AI data center optical interconnect demand is undergoing an architectural transition from pluggable modules to NPO/CPO. Who holds the chokepoints in the supply chain, who has opportunities, and where are the supply-demand gaps — this article breaks down the silicon photonics value chain over the next three years from a component technology perspective.
I. The Demand Baseline
The most frequently cited 2030 AI capex figure is $3 trillion, derived from a five-step chain (GDP anchoring → gross margin conversion → physical ceiling verification). But every parameter picks the most optimistic value — 10% GDP contribution is 20–50x the IMF estimate, and 75% gross margin is a four-year-out target, not today's reality. Cross-checking with four independent methods yields a convergence range of approximately $1–2.5 trillion/year.
For the optical communications supply chain, what matters is the implication: even at the most conservative end, total demand still has upside relative to today. Optical interconnects currently account for roughly 5–6% of computing center costs (cross-verified with industry sources), corresponding to a global optical module and component market of approximately $20–25 billion/year (LightCounting / TrendForce 2025–2026 estimates). If capex reaches $1–2.5 trillion by 2030 and optical interconnect's cost share breaks through from 5–6% toward 10% (driven by scale-up opticalization), the optical communications market could expand to $50–100 billion/year — doubling to tripling from current levels.
This is not a shrinking market. What follows is a supply-demand breakdown within this market.
II. Component Technology Breakdown: Chokepoints from Electrical to Optical
The bill of materials for an optical module can be divided into three concentric layers: electrical chips (DSP/TIA/Driver/SerDes), optical chips and light sources (EML/silicon photonics wafers/CW lasers), and materials and packaging (TFLN/glass substrates/NPO-CPO processes). The barriers differ across layers, domestic substitution progress differs, and the impact on industry structure differs.
Starting from the highest-barrier layer.
2.1 Electrical Chips: Highest Barriers, Hardest for Domestic Substitution
DSP. Only three companies can deliver 1.6T optical module DSPs at volume globally: Marvell, Broadcom, and NVIDIA. Credo has 800G products but is still catching up on 1.6T. MaxLinear quotes roughly half of Marvell's price, but its production scale is far from the million-unit level.
The DSP accounts for one of the highest BOM shares of any single component in an optical module. Its barriers come from two compounded factors: design complexity (a 1.6T DSP respin costs roughly $100 million), and manufacturing process (requiring successive generations of advanced-node refinement). Chips are made through iteration — without manufacturing capability you can't hone design capability, without design capability you can't win customers, and without customers you never get another iteration cycle. This loop is extremely hostile to latecomers.
More subtle is the pricing power dynamic. According to industry sources, for the same DSP model, the purchase price paid by top-tier module makers versus second-tier players differs by roughly 30–40%. This is not a volume discount; it stems from structural cost advantages that top-tier manufacturers obtain by locking in chip suppliers (Innolight → Marvell, Eoptolink → Broadcom, TFC → NVIDIA). Even when second-tier companies win orders, turning a profit is extremely difficult.
TIA and Driver. For 1.6T modules that pass top-tier customer validation, essentially only Marvell's TIA qualifies. "Low value in BOM but a critical bottleneck" — the TIA accounts for far less BOM share than the DSP, but without it the entire chain stops. On the Driver side, MACOM is the exclusive Driver/TIA supplier for NVIDIA's in-house modules, while Marvell has a head start in the merchant market.
SerDes. 112G will be constrained for a while; 224G power consumption can't be brought down; and anything beyond that will definitely hit a wall. SerDes is the fundamental ceiling on electrical chip performance — it determines how fast the DSP can run and therefore the upper speed limit of the optical module.
Trend assessment: electrical chip chokepoints are hardening. Process and design barriers rise rather than fall as speeds increase, making domestic substitution the toughest challenge. No breakthrough path is visible in the near term.
2.2 Optical Chips: Segmented Assessment
Electrical chips have the highest barriers, but the optical chip landscape is more complex — it requires a component-by-component, tier-by-tier assessment.
