Photonics Is the Future of Computing

TL;DR

• Why the build-out is happening: the run time over which a model has a 50% chance of completing a hard task keeps growing exponentially and now sits at roughly 24 hours, with no saturation visible in the data. The constraint is power — the US grid adds only a few percent of capacity per year, while the frontier-AI power curve smashes straight through it. A gigawatt is 1,000 racks; New York City is about 7 GW; that scale is becoming an average Texas data center.
• Interconnect, not compute, is the bottleneck: how GPUs and XPUs talk to each other is the principal challenge in scaling agentic AI workloads. At the limit of zero latency and infinite bandwidth, a GPU meters away is indistinguishable from a local one — that's strong scaling, thousands of chips acting as one.
• Measured payoff and hardware: same GPUs, same workload, 3× faster time to train; 3× pre-fill and 11× decode interactivity (tokens/sec/user) for inference. The M1000 moves 114 Tb/s (state of the art ≈ 10 Tb/s) over 1.6 Tb/s fibers (state of the art ≈ 0.2 Tb/s), and a single M1000 rack carries several petabits/sec of I/O.

Key Points

The build-out and the power ceiling (~00:24:50–00:27:30)

The world is undergoing a massive AI build-out: data centers in Texas and British Columbia now draw as much power as the largest cities on Earth. New York City is around 7 GW — soon a fairly average data center in a place like Abilene, Texas.

The justification is capability growth. Model run time — how long a model can work on a complex task with a 50% chance of finishing — has been growing exponentially and now sits near 24 hours. Harris sees no saturation in the data, and frames the whole build-out as compressing the time it takes to do very hard, very valuable work.

Supply can't keep up. US grid capacity grows a few percent per year; frontier AI power consumption blows past it. Nuclear is one answer (roughly ten 100 MW reactors per gigawatt data center), but he first wants the audience to feel the scale: a house is a few kilowatts, a rack is approaching a megawatt, a thousand racks is a gigawatt — and the conference hall they were sitting in would hold a few hundred racks.

Why interconnect: strong scaling and the "single brain" (~00:28:00–00:31:00)

Given ten minutes, Harris jumps to the conclusion: interconnect is the principal challenge. Lightmatter connects GPUs and chips with light, at 1.6 Tb/s per waveguide/fiber — "1,600 homes' worth of bandwidth in a single optical fiber." The results: 3× faster time to train on the same GPUs running the same workload, 3× on inference pre-fill, and 11× interactivity (tokens per second per user) on decode.

The underlying goal is systems that behave as a single brain — strong scaling in the computer-science sense. Ideally two chips on one workload give you two units of performance and a thousand chips give you a thousand; in practice most workloads aren't embarrassingly parallel and need synchronization and shared state. At the limit you want zero latency and infinite bandwidth, at which point GPUs many meters apart are indistinguishable from local ones. Lightmatter is approximating that perfect interconnect, and Harris expects the first systems where a thousand GPUs or XPUs act as one giant XPU within the next two years.

The hardware, and reliability (~00:31:00–00:32:50)

• M1000: billed as the fastest optical communication device in the world at 114 Tb/s, with optical fibers along the top and bottom edges moving 1.6 Tb/s each. State of the art is around 10 Tb/s per device and 0.2 Tb/s per fiber — "there's 10x's all over the place." The transatlantic cables connecting North America to Europe run about 200 Tb/s, less than two of these chips.
• Racks: a single M1000 rack carries several petabits per second of I/O — the entire world's internet traffic in one rack.
• Reliability: a major scaling challenge. A data center with 100,000 GPUs has on the order of 10 million links, and those links had better not produce errors — errors freeze training runs and crash inference. Lightmatter therefore builds entire validation data centers, operating hundreds to thousands of these platforms to prove they simply won't crash.

Quotes

"I don't know anybody who can work for 24 hours non-stop and have a 50% chance at solving a hard technical problem. Certainly I can't do it." (~00:26:00)

His framing of the time-horizon curve — the curve that motivates the entire build-out.

"It enables you to build gigantic computer systems that behave as a single brain." (~00:29:33)

Lightmatter's one-sentence answer to "what does photonics buy you?"

"The entire world's traffic, internet traffic, is a single rack from Lightmatter M1000 there on the left." (~00:31:54)

A comparison built to be remembered.