Neurophos hires a Kymeta founder to commercialize its 2028 photonic chip

Why it matters

Bowen is spending Neurophos' $110M round on the operators and facilities needed to turn photonics research into shipped silicon, with a 2028 deadline and no public customer benchmarks yet.

Patrick Bowen is adding an experienced metamaterials operator to Neurophos as the Austin chipmaker moves from laboratory work toward a photonic AI processor it says will ship in 2028.

Neurophos announced Thursday that Nathan Kundtz has joined as chief technology officer, alongside an expanded engineering leadership group, a new Sunnyvale office and larger headquarters in Austin. Neurophos also plans to open a low-rate initial production facility in Austin later in 2026.

Bowen, a Duke-trained metamaterials researcher, co-founded Neurophos with chief scientist Andrew Traverso around a physics problem that has become a financial problem for the AI industry: conventional processors require more power, manufacturing capacity and supporting infrastructure as models and workloads grow. Duke says Neurophos was spun out of professor David Smith's lab, where Bowen completed his doctorate.

Kundtz gives Bowen a lieutenant who has already tried to carry metamaterials from academic research into manufactured hardware. Kundtz founded Kymeta, which used metamaterials in electronically steered satellite antennas, and served as its CTO, CEO and president. He also founded Intellectual Ventures' Center for Metamaterials and most recently ran synthetic-data platform Rendered.ai.

That hiring choice is the clearest indication of what Neurophos needs after raising a large early round. Bowen has already assembled the scientific thesis and a design for shrinking optical computation onto silicon. Neurophos now needs to turn those ideas into chips that can be fabricated, packaged, installed in servers and delivered on a credible schedule.

Building the commercialization bench

Neurophos also promoted Michael Shebanow to senior vice president of engineering. Its team page credits Shebanow with a decade of GPU and CPU work at Nvidia, followed by engineering leadership at AMD, Samsung and Cadence. Preston Woo's remit has expanded from chief financial officer to chief financial and commercialization officer, placing customer and partnership work directly alongside Neurophos' capital planning.

Kevin Keller has joined as general counsel after legal roles spanning Wilson Sonsini, Micron, Microsoft, Amazon and Meta. The broader technical hiring group includes optics researchers from Magic Leap and Eikon Therapeutics, a board-design leader with experience at Intel, Qualcomm and Juniper, and an intellectual-property director focused on metamaterials, nanofabrication and semiconductors.

Bowen is also bringing Nigel Toon onto Neurophos' board. Toon co-founded Graphcore, one of the best-funded attempts to build an AI processor outside the dominant GPU architecture, and led it through its 2024 acquisition by SoftBank. His earlier semiconductor career included co-founding Icera, which Nvidia acquired in 2011.

Toon's presence gives Neurophos direct access to the lessons of selling a new processor architecture into a market where technical performance alone rarely settles the outcome. Chip challengers need software, manufacturing partners, systems integration and customers willing to commit workloads before volume production begins. Neurophos' Thursday announcement puts people around each of those problems, even as its first product remains two years from its stated delivery date.

Neurophos says its workforce has grown fourfold during 2026 and is scheduled to double again by year-end. A fourfold multiple is difficult to size because Neurophos has not published the starting or current headcount. An April update described the workforce as having nearly tripled since the January financing, using a different time window.

The geographic expansion follows the same commercial logic. Sunnyvale puts Neurophos near chip designers, suppliers and potential customers in Silicon Valley. Austin retains the headquarters, existing laboratory and prototyping operations, while the planned initial production facility would give Neurophos a place to establish repeatable assembly and testing before pursuing higher volumes.

A $110M wager on smaller optics

The financing behind the expansion is nearly eight months old. Neurophos announced a $110M Series A on January 22nd, bringing its reported total funding to $118M. Gates Frontier led the round, joined by M12, Aramco Ventures, Bosch Ventures, Carbon Direct Capital, Tectonic Ventures, Space Capital and a longer group of venture investors. Neurophos did not disclose a valuation.

The round funded an unusually ambitious product schedule. Neurophos is developing the Tulkas T100 Optical Processing Unit, which is designed to perform matrix multiplication with light on a CMOS-compatible photonic die. Neurophos says one million photonic elements fit inside a 5-by-5-millimeter area, compared with optical systems that traditionally occupied a laboratory bench.

The T100 is aimed initially at prefill, the compute-heavy stage in which an inference system processes a user's input before generating output tokens. Neurophos plans to package the photonic processor with an electronic companion die handling memory and input-output functions, allowing the module to fit into standard server infrastructure.

Neurophos claims roughly 300 trillion operations per second per watt for the T100, compared with roughly nine for Nvidia's B200 in its presentation. Other sections of Neurophos' website describe the efficiency advantage as 30x or 40x, while its January financing announcement claimed gains of up to 100x across performance and energy efficiency. Those comparisons remain Neurophos' own projections. Neurophos has not published independent benchmarks, production yields or customer deployments for the T100.

Bowen's commercial challenge is therefore measured in validation milestones rather than hiring multiples. The 2028 shipment target leaves time to finish the module, software and production process, while also giving competing optical-computing efforts time to mature.

Lightmatter's Envise likewise performs matrix multiplication using photonics, while Lightmatter has increasingly emphasized optical connections between chips. Ayar Labs raised $500M in March at a reported $3.75B valuation to expand production of optical interconnect technology. Those approaches address different parts of the computing system, but they compete for the same infrastructure budgets and engineering attention as data centers confront power and bandwidth constraints.

Neurophos is making the harder, broader bet that optics can become the compute engine itself. Bowen now has $110M, a second engineering base and executives who have built metamaterial products and alternative AI processors before. The next proof will come from manufactured hardware that outside customers can test against the processors already in their racks.