EML (Electro-Absorption Modulated Laser). Industry consensus identifies a roughly 30–40% supply gap that will persist at least through next year. High-end EML remains in chronic shortage; low-end EML may face oversupply. Overseas suppliers (Lumentum, Coherent) are prioritizing wafer capacity toward high-power and EML products, while low-power legacy products may be strategically phased out.
Silicon Photonics Wafers. Tower Semiconductor is the key foundry. Its visible 2027 capacity corresponds to approximately 80–100 million 800G-equivalent optical modules. What does this mean? Global demand for 800G + 1.6T combined in 2026 is roughly 65–100 million units (multi-source cross-estimate). In other words, Tower's visible capacity alone approaches satisfying global demand — but capacity ramp takes time, and high-end capacity is prioritized for long-term agreement customers. The new fab scheduled for 2028 production is, according to industry sources, "built almost entirely for NPO." Silicon photonics wafer shortages mirror the EML pattern — tight at the high end, loose at the low end, with "the chokepoint shifting every two quarters."
High-Power CW Laser Sources. These need to be assessed in two tiers. NPO mainstream solutions use built-in light sources requiring only 100–130mW — this tier already has mature domestic solutions. The truly constrained tier is the roughly 400mW class aligned with NVIDIA's external light source architecture: domestic leader Yuanjie has just produced samples and is undergoing customer validation, with volume shipments expected no earlier than next year. Currently this tier is mainly supplied by Lumentum and Coherent. One revealing signal: during the most heated period of the NPO vs. CPO debate, orders for high-power CW lasers from these two suppliers rose rather than fell — indicating that total demand is exploding.
Trend assessment: optical chip chokepoints are diverging. EML and high-end CW laser sources remain in persistent shortage, but silicon photonics chips themselves are not the bottleneck. "Moore's Law has failed in photonics manufacturing, and that is precisely what benefits domestic players most." Photonics manufacturing does not depend on advanced process nodes; the process stays in place, allowing latecomers to catch up. Five to eight years ago, Chinese companies already held seven of the top ten global optical module rankings, and all silicon photonics chips used in Innolight's 800G/1.6T modules are now fully self-developed.
2.3 Materials Side: Opportunities for Breakthrough
The first two layers — electrical chips and optical chips — have relatively fixed competitive landscapes. But the third layer (materials and packaging) is where variables capable of reshaping the landscape are emerging.
Thin-Film Lithium Niobate (TFLN). The fundamental limitation of traditional silicon photonic modulators is this: silicon's own electro-optic coefficient is low, so high-speed modulation relies on PN junction carrier dispersion effects — at the cost of high drive voltage and limited bandwidth. TFLN uses lithium niobate's electro-optic effect to achieve modulation directly — higher bandwidth, lower drive voltage, and lower insertion loss (referencing HyperLight product data from the Harvard spinout).
More critically, the industrialization barrier is extremely low. The ultimate bill-of-materials cost for an 8-inch wafer is roughly ¥2,000 RMB, with a full production line investment of about ¥100 million RMB. The two critical equipment types — ion implantation and hybrid bonding — already have domestic turnkey solutions. "The barrier is so low that individual entrepreneurs can enter the space." If TFLN achieves scaled adoption in 1.6T and above modules, the profit distribution of the modulator segment will be reshuffled.
Strained Germanium Substrates. Silicon's low hole mobility limits the performance ceiling of silicon-based modulators and photodetectors. Strained germanium uses lattice strain to raise hole mobility to compound-semiconductor levels — roughly 10× that of silicon. If volume production matures, it could enable silicon-based optoelectronic integrated chips to approach InP performance, reducing dependence on scarce III-V materials. Still in early R&D, but the direction merits attention.
Glass Substrates and TGV (Through-Glass Via). Traditional organic substrates (ABF carriers) suffer from high dielectric loss at high-speed signals, becoming a bottleneck for 224G and beyond SerDes. Glass substrates have a lower dielectric constant and lower loss, significantly improving signal integrity. Universities have already built 8-inch pilot lines. NVIDIA senior management recently visited TFC to discuss a "glass bridge" approach — using glass substrates to replace traditional organic substrates.
