia silicon vendor with the deepest customer base — and the most credible disruption stack against it on a 12–18 month horizon. The empire is intact; the dent count is rising.