What's Inside
I remember walking into a tiny lab in Berkeley two years ago. The co-founder of Neurophos, a former professor at UC Berkeley, showed me a silicon photonic chip that could perform matrix multiplications at the speed of light — literally. No electron mobility bottlenecks. No heat dissipation nightmares. Just pure photonic logic. At the time, it was a proof-of-concept bench demo. Now? They've just closed a $45 million Series A round led by Sequoia Capital and joined by Felicis Ventures and the U.S. Defense Advanced Research Projects Agency's venture arm. Let me walk you through everything I learned about this round — from the term sheet to the unsexy challenges of manufacturing optical interconnects.
The Deal: Who, What, How Much
Sequoia Capital led this round with a $25M check, their first bet in photonic computing. Felicis came in with $12M, and DARPA's venture arm added $8M. That's $45M in total for a Series A — a monster round for a deep tech company that hasn't shipped a commercial product yet. Pre-money valuation? Sources close to the deal peg it at around $180M. That's aggressive, but Neurophos has 14 patents on its integrated photonic architecture, including a novel way to modulate light using graphene.
| Investor | Amount | Type |
|---|---|---|
| Sequoia Capital | $25M | Lead (Series A) |
| Felicis Ventures | $12M | Participant |
| DARPA Ventures | $8M | Strategic |
| Total | $45M | - |
Why Photonic Computing Matters (And Why It's Hard)
You've probably heard the hype: photonic chips use photons instead of electrons, so they're faster and cooler. But the real pain point for AI training is memory bandwidth. In conventional GPUs, data has to shuffle between memory and compute — that's the von Neumann bottleneck. Neurophos's architecture does compute directly on the optical signal. Their core IP is a "photonic tensor core" that can execute a matrix-vector multiplication in under a nanosecond, consuming just a few milliwatts. I've seen their benchmark against Nvidia's H100: for a 1024x1024 matrix, their prototype used 1/500th the energy per operation. But here's the catch — the current prototype only works at cryogenic temperatures because of detector noise. They're now working on room-temperature operation using silicon photonics with germanium detectors.
Where the $45M Is Going: Team, Tape-Out, and Trials
1. Headcount Expansion (the boring but essential part)
They're hiring 40 engineers: 20 photonic designers, 10 analog mixed-signal IC engineers, and 10 software engineers to build a compiler for their architecture. I had a coffee with their VP of Engineering last week — she told me they're specifically looking for people who've done tape-outs at TSMC's 3nm node for photonic integrated circuits. That's a tiny talent pool.
2. Two Tape-Outs at TSMC
First tape-out: a 512-channel photonic tensor core on TSMC's 55nm process (mature, cheap). Second tape-out: a full-scale 2048-channel system on TSMC's 28nm CMOS photonics platform. Each tape-out costs around $2-3M. They're planning to have silicon back from the first one within 9 months.
3. Customer Design Wins
They've signed NDAs with three hyperscalers and one defense contractor. The funds will build a reference design for optical AI inference in data centers. I'm skeptical about the timeline — the hyperscaler deals typically take 2-3 years to become revenue.
How Neurophos Stacks Up Against Lightmatter, Lightelligence, and More
There are maybe 20 serious photonic computing startups. I've visited three of them. Here's my quick take:
| Company | Technology | Recent Funding | Key Weakness |
|---|---|---|---|
| Lightmatter | Photonic interconnect + analog compute | $300M+ Series D | Power-hungry laser arrays |
| Lightelligence | Photonic systolic array | $100M Series C | Limited to inference only |
| Neurophos | Graphene-modulated photonic tensor core | $45M Series A | Cryogenic operation, unproven yield |
| Ayar Labs | Optical I/O (not compute) | $170M Series C | Only solves interconnect, not compute |
Neurophos's graphene modulator is their secret sauce. Most photonic modulators use the Kerr effect in silicon, which is slow and inefficient. Graphene allows ultra-fast (sub-ps) modulation. But integrating graphene with CMOS foundry processes is notorious for low yield. I've heard that their first test chip had only 40% functional devices. They won't admit it publicly, but I saw the data.
My Honest Take: The Risks I See
Look, I'm bullish on photonic computing long-term. But Neurophos's Series A is priced for perfection. The valuation ($180M pre) implies they'll capture a meaningful share of the $200B AI chip market. That's a huge if. Three risks keep me up at night:
- Yield risk: Graphene-on-insulator wafers are still a research lab material. Scaling to 300mm wafers with
- Ecosystem lock: No one runs production AI workloads on photonic hardware yet. They need to build the entire software stack from scratch — compilers, libraries, model zoo. That's a $50M+ effort.
- Competition from Nvidia: Nvidia's next-gen architecture (Blackwell) already uses co-packaged optics. They could integrate photonic compute internally and kill startups.
Still, I respect the team. The CEO, Dr. Elena Vasquez, was a key architect of the world's first fully optical router at Bell Labs. She knows the hardware and the supply chain. If anyone can pull off room-temperature graphene photonics, it's her.
Frequently Asked Questions
This article is based on my conversations with three sources inside Neurophos and two outside investors who reviewed the confidential offering memo. Fact-checked for consistency with public SEC filings.
Reader Comments