Trend assessment: the acoustic filter industry proved this once before. From Murata's monopoly to domestic substitution, the core driver was material switching. The silicon photonics industry may be undergoing the same transition: material changes alter manufacturing barriers, and manufacturing barrier changes alter competitive landscapes.
2.4 Packaging and Interconnects: Yield Decides the Winner
| Route | Current Yield | Expected Next Year | Key Challenges |
|---|---|---|---|
| Pluggable | ≥95% | Stable | Power pressure at higher speeds |
| NPO | <60% | 80–90% | Socket signal integrity, impedance, stress |
| CPO | Low | Gradually improving | TSMC process yield ramp |
Every route competition ultimately resolves on the yield curve.
NPO's current yield is below 60%, meaning that for every 100 optical engines produced, over 40 have something go wrong at some stage — coupling deviation, solder defects, socket contact failure. Next year it's expected to reach 80–90%. This is the dividing line for NPO economics — and it's not just a number.
Why is 80% the critical inflection point? The cost structure of optical module assembly follows a classic yield curve: every 10 percentage points of yield improvement increases effective output by roughly 17% (1/0.6 → 1/0.7). But the more important effect is non-linear: fixed costs (equipment depreciation, production line construction) account for a disproportionately large share of total cost, and during the yield ramp-up phase, those fixed costs are borne entirely by the first N units. When yield moves from 60% to 80%, the fixed-cost allocation per good unit shifts from "1/0.6" to "1/0.8" — a ~25% reduction. And beyond 80%, every additional 5 points of yield improvement produces an even sharper cost drop, because the breakeven point has already been crossed.
This means NPO is uneconomical in 2026 but may suddenly become economical in 2027 — and the non-linear cost curve means this transition could happen within a single quarter. Module makers that invest in NPO socket design and precision assembly equipment in 2026 — even if they lose money on early deliveries — will face a cost advantage reversal in 2027. Those who wait until NPO economics are confirmed will confront an already-matured supply chain and locked-in customer relationships.
NPO's serviceability is often questioned. But NVIDIA's "external laser + replaceable optical assembly" architecture has already thought through the maintenance interface — what remains is primarily a yield problem. That said, "everyone assumes you just plug it in, but that's not how it works." NPO sockets involve high-speed differential signals (112G/224G) transmitting through mechanical contact surfaces — impedance matching, contact stress, and mechanical stability after thermal cycling are all hard engineering problems. These cannot be solved by "production-line tuning"; they require incorporating the socket into signal-link simulation from the design stage.
Trend assessment: the packaging segment represents an opportunity window for coupling, bonding, and inspection equipment vendors. Regardless of whether CPO or NPO ultimately succeeds, demand for precision manufacturing equipment is growing.
III. Industry Structure: A Technology-Driven Reshuffle
The chokepoint analysis at the component level reveals barrier heights. But barriers are only part of the picture — who is surmounting them and who is being crushed by them requires looking at the direction of industrial organization.
3.1 Route Divergence: Two Models of Industrial Organization
Co-Packaged Optics (CPO) is, in essence, integration — it requires a single dominant player who simultaneously controls AI chip design, advanced packaging processes, and optical integration. Globally, only one or two groups fit this profile: NVIDIA plus TSMC. CPO is not a standalone technology; it is system-level integration — the AI chip and the optical engine are interconnected on the same advanced packaging substrate, physically inseparable.
Near-Packaged Optics (NPO) is, in essence, decoupling — the optical engine and the switch ASIC are independently packaged and connected through a standardized socket. Each segment can be supplied by different specialized companies; the system is serviceable and replaceable.
This is not purely a technology choice. The US is pursuing CPO because it lacks a massive optical module supply chain comparable to China's, but possesses a "master integrator" in NVIDIA + TSMC. China is pursuing NPO because it commands 70% of global optical module production capacity, yet "suffers from a severe shortage of electrical chips" — without an advanced packaging and high-end electrical chip ecosystem, blindly pursuing CPO would be suicidal.
