Technical Deep Dive
The optical module revolution is fundamentally a physics problem solved by engineering. At its core, the challenge is moving data between GPUs in a cluster faster than the GPUs can process it. For NVIDIA's H100 and B200 clusters, the interconnect bandwidth required per GPU has jumped from 400 Gbps to 1.6 Tbps in just two generations. Copper cables, limited by signal integrity over distances beyond 3 meters, simply cannot sustain these speeds without unacceptable power and latency penalties.
Enter silicon photonics (SiPh). Traditional optical modules used discrete components: a laser diode, a modulator, a photodetector, and control electronics, each manufactured on different material platforms (InP, LiNbO3, SiGe). SiPh integrates these onto a single silicon substrate using CMOS-compatible fabrication processes. This reduces cost, size, and power. The key breakthrough is the micro-ring resonator modulator, which can achieve 100 Gbps per wavelength with sub-100 femtojoule/bit energy consumption—a 10x improvement over Mach-Zehnder modulators.
| Parameter | Traditional Discrete Module (400G) | Silicon Photonics Module (800G) | Co-Packaged Optics (1.6T prototype) |
|---|---|---|---|
| Power per port | 12W | 7W | 4W |
| Reach (max) | 2km (SMF) | 10km (SMF) | 100m (MCF) |
| Latency (per hop) | 350ns | 150ns | 50ns |
| Bandwidth density | 1.6 Tbps/inch | 3.2 Tbps/inch | 12.8 Tbps/inch |
| Cost per Gbps | $0.80 | $0.35 | $0.12 (est.) |
Data Takeaway: Silicon photonics delivers a 2x improvement in power efficiency and bandwidth density over traditional modules, but co-packaged optics (CPO) promises a 4x leap across all metrics. The cost per Gbps drop from $0.80 to $0.12 is the economic driver for mass adoption.
A critical open-source reference point is the OpenLight platform (GitHub: openlight-photonics), which provides a PDK for designing SiPh circuits with integrated lasers—a major hurdle since external lasers were previously required. The repository has seen 2,300 stars and active contributions from 15 companies, indicating the community's push toward democratizing photonic design.
The next frontier is co-packaged optics (CPO) , where the optical engine is integrated directly into the GPU package or interposer. Broadcom's 51.2 Tbps CPO switch, announced in late 2025, uses 32 optical engines each delivering 1.6 Tbps. This eliminates the pluggable module interface, reducing signal loss and power by another 30-50%. The engineering challenge is thermal management: lasers are temperature-sensitive, and GPU hotspots can exceed 100°C. Advanced thermal solutions using microfluidic cooling are being developed, but reliability remains unproven at scale.
Key Players & Case Studies
The optical module market is dominated by a handful of players who have pivoted aggressively to AI. The most notable is Innolight (formerly Zhongji Innolight), which has become the poster child for the 1000% surge. Their 800G OSFP modules, based on SiPh, are the primary interconnect for NVIDIA's H100 clusters. Innolight's strategy: secure exclusive supply agreements with hyperscalers (Microsoft, Meta, Google) and invest heavily in automated manufacturing to achieve 60% gross margins—unheard of in traditional optics.
| Company | Market Cap (2024) | 12-Month Return | Key Product | Customer Concentration |
|---|---|---|---|---|
| Innolight | $120B | +1,200% | 800G OSFP SiPh | 70% to NVIDIA/Meta |
| Coherent Corp. | $45B | +350% | 1.6T FR4 | 40% to hyperscalers |
| Lumentum | $28B | +280% | VCSEL arrays | 50% to Apple/Amazon |
| Cisco (via Acacia) | $18B (optics unit) | +150% | CPO switches | Internal only |
Data Takeaway: Innolight's 1,200% return is not just about technology; it's about strategic positioning as the sole qualified supplier for NVIDIA's next-gen clusters. The high customer concentration is both a strength and a vulnerability.
A notable case study is Marvell Technology, which acquired Inphi in 2020 for $10 billion. Inphi's electro-optics platform (PAM4 DSPs) became the backbone for 400G and 800G modules. Marvell's recent 1.6T DSP, the Alaska C, integrates a 7nm DSP with a SiPh modulator, achieving 8W per port—a 20% improvement over competitors. Their strategy is to sell the DSP as a chipset to module makers, creating a royalty stream that scales with volume.
On the research side, Prof. John Bowers at UC Santa Barbara (founder of Aurrion, acquired by Juniper) has pioneered heterogeneous integration of III-V lasers on silicon. His group's recent demonstration of a 64-channel DWDM laser array on a single chip (published in Nature Photonics, 2025) could enable 6.4 Tbps per fiber pair, a 4x increase over current WDM systems.
