ZR is the meat, Coherent Lite is the bone. 2026 is the year ZR generates cash; 2028 is when Coherent Lite truly ramps. Three signals could reverse everything: whether Amazon commits, how fast LPO/CPO win sockets, and production-grade demos at OFC 2027. Taiwan's real opportunity lies in upstream components, not the DSP.

As copper can no longer keep up with the data movement demands of AI clusters, Marvell answers with its Photonic Fabric optical platform. This is not just a speed upgrade, but an architectural shift that breaks through the memory wall and the power bottleneck.

Marvell's 2.5D heterogeneous integration uses a via-last TSV process to free silicon photonics engines from wire bonds, keeping clean signal integrity at 100 Gbaud while pushing energy efficiency toward 5 pJ/bit.

Marvell's 3nm FinFET 212.5Gbps SerDes hits 54.7dB reach and 2.05pJ/bit, the longest-reach and most efficient published at this rate. A shared PLL plus heavy DSP equalization breaks the old rule that reach costs power, giving 1.6T racks structurally more room in cable length and power budget.

Earnings, timelines and shipment volumes all confirmed the optical bull case in the same week, yet Taiwan's silicon photonics stocks hit limit-down. This is a valuation correction, not a demand alarm, and the line between the two hinges on one sentence in next week's Broadcom earnings call.

At FMS 2026 Marvell unveiled a trio of memory products, and the market read it as another wave of AI launches. Put them in the context of four acquisitions, though, and the real play is owning the entire "data movement layer" — memory is just the first toll. We break down the map, the moat and the supply chain positioning.

Demand broke records this week, and so did costs. The same week NVIDIA posted $96.2 billion in revenue, it told customers prices would rise 15%, and Coherent paid upfront to lock in three years of InP capacity. In 2027, optical module makers will win or lose not on whether they can build 1.6T, but on whether they have secured substrates and how much BOM budget is left for optics.

Marvell raised FY28 revenue to $16.5B, lifted interconnect growth from 50% to 70%+, and unveiled a three-pillar NVIDIA partnership spanning silicon photonics, NVLink Fusion, and AI-RAN. The real signal is a structural narrative upgrade: from a DSP vendor riding AI CapEx to a full-stack scale-up interconnect provider.

Jensen Huang declared that "compute has been disaggregated"; Matt Murphy supplied the missing physical timeline. AI's next bottleneck is connectivity, and the deciding factor is a copper wall moving into the rack. Every doubling of bandwidth halves copper reach — at 400G/lane, copper can't even span a single rack, and physics forces CPO from slideware into shipments.

Silicon modulators are stuck at 3,000–5,000 µm and 60 GHz by the diffraction limit, short of the 3.2T AI needs. Marvell's plasmonic modulator shrinks to ~10 µm, reaches 1 THz, and can be added at the back end of existing SiPh processes. It's a key piece for extending SiPh beyond 3.2T — but volume yield is the real test.

Four industry heavyweights on one stage reached an unromantic consensus: demand isn't a bubble, technology isn't heading for a revolution, and the real battlefield is manufacturing. For Taiwan's supply chain, that means the winners will be decided by yield, process, and automation.

Marvell added $1.5B to its FY2028 outlook in a single quarter, and the largest piece came from scale-up optics, a category sized at only ~$300M last quarter. The $120B Google warrant looks huge, but management says pre-FY2028 volume is already in the numbers; the real step-up lands in FY2029.

Why did Marvell pay up for Celestial AI? The answer is the SiPh interposer, which turns the PIC into an active interposer between GPU and HBM: the endgame architecture for scale-up CPO. Lightmatter runs an independent track, and neither uses TSMC. Part 4 of the series unpacks this underrated long-term axis.

Marvell's $3.25B Celestial AI deal is not another CPO startup buy. What it really paid for is Photonic Fabric, which brings light into the center of the die and breaks the memory wall: a ticket to whatever replaces copper inside the package.

2026 is the breakout year for 1.6T. Amid EML shortages and a split over the 3.2T path, Broadcom brought 50 million device-hours of CPO stability data, while Marvell and OpenLight showed breakthroughs in heterogeneous SiPh integration and coherent technology.

As NVIDIA GPU rack power soars from 120 kW to 600 kW, interconnect technology faces a sweeping overhaul. From Andy Bechtolsheim's XPO vision to Dave Welch's THz dielectric waveguides, who will survive the 1.6T/3.2T race?