Signal for domestic CPO activation: It is not a yield number breakthrough; it is the emergence of a domestic "master integrator" — that is, a company that simultaneously possesses its own AI chips, advanced packaging integration capability, and optical integration capability. For most domestic GPU companies, "before talking about CPO, talk about whether you'll still be in business." If the chip itself lacks competitiveness, no optical interconnect solution — however advanced — earns you a seat at the table.
Demand-side evidence also supports the NPO route: Amazon and Meta have signed long-term agreements for NPO with volume ramping next year; NVIDIA is testing 3.2T NPO, Amazon is testing 6.4T NPO; Tower's 2028 new fab is built essentially for NPO.
3.2 Whose Hand Is Strengthening, Whose Hand Is Weakening
NVIDIA is evolving from a chip supplier into an optical system integrator. It doesn't just sell GPUs — it is manufacturing its own optical modules (future share may exceed 50%), replacing external DSP procurement with in-house designs, and locking up upstream high-power laser capacity by investing $2 billion each into Coherent and Lumentum. $4 billion is not a financial investment; it is about securing delivery and exclusivity. NVIDIA's "external laser + replaceable optical assembly" architecture effectively pulls a portion of the optical module industry's profit and decision rights from module makers into its own ecosystem.
Cloud providers are fighting for upstream bargaining power. Meta is paying Corning $6 billion to source glass substrates directly, bypassing the entire intermediary chain. Google has long self-developed TPUs and OCS (Optical Circuit Switching), keeping optical interconnect control in its own hands. The party that pays does not want intermediaries skimming profits — "whoever decides, whoever pays" — and when optical interconnect costs rise from 5–6% to 10% of computing center capex, the paying party will only care more.
China's top-tier module makers hold their ground through chip supplier lock-in. Innolight locks in Marvell, Eoptolink locks in Broadcom, TFC locks in NVIDIA. For the same DSP, top-tier firms buy 30–40% cheaper than second-tier players. This price gap is not a simple "volume discount" — it is a structural cost advantage. Chip vendors allocate limited capacity to long-term order partners who already have customer qualification. Top-tier module makers operate a virtuous cycle: "chip resource lock-in → cost advantage → customer qualification → larger orders → stronger lock-in."
Second-tier module makers are being squeezed from both sides. One quote from the industry captures the situation sharply: "Even if you give me an order, I still can't make money." Without chip resource lock-in or a top-tier customer qualification cycle, the same DSP costs 30–40% more on procurement, leaving essentially zero profit margin. The pressure comes from both directions: upstream, chip vendors prioritize capacity for top-tier customers — second-tier players face higher procurement prices or cannot access the latest models; downstream, the cloud provider qualification cycle runs 2–3 years, and failing to enter the qualification list means no bidding eligibility at all. Hundreds of companies make 800G modules, but only a handful of top-tier firms can actually turn a profit.
3.3 Two Simultaneous Movements: Integration and Decoupling
Two opposing movements are happening in this industry at the same time:
Upstream is integrating. Barriers in electrical chips (DSP, TIA) rise with speed increases. The number of suppliers capable of delivering 1.6T at volume has never increased — it is actually shrinking. NVIDIA's in-house DSP replaces some of Marvell/Broadcom's addressable market, but the set of external options does not expand. In optical chips, high-end EML and 400mW CW laser source capacity is concentrated among a small number of overseas suppliers (Lumentum, Coherent). The upstream is characterized by "a shrinking supplier count and rising individual bargaining power."
Midstream is diverging. The technology barrier for module assembly is declining (standardization is increasing), but the commercial barrier is rising (2–3 year qualification cycles, non-replicable chip supplier lock-in). The result: top-tier concentration rises, while the number of tail players increases but without profitability. "Hundreds of module makers produce 800G, but only about ten optical module companies are listed on A-shares, and roughly one-third of them are unprofitable." By comparison, there are about 80 listed LED companies and about 40 listed PCB companies — the optical module segment shows notably higher concentration and "top-tier profitability" than comparable industries.
Downstream is thriving. Coupling equipment, bonding equipment, inspection equipment, optical assembly — regardless of whether CPO or NPO ultimately succeeds, demand for precision manufacturing equipment and advanced packaging is structurally inelastic. As NPO yields climb from under 60% toward 80–90%, massive engineering investment and equipment upgrades will be required.
