LIGHTMATTER

Three MIT PhDs — Nicholas Harris, Darius Bunandar, and Thomas Graham — spent years researching silicon photonics at MIT and realized that light-based computing could solve AI's biggest bottleneck: energy consumption. Training large AI models consumes megawatts of power, and most of that energy is wasted as heat in electronic interconnects.

Photonic interconnects could reduce that energy consumption by 10x while increasing bandwidth. They founded Lightmatter in 2017 to turn their MIT research into commercial products.

Builds photonic (light-based) computing chips and interconnects for AI data centers. Instead of using electrons to process data (like traditional chips), Lightmatter uses photons — particles of light — which can carry more data faster and with dramatically less energy.

The company sells both photonic AI accelerator chips and Passage, a photonic interconnect fabric that connects GPU clusters at the speed of light. Revenue comes from chip sales and data center infrastructure deals.

Passage (photonic interconnect fabric for connecting AI chips at the speed of light), Envise (photonic AI accelerator chip for neural network inference), and software tools for programming photonic hardware. Passage is the lead product — it plugs into existing GPU clusters and dramatically improves data transfer speeds between chips while cutting energy consumption.

Passage — the photonic interconnect product — is the entry point. Rather than asking customers to replace their entire compute stack, Lightmatter offers Passage as an upgrade to existing GPU clusters.

It connects Nvidia GPUs using light instead of copper, reducing latency and energy consumption. Once customers adopt photonic interconnects, upgrading to full photonic computing becomes a natural next step.

The strategy is brilliant: sell the bridge before selling the destination.

Photonic computing has been "5 years away" for 30 years. Convincing data center operators to adopt fundamentally new physics — replacing electrons with photons — requires proving reliability at scale, which takes time.

Manufacturing photonic chips at commercial volumes is also extremely difficult. And Lightmatter has to convince customers to change their entire data center architecture, not just swap out a chip.

Viking Global Investors, GV (Google Ventures), Hewlett Packard Enterprise, SIP Global Partners, and Spark Capital have backed Lightmatter.