Two facts to hold side-by-side. Broadcom is the AI silicon company with the deepest dominance after Nvidia — Google’s TPU partner, the dominant data-centre switch chip (Tomahawk 6 at 1.2 T per port), one of two real EML (electro-absorption laser) suppliers worldwide, and roughly 90% of the high-speed pluggable-DSP market outside the Nvidia ecosystem. By volume and by surface area, no one else is close.
And: Broadcom is also the company facing the most credible incumbent-disruption pressure in the sector right now. Three vectors at once — MediaTek peeling off a slice of Google’s TPU work, Marvell winning AWS Trainium and a Google FFN socket back from Broadcom, and a CPO (co-packaged optics) transition that would erase a meaningful share of Broadcom’s own pluggable-module business.
This piece walks both halves: the four-pillar empire as it stands, then the three pressure vectors against it.
Companion pieces: Where Marvell Sits in the AI Optical Stack (the chips just upstream of these), What’s Cooling the AI Build-Out (the cage layer that holds them), and What AI Hardware Is Running Out Of (the broader scarcity map).
Pillar 1 — Custom ASIC: the Google base, the OpenAI bet
The fastest-growing slice of Broadcom’s AI revenue today is custom ASIC design — building bespoke accelerators for one hyperscaler at a time. Two customers carry it.
Google. Broadcom has been Google’s TPU design partner across generations and is still the partner on the current TPU v7 (training) and the TPU v8 family (training + inference variants). The division of labour is: Broadcom owns the front-end and back-end physical design, including the high-speed SerDes; TSMC owns the fab and the package; Google keeps a large portion of architectural design in-house. It’s the most stable revenue stream Broadcom has in AI — the base under everything else.
OpenAI. The newer, louder line item. Broadcom is the exclusive design partner for OpenAI’s in-house training ASIC — OpenAI’s attempt to build silicon supply independent of Nvidia. Program scale, as currently guided: about 2,000 racks, with rack-level cost around $3 M, putting the total project somewhere in the $6–8 B range. The chip is currently in NPI (new product introduction); volume ramp starts 2H 2026.
The OpenAI line is what the market is repricing right now — more on that below.
Pillar 2 — Switch silicon: the moat that still holds
Broadcom’s Tomahawk and Trident families are the data-centre Ethernet switch chips. They’re inside almost every brand-name switch the industry buys — Cisco, Arista, Huawei, H3C. The current generation, Tomahawk 6, delivers up to 1.2 T per port and is the hardware backbone that makes 1.6T and 3.2T optical-module deployments economic.
Marvell is the credible challenger here too, but Broadcom’s first-mover share and incumbent vendor relationships keep the lead large. Of the four pillars, this is the one where Broadcom’s position looks most secure on a 2–3 year horizon.
Pillar 3 — Optical chips: EML and DSP, the bottleneck duopoly
On the optical-module side, Broadcom sits at the two most-strained points in the supply chain.
EML (electro-absorption modulated laser) chips are the light source inside 800G/1.6T modules. Globally, the two real volume suppliers are Lumentum and Broadcom. Broadcom’s EML production is stuck on a mix of 2- and 3-inch wafers — the metro of capacity extension is small and slow. The 200 G EML grade (the part 1.6T modules need eight of each) is the sharpest shortage, with lead times that stretch to a year. Eoptolink (新易盛) alone has booked more than half of Broadcom’s EML capacity; Innolight and other top module makers take the rest.
DSP (digital signal processor) chips — the brain of a high-speed optical module — are the other socket. Outside the Nvidia ecosystem (i.e. CSP-spec pluggables and the broader long tail), Broadcom holds approximately 90% share. Marvell is the live competitor, but the consensus on both power consumption and link performance leans toward Broadcom at the current generation. The bottleneck on this socket is TSMC’s 3 nm capacity, not design — Broadcom and Marvell compete for the same overbooked allocation, and what each can ship at 1.6T is mostly a function of how much 3 nm wafer-start they pulled.
Pillar 4 — CPO: the catch-22
The fourth pillar is the one Broadcom is publicly pushing hardest and quietly making the slowest progress on. CPO (co-packaged optics) collapses the optical engine onto the same substrate as the switch ASIC — no pluggable module, no cage. Broadcom’s technical bet is a Mach–Zehnder modulator route inside an open ecosystem — the architectural opposite of Nvidia’s closed, vertically integrated CPO stack.
The publicly indicated volume window is 2H 2027 to 2028, which is six to twelve months behind Nvidia’s plan. Today the work is at prototype and pilot scale, no formal shipments. The pacing problem is partly capacity: CPO at the latest node needs the same TSMC packaging slots Nvidia is consuming first, and Broadcom’s priority inside TSMC is materially below Nvidia’s.