3.4 An Additional Layer: Global vs. China Divergence
The same route question may have two different answers in the global market versus the Chinese market.
Globally, the three system platform players — NVIDIA (GPU architecture) + Broadcom (switch ASIC) + TSMC (advanced packaging) — hold enormous pricing power. Optical module makers are "spec-guided contract manufacturers" in this system: NVIDIA defines architecture and specifications, TSMC defines process and yield, and module makers execute to drawings. Profit margins are compressed between manufacturing cost and a thin engineering premium.
In China, procurement logic differs. Cloud providers and state-owned enterprises prioritize power reduction and cost savings over matching the global cutting edge in absolute performance. This implies two things: first, domestic module makers do not need to keep pace with NVIDIA's latest specifications every generation — 1.6T can arrive six months to a year later, as long as power and cost are under control; second, value allocation may shift back toward chip suppliers — because customers are willing to pay a premium for power reduction, and the core of power reduction often lies at the chip level.
For domestic module makers, this is both a buffer and a constraint. The buffer: no need to participate in the global frontier arms race; they can follow behind and scale up mature products. The constraint: profit margins are squeezed by chip suppliers, and if they stay out of frontier iteration long enough, the technology gap keeps widening.
A signal worth noting: an NPO advancement organization led by Huawei has recently been established. This means China is attempting to define NPO standards at its own pace — not following NVIDIA's CPO route, but building technical specifications suited to the capabilities of China's optical communications supply chain within the NPO framework. If this route gains procurement support from domestic cloud providers, China's optical communications supply chain will form a relatively self-contained supply-demand loop within the NPO ecosystem.
IV. Market Assessment: Where the Light Is, and Who Pays
The previous three sections analyzed the technology landscape and competitive structure of the supply chain. The final question: what do these structural changes mean? Where is the optical market? Who pays? Which way does the supply-demand balance tilt?
4.1 Where is the Optical Market?
Optical interconnect demand is not a homogeneous market; it has three tiers with different growth rates:
Scale-out (horizontal expansion): largest volume, steady growth. Switch interconnects within AI clusters, currently dominated by 800G and transitioning toward 1.6T. This is the main body of the current optical module market — approximately 70%+ of global optical module shipments (TrendForce estimate). Growth follows AI cluster scale expansion linearly — not the fastest, but the largest volume. The 1.6T transition will push single-mode ASP up 30–50%.
Scale-up (vertical expansion): fastest growth, largest increment. GPU interconnects are approaching the bandwidth and power ceiling of copper. The opticalization of NVIDIA's NVLink is the single largest incremental driver over the next three years — if scale-up opticalization lands in 2027–2028, this tier alone will create a multi-billion-dollar market that does not exist today. This is also the primary driver of optical interconnect's cost share breaking through from 5–6% toward 10%.
Scale-across (inter-building interconnect): steady growth. When AI clusters are distributed across multiple buildings due to power and space constraints, DCI (Data Center Interconnect) modules are a structural necessity. Marvell's integrated coherent DSP solution holds a first-mover advantage in this segment. Growth is slower than scale-out and scale-up, but highly certain.
Assessment: whichever interconnect tier gets opticalized first is the largest incremental market. Scale-up opticalization is the key variable over the next three years.
4.2 Shifting Customer Landscape
The current buying structure is dominated by cloud providers. "Whoever decides, whoever pays" — the higher optical interconnect costs rise as a share of computing center capex, the more the paying party cares.
But this structure is being challenged from multiple directions:
- NVIDIA's role is evolving from supplier to integrator. It doesn't just sell GPUs — it manufactures its own modules, locks up laser capacity, and defines interconnect architectures. $4 billion invested in Coherent + Lumentum is not a financial investment; it is about securing supply and exclusivity. When NVIDIA's in-house module share exceeds 50%, it effectively becomes the world's largest optical module "buyer-manufacturer" — procuring from itself while defining market standards.
- Cloud providers are extending upstream. Meta sources glass substrates directly from Corning for $6 billion; Google self-develops OCS. The paying party does not want intermediaries skimming profits.
