The New Logic of Supply Chain Sovereignty

Coordination Through Capital

TL;DR

• By committing billions to Lumentum and Coherent, Nvidia is assuming the financial risk of a vertically integrated InP manufacturer without the control of owning the plants.

• The moat for these components is not software or talent. It is decades of tacit knowledge baked into specific physical facilities and equipment.

• Nvidia had to act to secure supply. AI demand is projected to consume potentially the entire current global supply of InP.

• The deal stratifies suppliers and is reminiscent of the automotive tiered supply chain structure. Unfunded suppliers making similar bets absorb all the downstream risks without the upfront compensation.

• Component suppliers that have the right technologies for the AI buildout should take note and move quickly: extract returns upfront, protect institutional knowledge, pursue diversification, and quarantine AI financials from the rest of the business.

In early March, Nvidia crossed a line the “fabless” model was supposed to prevent. By announcing simultaneous multibillion dollar strategic partnerships with Lumentum and Coherent, Nvidia moved to secure the development and manufacturing of indium phosphide (InP)-based lasers. This was not merely a procurement move. It was a recognition that in the AI age, the bottleneck has moved beyond the silicon to the optical interconnects that link it.

InP-based lasers are the core light source in next-generation co-packaged optics (CPO) interconnects. With technology from its NeoPhotonics acquisition, Lumentum holds a clear edge in the narrow-linewidth continuous-wave laser technology CPO requires. These slivers of InP chips will gate the performance of the next generation of AI scale-up and scale-out networks.

Nvidia has already architected its Spectrum-X and Quantum-X switches around Lumentum’s InP source and TSMC’s CoUPE platform. Coherent, meanwhile, has a world-class six-inch InP fab and is Lumentum’s primary challenger.

Each deal carries an identical structure: a $2 billion equity investment, a multibillion dollar multiyear purchase commitment, future capacity access rights, collaborative R&D, and funding for new U.S.-based manufacturing (here and here). Both are explicitly nonexclusive.

Capitalizing the Market Needed

The InP deals confirm a divergence in how AI hardware makers treat different classes of technology. In December 2025, Nvidia effectively “acqui-hired” Groq by licensing its technology and absorbing its key talent (here). That model works because the asset is almost entirely human capital. Process knowledge lives in the minds of engineers and transfers via employment contracts.

Electronic hardware and components are different. The true moat in InP manufacturing is tacit knowledge. It lives in the decades-long co-evolution of engineers, MOCVD reactors, and specific process calibrations. This intelligence is inseparable from the physical facility. While software talent is mobile, InP expertise is bound to its institutional history. Nvidia’s decision to fund these partnerships rather than build in-house was a recognition of this physical reality.

One might argue that simple prepayments, such as those used with TSMC, would suffice. Indeed in some component sectors they can. Samsung’s recent negotiating behavior illustrates the new dynamic. For decades, Samsung absorbed cyclical risk on its customers’ behalf. AI changed that. Samsung, with diversified DRAM and NAND customers, can now restructure relationships on their own terms, demanding prepayments for unbuilt fabs, locking customers into long-term agreements, and bundling HBM access with foundry commitments.

Lumentum and Coherent have no such leverage. Today, InP is a niche business. There are no credible high-margin and high-volume applications outside the AI sector. Automotive LiDAR is emerging and could offer volume, but not the margins. Consequently, Nvidia was forced to capitalize the market it required. Lumentum’s new fab is now a physical manifestation of Nvidia’s capital, while Coherent has been handed the resources to close the technical gap.

Asset Light, Liability Heavy

AI’s rapid innovation cycle is dismantling the “fabless” myth, the idea that value resides solely in software and design. We have entered an era of supply chain sovereignty. At the cutting edge, external suppliers, on their own, cannot match the pace of development. Qualifying and scaling physical production has become part of the competitive moat.

RCD Advisors has tracked this shift in our consulting work. In AI’s New Iron Age, we explored how technology leaders in the 1970s and 1980s internalized their supply chains. Those companies understood that managing system-critical components through a transactional merchant market was insufficient. It has come full circle. Today, AI servers are direct functional descendants of those Big Iron mainframes. Both have high-value silicon, proprietary interconnects, rack-level integration, and the same concentration among a handful of customers.

