The Light Inside the Machine — How Photonics Is Becoming the Nervous System of AI

Every time you use an AI model, something invisible happens inside a data center. Thousands of chips talk to each other at extraordinary speed, moving enormous amounts of data in milliseconds. For decades, that communication happened over copper wire. Now, it’s happening over light. And the race to bring light all the way down to the chip itself is one of the most important technology battles of this decade.

What is opto-electronic technology — and why does it matter now?

An opto-electronic device converts between electricity and light. In the context of AI data centers, it means placing that conversion as close as possible to the AI chip itself — so data travels between chips not as electrical signals through copper, but as pulses of light through tiny glass channels called waveguides.

This matters now because AI has created a data movement crisis. The chips themselves are extraordinarily fast. The bottleneck is no longer computation — it’s communication. Moving data between chips fast enough to keep the processors fed. Copper, which has handled this job for decades, is struggling to keep pace.

The problem with copper — and where optical fits in

At the speeds modern AI demands, copper runs into three hard limits. Signal degrades with distance — a 200 Gbps signal travels perhaps one metre on copper before degrading; at 800 Gbps, that shrinks to centimeters. Power consumption — NVIDIA’s NVL72 system uses approximately two miles of copper cabling for 72 GPUs, all of it generating heat. Density — AI clusters now draw 100+ kilowatts per rack, and copper adds to that burden.

Light is faster, carries dramatically more data, and produces almost no heat in transit. But optical technology has historically operated at large scale — between cities, between buildings. The challenge is bringing it closer, all the way to the chip.

Scale	Distance	Status
Long-distance	Cities, continents	Mature, decades-old
Data centre racks	Metres	Well established, upgrading to 800G/1.6T
Chip-to-chip (CPO)	Millimetres	The new frontier — race underway

Co-packaged optics — placing an optical engine right next to the AI chip on the same substrate — is where the most intense competition now lives. The optical engine is not a cable. It is a component roughly the size of a postage stamp, containing a laser, modulator, waveguides, detectors, and control electronics, sitting millimetres from the GPU.

Question — Abhilash Gopinath

I had no idea AI data centers consume water at this scale. How does water get involved — and what does photonics have to do with it?

Answer

Every AI chip generates heat that has to be removed or the chips fail. The dominant method is evaporative cooling — water absorbs heat from the servers and evaporates into the air, consuming the water permanently in the process. A study by the Houston Advanced Research Center found that data centres in Texas alone will use 49 billion gallons of water in 2025, rising to a projected 399 billion gallons by 2030 — equivalent to draining the largest reservoir in the US by more than 16 feet in a single year. This is where photonics connects directly: copper interconnects generate heat; optical interconnects generate almost none. Every step towards replacing copper with light at the chip level reduces the cooling burden — and with it, the water consumption — of the data centres powering the AI era.

The supply chain — who makes what

Silicon photonics is not a single company’s product. It is a layered supply chain involving dozens of specialized players:

Layer 1 — Component suppliers
POET, Marvell (Celestial AI), Coherent, Lumentum, MACOM, Sivers — make the optical engines, lasers, and waveguides

↓

Layer 2 — Module makers
Foxconn, Luxshare, Lessengers — integrate optical engines into pluggable transceiver products

↓

Layer 3 — Network and server OEMs
NVIDIA, Marvell, Cisco, Arista, Dell, HPE, Supermicro — build the switches and servers that go into data centres

↓

Layer 4 — Hyperscalers
NVIDIA, Microsoft, Google, Meta, Amazon, Oracle — deploy at scale in AI data centres globally

The major competitive forces span this entire chain. Broadcom ($54B revenue) and Marvell ($50–60B market cap) design networking chips and have both moved aggressively into optical interconnects. Marvell in particular acquired Celestial AI — a well-funded optical startup — for up to $5.5 billion in late 2025, signalling how seriously the industry’s giants view this space. NVIDIA controls what gets co-packaged alongside its chips and consistently prefers large, established suppliers. At the other end, POET Technologies sits at Layer 1 — technically interesting, commercially precarious, and three handoffs from the end customer.

What happened between Marvell, Celestial AI, and POET

In April 2023, POET announced that Celestial AI had placed initial production purchase orders for its optical engines. For a company with under $1 million in trailing revenue, this was meaningful validation. In late 2025, Marvell acquired Celestial AI for $5.5 billion. POET’s customer became a division of a $50 billion competitor — but the orders remained in place. Marvell’s engineering team inherited the supplier relationship and chose to continue it.

Then, in the week before April 23, 2026, POET’s CFO gave an interview to the financial social network Stocktwits where he confirmed the $5 million order was from Celestial AI — now Marvell — and hinted the order had grown beyond the original figure. Purchase order details are typically confidential under NDAs. Major customers do not want competitors knowing what they are buying or in what volumes.

On April 23, 2026, Marvell formally cancelled all purchase orders with POET, citing a breach of confidentiality obligations. Four days later, POET issued a press release confirming the cancellation. The stock fell 46% in a single day — the largest decline in the company’s history.

What this doesn’t change: Celestial AI placed a $5 million order with POET in the first place. Marvell’s engineers reviewed the relationship after the acquisition and chose to continue it. The technology is real and has been commercially validated. The optical interconnect market is projected to reach $12 billion or more by 2030, and the structural forces driving it — AI’s appetite for bandwidth, copper’s hard limits, the environmental pressure to reduce data center heat — are unchanged. The question for POET is not whether the market exists. It is whether the company can rebuild trust, replace lost revenue, and win the design relationships that would make the technology matter at scale.

Sources: POET Technologies press releases · Bloomberg — AI data centres and water · Lincoln Institute of Land Policy · Environmental and Energy Study Institute