- Model companies as a potential wildcard. If a "token factory" business model emerges — model companies bypassing cloud providers to deploy inference capacity directly — the buying structure shifts yet again. The precondition is that model companies' inference revenue is sufficient to cover data center depreciation and operating costs; the current threshold is around 75% inference gross margin (Anthropic's 2028 target). If token prices fall faster than costs, the "token factory" is uneconomical; if the reverse, model companies have every incentive to build their own inference infrastructure and order directly from laser and silicon photonics wafer suppliers. This scenario in 2026 is only a possibility, but it may be the biggest variable reshaping the buying structure after 2028. "In the end, it's a power game over who controls the token exit."
- The survival question for domestic GPU companies. For most domestic GPU companies, "before talking about CPO, talk about whether you'll still be in business" — if the chip itself lacks competitiveness, no optical interconnect solution matters.
4.3 Supply-Demand Summary
Back to the original question: will silicon photonics capacity be oversupplied?
Segmented view:
- Pluggable modules: possible oversupply. China's production capacity is large, standardization is high — "any standardized product that China can manufacture at scale will inevitably face oversupply."
- Upstream optical chips, EML, silicon photonics wafers: persistent shortage. Gap of roughly 30–40%, persisting at least through next year. "The chokepoint shifts every two quarters."
- CPO capacity: far from sufficient. Yield is still low; volume production is at least 2–3 years away.
- Electrical chips (DSP/TIA): tightest bottleneck. Three-company monopoly with rising barriers.
Why is "the fear not oversupply, but becoming obsolete right when you've built it"? Because the technology generation cycle in silicon photonics is accelerating — the window from pluggable to NPO to CPO may be only 3–5 years. A company investing in an 800G pluggable line today may face NPO substitution pressure by 2028. Investment logic is replacing consumption logic: the competition is no longer about how cheap each module is, but whether you show up at the right technology node at the right time.
In this landscape, "getting a seat at the table" comes before parameter leadership. No top-tier cloud provider has ever had more than five optical module suppliers. The top-tier manufacturers obtain structural cost advantages through chip supplier lock-in. For every entrant — solve the qualification problem first, then debate the merit problem.
Closing: Three Layers of Judgment
Returning to the opening question: optical interconnect is transitioning from pluggable to NPO/CPO — what does this mean for the supply chain?
First layer: not every segment benefits simultaneously. Electrical chip barriers are rising, optical chip chokepoints are diverging, and the materials side is producing disruptors. The biggest beneficiaries are those standing at technology inflection points — if TFLN matures, the modulator segment gets reshuffled; if NPO yield crosses 80%, the NPO economics inflection point arrives; if scale-up opticalization lands, optical interconnect's cost share doubles. Each inflection point is a window.
Second layer: the route debate is not about technology, it's about industrial organization. Whether CPO or NPO wins does not depend on whose parameters are better, but on which model of industrial organization suits its soil. America's integration route requires a "master integrator"; China's decoupling route requires ecosystem collaboration. Both answers can be correct simultaneously.
Third layer: the biggest risk is not oversupply, it's mismatch. In a window of accelerating technology generations — pluggable to NPO to CPO may span only 3–5 years — the most expensive outcome is not building something nobody wants, but building something that gets replaced just as you finish. Investment logic has already replaced consumption logic. The survivors will not be the cheapest, but those standing on the right node.
Beneath these three layers of judgment, there is only one word: positioning.
Data sources: IMF World Economic Outlook (April 2026); Gartner IT Spending Forecast (April 2026); BloombergNEF AI Data Center Report (March 2026); Epoch AI Data Insights (May 2026); Futurum Group AI Capex 2026; LightCounting / TrendForce optical module market estimates 2025–2026; Forbes / Sacra / SemiAnalysis model company financial analysis; DeepFundamental optical module market analysis; ASML 2025 Annual Report and Q1 2026 Earnings; CNBC / BusinessWire / Meta / Tower Semiconductor / Coherent / Lumentum official announcements. Industry supply-demand data cross-verified with industry sources and not independently audited. This is not investment advice.
Data as of: July 15, 2026.