Copper has a physics problem. At the interconnect densities required to link hundreds of thousands of GPUs inside a modern AI training cluster, copper wiring generates heat, loses signal integrity, and consumes power at rates that create a hard ceiling on scalable performance. This isn’t a cost optimization story. It’s an engineering constraint that gets worse as clusters grow larger.
That constraint is what makes NVIDIA’s photonics investment strategy legible. Since March 2026, NVIDIA has reportedly committed approximately $6.5B across five companies working on optical interconnect technology, according to TNW and GuruFocus. The investments aren’t product acquisitions. They’re supply chain capacity reservations, securing access to the materials and manufacturing infrastructure that will produce optical components at the volumes NVIDIA’s roadmap requires.
Before examining the investment map, one disclosure note: the individual investment figures as reported sum to approximately $7B, not the $6.5B headline total. The discrepancy, $500M, is unresolved from the available reporting. It likely reflects deal structures where some commitments aren’t fully disclosed, or where NVIDIA’s participation in a round is smaller than the round’s total size. Both the $4B figure from EE Times and the $6.5B figure from TNW and GuruFocus are presented here with their respective sourcing. Readers should treat individual figures as reported approximations, not confirmed primary disclosures. NVIDIA has not made an official consolidated investment announcement available in this reporting package.
The Investment Map
Five companies. Each fills a different position in the optical interconnect stack.
Coherent and Lumentum, reportedly approximately $2B each, are the laser source suppliers. Coherent and Lumentum are among the few companies with the manufacturing scale to produce indium phosphide (InP) lasers at volumes relevant to hyperscale AI infrastructure. InP is the semiconductor material that enables high-speed, low-power optical transmission at the wavelengths used in data center interconnects. By securing capacity commitments from both major suppliers, NVIDIA is, per analysts cited by EE Times, establishing preferential access to a constrained supply before competitors can do the same.
Marvell, reportedly approximately $2B, occupies the integration layer. Marvell specializes in co-packaged optics, the architecture that integrates optical components directly with compute dies rather than using separate optical modules. Co-packaged optics reduce the copper traces required between components, which is precisely the bottleneck the broader investment addresses. Marvell’s involvement suggests NVIDIA isn’t just buying laser supply, it’s investing in the integration technology that makes those lasers useful at the chip level.
Corning, reportedly approximately $500M in a long-term manufacturing agreement, is the fiber supply layer. Corning is the dominant supplier of optical fiber globally. A long-term manufacturing agreement here functions as a volume reservation: NVIDIA secures a priority position in Corning’s production capacity as fiber demand scales with optical interconnect adoption.
Ayar Labs, participation in a reported $500M Series E, represents the chip-level optical I/O layer. Ayar Labs is building optical I/O technology that embeds light transmission directly into silicon chips, eliminating the electrical-to-optical conversion step that currently adds latency and power consumption. This is the furthest out on the technology maturity curve of the five investments.
NVIDIA Photonics Strategy: Risk Assessment
AI Infrastructure Photonics: Who's Positioned Where
The combined picture: NVIDIA has capital commitments at the laser source layer, the integration layer, the fiber layer, and the chip-level optical I/O layer. That’s vertical coverage of the photonics supply chain from raw component to assembled interconnect architecture.
The Constraint That Capital Doesn’t Solve
InP fabrication is the part nobody mentions in the investment announcements.
Indium phosphide wafer fabrication faces severe global manufacturing constraints, per EE Times’ analysis. InP fabs are expensive to build, slow to expand, and require specialized expertise that doesn’t exist at the same depth as silicon semiconductor manufacturing. Unlike silicon wafer production, InP manufacturing hasn’t benefited from decades of yield improvement and capacity expansion driven by consumer electronics volumes. The result: InP laser production capacity is tight relative to the demand that hyperscale AI build-outs would require if photonic interconnects achieve broad adoption.
NVIDIA’s investments in Coherent and Lumentum don’t expand InP fab capacity, they secure priority access to existing and near-term capacity. That’s strategically valuable. It’s also a ceiling. If InP yield rates don’t improve, or if InP fab expansion timelines slip, even a well-positioned NVIDIA faces supply constraints on the components its photonics strategy requires. The 2028 window is the one analysts are watching: that’s the timeframe in which co-packaged optics and chip-level optical I/O need to demonstrate production viability at scale. Capital can accelerate. It can’t substitute for manufacturing process maturity.
Who Doesn’t Have This
The competitive read on NVIDIA’s supply chain strategy requires identifying who’s on the outside of these agreements.
AMD and Intel both have AI accelerator roadmaps that will eventually require the same optical interconnect infrastructure NVIDIA is locking in now. Neither company has announced comparable photonics investment commitments at this scale. Hyperscale cloud providers, AWS, Google, Microsoft, Meta, have their own photonics R&D programs, but the question is whether those in-house efforts can match the supplier capacity NVIDIA is reserving through external investment.
Analysts suggest, per EE Times’ framing, that the investment pattern functions as a supply chain moat: competitors who want InP laser capacity from Coherent and Lumentum, co-packaged optics from Marvell, or fiber from Corning may find that NVIDIA’s long-term agreements have priority claims on available production. That’s a structural advantage in AI infrastructure procurement that doesn’t show up in any benchmark comparison. It shows up in 2028 when clusters need to scale past the copper ceiling and the supply of optical components determines who can build them.
What to Watch
Analysis
The EE Times $4B figure vs. the TNW/GuruFocus $6.5B figure likely reflects different reporting windows, not an error in either publication. EE Times may have captured earlier-announced commitments; TNW and GuruFocus appear to include the full three-month consolidated total. Both figures are worth tracking, if NVIDIA makes a primary disclosure, comparing it against both journalism figures will clarify which reporting was more complete.
What to Watch
Three signals matter over the next 18 months. First: whether NVIDIA or any of the five partner companies makes an official investment disclosure that resolves the $6.5B vs. $7B discrepancy and confirms individual deal terms. Journalism-corroborated figures are useful context; primary disclosure is the basis for any investment or procurement analysis.
Second: InP fab expansion announcements from Coherent and Lumentum. Capital investment precedes capacity expansion; the operational indicator of whether NVIDIA’s strategy is working is whether these companies announce fab build-outs, not just investment receipts.
Third: co-packaged optics production timelines from Marvell. This is the integration layer that makes optical interconnects practical at the chip level. Slippage here would indicate that the photonics transition timeline is longer than current AI infrastructure roadmaps assume. For context on how this fits into broader AI infrastructure trends, see the HBM Supercycle brief and the SK Hynix iHBM brief, each identifies a different physical constraint that AI infrastructure scaling is running into simultaneously.
TJS Synthesis
NVIDIA’s photonics investment strategy is legible as a two-part move: solving a real engineering constraint (copper at hyperscale density fails) while creating a supply position that competitors can’t easily replicate (InP laser capacity is finite and now partly reserved). Both parts are credible. Neither is guaranteed.
The InP yield risk is genuine and unresolved by investment capital. The $6.5B doesn’t build new fabs, it buys priority access to existing and near-term capacity. If the broader AI infrastructure build-out accelerates faster than InP manufacturing can scale, NVIDIA will be better positioned than competitors, but everyone faces constraints. Enterprise infrastructure teams planning AI cluster expansions beyond 2027 should be tracking photonic interconnect availability as a procurement variable, not a background technology story. The transition from copper to light is coming. The question is whether it arrives on a timeline that matches current AI infrastructure build plans, and that question won’t be answered by investment announcements.