Patrick was involved in many of the early Meta spinouts. Received TS from Gates Frontier; technically oversubscribed. Michael: 4x founder in Silicon Photonics with ~$1B aggregate exit value, involved with 30–40 silicon-photonics startups.

Headline: take on NVIDIA head on. Market framed as $14B AI inference at 37% CAGR (pre-ChatGPT); power can be over 50% of datacenter opex; neural nets growing much faster than Moore's law. Most prominent customer engagement is James Hamilton from Amazon — the #1 buying-decision driver is speed per unit dollar, energy second.

Neurophos technology uses in-memory compute via a metasurface processor: 8,000x denser than silicon photonics elements, lower loss because it's a single-scattering-event processor, even power distribution across the array. Analog compute solves the speed-of-compute problem. Purpose of the round: 12-month IMEC process program then Series A to develop the electronic circuit.

Team notes: Dave Baker likely to join. Board: Nathan Kundtz, Chris Allegro, David Smith. Michael pressed on tape-out experience across all three technologies (silicon photonics, metasurface array, control) and whether Patrick has taped out at IMEC ratios. VP of Engineering candidate from IMEC not yet fully onboard.

Michael's dollar concerns: typical ASIC tape-out over $100M on its own; IMEC is a prototyping facility so transfer elsewhere is required; that transition is slow. "We're talking $100M–$200M to get to market." Everyone Michael knows who has done commercial development at IMEC says it takes too long.

Software strategy: needs fundamentally different compilers, not just PyTorch primitives. PCI/HBM bandwidth question: even if PCI is solved, how do you keep the GEMM fed? NVIDIA already maxes out HBM bandwidth.

Customer strategy: hyperscalers self-select for low risk tolerance and won't buy from a startup on a 5-year timeline; wedge customer unclear. High-frequency trading floated but not scoped — every millisecond of latency is worth incredible amounts of money. Michael strongly encouraged Patrick to spend a large portion of his time speaking to customers now.

Founding ownership: Patrick 30%, Tom 10%, Andrew 6.5%. Michael's overall read: "This is going to be very expensive; that's just development, that's not ramping production." Gates does not invest for economic return out of this fund; family office follow-on is possible. Assigns 10–20% odds "if they're right about all those things" — and calls the story genuinely interesting.

Send me the invite and I'll try to join if I can. This doesn't sound investible to me, looking at their answers to your questions. What they're saying varies from not addressing the question to actually incorrect to missing the most important points to totally nonsensical.

Just as a starting point, the reason that GEMM units aren't most of the area on a die for AI is that the computations are limited by other things. Keeping the GEMMs fed is one of the major challenges, so there's a balancing act. Building a giant GEMM engine is all well and good, but how are you going to keep it fed? If this card sits in a PCIe socket, the PCI-E bandwidth is going to limit how much data can be fed into the GEMM, which will be the speed-limiting activity.

And customers do very much care about power, insofar as you have to fit into their existing infrastructure. Or you're dead.

Who is the customer? Hyperscalers who already have their own hardware? They're among the hardest places in the world to sell into as a startup. Like specifically who is the real customer?

Also, the idea of complete generality is foolish in the extreme. Inference, video, and training are completely different markets, running on completely different hardware. Which one are they targeting for a real advantage on a real algorithm? Have they even done this analysis? Doesn't sound that way if they're just talking about TOPS on a GEMM.

Hi Miles, apologies for the delay — we've been scrambling through round details.

Technology evolution: We realized that a product in the AI datacenter inference space must (1) be general enough to address all existing and future hypothetical AI algorithms, and (2) the #1 KPI for customers will be speed per dollar, not speed per watt. While people talk a big game about energy efficiency, in today's race to capture the applications that will scale, the service with the most and most-affordable inference wins the customer accounts — that drives hardware buying decisions. That said, you cannot go 100x faster without improving fundamental energy efficiency, because 100x speed would imply 100x power. As we came to understand this, we rearchitected the product around both goals. Current vision: a general matrix multiplication (GMM) engine starting at 50,000 TOPS and 100 TOPS/W, with a roadmap to 1,000,000 TOPS with >1,000 TOPS/W. We are able to achieve this because we have fundamentally reinvented optical compute-in-memory (using a metamaterial design technique) that is 8,000x denser than state of the art.

Product status: We have built a bench-top experiment doing optical processing using off-the-shelf components. This experiment is close to breaking the Landauer Limit on energy efficiency — to the best of our knowledge, the first time in history. We will be publishing this experimental work soon. In parallel, we have been working with IMEC to refine the design of our metasurface to be compatible with existing CMOS fabrication process rules (PDK). Four iterations in, we have reduced the number of masks/layers, minimum feature sizes, relaxed tolerance, and ended at a stackup that uses only existing-process compatible materials. We are aligned to kick off this fabrication contract with IMEC at close of round via LOI.

Customer feedback: The most impactful customer feedback comes from conversations with James Hamilton, who reports directly to Andy Jassy at Amazon and is responsible for scaling AWS. He is considering an Advisory position with Neurophos, doing diligence on CIINA & COI.

IP Portfolio: We licensed the Intellectual Ventures ISF portfolio on metamaterials, which encompasses broad metamaterials patents plus specific work developed at ISF. We have also filed 7 assets since inception, all at various stages from application to granted (with 3 more currently on the way).

Term Sheet: $7M round on $12M pre. Some flexibility on total raise. Frontier wants a minimum of $4M; our lead from the incubation round wants a minimum of $1M. Another VC in final stages of diligence for $1M; balance to invite value-add smaller-check investors.

Use of Funds: Primary use is an 18–12 month process development program at IMEC to fabricate, verify, and experimentally measure a 1MB in-memory compute metasurface. Reference points: the Google TPU has a 64KB in-memory processor occupying much larger die area than our metasurface will; Lightmatter's core is only a 4KB in-memory processor. To our knowledge this metasurface will be the largest, highest-throughput, and most energy-efficient in-memory-compute ever. The remaining two milestone sets cover the photonic integrated circuit subsystems, optical lens assembly, digital/data-transfer subsystems, and SW & API stack.

Supply chain: Entirety of the platform can be sourced from any CMOS foundry capable of <22nm nodes. Obvious 800-lb gorillas are Global Foundries and TSMC. IMEC programs are explicitly designed to transfer process knowledge and IP to big foundries at completion; leveraging IMEC to get our foot in the door of GF and TSMC will shape long-term supply relationships.