Neurophos, the Austin-based photonic AI chip startup, announced on September 10, 2026 that it has opened a new office in Sunnyvale, California and added eight leadership hires as it moves to commercialize metasurface-based optical processors for data centers. The expansion follows the company’s oversubscribed $110 million Series A round, announced in January 2026 and led by Gates Frontier, the investment vehicle associated with Bill Gates, according to Cooley, the law firm that advised Neurophos on the deal.
The announcement lands at a moment when the AI hardware industry is openly searching for an alternative to the electron. GPUs from Nvidia and AMD still do the vast majority of AI training and inference work worldwide, but power and heat limits are pushing hyperscalers to fund a small group of photonic and optical computing challengers. Neurophos is one of the best-capitalized of that group, and its Silicon Valley expansion signals the company is trying to move from a research-stage curiosity to a vendor that data centers can actually order from.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Neurophos Announced on September 10, 2026
According to Neurophos’ own press release distributed via PR Newswire, the company is opening a new office in Sunnyvale, California, while also expanding its headquarters footprint in Austin, Texas, where it has already begun planning a Low-Rate Initial Production facility. The Sunnyvale site puts Neurophos inside the same geography as Nvidia, Lightmatter, and most of the venture firms that backed its Series A round.
The company said it has scaled its workforce 4x during 2026 and is on track to grow headcount by another 2x before year-end, an eightfold expansion in a single calendar year if the target holds. Dr. Patrick Bowen, Neurophos’ CEO and co-founder, framed the announcement as a step toward turning a lab-grade breakthrough into shippable hardware. “By scaling our team and expanding our geographic footprint, we are positioning ourselves to lead the next generation of high-performance computing and deliver on our mission to bring very large orders of metasurface-based photonic AI chips to market,” Bowen said, according to the company’s press release.
Bowen also tied the office opening directly to the company’s hiring push, saying: “The addition of these industry-leading experts and the opening of our Silicon Valley office are crucial steps in accelerating Neurophos’ commercialization efforts.”
Inside the $110 Million Series A That Set Up This Moment
The Series A that funded this expansion closed in January 2026 and was, according to Cooley’s deal coverage, oversubscribed at $110 million, bringing Neurophos’ total funding to $118 million since its founding. Gates Frontier led the round. The participant list is unusually broad for a Series A of this size: M12 (Microsoft’s venture fund), Carbon Direct Capital, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, Silicon Catalyst Ventures, Morgan Creek, Alumni Ventures, DNX Ventures, Geometry, Gaingels, Mana Ventures, and MetaVC Partners all participated, per Cooley and corroborating coverage from optics.org.
That investor mix tells its own story about who wants a stake in post-GPU compute. Microsoft’s M12 fund and Aramco Ventures both have direct exposure to the cost of running AI data centers at scale, while Bosch Ventures brings a semiconductor-manufacturing lens. Gates Frontier’s decision to lead is notable given Bill Gates has been a long-time backer of energy and climate-adjacent deep tech, and photonic computing’s central pitch is lower power draw per operation than an electronic GPU.
M12 characterized the round on LinkedIn as part of an effort to “bring photonic computing into the heart of AI data centres,” listing the same investor syndicate reported by Cooley. Neither Neurophos nor its investors have disclosed a post-money valuation for the round.
Who’s Leading Neurophos Now: The New Executive Bench
The September 10 announcement named eight new or newly promoted leaders, the clearest sign yet that Neurophos is staffing up for manufacturing and sales, not just research. Dr. Nathan Kundtz joined as Chief Technology Officer, Dr. Michael Shebanow was promoted to Senior Vice President of Engineering, Kevin Keller joined as General Counsel, and Preston Woo’s role was expanded to Chief Financial Officer and Chief Commercialization Officer, according to the company’s press release.
Rounding out the announced hires are Justin Flanagan as Director of Intellectual Property, Dr. Kristina Uhlendorf as Technical Fellow for Optics, Dr. Josh Isaacs as Director of Optics, and Dr. Chris Pan as Director of Board Design. Nigel Toon also joined Neurophos’ board of directors.
A CTO Change at a Critical Moment
Kundtz’s arrival as CTO is a change at the top of the technical org chart. Earlier coverage of Neurophos, including reporting from optics.org around the January funding round, identified Hod Finkelstein as the company’s CTO at that time. The September leadership slate suggests Neurophos brought in outside technical leadership specifically for the commercialization phase, on top of the team that built the original metamaterial modulator technology.
Why Austin Keeps the Headquarters While Sunnyvale Gets the New Office
Neurophos was founded and remains headquartered in Austin, Texas, where the company has also begun planning what it calls a Low-Rate Initial Production facility, per its own press release. That keeps the manufacturing and scale-up work anchored in Texas, a state that has become a magnet for chip fabrication investment over the past several years.
