Nvidia’s Optical Engine Ramp Lands in POET’s Wheelhouse

Nvidia’s Optical Engine Ramp Lands in POET’s Wheelhouse

The bottleneck in AI has moved from compute to connectivity — and the supply chain is repricing around it

For two years, the AI trade has been a story about compute: more GPUs, bigger clusters, larger memory stacks. That story hasn’t ended, but the constraint has quietly migrated. The new ceiling isn’t how fast a chip can think — it’s how fast data can move between chips without melting the power budget. That is an optics problem, and the past two weeks have produced two data points that put a fine point on it.

The first came from Nvidia’s own roadmap. According to a FundaAI report circulated by Citrini analyst “Jukan,” Nvidia’s NVL576 passive optical co-packaging architecture nearly doubles the density of optical engines and optical devices inside the rack. The report pegs the 3.2T optical-engine configuration per GPU rising from roughly 2.25 to roughly 4.0 — a 78% increase — and projects Rubin Ultra optical-engine demand at approximately 12 million units. Critically for our readers, the same note named the direct beneficiaries of that ramp: Lumentum, Coherent, and POET Technologies.

The second came from further up the stack. On June 2, Sweden’s Sivers Semiconductors announced a strategic collaboration with GlobalFoundries $GFS to integrate its laser arrays into GF’s silicon photonics platform and its SCALE co-packaged “light engine” architecture — explicitly aimed at co-packaged optics (CPO) and linear pluggable optics (LPO) for AI data centers. In the same breath, the trade press noted Nvidia is now ramping production of its Vera Rubin platform, which incorporates Spectrum-X Photonics CPO-based switching. GlobalFoundries framed the opportunity bluntly: it is targeting a $25 billion pluggable-optics market by 2030, with silicon-photonics revenue expected to roughly double in 2026.

Taken together, these aren’t two unrelated press releases. They’re the same signal from two altitudes: the hyperscale architecture (Nvidia) and the foundry/component layer (GF, Sivers) are both reorganizing around dense, co-packaged optical engines — the exact product category POET has spent the last decade building toward.

Why density is the whole game

The reason “nearly doubling density” is more than a spec-sheet flex comes down to physics and economics. As GPU clusters scale, the copper interconnects that move data between them hit a wall on reach, bandwidth, and power. Optics solve reach and bandwidth — but only if the optical engine can be made small enough, cool enough, and cheap enough to sit right next to the processor. Every doubling of optical-engine density per GPU is, in effect, a doubling of the addressable unit volume for whoever can supply those engines at scale.

That is precisely POET’s pitch. The company describes itself as a leader in highly integrated optical engines and light sources for AI networks, and its differentiation is structural: the POET Optical Interposer integrates lasers, drivers, photodiodes, and multiplexing into a compact, wire-bond-free module assembled at wafer scale. Smaller, lower-power, and cheaper to manufacture is not a marketing slogan in a density-constrained world — it’s the spec the architecture is now demanding.

POET’s product stack maps onto the demand

What makes the analyst call more than noise is that POET already has commercial product lines pointed directly at this shift:

• Teralight — POET’s 1.6Tbps transmit-and-receive optical engines, which earned an Elite Score at the 2026 Lightwave Innovation Reviews.

• Blazar and Starlight — external light-source products demonstrated at OFC 2026, explicitly designed to power co-packaged optics and chip-to-chip optical links, positioned as a wafer-level alternative to traditional DFB-laser solutions.

• Infinity — the 800G optical-engine line that drew a $5M-plus production order from a systems integrator in late 2025.

• The Lumilens supply agreement — a $50 million initial purchase order for Electrical-Optical Interposer (EOI)-based engines, structured to potentially exceed $500 million in cumulative purchases over five years, with a roadmap running from 800G/1.6T pluggables into Near-Package Optics and Co-Packaged Optics.

In other words, when the FundaAI note lists POET beside Lumentum and Coherent, it isn’t reaching. POET sits in the same supply category the Sivers-GF announcement is enabling — the co-packaged light engine — and it has the design wins, the product awards, and a nine-figure commercial framework to point to.

What to watch, and what to discount

The constructive case is clear, but discipline matters. A few caveats belong in any honest read:

• The NVL576 density figures and the 12-million-unit Rubin Ultra projection are an analyst estimate drawn from a third-party report, not Nvidia guidance. Treat them as directional, not gospel.

• $POET’s largest commercial milestones remain forward-looking. The Lumilens program targets engineering samples in late 2026 and production ramp in 2027, contingent on successful qualification and manufacturing scale-up. Revenue recognition follows execution, not announcements.

• $POET recently raised US$400 million in a registered direct offering, which strengthens the balance sheet for exactly this ramp — but also reminds investors this is still a company funding its way toward scale, not harvesting it.

The signal worth holding onto is the directional one. The AI bottleneck has moved to the optical layer, the largest players in the chain — Nvidia at the architecture level, GlobalFoundries and Sivers at the foundry level — are repricing around dense co-packaged optics, and the analyst community is now naming POET in the same sentence as the incumbents. For a company whose entire thesis was built on the bet that this day would arrive, the question is no longer whether the market shows up. It’s whether POET can convert design leadership into delivered volume on the timeline the hyperscalers are setting.