Industry Impact & Market Dynamics
The optical module market is projected to grow from $12 billion in 2024 to $65 billion by 2028, according to industry estimates. This growth is driven by three structural shifts:
1. From copper to optical at every level: Even within racks, optical interconnects are replacing copper for distances as short as 1 meter. The cost crossover point has shifted from 10m to 3m due to copper's power penalty at 800G+.
2. The 'optics-as-a-service' model: Companies like Innolight and Coherent are now offering subscription-based contracts where hyperscalers pay per Gbps per month, rather than buying modules outright. This converts a $500 module into a $15/month revenue stream over 3 years, with 70% gross margins. The market is pricing these recurring revenues at 15-20x multiples, versus 5-8x for hardware sales.
3. Vertical integration: NVIDIA is acquiring optical module startups (e.g., Lightmatter, Ayar Labs) to secure supply chains. This is driving M&A premiums of 10-15x revenue, further inflating valuations.
| Metric | 2023 | 2024 | 2025 (est.) | 2028 (est.) |
|---|---|---|---|---|
| Global optical module revenue | $11.2B | $18.5B | $32B | $65B |
| % from AI clusters | 25% | 45% | 60% | 75% |
| Average module speed | 400G | 800G | 1.6T | 3.2T |
| Number of 10k+ GPU clusters | 12 | 38 | 120 | 500 |
Data Takeaway: The market is doubling every 18 months, with AI clusters accounting for an increasing share. The average module speed is doubling every 2 years, driving ASPs up despite volume growth.
The geopolitical dimension is critical. China's optical module manufacturers (Innolight, Eoptolink, Accelink) control 60% of global production capacity. US export controls on advanced lithography tools (ASML EUV) do not directly affect SiPh fabrication, which uses 130nm-45nm nodes. However, restrictions on GaAs and InP substrates could limit high-end laser production. This has led to a 'dual supply chain' strategy: US companies are investing in domestic SiPh fabs (e.g., GlobalFoundries' SiPh platform), while Chinese firms are stockpiling laser chips.
Risks, Limitations & Open Questions
Despite the euphoria, significant risks exist:
- Thermal reliability: Lasers degrade faster at high temperatures. In GPU clusters running at 100% utilization, ambient temperatures can reach 70°C. Current SiPh lasers have a mean time to failure of 50,000 hours at 70°C, versus 200,000 hours at 25°C. This could lead to premature module failures in large clusters, causing costly downtime.
- Supply chain bottlenecks: The global supply of high-speed DACs (digital-to-analog converters) and TIAs (transimpedance amplifiers) is constrained. Companies like MaxLinear and Semtech are running at 100% capacity, with lead times exceeding 52 weeks. Any disruption could stall cluster deployments.
- Overvaluation risk: At 50-100x forward earnings, optical module stocks are pricing in perfection. If AI model scaling slows (e.g., if scaling laws hit diminishing returns), demand could plateau. A 20% demand drop would trigger a 50%+ correction in these high-beta stocks.
- Technological disruption: Direct laser modulation (using VCSELs for 100G per lane) could bypass SiPh entirely for short-reach links. Companies like Lumentum are pushing 100G VCSEL arrays that are cheaper and simpler than SiPh, potentially disrupting the 800G market.
AINews Verdict & Predictions
The optical module revolution is real, but the market has overshot in the short term. Our analysis yields three clear predictions:
1. Consolidation within 18 months: The top 3 players (Innolight, Coherent, Lumentum) will acquire 10+ smaller startups, driving further multiple expansion. Expect a $50B+ acquisition by a hyperscaler (Microsoft or Amazon) of a module maker to secure supply.
2. CPO will be the next battleground: By 2027, 30% of new GPU clusters will use co-packaged optics. The first mover (likely Broadcom or NVIDIA) will capture 60% market share and see a 300%+ stock surge. Investors should watch Ayar Labs (backed by NVIDIA) and Lightmatter for IPOs.
3. The 'optics-as-a-service' model will become standard: By 2028, 50% of optical module revenue will be recurring. This will lift industry-wide valuation multiples from 10x to 25x EBITDA, supporting a $200B total addressable market.
The bottom line: The trillion-dollar valuation is justified by structural demand, but the current price implies zero margin for error. The smart money is on CPO and recurring revenue models, not on chasing the 1,000% gainers. The real winners will be those who own the optical engine inside the GPU package, not those selling pluggable modules.