The Sunnyvale office serves a different purpose: proximity to customers, investors, and talent. Silicon Valley remains the default location for engineers who want to work on frontier AI hardware, and it puts Neurophos closer to the hyperscalers and chip buyers who would need to evaluate its optical processors before placing orders. Splitting headquarters functions from a commercial and engineering satellite office is a familiar playbook for hardware startups that outgrow a single-city footprint, though Neurophos has not disclosed exact staff counts for either site.
The Technology Bet: Metasurface Photonic Chips Explained
Neurophos was founded by Patrick Bowen, who previously researched metamaterials at Duke University and founded Metacept Systems before starting Neurophos, according to optics.org’s reporting. The company’s core technology is a metasurface: a flat, engineered material that manipulates the wavefront of light, built using standard silicon-foundry materials – Neurophos says it reached first silicon in May 2025 using a standard CMOS process, according to YesPress. Applying an electric field to the surface changes its optical properties, which is what makes the chip reprogrammable rather than fixed at manufacture time.
That metasurface sits inside what the company describes as a compute-in-memory processor architecture, fed by high-speed silicon photonics, and aimed specifically at matrix-matrix multiplication, the operation that dominates neural network inference. The flagship product built on this architecture, the Tulkas T100, reportedly packs about 1 million processing elements, with its optical processing elements operating at roughly 56 GHz, according to YesPress. Optics.org describes the resulting optical modulators as achieving “four-order-of-magnitude level of miniaturization” compared with prior approaches, meaning they are roughly 10,000 times smaller than earlier photonic modulator designs.
The Exaflop and Efficiency Claims
Neurophos has said its metamaterial-based modulators, combined at scale, are designed to support well over a quintillion operations per second, an exaflop-class compute density, and the company has told optics.org it is targeting up to a 100x improvement in power efficiency over current GPUs; its flagship Tulkas T100 system is specifically targeted at roughly 470 PFLOPS of throughput within a 1-2 kW power envelope, according to YesPress. Bowen has framed the pitch around Moore’s Law slowing down just as AI compute demand keeps climbing. On the company’s own site, he put it this way: “Our breakthrough in photonics unlocks an entirely new dimension of scaling – by packing massive optical parallelism on a single chip,” according to Neurophos’ website.
None of these performance figures are independently benchmarked yet. They are company claims about a chip still headed toward commercial production, and readers should treat them the way they’d treat any pre-shipment hardware roadmap: directionally informative, not verified.
From Duke University Lab to AI Data-Center Contender
Neurophos’ path to a $110 million Series A started small. The company raised a $7.2 million seed round in December 2023 to commercialize metamaterial and optical AI inference chips for data centers, positioning itself early as a metamaterials spinout rather than a traditional silicon photonics company. That distinction matters: most optical computing startups build on top of conventional silicon photonics foundry processes, while Neurophos’ metasurface approach is closer to materials science than chip design in its origins.
Three years later, the company has gone from a university-adjacent seed-stage startup to a Series A company backed by Microsoft’s venture arm, Saudi Aramco’s investment arm, and Bill Gates’ Gates Frontier, with a roadmap that according to Cooley’s coverage of the round includes datacenter-ready optical processing unit modules, a software stack, and early-access developer hardware. Optics.org reports Neurophos is targeting initial customer evaluations in 2026, with first commercial systems and a production ramp targeted for early 2028, according to Built In Austin.
Neurophos vs. the Photonic AI Chip Field
Neurophos is not the only well-funded startup betting that light, not electricity, is the way past the GPU’s power and heat ceiling. Lightmatter, Celestial AI, and Lightelligence have all raised far more capital and are further along commercially, which puts Neurophos’ September expansion in context: it’s catching up on funding scale, not leading it.
| Company | Total Funding | Latest Round | Approach |
|---|---|---|---|
| Neurophos | $118 million | $110M Series A, Jan. 2026, led by Gates Frontier | Metasurface modulators + compute-in-memory optical processor |
| Lightmatter | $850 million | $400M Series D at a $4.4B valuation, Oct. 2024, led by T. Rowe Price | Silicon photonics interconnect and photonic compute |
| Celestial AI | Approx. $594 million | Multiple rounds across 8 financings | Optical fabric (“Photonic Fabric”) for chip-to-chip and memory interconnect |
| Lightelligence | $247 million | Series C, Sept. 2025; Hong Kong IPO debut surged 383%, April 2026 | Photonic arithmetic units for AI inference |
The funding gap is stark: Lightmatter alone has raised roughly seven times what Neurophos has banked to date. But Neurophos’ pitch is different in one respect – its metasurface approach claims a more radical miniaturization of the optical modulator itself, rather than optimizing interconnect fabric or arithmetic units built on more conventional silicon photonics. Whether that translates into a manufacturing or cost advantage is something none of these companies has proven at volume yet.
