A version of this post previously appeared in a LinkedIn newsletter

All eyes this week on are Taiwan, where every major semiconductor company is announcing.. well… more silicon. Compute is sexy and interesting, but it’s only as useful as the fabric connecting it all together.

Copper interconnects, which have served as the backbone of high-performance computing for decades, are running out of runway. At 200 Gb/s per lane, copper’s effective reach in active electrical cables drops to roughly three meters, making large-scale, multi-rack GPU clusters physically difficult to build and prohibitively expensive to operate. Moving electrons at that density over those distances consumes power at a rate that data center operators can no longer absorb.

The numbers are stark. A conventional 1.6 Tb/s pluggable optical transceiver draws about 30 watts, with the digital signal processor inside consuming more than half of that power budget. Multiply that across tens of thousands of switch ports in a hyperscale AI factory, and the power consumed by networking infrastructure alone runs into the tens of megawatts per facility. 

For an industry already straining against grid capacity and cooling constraints, networking has become as much of a bottleneck as compute.

NVIDIA clearly recognizes the implications and is actively working to replace electrons with photons across the entire connectivity stack, from co-packaged optics at the switch level to optical fiber cabling throughout the rack and facility.

The transition is neither cheap nor simple, but from NVIDIA’s perspective, it’s too important to leave to the broader market to figure out. This is why, between March 2025 and May 2026, the company committed more than $4.5 billion in investments and partnerships to secure the supply chain for that photonic future. 

The moves span silicon photonics chip manufacturing, laser sourcing, fiber production, system integration, and domestic manufacturing capacity. Taken together, they represent one of the most deliberate and comprehensive infrastructure supply chain strategies the semiconductor industry has seen in years.

GTC 2025: Spectrum-X and Quantum-X Photonics Switches

NVIDIA’s focus on photonics was unmistakable at its GTC 2025 conference, when CEO Jensen Huang announced the Spectrum-X Photonics and Quantum-X Photonics silicon photonics networking switches, alongside a named ecosystem of suppliers spanning foundry, packaging, lasers, fiber, connectors, and assembly. 

Jensen characterized the development as “the fusion of electronic circuits and optical communications at massive scale.”

The two switch platforms share the foundational photonics architecture but serve different networking protocols and deployment timelines:

• Quantum-X Photonics (InfiniBand): Delivers 144 ports at 800 Gb/s using 200 Gb/s SerDes, with a liquid-cooled design to manage the thermal load of onboard silicon photonics. Scheduled for availability in the second half of 2025. Each Quantum-X CPO package integrates 18 silicon photonics engines, enabling 324 optical connections and 288 data links from 36 laser inputs, with six detachable optical subassemblies each providing 4.8 Tb/s of aggregate throughput.
• Spectrum-X Photonics (Ethernet): Delivers configurations ranging from 128 ports at 800 Gb/s (100 Tb/s total) to 512 ports at 800 Gb/s (400 Tb/s total). A four-ASIC chassis design delivers 409.6 Tb/s of aggregate bandwidth. Each switch package integrates 36 silicon photonics engines (32 active, four reserved for redundancy) in a single compact footprint. Scheduled for the second half of 2026.

Compared with conventional pluggable transceivers, NVIDIA claims the CPO architecture delivers 3.5x lower power consumption (from 30W per 1.6T port to 9W), 63x higher signal integrity, 10x better network resiliency, and 1.3x faster deployment.

The 4x reduction in the number of lasers required, relative to an equivalent pluggable transceiver deployment, is a central efficiency driver, achieved by using external laser sources that feed multiple silicon photonics engines. This differs from more traditional approaches that embed a laser in every transceiver.

Co-packaged optics eliminates the DSP retimer, which accounts for most of the power draw in a pluggable transceiver. By placing the silicon photonics engine directly on the same substrate as the switch ASIC, the electrical path between them shrinks from centimeters to millimeters, reducing signal loss by approximately 4 dB and eliminating the need for power-hungry signal conditioning.

When a pluggable transceiver fails in a conventional deployment, recovery requires manual intervention that takes hours. NVIDIA’s CPO architecture uses a simpler component structure with a lower failure probability, and its detachable optical subassemblies support field serviceability without taking the entire switch offline.

NVIDIA’s Optics Supply Chain

NVIDIA’s silicon photonics strategy requires a vertically integrated supply chain that did not previously exist at the required scale or to the required specifications. At GTC 2025, NVIDIA named 11 ecosystem partners, each covering a functional layer of the CPO stack. 

NVIDIA’s photonics supply chain is divided into three functional tiers:

Photonic Integrated Circuit Manufacturing

TSMC fabricates silicon photonic integrated circuits for NVIDIA using the COUPE process, integrated with SoIC-X 3D chip stacking. This foundry capability did not exist in a production-ready form before this collaboration.

TSMC has committed to completing COUPE validation with support for small-form-factor connectors by the end of 2025 and to full integration into CoWoS packaging by 2026. Future platform iterations target bandwidths of 6.4 Tb/s and, as the technology matures, eventually 12.8 Tb/s.

