$9.5 Billion in 6 Months: The Photonic Interconnect Bet That Dwarfs the Market It's Replacing

Nine-point-five billion dollars. That is the approximate sum that NVIDIA, AMD, Marvell, and their co-investors committed to photonic interconnect technology between October 2025 and March 2026, according to a tally of disclosed deals. For context, the entire global silicon photonics market generated $1.8 billion in revenue in 2025, per Global Market Insights. Committed capital is 5.3 times the annual market. When the bets are five times bigger than the pot, something fundamental has shifted.

Here is what happened, in chronological order. In October 2024, Lightmatter raised $400 million at a $4.4 billion valuation to build photonic interconnects for AI data centers. In November 2025, Celestial AI raised $250 million for its Photonic Fabric technology, followed weeks later by Reuters reporting that Marvell was in advanced acquisition talks at a price north of $5 billion. On March 2, 2026, NVIDIA announced $4 billion in combined investments in Lumentum Holdings and Coherent Corp., $2 billion each, with attached multibillion-dollar purchase commitments. Two days later, Ayar Labs closed a $500 million Series E at a $3.75 billion valuation, with both NVIDIA and AMD signing on as strategic investors alongside Sequoia Capital and the Qatar Investment Authority.

Five deals. Six months. Combined disclosed capital: approximately $9.5 billion, not counting the Marvell-Celestial acquisition price, which would push the total well above $14 billion. Nobody in the industry is debating whether to switch from copper to light. It is negotiating the terms.

Why Copper Is Dying at AI Scale

Electrical signals traveling through copper traces between chips face three physics constraints that worsen at scale. First, signal degradation: copper interconnects lose signal integrity beyond approximately 2-3 meters, requiring signal repeaters that add latency and power. Second, heat density: higher currents through thinner copper traces generate resistive heating that compounds data center cooling costs. Third, bandwidth ceilings: NVIDIA's Rubin GPU architecture, its most advanced, supports 28.8 terabits per second per package on copper NVLink interconnects.

Ayar Labs' TeraPHY optical chiplets, by contrast, support more than 200 terabits per second aggregate bandwidth per package. That is roughly seven times the Rubin copper figure. Energy math is similarly lopsided: Ayar claims 4 to 20 times more compute throughput per watt compared to copper, depending on the configuration and distance.

Lightmatter's newly announced Passage L20 puts concrete numbers on the power equation: 6.4 terabits per second of optical bandwidth at a maximum thermal design power of 30 watts. That works out to roughly 213 gigabits per second per watt. A comparable copper interconnect running at similar bandwidth would consume an estimated 200-300 watts at the subsystem level, yielding roughly 96-144 Gbps per watt. Today's optical advantage is 1.5 to 2.2 times at the component level. At 100,000-GPU cluster scale, where interconnect power can represent 30-40% of total system power draw, the savings compound into megawatts.

Metric	Copper (NVLink 5)	Optical (Ayar Labs)	Optical (Lightmatter L20)
Bandwidth per package	28.8 Tbps	200+ Tbps	6.4 Tbps per module
Max reach without repeaters	~2-3 m	Meters to kilometers	Meters to kilometers
Throughput per watt	~96-144 Gbps/W	4-20x copper (claimed)	~213 Gbps/W
Interface standard	NVLink 5	UCIe	212.5G PAM4 SerDes
Production status	Shipping (Rubin)	Volume production 2027-2028	Sampling late 2026

The Investment-to-Market Ratio: A Pattern From History

A 5.3x ratio of committed investment to annual market revenue is unusual but not unprecedented. When NVIDIA committed heavily to CUDA tooling and GPU-accelerated computing between 2008 and 2012, the addressable GPU computing market was a fraction of NVIDIA's R&D spending on the software stack. That bet was not that the current market justified the investment. It was that the investment would create the market. Photonics today follows the same pattern.

Global Market Insights projects the silicon photonics market will grow from $2.3 billion in 2026 to $17.8 billion by 2035, a 25.3% compound annual growth rate. Intel holds roughly 21.5% market share today. Today's top five players (Intel, Cisco, Broadcom, Lumentum) hold a combined 58.6%. But the investment dollars flowing in are not going to the incumbents. They are going to startups (Ayar Labs, Lightmatter, Celestial AI) and to dedicated manufacturing capacity (Lumentum, Coherent). NVIDIA is building a parallel optical supply chain the same way it built a parallel AI software platform with CUDA.

Jensen Huang said it plainly in the Coherent announcement: the goal is "AI infrastructure at unprecedented scale, speed, and energy efficiency." Not a product improvement. An infrastructure replacement.

