Marvell Optical Interconnects, September 2026: Industry-First 2nm Demo Targets AI Data Centre Scale

At ECOC 2026, Marvell demonstrated the industry's first 2nm optical chips supporting 400G-per-lane links and a 102.4-terabit co-packaged optics platform, the company's most technically advanced bid yet to anchor the interconnect layer of hyperscale AI data centres.

When Marvell took the floor at ECOC 2026, it brought what it described as the industry's first 2-nanometre optical technology demonstrations: a 400G-per-lane photonic link and a 102.4-terabit co-packaged optics platform aimed at the next generation of AI data centres. The timing was deliberate. As GPU clusters have scaled to tens of thousands of accelerators, the constraint has migrated from raw compute toward communication, and the energy consumed by copper-based electrical signalling has hit what engineers call the power wall. Moving data-centre interconnects to 2nm silicon photonics, Marvell argues, is among the few credible routes to expanding AI infrastructure capacity without pushing power consumption beyond what grids can absorb.

The optical pivot predates the current AI spending wave but has sharpened considerably as the stakes rose. Coverage through mid-2026 consistently noted that optics revenue was still driving Marvell's AI business more than its custom ASIC pipeline — a distinction that gets obscured in the broader conversation about bespoke silicon. At ECOC, Marvell demonstrated 3.2-terabit-per-second data-centre links at the 2nm node, with co-packaged optics designed to close the physical gap between transceiver and switch silicon. Whether those demonstrations translate into volume production is the operative question; Marvell is simultaneously reported to be pressing GlobalFoundries to accelerate its wafer production ramp for custom AI networking and compute silicon, a signal that supply-chain execution — not technology invention — has become the binding constraint.

Running parallel to the optics story is the still-developing relationship with Google. Marvell's CEO described a reported $12 billion custom-silicon agreement with Google as game-changing and said actual revenue could prove far larger than analysts currently model. RBC reportedly raised its fiscal-year 2029 AI revenue estimate for Marvell by at least $2 billion on the strength of that partnership; JPMorgan characterised the same deal as representing a $120 billion long-term revenue opportunity. Yet when reports surfaced in August 2026 that Google's custom AI revenue contribution might be deferred beyond 2029, the stock reportedly lost 10.3%, erasing roughly $22 billion in market capitalisation — a sharp demonstration of how acutely the company's near-term valuation is bound to the timing of a single customer relationship. Separately, Marvell raised its two-year revenue outlook to $30 billion and projected its data-centre semiconductor segment would reach $18 billion by 2028, both figures contingent on AI infrastructure spending holding its current pace.

The financial structure supporting those projections is leaner than it might appear. In fiscal year 2026, Marvell reported capital expenditure of $354 million — a 24.4% increase from $285 million in fiscal year 2025, against revenue of $8.19 billion, implying a capex intensity of 4%. Among the chip companies in Slicast's peer set, that places Marvell ninth out of fifteen by intensity and twelfth out of fourteen by absolute spend, with Micron leading both rankings. The light footprint reflects Marvell's fabless design model, but it transfers production-capacity risk to foundry partners — the same partners it is reportedly pressing to accelerate. Piper Sandler initiated coverage in September 2026 with a $270 price target anchored to the AI custom-silicon pipeline, and shares gained roughly 5% on the initiation. The company declared a quarterly dividend of $0.06 per share at end-September, payable October 29. Its 10-Q filed in late August 2026 flagged customer-concentration risk as material, noting that major customers account for an increasing share of total revenue — a category that now prominently includes Google.

The competitive framing has also shifted. Some analysts have begun positioning Marvell as the anti-Broadcom in custom silicon and network connectivity. Broadcom faces competitive pressure as Google deepens its Marvell relationship. Qualcomm's reported entry into AI data-centre silicon opens a new front. Nvidia's reported $5.5 billion strategic investment across MediaTek and Marvell, noted in August 2026, adds further context: the ecosystem's dominant GPU maker appears to have a direct interest in maintaining competitive alternatives in the interconnect and custom-silicon layer beneath its own chips.

Three signals will determine how much of September's optimism holds. First, Google revenue timing: the difference between recognition in 2028 versus 2030 is the difference between a near-term earnings catalyst and a multi-year holding thesis, and any public clarification from either company would reprice the stock. Second, the optics-to-custom-silicon mix: if the 2nm demonstrations progress toward commercial tape-out and volume production, they could rebalance a revenue base that has historically leaned more toward photonics than bespoke accelerators, potentially narrowing the competitive gap with Broadcom. Third, foundry throughput: Marvell's capacity to fulfil hyperscaler demand depends entirely on its foundry partners' ability to ramp, and any supply shortfall would constrain upside regardless of the design-win pipeline. The ECOC demonstrations are a genuine technical advance in a field where genuine advances remain rare. They are not, by themselves, a shipment.