Laser and Optical Components

Coherent and Lumentum are the two primary suppliers of lasers and optical components, with roles initially defined at GTC 2025 and dramatically deepened by recent (March 2026) financial commitments between NVIDIA and the companies.

At GTC, Lumentum provided lasers for the Spectrum-X product while Coherent collaborated on silicon photonics components.

Browave, SENKO, TFC Communication, and Corning handle fiber array units and optical connectors that link the external laser sources to the silicon photonics engines and the switch front panel.

Packaging, Assembly, and Integration

Beyond the optics itself, NVIDIA relies a broader ecosystem of packaging, assembly, and integration partners to bring its switches to market:

• SPIL handles advanced semiconductor packaging for the NVIDIA CPO multi-chip module, including wafer bumping, assembly, and test.
• Fabrinet provides contract manufacturing for CPO assemblies and sub-assemblies.
• Foxconn handles system-level CPO integration into switch chassis.
• Sumitomo Electric provides additional fiber and connector components.

NVIDIA’s $4 Billion Laser Bet (March 2026)

On March 2, 2026, NVIDIA announced simultaneous strategic investments in Coherent and Lumentum, committing $2 billion to each company. Both multiyear, nonexclusive agreements include a multibillion-dollar purchase commitment from NVIDIA and future access and capacity rights to advanced laser and optical networking products.

The investment capital is explicitly earmarked to support R&D, future capacity, and operations as both companies build out U.S.-based manufacturing capabilities.

Lumentum had already begun expanding U.S. manufacturing capacity for CPO-relevant lasers before the investment. NVIDIA’s capital secures capacity commitments on that roadmap and funds further expansion, including a new domestic fabrication facility.

The company’s CEO, Michael Hurlston, described the deal as reflecting a “shared commitment to advancing the optics technologies that will power the next generation of AI infrastructure.”

The Coherent agreement extends and deepens a 20-year supplier relationship. Coherent’s own 8-K filing confirms that the partnership expands NVIDIA’s access to five additional product families for co-packaged optics, in addition to the laser components already in use.

CEO Jim Anderson described the relationship as underscoring “Coherent’s role as a key enabler of next-generation AI data center infrastructure.”

The logic behind the laser investments becomes clear in the context of CPO architecture. Conventional pluggable transceivers embed a laser in each port module, creating a distributed laser inventory that scales linearly with port count.

NVIDIA’s CPO design uses external laser sources shared across multiple silicon photonics engines, dramatically reducing the laser count relative to the port count. That architecture creates a concentrated, high-specification laser supply requirement that Coherent and Lumentum must meet at production scale.

The moves show NVIDIA underwriting the manufacturing infrastructure required to supply those components at the volumes that million-GPU AI factories will demand.

Corning and the Fiber Backbone (May 2026)

In May 2026, NVIDIA and Corning announced a multiyear commercial and technology partnership spanning the optical fiber and connectivity layers of AI data center infrastructure.

Corning is the inventor of low-loss optical fiber, having pioneered the technology in 1970, and currently supplies fiber cabling to all major hyperscale data center operators.

The deal saw NVIDIA purchase $500 million in warrants for Corning shares, with the structure including a pre-funded warrant for up to 3 million shares at a nominal price and a traditional warrant to purchase up to 15 million additional shares at $180 per share.

The total potential investment value reaches $3.2 billion if all warrants are exercised, making this NVIDIA’s largest single optics commitment by potential financial exposure.

Corning’s obligations under the partnership are concrete and measurable:

• Build three new advanced manufacturing facilities in North Carolina and Texas dedicated to optical networking products for AI data centers
• Increase its U.S.-based optical connectivity manufacturing capacity by 10x
• Expand U.S. fiber production capacity by more than 50%
• Create more than 3,000 new American manufacturing jobs

The Corning partnership addresses a part of the infrastructure stack that co-packaged optics does not replace: the optical fiber interconnects running between racks, between rows, and across data center floors. As AI factories scale from hundreds to hundreds of thousands of GPUs, the volume of fiber required increases dramatically.

Today’s AI clusters already use fiber to connect GPU racks at the scale-out layer, where distances make copper impractical. Corning’s Flow Ribbon Technology cables, which use compressed air to rapidly install into microduct infrastructure, are already deployed in major AI data centers.

The partnership with Corning allows NVIDIA to secure preferential supply and to drive technology development for the next generation of those products.

Scale-Out First, Scale-Up Follows: The Deployment Roadmap

Scale-out networking connects separate GPU clusters within and across facilities, over distances that already require optical fiber. Copper has largely been phased out of the scale-out layer in production AI deployments. NVIDIA’s Spectrum-X Photonics (Ethernet) and Quantum-X Photonics (InfiniBand) switches target this layer, bringing CPO to the switch ASIC itself rather than relying on pluggable transceivers for electrical-to-optical conversion. 

NVIDIA’s Quantum-X Photonics switch was the first to market, released in 2025. The Spectrum-X Photonics switch is scheduled for the second half of 2026. Both are designed for AI factory scale-out: connecting GB300 and Vera Rubin racks across multi-building campuses.