The fabless model was explicitly designed to shed supply-side capital investment exposure. Offload fixed costs. Reduce cyclical risk. Delegate manufacturing execution to the foundry. Compete on design and demand aggregation. But Nvidia is now funding supplier fabs, technologists, and capacity sized around its own roadmap.

The consequence is that Nvidia assumes the financial risk profile of a vertically integrated manufacturer while forfeiting some of the managerial autonomy of one. Nvidia carries the pre-paid exposure of a plant owner but lacks the internal governance levers to manage technology or market shifts. The result is Nvidia remains asset light, yet is increasingly liability heavy.

The logic of a buyer investing directly in a critical supplier is not new. RCD Advisors explored this continuum in Overhauling Modularization with Merchant SIP Processes. The automotive industry committed tens of billions into dedicated Li-ion battery joint ventures in the early 2020s. Apple invested nearly $3 billion in Corning since 2017 and made prepayments to both Lumentum and Finisar to develop the VCSEL supply base for Face ID. In 2012, Intel, TSMC, and Samsung collectively invested $5 billion in ASML alongside $1.7 billion in NRE funding to accelerate EUV development.

In each case, a technology change created a dependency. The system leader required a differentiated component that sat outside its core competency and could only be developed by the supplier. Capital followed that dependency.

At the AI frontier, system leaders depend on their suppliers’ technology more than ever. There is no rest for the pick makers during the gold rush. For now, the major AI hardware makers must incentivize the ecosystem without taking on the organizational burden or regulatory scrutiny of formal integration.

An Offer You Can’t Refuse

The market opportunity for InP laser chips in co-packaged optics is seductive. As AI data centers scale toward millions of GPUs, every accelerator will need optical interconnects for both scale-out networking and scale-up GPU-to-GPU links. RCD Advisors estimates a total addressable market of roughly $7 billion by 2030 across three technology scenarios, assuming 32 million AI accelerators deployed globally. We assume die volume and price per chip move in opposite directions across scenarios, allowing the math to converge. See the table below for other assumptions.

In wafer terms, the three scenarios translate to 84,000 to 278,000 six-inch equivalent InP wafers per year. RCD Advisors estimates total global InP wafer production today, across all applications, at roughly 250,000 six-inch equivalent wafers per year. CPO lasers alone could consume the current worldwide capacity of InP.

Half of that TAM, roughly $3.5 billion, is Nvidia’s projected volume. A large portion of that 2030 value is prepaid with the $2 billion deal. Nvidia’s actual procurement cost is lower. Nvidia is effectively buying down its own TAM by funding the capex and development overhead that would otherwise be embedded in the per-die ASP. That alone explains the deals with Lumentum and Coherent. 

All parties have emphasized that the deals are nonexclusive and they are free to supply others. But this is a red herring, or at most a concession to regulatory scrutiny. Nvidia has captured the merchant market not by banning competitors, but by buying its place at the front of two different lines.

There won’t be many other organizations getting in either of those lines. Broadcom, supporting Google’s TPU, already operates an InP fab in Fort Collins, the legacy of Avago. Huawei has domesticated its entire stack through Source Photonics and Hisense, building a parallel Chinese ecosystem effectively immune to Western export controls. AMD and Marvell, also relying on a “laser-fabless” model, cannot risk their roadmaps on residual capacity from their main competitor. They will almost certainly pursue deals with remaining suppliers like Infinera, Mitsubishi, Sumitomo, or Sivers, or change the CPO laser technology entirely. Lumentum and Coherent are now economically tied to Nvidia. In practice, there aren’t real alternatives.

Nvidia’s upfront funding changes the risk profile for each supplier, but it also locks down future margins. Suppliers who extract their returns upfront can leverage AI-funded InP development into adjacent markets or by bundeling other capability (package assembly). Those who don’t will find themselves dependent on a single customer with lousy margins.

Unfunded suppliers, making bets from their own free cash flow for a single AI customer, are at an enormous disadvantage. The automotive industry offers a forensic warning. Delphi and Visteon were 80% dependent on their former parent automakers for revenue, but were saddled with thin margins and little breathing room for investments. When both went bankrupt in 2009, the OEMs were forced to backstop them to prevent total collapse.

The Next Targets for Sovereign Investment

In a post-Moore’s Law world, the primary constraints on AI scaling have moved from the transistor to the rack. Interconnects, power delivery, and thermal management now form the binding constraints.