The deeper problem isn’t timing — it’s structural. Broadcom is the EML and DSP supplier whose products CPO is built to replace. Every successful CPO deployment is one fewer pluggable module with a Broadcom EML inside and a Broadcom DSP next to it. Push CPO hard, you cannibalize your most-profitable legacy product. Push it slow, you cede the next generation to Nvidia. The observable behaviour is closer to follower-with-hedge than to disruptor-with-conviction — which is exactly what the financial conflict predicts.
Pressure 1 — MediaTek peels off Google inference
For most of the TPU’s history Broadcom was, in practice, the only design partner Google considered. That ended in 2025–2026. Google is actively diversifying its TPU supply chain and has shifted part of its inference-class TPU work — specifically the V8E series — to MediaTek.
MediaTek’s case is the textbook one: lower price, more flexible customisation service. Google’s case is also textbook: it wants to stop paying what its procurement team calls the “Broadcom premium,” and it wants to be optionable across vendors. The market read the news as exactly what it is — the first dent in the assumption that Broadcom’s position inside Google’s TPU program was structurally unassailable. Training (v7, v8) is still Broadcom; inference (v8E) is partially MediaTek. The question for 2027 is whether the next TPU’s training tier stays with Broadcom or whether the MediaTek wedge widens.
Pressure 2 — Marvell takes Trainium and a Google FFN socket
The second pressure vector is from the company you’d most expect.
On the AWS side, Marvell has won the Trainium 2/3 ASIC design service contract from Broadcom. AWS’s in-house accelerator program now flows through Marvell, not Broadcom — a direct head-to-head loss in the same design-services market Broadcom has used as its growth proxy.
On the Google side, Google has hedged by bringing Marvell in as the designer of a chip targeted at a specific compute pattern — the FFN (feed-forward network) workloads at the centre of modern inference graphs. Marvell here is “strategic backup,” not lead, but the pattern is identical to the MediaTek arrangement above: Google is making sure no single design partner can hold its roadmap hostage.
The two news items together rewrite the AI ASIC design-services market from “Broadcom and everybody else” to “Broadcom and Marvell, with MediaTek on the inference flank.”
Pressure 3 — OpenAI’s conservative 2026 guidance
The most immediate financial pressure is OpenAI itself. Broadcom flagged OpenAI as a top-five customer; the market priced in a steep ramp on the back of the in-house ASIC program. The 2026 unit guidance the program is currently carrying has come in well below that expectation — conservative enough that consensus near-term revenue estimates for Broadcom were pulled down on the same news cycle.
Two things behind the guidance number.
Scale relative to Nvidia is small. OpenAI’s Broadcom program is $6–8 B total. OpenAI’s Nvidia spend is reported on the order of $30 B per year. Broadcom is one supplier in a portfolio, not the centre of it.
OpenAI is deliberately multi-vendoring. Its compute buildout now spans Nvidia, AMD, Broadcom, and Arm-based options, plus its own custom ASIC. That diversification is sensible product strategy and it also directly weakens Broadcom’s pricing power — OpenAI can credibly shift any single line. Broadcom’s upside on the program is bounded by how much of OpenAI’s aggregate compute it can plausibly capture, and the answer is currently “a slice, not the whole.”
The shape
- Strengths. Switch-chip dominance (Tomahawk 6, 1.2 T/port); the deep TSMC manufacturing relationship; EML + DSP near-bottleneck position; the Google TPU base.
- Live pressures. MediaTek on inference TPU (V8E); Marvell on AWS Trainium + Google FFN; OpenAI conservative on 2026 ASIC volumes; multi-vendor sourcing reducing pricing power across the board.
- The catch-22. CPO progress is materially behind Nvidia, partly because of TSMC priority and partly because every CPO win erodes Broadcom’s own EML + DSP + pluggable-module book.
What to watch over the next two quarters:
- CPO execution. Any 2H 2026 milestone on the Mach– Zehnder open-CPO route would change the catch-22 calculus — or, if missed, cement the “follower” reading.
- Google TPU retention. Does Google’s next training generation stay with Broadcom, or does the MediaTek wedge widen from inference into training?
- OpenAI iteration. If real demand for the in-house training ASIC outruns the conservative current guidance, that single line item can move the entire near-term P&L print.
One-line read on Broadcom’s AI position: still the most-dominant non-Nvidia silicon vendor, with the deepest customer base in the sector — and the most credible disruption stack against it on a 12–18 month horizon. The empire is intact; the dent count is rising.
Figures here are drawn from current supply-chain research (industry Q&A notes and sell-side reports through mid-2026) and stated as the thesis, not independently re-derived.
Adjacent reading: Where Marvell Sits in the AI Optical Stack (the other half of the duopoly); What’s Cooling the AI Build-Out (the cage layer above the EML / DSP sockets); What AI Hardware Is Running Out Of (the broader scarcity map — including the 200G EML cliff and Coherent / Faraday rotator monopoly Broadcom doesn’t touch).