Scale-up networking is more demanding and more consequential. NVLink, NVIDIA’s proprietary scale-up interconnect, connects GPUs within a rack and across adjacent racks, forming a shared memory domain.

The current Blackwell-generation NVL72 rack uses NVLink over a passive copper backplane within the rack, a practical design because distances remain under two meters.

The Vera Rubin NVL144 (Kyber) architecture supports 144 GPU packages in a single rack, where copper remains viable. Vera Rubin Ultra NVL576 scales to 576 GPUs across eight racks and will use a combination of copper and direct optical NVLink connections.

At this scale, copper’s physical limitations in signal integrity and reach force the transition.

The Feynman generation, targeted for 2028, is when CPO enters NVLink at volume. NVIDIA has confirmed that Feynman systems will be available with either copper or co-packaged optical NVLink interconnects.

The 2028 timeline gives NVIDIA two full platform generations to validate CPO in scale-out networks before deploying it in the more critical, lower-latency scale-up environment.

The investments in Coherent, Lumentum, and Corning give NVIDIA the laser and fiber supply chain that Feynman CPO’s NVLink will require.

Competitive Landscape

NVIDIA’s CPO push accelerated a market transition Broadcom has advocated since 2021. Broadcom’s Tomahawk and Jericho switch families have been on the CPO roadmap for several years, and the company projects that the CPO market will generate meaningful revenue in the 2027-2028 timeframe. NVIDIA’s entry validates that roadmap and compresses the commercialization timeline.

Marvell also has a strong play in this space. The company acquired Celestial AI specifically to offer CPO solutions and projected CPO revenue to reach a $500 million ARR in its fiscal Q4 2028, before doubling to $1 billion by fiscal Q4 2029.

Marvell and NVIDIA also announced a strategic partnership that connects Marvell to NVIDIA’s AI factory ecosystem through NVLink Fusion. This creates a complex competitive dynamic, as Marvell builds its own CPO switches while also partnering with NVIDIA on interconnect standards.

Credo Technology similarly added silicon photonics to its portfolio through an acquisition, expecting $500 million in optical revenue in fiscal 2027.

NVIDIA’s decision to develop its own CPO switch architecture rather than purchase CPO-enabled ASICs from Broadcom or Marvell means it is competing at the silicon level, not just the system level. This is a significant commitment that requires sustained investment in photonics engineering capabilities NVIDIA previously lacked at this scale.

For the optical transceiver market, NVIDIA’s CPO strategy introduces a measured disruption. NVIDIA has been explicit that pluggable optical transceiver technologies will continue alongside CPO, with Coherent, Eoptolink, Fabrinet, and Innolight named as pluggable transceiver partners in the GTC 2025 announcement.

The transition from pluggable to co-packaged optics will span multiple product generations, and organizations that have invested heavily in pluggable infrastructure will not be forced to migrate.

The 2027 timeframe, when both NVIDIA and Broadcom are expected to have 200G per-lane CPO switches in production, is when the market begins making structural decisions about which architecture to standardize on for new builds.

Analyst’s Take

NVIDIA’s optical strategy is, at its most fundamental level, an infrastructure prerequisite play. The company’s GPU roadmap requires networking infrastructure that does not yet exist at commercial scale, so NVIDIA has chosen to fund its creation rather than wait for the market to build it independently.

The $4.5 billion the company committed between March 2025 and May 2026 is the cost of ensuring the supply chain can support million-GPU AI factories within the planning horizons of major hyperscale customers.

The Corning partnership highlights the optical opportunity in terms that extend beyond photonics specialists. Corning’s fiber is used in every major AI data center operating today, and the May 2026 agreement locks NVIDIA into Corning’s domestic expansion roadmap while giving NVIDIA upside exposure through warrants if Corning’s revenue grows in line with AI infrastructure demand.

It also positions NVIDIA ahead of potential supply constraints as Meta‘s $6 billion, five-year fiber commitment to Corning (announced in January 2026) and similar hyperscale agreements compete for Corning’s constrained U.S. manufacturing capacity.

NVIDIA’s warrant structure and three dedicated new facilities ensure NVIDIA-specific demand is served before general market allocation. It’s a smart play by the company.

For enterprise IT and infrastructure buyers, the near-term watch items are straightforward. Quantum-X Photonics InfiniBand switch availability in the second half of 2025 establishes the first production proof point for NVIDIA’s CPO architecture, while Spectrum-X Photonics Ethernet in the second half of 2026 brings CPO to the broader Ethernet-based AI factory market.

Organizations planning data center builds or expansions in 2027 and beyond should treat optical-readiness, including fiber density, connector standards, and transceiver roadmap flexibility, as first-order planning considerations rather than as future concerns.

The deeper strategic inflection arrives with Feynman NVLink CPO in 2028. If CPO delivers on its power efficiency and resiliency claims in production-scale-out networks over 2026 and 2027, NVIDIA will have the validation it needs to push CPO into NVLink and fundamentally change the economics of GPU scale-up.

At that point, the architectural advantage of building AI factories around photonic interconnects, from rack to campus, becomes the default. The companies in NVIDIA’s supply chain and the data center operators who have built optical-ready infrastructure will scale without constraint. Everyone else will retrofit.