The Rubin Ultra Era: Reshaping the $154 Billion Optical Interconnect TAM and the Ascendancy of CPO Value

Core Thesis

Networking is overtaking compute silicon as the fastest-expanding value pool in AI infrastructure. After dissecting Nvidia’s hardware roadmap from GB300 to Rubin Ultra, we arrive at a non-consensus conclusion: the total optical value per compute unit will balloon 29x, from $315,000 to $9.4 million. This propels the total addressable market from roughly $15 billion in the GB300 generation to $154 billion in the Rubin Ultra era (primarily reflecting 2028) — 69% of which comes from GPU-to-GPU scale-up fabrics, a domain where optical solutions were virtually absent before.

Within this $154 billion market, Co-Packaged Optics (CPO) will capture 59%, equivalent to approximately $91 billion. This is not a simple product refresh. It is a structural transformation in which optical connectivity moves from a supporting role in switch-to-switch links to the core backbone of high-speed GPU interconnects. Supplier competitive moats, cost structures, and the rules of value distribution will be rewritten. In a single rack configuration for Rubin Ultra, Innolight could realize an EPS uplift of RMB 28.6 — exceeding many companies’ full-year 2025 profits.

What the Market Underestimates

First, the speed of the tipping point at which CPO shifts from an optional replacement to a mandatory architecture. The market generally treats CPO as a distant threat to pluggable transceivers, ignoring a physical constraint: when single-rack scale-up bandwidth demand reaches 512 ports × 3.2 Tb/s, pluggable solutions hit a ceiling on power consumption and signal integrity. In the two-layer scale-up fabric of Rubin Ultra NVL576, the inter-rack L2 layer is forced to adopt CPO — not because it is cheaper, but because no other approach can simultaneously meet power and bandwidth density requirements. Here, CPO’s TCO advantage is no longer a nice-to-have; it is a barrier to entry. This logic will rapidly push CPO penetration in scale-out from zero to 29% (per our adoption assumptions), and to 40% in scale-up.

Second, optical connectivity’s jump from an accessory to the dominant value carrier. In GB300, the transceiver attach rate per GPU was only 3.0, with the scale-up portion using pure copper. By Vera Rubin, that ratio climbs to 6.0. By Rubin Ultra, the scale-up layer introduces 4.5 optical engines per GPU, and scale-out adds 0.9. Optics no longer passively follow GPU unit volumes; they actively open an incremental TAM as a growth engine in their own right. In a single CPO switch bill of materials, optical engines account for 43% ($32,400), far exceeding the switch ASIC at 16% — optics become the single largest value category, upending traditional switch cost structures.

Evidence Chain

Quantifying the value leap. Based on Nvidia’s GTC 2026 roadmap and cross-verified supply chain data, we modeled costs across five hardware generations. From GB300 NVL72 to Rubin Ultra NVL576, the per-rack value trajectory is clear: scale-out rises from $175,000 to $366,000, and scale-up jumps from $140,000 (all-copper) to $803,000 (copper + CPO). Key verifiable assumptions: 1.6T transceiver ASP of $800, 3.2T at $1,600, CPO optical engines priced at parity for the same data rate; Rubin Ultra unit assumption of 16,500 compute units.

Substantive breakthroughs in CPO architecture and silicon photonics industrialization. Nvidia’s Quantum-X/Spectrum-X Photonics switches will commercialize in early 2026 using micro-ring modulators; Broadcom’s 102.4T Davisson switch sampled in October 2025. Both are building closed but clearly defined supplier ecosystems. Meanwhile, silicon photonics solutions deliver a bill-of-materials cost 32% below EML-based designs for 1.6T transceivers, a 20% lower selling price, and a 7-percentage-point higher gross margin. This cost scissor is the economic foundation for silicon photonics penetration to surge from 6% in early 2024 to an expected 46% in 2028. It will determine who captures the largest optical engine value in the CPO wave.

Hard constraint in light source supply. EMLs and CW lasers share InP substrate capacity, an area squeezed by equipment lead times and geopolitical export controls. VPEC plans to expand its InP MOCVD tool count from 60 to 64 in the second half of 2026; Lumentum committed 40% capacity expansion over the same period. Yet against the explosive per-rack CW laser demand in Rubin Ultra, the supply gap may persist into the second half of 2028. Prysmian’s fiber orders are already fully booked for 2026, with customers willing to sign long-term framework agreements at current pricing — the fiber and cable infrastructure bottleneck cannot be ignored. This hands sustained pricing power to suppliers like Sumitomo and Furukawa, which combine FAU and laser capabilities.

Key Risks

The technology feasibility query chain. CPO packages optical engines together with expensive switch ASICs or GPUs. A single engine failure could scrap the entire chip. In a pluggable world, you simply swap a module. Hyperscaler customers place extreme demands on serviceability, and this risk could trigger strong conservative sentiment in early deployments, delaying the rate of penetration improvement.

Path dependency and vendor lock-in. Nvidia’s and Broadcom’s CPO ecosystems are relatively closed. A supplier that picks the wrong technology path — in engine pluggable standards, fiber array unit specifications, or light source packaging — risks being locked out for an entire generation. All our TAM calculations are based on two specific architectural assumptions, “Spec A” or “Spec B.” A minor adjustment could reallocate billions of dollars of value between pluggable and CPO.

Single-customer concentration and the amplification of demand swings. The $154 billion TAM rests critically on Rubin Ultra unit volumes. If a shift from AI training toward inferencing reduces the need for extreme-scale clusters, or if global hyperscaler capex cadence slows, the heavy upfront CPO investment will face a protracted digestion period. The time lag between supplier valuations shifting from P/E to realized earnings could become a significant share price headwind.

Investment Implications

The core trade for this market expansion is not a zero-sum bet on “CPO versus pluggable” but rather capturing structural increments within the swelling total value pool. Innolight holds leading positions in both pluggable and CPO, offering the greatest EPS sensitivity (RMB 28.6 per share under Rubin Ultra Spec B). TFC Optical’s incremental contribution from FAU and optical engines starts practically from zero, making its earnings most sensitive to CPO penetration. Sumitomo and Furukawa sit at the cross-section of light source and FAU supply, giving them the highest scarcity value among upstream nodes in the CPO supply chain.

Key risk-reward observation points: commercial shipment validation of Nvidia CPO switches in the second half of 2026, measured power consumption data for silicon photonics at 3.2T, and major hyperscalers’ final evaluations on the “serviceability vs. TCO” trade-off. These catalysts will determine how much of the $154 billion TAM translates into revenue and how much remains on paper.

Appendix: Key Data Compression Table

Table 1: Optical Value Leap per Compute Unit by Generation

Table 2: Core Supplier EPS Sensitivity under Rubin Ultra NVL576

Source: Goldman Sachs estimates. Key assumptions as noted in ASP and unit volume parameters above.