Market Impact: What This Means for the AI Hardware Supply Chain
Every dollar that flows into a Nvidia GPU alternative is, at least in theory, a dollar of future demand Nvidia and AMD don’t automatically get to keep. That’s still a distant scenario for Neurophos specifically, given it has no shipping product, but the pattern across the sector is real. Nvidia’s roughly 90% share of GPU shipments gives hyperscalers a strong incentive to fund alternatives, if only to build negotiating leverage, even before those alternatives are commercially viable.
The timing also lines up with a broader AI infrastructure spending wave. TSMC’s recent record revenue, driven by the AI chip boom, shows how much capital is currently chasing any technology that promises to relieve pressure on advanced packaging and leading-edge fabrication capacity, both of which are consumed heavily by GPU production. A photonic chip that genuinely needs less advanced-node silicon real estate per unit of compute would ease that specific bottleneck, which is part of why chipmakers like Bosch, which invests through Bosch Ventures, have money in this round.
There’s a parallel data-center power story here too. AI data centers are running into grid capacity constraints in multiple U.S. regions, and Neurophos’ entire pitch rests on using dramatically less power per operation. If that claim holds up in real deployments rather than lab demonstrations, it would matter to utilities and data-center operators well beyond the AI chip market itself.
Historical Context: The Long Search for a Post-GPU AI Chip
Optical and photonic computing research is not new; it dates back decades in academic labs. What’s changed is the economic incentive. As long as GPUs kept getting cheaper and more efficient generation over generation, there was little commercial pressure to fund an entirely new compute substrate. That calculus shifted once AI training and inference workloads started outpacing the historical rate of GPU efficiency gains, a dynamic that’s been visible across the industry as companies like Intel try to reposition their chip roadmaps around AI-specific demand rather than general-purpose compute.
Neurophos’ own funding trajectory reflects that shift in investor appetite. A $7.2 million seed round in December 2023 was a modest, research-adjacent bet. Just over two years later, the same underlying technology attracted a $110 million round from a syndicate that includes two of the largest cloud and chip-adjacent investors on the planet, Microsoft and Aramco. That’s a large jump for a company that, per its own timeline, doesn’t expect a production ramp until 2028.
Investor Voices and Company Statements on the Expansion
Beyond Bowen’s own comments, Microsoft’s M12 fund framed its rationale for backing the round in explicitly infrastructure-focused terms on LinkedIn, describing the investment as a bet on bringing photonic computing directly into AI data centers, while listing the same investor syndicate reported by Cooley. That framing matches how Bosch Ventures, Aramco Ventures, and Carbon Direct Capital have positioned similar bets elsewhere: not as a wager on a single startup winning, but as insurance against a future where GPU-only scaling runs into a wall.
Cooley’s own coverage of the round, drafted from its position advising Neurophos on the transaction, described the new funding as set to “accelerate delivery of Neurophos’ first integrated photonic compute system, including datacenter-ready OPU modules, a full software stack, and early-access developer hardware,” a roadmap that puts real hardware in outside hands well before the 2028 production ramp optics.org has reported.
Neurophos’ 2026 Growth in Numbers
| Milestone | Detail | Date |
|---|---|---|
| Seed round | $7.2 million | December 2023 |
| Series A | $110 million, led by Gates Frontier | January 2026 |
| Total funding to date | $118 million | As of Sept. 2026 |
| Workforce growth (2026) | 4x scaled so far; targeting another 2x by year-end | Reported Sept. 10, 2026 |
| New leadership hires/promotions | 8 named executives and one board seat | Announced Sept. 10, 2026 |
| New office | Sunnyvale, California | Opened Sept. 2026 |
| Planned production ramp | Low-Rate Initial Production, Austin, TX | Targeted 2028 |
Risks and Skepticism: Can Photonic Chips Actually Ship?
The gap between Neurophos’ claims and independently verified results is the central risk here. No third party has published benchmark data confirming the company’s stated exaflop-class throughput or its 100x power-efficiency target against current-generation GPUs. Every optical and photonic computing company in this space, including better-funded rivals like Lightmatter and Celestial AI, has made similarly bold efficiency claims that have yet to be tested at the scale hyperscalers actually deploy.
The manufacturing challenge isn’t unique to Neurophos either. Meta’s own custom AI chip effort shows how long it can take even a company with far deeper pockets to move a novel accelerator from design to production at scale.
There’s also a manufacturing risk specific to metasurfaces. Reprogrammable optical components built from novel materials tend to be harder to yield consistently at foundry scale than conventional CMOS logic, and Neurophos’ 2028 production-ramp target leaves roughly two years for the company to move from developer hardware to volume manufacturing. Missing that window would put Neurophos behind rivals that are already shipping interconnect and inference products commercially.