When a component dictates system performance, the merchant market structure loses its advantage. Technical interdependencies become the primary source of differentiation. Transactional purchasing gives way to something closer to a partnership, whether the contract says so or not.

“Partnership” technologies share four traits:

• Sudden Dependency: A rapid technical change creates a capability gap that sits outside the system architect’s core competency.

• System Differentiation: The technology directly shapes performance: bandwidth, latency, power density, or scaling behavior.

• Market Concentration: Only a small number of qualified suppliers can deliver the required performance at volume. A small number of buyers limits the suppliers’ ability to diversify their revenue base.

• Lock-In: No viable alternative exists at the required performance level, or co-development has embedded the supplier so deeply that switching costs are too high.

These four conditions move a transactional supply chain toward a vested relationship and attract the capital that underwrites it.

The vast majority of components in an AI server do not meet these criteria. The electronic supply chain remains largely competitive and well capitalized. Only a handful of components are coupled tightly enough to define differentiated performance.

For example, Nvidia’s 800V ecosystem development sets a common rack-level power architecture that a broad supplier base is designing into. Nvidia is unlikely to take equity positions in any of them, though pockets of asset specificity run deeper in the stack than the rack-level architecture suggests.

Within that stack, some components meet some criteria. Isolation and protection components optimized for AI rack profiles have limited alternative markets and real asset specificity. But the supply base is large and competitive, and lock-in risk is low. No orchestrator will fund development or expand capacity on their behalf. Design wins accrue to first movers who self-fund into the specification. That is an opportunity, not a candidate for supply chain sovereignty.

InP sat in that sweet spot across all four traits. The framework points toward other candidates, though the list will evolve as AI architecture does. Diamond heatspreaders and embedded power conversion technologies look ripe today. Both are emerging from rapid technical change that system architects cannot address internally. Both directly shape thermal and power performance at the rack level. Both have supply bases concentrated among two or three viable suppliers with limited ability to diversify. And for both technologies, the co-development required to optimize them to a specific architecture makes switching prohibitively expensive.

The Supplier Playbook

The AI hardware market is overtaking the rest of the electronics industry, but it hinges on hyperscaler capex and investor sentiment. The risk is real but so is the return. Lumentum and Coherent extracted their returns before the InP market stratified. If a component supplier discovers they have technology that meets these criteria, they should move quickly on four operational imperatives:

1. Keep the Recipe in the Building. The true moat in electronics and semiconductor manufacturing is not any single engineer’s knowledge. It is the decades-long co-evolution of process calibrations, equipment history, and institutional memory. A buyer can hire away talent. It cannot hire away historical experience and tuned equipment. Suppliers who bind their expertise to the physical asset force the buyer to fund the whole operation rather than poach the best people.

2. Make the Buyer Share the Risk. Capacity built primarily to a single buyer’s specifications carries real concentration risk. If the buyer changes course and strands your equipment, the contract should make them pay for it. That means change-of-spec penalties, minimum volume commitments, and exit fees sized to the actual stranded cost. The buyer who created the lock-in should underwrite it.

3. Make Bets on Diversification. A supplier’s negotiating power depends on having options. For InP specialists, the Optical Circuit Switch business represents a defensible independent revenue stream. Automotive LiDAR offers volume even at lower margins. These markets may never match AI datacenter returns, but they tilt the stakes when contracts get renegotiated. A supplier with one customer has no leverage. A supplier with three in different sectors has some bargaining position.

4. Quarantine the AI Business. Keep the AI-specific financials separate from the rest of the operation. A clean P&L positions the division for divestiture or acquisition if the cycle turns. Suppliers who blur the lines will find themselves absorbing AI’s volatility across the entire business. Those who keep it contained will have options.

Server makers are no longer buying parts. Across interconnects, power, and cooling, they are architecting a new Iron Age. Critical upstream technologies are being pulled into their orbit through capital and contracts rather than formal ownership. The result is vertical integration without its formal structure: control without ownership, obligation without authority.

For suppliers, the lesson is not loyalty but positioning. Protect the recipe. Extract the returns upfront. Build revenue in different sectors. Quarantine the AI financials. Those that do not will wake up as someone else’s cost center.

If you find these posts insightful, subscribe above to receive them by email. If you would like to learn more about our consulting practice and how we assist organizations in the Tech hardware supply chain, please get in touch with us at info@rcdadvisors.com.