A 400 Gbps signal dies after about 1.25 meters of copper. Every time you double the per-lane speed, the reach halves. Photonics does not care — it will carry the same signal for kilometers.
That single number is the whole deal. Marvell agreed on November 2 to buy Lumentum in an all-stock transaction with an implied enterprise value of $115 billion, roughly $1,132 per Lumentum share, a 23.9% premium to the unaffected October 28 close. Lumentum holders get 4.5 Marvell shares each and will own about 33% of the combined company.
Last week we traced the substrate and interconnect supply chain through the packaging layer. This week the first real consolidation bet landed on optical interconnect, and in the same seven days Nvidia quietly told us where the physical layer is actually breaking.
Marvell is not buying an optics business. It is buying indium phosphide fabs.
The honest framing is that Marvell is not buying an optics business. It is buying indium phosphide fabs.
Marvell's optical franchise runs on PAM4 DSPs, drivers, TIAs and silicon photonics. Linear-drive optics and co-packaged optics are the two architectures explicitly designed to take the DSP out of the link. The laser stays in every one of them.
That is the hedge inside the deal, and Counterpoint Research's Gareth Owen says it plainly: "Whatever happens with CPO, links still need lasers, and now Marvell will make them." Marvell bought Celestial AI for about $3.5 billion in February 2026 for near- and co-packaged optical interconnect that uses a fraction of pluggable power. Some of those designs delete Marvell's own DSP. The laser survives them all.
Lumentum is sold out through early 2029, and a new InP fab takes years to build. Neil Shah of Counterpoint: "Marvell is buying guaranteed access to a resource its competitors depend on." That is a supply-side wall, not a demand story.
Now the counter-evidence, from the same week, and it cuts both ways.
Nvidia cut Rubin's scale-up NVLink bandwidth from a preliminary 3.6 TB/s to 3.0 TB/s — a 17% reduction — and dropped HBM4 per-pin rates from 10.75 to 9.375 Gbps, taking per-GPU memory bandwidth from 22 to 19.2 TB/s, an 11% hit. NVL72's aggregate scale-up fabric fell from 260 TB/s to 216 TB/s.
At the same time, scale-out networking rose from 0.4 to 0.45 TB/s per GPU, and 28.8 to 32.4 TB/s per rack.
Read those two moves together. Nvidia is rationing bandwidth inside the rack-scale domain and buying more outside it. Scale-up stays electrical where it can; scale-out, which has to leave the rack anyway, gets the headroom. The physical layer is setting the roadmap, not the marketing calendar.
For Marvell, that is a complicated signal. If scale-up stays on copper longer than the optics camp expects, the rack-boundary crossing slips right — and Marvell has handed over a third of the company for a close that does not arrive until the first half of 2028. Marvell shares fell about 22% and Lumentum about 8% after talks leaked.
The efficiency problem is asymmetric. Lumentum brings roughly $5 billion in annualized revenue to Marvell's $12.6 billion, less than a third of the combined top line, for a third of the equity. Marvell promises more than $1 billion in annual run-rate cost synergies.
The comparison everyone reaches for is Broadcom, which pairs custom silicon with its own InP fab. The combined Marvell would go further — DSPs, drivers, TIAs, silicon photonics, switching, custom XPUs, Celestial's photonic fabric, CW and EML lasers, external light sources, datacom and coherent transceivers, and Lumentum's R300 optical circuit switch, which uses MEMS mirrors to cut intermediate packet processing and optical-electrical-optical conversion entirely.
Supplying every active part of the optical link is a strong position. It is also a customer problem.
Innolight, Eoptolink and other module makers buy Marvell DSPs and Lumentum EMLs. After close, both come from a company that also sells modules. Counterpoint expects them to dual-source harder — moving DSP sockets to Broadcom or Credo and qualifying more EML and CW suppliers in Japan and China. Pricing leverage and customer attrition arrive on the same day.
Then there is Beijing, which effectively holds the light switch. The 2028 close date is itself the tell: long reviews in several jurisdictions, and the China market regulator SAMR has held up US chip deals before. US moves against Chinese optics, like the FCC's proposed limits on Chinese transceiver imports, give Beijing every reason to take its time.
And there is the Nvidia position, which is the strangest fact in the file. Nvidia invested $2 billion in each of Lumentum and Coherent this year. In an all-stock deal, its Lumentum shares convert into one larger Marvell stake.
So Nvidia becomes both the deal's largest beneficiary and its most hedged counterparty. Its key laser supplier now sits inside a company that helps AWS, Microsoft and others build XPUs that compete with Nvidia GPUs. It will not fight this in public. It will lean on Coherent.
That matters because the whole thesis is a timing bet. Marvell is paying 2026 prices for a 2028 crossing, and the physics sets the date.
Copper holds 400 Gbps to 1.25 meters and halves from there. Photonics is largely immune and can push the signal across an entire aisle while cutting per-rack power and cooling. Jensen Huang avoided optics in 2024's Blackwell racks because they would have added 20 kW to a system already drawing 120 kW under load. The power math has since moved.
But Nvidia's revisions show the other half: when the electrical layer can't deliver the spec, the spec changes. Rationing bandwidth inside the rack is a deliberate slowing of exactly the transition Marvell is financing.
There is also an execution risk that no spreadsheet captures. Marvell is a merchant silicon house — CMOS, SerDes, DSPs, design wins. Lumentum is a compound-semiconductor manufacturer running InP process technology.
Those are different engineering cultures with different cycle times, different capital intensity, and different definitions of a good day. The stated plan is heterogeneous integration and deeper co-design across electrical and optical. That is a multi-year integration before a single dollar of synergy.
Marvell's total addressable market claim is $550 billion. It spent a third of itself to enter it.
The copper wall is real. The physics does not negotiate. But the wall is not a date on a calendar — it is a function of per-lane yield, power budgets and how long hyperscalers tolerate electrical scale-up. Marvell is not buying the wall. It is buying the laser that survives on the far side of it, and betting that the far side arrives on schedule, at a price set before Nvidia revised the spec down.