Finally, the CTO change disclosed alongside this expansion is worth watching. Bringing in Dr. Nathan Kundtz to replace or supplement earlier technical leadership during a scale-up phase can be a healthy sign of a company maturing its engineering organization, but it can also signal friction over execution timelines. Neurophos has not commented publicly on the reasons behind the change.
What Comes Next: 5 Predictions Through 2028
- Neurophos will likely announce its first named pilot customer or design partner within the next two to three quarters, given the “early-access developer hardware” language in its own funding announcement.
- Expect at least one additional funding round, a Series B, before the 2028 production ramp, given how capital-intensive photonic chip manufacturing tends to be relative to seed and Series A budgets.
- Competitive pressure from Lightmatter and Celestial AI will likely intensify as all three companies chase the same hyperscaler customer base, potentially compressing Neurophos’ commercialization timeline.
- Sunnyvale hiring will probably outpace Austin hiring over the next year, as customer-facing and software roles are typically easier to fill in the Bay Area than specialized optics manufacturing talent.
- Independent, third-party benchmarks of Neurophos silicon are unlikely to appear before 2027 at the earliest, meaning the company’s efficiency claims will remain unverified by outside labs for some time.
Why This Matters Beyond One Startup’s Funding Round
Neurophos is a $118 million company competing against rivals that have raised hundreds of millions more. That imbalance alone makes this week’s news more about validation than dominance: it tells other investors, and other potential customers, that a syndicate including Microsoft, Aramco, and Bill Gates’ Gates Frontier believes metasurface photonics is worth funding at this stage. Whether that belief survives contact with actual manufacturing yields and hyperscaler benchmarking is the story that will play out over the next two years, right through Neurophos’ targeted 2028 production ramp.
For now, the concrete facts are these: a new Sunnyvale office, eight named leadership additions, a fourfold-and-counting headcount increase in 2026, and $118 million in total funding behind a chip that has not yet shipped to a paying customer. The rest, including whether Neurophos closes the funding gap with Lightmatter or falls further behind as rivals like well-capitalized AI companies keep raising at multi-billion-dollar valuations, is still an open question.
Frequently Asked Questions
What did Neurophos announce on September 10, 2026?
Neurophos announced it opened a new office in Sunnyvale, California, expanded its Austin, Texas headquarters, and added eight new or newly promoted executives, following its $110 million Series A round from January 2026.
How much funding has Neurophos raised in total?
Neurophos has raised $118 million in total, including a $7.2 million seed round in December 2023 and an oversubscribed $110 million Series A in January 2026 led by Gates Frontier, according to Cooley, the law firm that advised on the deal.
Who invested in Neurophos’ Series A?
Gates Frontier led the round, with participation from M12 (Microsoft’s venture fund), Carbon Direct Capital, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, Silicon Catalyst Ventures, Morgan Creek, Alumni Ventures, DNX Ventures, Geometry, Gaingels, and MetaVC Partners.
What does Neurophos actually make?
Neurophos develops photonic AI chips built around reprogrammable metasurface optical modulators, paired with a compute-in-memory processor architecture, aimed at accelerating the matrix-matrix multiplication operations used in AI inference and training.
Who leads Neurophos, and who was just hired?
Dr. Patrick Bowen is CEO and co-founder. New leadership named September 10, 2026 includes Dr. Nathan Kundtz as CTO, Dr. Michael Shebanow as SVP of Engineering, Kevin Keller as General Counsel, and Preston Woo as CFO and Chief Commercialization Officer, plus Justin Flanagan, Dr. Kristina Uhlendorf, Dr. Josh Isaacs, Dr. Chris Pan, and new board member Nigel Toon.
How does Neurophos compare to Lightmatter and Celestial AI?
Neurophos, at $118 million total funding, has raised far less than Lightmatter ($850 million total, $4.4 billion valuation) or Celestial AI (roughly $594 million). Lightelligence has raised about $247 million and completed a Hong Kong IPO in April 2026. All four companies are pursuing photonic or optical approaches to AI compute, but with different architectures.
When will Neurophos chips actually be available?
According to optics.org’s reporting, Neurophos is targeting initial customer evaluations in 2026, with first commercial systems and a production ramp planned for early 2028, per Built In Austin, at a Low-Rate Initial Production facility in Austin, Texas.
Why did investors like Microsoft and Aramco back a photonic chip startup?
Both companies have direct exposure to the power and cost constraints of running large AI data centers. Backing an alternative compute architecture, even at an early stage, gives them insight into and potential leverage over technology that could eventually reduce dependence on GPU-based infrastructure.
