● dense/○ explainer
aileena·machina
markets · 2026 · jun

Marvell owns one chip in the optical stack.

The TIA. 80% of 800G single-mode, the only 1.6T volume player today. Above it: a DSP slugfest with Broadcom.

— tldr —

owns TIA. fights Broadcom on DSP. absent from Google’s 2.4T coherent and the Nvidia–TSMC 3.2T CPO main line.

Marvell’s AI story is one specific layer of the optical supply chain — the chips between the photon and the data path — and how it owns, shares, or has been locked out of each box at that layer.

what a TIA is

fiberlaser · EMLPDTIADSP→TIA — MarvellDSPASICswitch ASIC
1
2
3
4
5

the optical receive path · top to bottom

① fiberabsent from 2.4T
Google coherent
② laser · EMLthe upstream
choke point
③ Marvell owns this80% 800G single-mode · only 1.6T volume player
④ DSP fightBroadcom holds
~90% non-Nvidia
⑤ switch ASICabsent from NVIDIA-TSMC
3.2T CPO
↓ direction of light
co-architects NPO w/ AWS ↗

The fiber drops light onto a PD (photodetector); the PD turns it into microamps. The TIA (transimpedance amplifier) boosts that signal and hands voltage to the DSP. PD → TIA → DSP.

in plain wordstiny preamp
between the
photodetector
and the brain.

Small, deeply embedded, hard to second-source. That’s why winning the TIA socket is sticky — and Marvell has won it.

the TIA monopoly

800G single-mode TIA: Marvell ~80%, Semtech ~20%. 800G multi-mode TIA: Macom ~50%, Marvell + Semtech split the rest. Almost all of the 800G TIA growth is single-mode — where Marvell is the nearest competitor.

800g sharesingle-mode is
the dollar that
grows. Marvell
takes most of it.

1.6T TIA: Marvell is the only company in volume today. Macom and Semtech are still sampling; ramp expected 1H 2027 at the earliest. Every 1.6T module shipped in 2026 carries a Marvell TIA by default.

1.6t cliffyear-plus lead.
the strongest
competitive seat
in the whole
AI build-out.
↓ one step down the receive path ↓

the DSP slugfest

The DSP is the brain of the module: equalisation, FEC (forward error correction), framing. At 800G and 1.6T it’s leading-edge silicon — TSMC 3 nm.

dspdecodes the
light. Fixes
bit errors.
Frames the data.

Outside Nvidia, Broadcom holds ~90% of the pluggable DSP market. Marvell is the credible challenger but materially smaller. The fight today isn’t design — it’s the same overbooked 3 nm capacity gating both sides.

3 nm allocationwhoever pulls
more wafers
wins.

white-box switches

Camp 1 — Nvidia: closed bundle, CSPs can’t customise. Camp 2 — Broadcom + Marvell: sell switch silicon to hyperscalers, which build customised white-box switches via JDM/ODM partners.

analogywhite-box switch
= bare-metal PC.
Marvell + Broadcom
are the CPU.

where Marvell is absent

Google’s 2.4T coherent module. Core DSP is Google’s own — built by an in-house team assembled out of former Inphi engineers. Marvell is not the supplier.

absence 1Google built its
own DSP. Marvell
out.

The Nvidia–TSMC 3.2T CPO main line. The optical-engine package on the substrate is controlled by Nvidia and TSMC. Third-party module specialists, Marvell included, have no direct seat at this layer.

absence 2the densest
Nvidia CPO box
has no Marvell.
↓ the bridge generation ↓

NPO — Marvell as co-architect

NPO (near-packaged optics) is the bridge between today’s pluggables and full CPO. The optical engine moves closer to the switch ASIC but doesn’t fully integrate onto the substrate.

npooptics gets close
to the chip —
but not inside
the package.

NPO is DSP-less. Signal recovery moves to SerDes blocks paired with enhanced drivers and TIAs. The dominant SerDes-IP suppliers are Marvell and Broadcom. Among overseas players, Marvell and Amazon lead the NPO push.

silver liningno DSP socket?
Marvell still
owns the SerDes
underneath.

one-line read

Technology lead in TIA, structural pressure in DSP, co-architect in NPO. The TIA monopoly buys time; the DSP fight resolves next; NPO decides whether the SerDes-IP socket replaces what the DSP socket used to be.

← dispatch