What the Lab Results Actually Show

While the industry pours billions into photonic interconnects, a separate research track is pushing light even further. Researchers at Xidian University in China published results in Optica in March 2026 demonstrating what they call the first large-scale programmable incoherent photonic neuromorphic computing system. Their system is built from two chips: a 16-channel photonic neuromorphic processor with 272 trainable parameters, paired with a distributed feedback semiconductor laser array.

What matters is that both linear and nonlinear computation happen in the optical domain. Previous photonic neural networks could handle the linear matrix multiplications that make up most of a neural network's work, but they had to convert signals back to electronics for the nonlinear activation functions. That conversion was the bottleneck, the reason photonic computing stayed in the lab. Xiang's team eliminated it.

"Previously, the nonlinear steps that make learning and decision making possible required the signal to be converted back into electronic signals," said research team leader Shuiying Xiang. "This adds delay and undercuts the speed and energy advantages of photonics."

To be clear: 272 parameters is microscopic compared to the billions of parameters in production AI models. This is a proof of concept, not a product. But it is the first demonstration that all-optical neural computation works end-to-end. If this architecture scales, the implications go beyond interconnects to the compute itself.

The Strongest Case for Copper

Copper is not going quietly. MaxLinear unveiled the Annapurna 224G retimer in March 2026, designed to extend copper reach in AI data centers. Marvell's own active copper cable designs target intra-rack connections where photonics remains overkill. For distances under one meter, inside a single server chassis, copper is cheaper, simpler, and well-understood. Nobody is replacing the traces on a motherboard with fiber optics.

Installed base matters too. Every existing data center runs on copper. Transition costs are real: new transceivers, new cables, new switch architectures, retrained technicians. Lightmatter's vClick Optics technology addresses some of this by enabling a "shift left" approach, integrating photonic testing earlier in the fabrication process, but volume production is still at least a year away. Ayar Labs projects volume production for 2027-2028. Until photonic modules ship at scale and at competitive per-port prices, copper retains the incumbency advantage that kept it dominant through the last three interconnect transitions (parallel to serial, copper to active copper, single-lane to multi-lane).

The strongest counterargument: copper has been declared dead before. In 2015, the consensus was that 100G Ethernet would require optics everywhere. Instead, copper pushed to 400G through PAM4 modulation and better signal processing. Every time the industry predicts a copper wall, engineers find another way to push electrons faster. Maybe 1.6T copper retimers will do the same.

What We Did Not Prove

This analysis aggregates disclosed deal values. The Marvell-Celestial AI acquisition has not been officially confirmed or priced. NVIDIA's purchase commitments to Lumentum and Coherent are multiyear and volume-dependent; the $4 billion equity investment is firm, but the total economic commitment could be substantially higher or lower depending on procurement volumes. Ayar Labs' throughput-per-watt claims are company-reported and have not been independently verified at production scale. The Xidian University results involve 272 parameters. Scaling photonic neural networks to useful size faces unsolved challenges in fabrication yield, thermal stability, and programmability. The 100,000-GPU cluster power analysis uses estimates; actual interconnect power allocation varies by architecture and workload.

What You Can Do

If you work in data center infrastructure, the signal is clear: start qualification testing with at least one photonic interconnect vendor now. Lightmatter's L20 samples in late 2026. Ayar Labs is in design-in phase with reference designs from Alchip and Global Unichip Corp. Waiting for volume production to begin evaluation means arriving 12-18 months late to a supply chain that will be allocation-constrained.

If you are an investor, watch the Marvell-Celestial AI deal closely. A confirmed acquisition at $5 billion or more would validate the market's pricing of photonic companies at 2-3x their most recent private valuations in under a year. The silicon photonics market is projected to grow at 25.3% CAGR through 2035, but the investment pattern suggests the market may front-load faster than consensus forecasts.

If you are a chip architect, the UCIe standard is the integration point. Ayar Labs' TeraPHY chiplets use UCIe to co-package directly with GPUs and accelerators. Designing your next-generation chip package with UCIe optical I/O slots is cheap insurance against a transition that now has $9.5 billion of committed capital behind it.

The Bottom Line

The semiconductor industry does not commit $9.5 billion to a technology in six months based on optimism. It does so when the alternative, in this case continuing to scale copper interconnects, has hit physical limits that no amount of signal processing can circumvent at the scale AI demands. Light is faster, cooler, and reaches further. It is also, for now, more expensive and less proven at volume. The next 18 months will determine whether the capital already committed is enough to cross the production valley, or whether photonics joins the long list of technologies that won the physics argument but lost the manufacturing one. Given who is writing the checks, I would not bet against the light.