Four billion dollars in a single week on two photonics companies and one optical startup. NVIDIA is buying its way into the physical layer that carries data between GPUs at the speed of light — because the copper wires inside AI data centers just hit a wall.
On March 2, NVIDIA announced two billion dollars in Lumentum and two billion in Coherent — equity stakes plus multibillion-dollar purchase commitments for advanced laser components. The next day, Ayar Labs closed a five-hundred-million-dollar Series E led by Neuberger Berman, with NVIDIA and AMD among the investors, at a valuation of three point seven five billion. Lumentum surged twelve percent. Coherent rose fifteen. The combined signal: four and a half billion dollars committed in forty-eight hours to companies that make things out of light.
This journal has tracked the largest capital expenditure cycle in technology history. The Foundation covered the six-hundred-and-fifty-billion-dollar bet. The Crowding Out followed the memory squeeze. The Hard Hat documented forty-one billion in construction. The Grid traced the parallel power infrastructure being built behind the meter. Each entry followed a different layer of the same build-out.
Every infrastructure project has the same structural pattern: the bottleneck never stays where you expect it. It migrates.
The Copper Wall
In 2023, the constraint was GPUs. NVIDIA could not manufacture H100s fast enough. Demand outstripped every production forecast. Then TSMC's CoWoS advanced packaging became the bottleneck — the GPUs could be designed but not assembled at scale. NVIDIA secured capacity commitments. The constraint moved again.
Now it has moved deeper into the physical stack: the wires connecting GPUs inside the data center.
AI training requires thousands of GPUs to synchronize constantly — sharing gradients, exchanging intermediate results, coordinating across a cluster that functions as a single machine. NVIDIA plans to scale from seventy-two to five hundred and seventy-six GPUs per system by 2027. The compute scales. The wiring does not.
At two hundred and twenty-four gigabits per second per lane — the current standard — passive copper cables can reach less than one meter before signal degradation makes them unreliable. Pushing further requires active signal conditioning: equalizers, amplifiers, retimers. Per-port power consumption hits thirty watts. Electromagnetic interference worsens with density. Insertion loss increases with frequency, forcing cables to be shorter and thicker at precisely the moment when clusters need to be larger and denser.
The GPUs are fast. The copper between them is not. At terabit-per-second data rates, the physics of electrical interconnects becomes adversarial.
What Light Solves
Co-packaged optics eliminates the electrical path between switch and transceiver. Instead of sending electrical signals across a circuit board to a pluggable module on the front panel, CPO integrates the optical engine directly onto the switch package. The conversion from electricity to light happens millimeters from the silicon die instead of centimeters away.
The numbers are not incremental. Per-port power drops from thirty watts to nine — a three-and-a-half-times improvement. Signal integrity improves sixty-four-fold. Resilience improves tenfold. NVIDIA's Quantum-X InfiniBand switches, built on co-packaged optics, deliver one hundred and fifteen terabits per second across one hundred and forty-four ports at eight hundred gigabits each. They are shipping now. The Spectrum-X Photonics Ethernet switches follow in the second half of 2026.
Broadcom is already in production. Its Tomahawk 6 Davisson — the industry's first hundred-terabit Ethernet switch with co-packaged optics — began shipping to early access customers in October 2025. Two design philosophies are competing: Broadcom permanently bonds the optical engines to the substrate, trading field replaceability for integration simplicity. NVIDIA uses detachable optical sub-assemblies, allowing failed optics to be swapped without replacing the entire switch. The architectural debate echoes through every infrastructure transition — modularity versus integration, serviceability versus density.
The startups tell the same story from a different angle. Ayar Labs builds optical interconnect chiplets — its TeraPHY engines claim four to twenty times more computing throughput per watt than copper. Lightmatter has raised eight hundred and twenty-two million at a four-point-four-billion-dollar valuation for three-dimensional photonic interconnects. Celestial AI raised five hundred and ninety-four million for silicon photonics memory and interconnect. The venture market has placed over two billion dollars on the thesis that light replaces wire.
The Supply Chain Play
The four billion dollars is not primarily a financial investment. It is a supply chain lock-up.
The critical bottleneck component is the indium phosphide laser — the light source at the heart of every optical transceiver and CPO engine. Current demand for InP lasers exceeds supply by a factor of two. There are only a handful of fabrication facilities in the world capable of manufacturing them at the required quality and volume. NVIDIA's purchase commitments at both Lumentum and Coherent secure priority access to capacity that does not yet exist, while the equity investments fund the U.S.-based manufacturing expansion required to build it.
The structure is familiar. When CoWoS packaging constrained H100 production, NVIDIA committed capital to expand TSMC's capacity. The playbook repeats: identify the binding constraint, secure supply before competitors recognize the bottleneck, fund the capacity expansion that resolves it. Dual-vendor, nonexclusive, but with priority access rights. The deals ensure NVIDIA gets its components first.
The geopolitical dimension adds another layer. China controls a significant share of global indium supply — the rare metal from which indium phosphide is synthesized. NVIDIA's emphasis on U.S.-based manufacturing at both Lumentum and Coherent is not incidental. The AI supply chain's geopolitical surface area keeps expanding: TSMC in Taiwan for fabrication, SK Hynix in South Korea for memory, and now indium sourcing that traces back to Chinese mines. Each layer deeper into the stack introduces a new geographic dependency.
The Migration
The optical interconnect market for AI data centers reached ten billion dollars in 2025. It is projected to exceed thirty-one billion by 2033. The co-packaged optics sub-market — two billion today — is forecast to reach twenty-five billion over the same period, growing at forty percent annually. These are not speculative projections about a technology that might matter. They are tracking curves for hardware that is already shipping.
But the deeper pattern is not in the market size. It is in the direction of the migration.
Each bottleneck NVIDIA has resolved has been further down the physical stack and involved a smaller, more specialized industry. GPUs are designed by thousands of engineers and fabricated in facilities that cost tens of billions. Advanced packaging is done by a handful of foundries. Indium phosphide lasers are made by a few dozen companies worldwide. The supply constraints get harder to resolve precisely because the industries making the critical parts are smaller, more concentrated, and less capitalized. A four-billion-dollar investment in photonics is transformative to Lumentum and Coherent in a way that a four-billion-dollar investment in TSMC would not be.
NVIDIA began as a graphics card company. Then a GPU compute company. Then a networking company when it acquired Mellanox for seven billion in 2020. Now it is becoming a photonics company — not by building lasers itself, but by financing the entire supply chain required to build them.
The trajectory is not toward more compute. It is toward controlling every physical layer required to move data between compute. The bottleneck migrates. NVIDIA follows it. And now the constraint is literally about light — how fast data can travel between machines, through glass fibers thinner than a human hair, at three hundred million meters per second.
The bet is no longer just on intelligence. It is on the fiber that carries it.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)