The $10T AI Buildout Has a Photonics Problem
Summary
AI’s next bottleneck is the fabric connecting GPUs, not merely the GPUs themselves. Models no longer fit on a single GPU, while agentic systems add latency-sensitive calls across servers; Yannick De Koninck says this demands a “much flatter network” with denser connections. His sharpest framing: roughly half an AI factory is GPUs, while “the other half is the technology to interconnect these GPUs.”
Photonics becomes essential because copper is running out of steam just as AI factories need dramatically more internal traffic. Optical interconnects offer much higher data-transfer bandwidth, while Herwig Van Hove frames data-center energy consumption as the “killer app” for a photonics paradigm shift decades in the making.
The investable scarcity sits across the manufacturing chain, not in photonic science alone. TSMC, GlobalFoundries, and Tower have invested in silicon-photonics wafers, but “if you deliver a wafer to NVIDIA, that’s not what they ordered”—NVIDIA orders a final chip that still requires lasers, tooling, and substrates. Throughput and yield remain inadequate, raw materials are concentrated among a handful of suppliers, and the East-West divide compounds the risk.
NVIDIA’s escalation turned a long-anticipated transition into an immediate capacity problem. Van Hove dates the demand shock to ChatGPT’s viral release in November ’22, followed by louder signaling from NVIDIA in ’24 and ’25 and investments in Lumentum and Coherent in ’26. His claim that NVIDIA invested “4× the total market size” in the first six months of the year captures the mismatch: “The innovation is ready,” but the supply chain is not.
Themaa is using an acquired European foundry to compress a new-facility build into a 2027 production target and 2028 full ramp. Located 20 kilometers from imec, the facility combines ready infrastructure with photonics and industrial talent. Customers are requesting volumes beyond Themaa’s “wildly ambitious projections,” although tool availability, machine performance, and substrates remain constraints.
Europe’s photonics opportunity is credible, but execution culture is the test. Belgium has unusually deep research density—Van Hove says the University of Ghent produced 35% of 4,000 global peer-reviewed white papers on photonics from 2000 to 2015—yet Europe has not consistently commercialized its own innovation locally. Themaa wants to pair that technical base with a fail-fast, decisive operating culture modeled partly on the approach De Koninck described at NVIDIA.
Deep dive
1. AI factories turn interconnects into core compute infrastructure
Molly O’Shea frames the transition through Tony Kim’s estimate of a $10 trillion infrastructure rebuild: legacy data centers consume too much power and water and were not architected for today’s parameter-heavy AI workloads. The guests narrow the problem to the rapidly growing traffic inside those facilities.
De Koninck’s mechanism is straightforward: models “no longer fit on a single GPU,” so computation depends on GPU-to-GPU communication. Copper interconnects cannot keep pace as processors demand more data at higher rates, pushing the industry toward optical links with substantially greater bandwidth.
Agentic AI intensifies the shift because a model may call a tool running on another server and wait for its result. That makes latency between calculations critical and requires replacing deep, hierarchical switch trees with a “much flatter network”—more direct links among GPUs and switches, and a much denser fabric than a traditional hyperscale data center.
His closing hot take makes the market legible: “Half of that is GPUs, but the other half is the technology to interconnect these GPUs.” Optics and photonics therefore address infrastructure of comparable importance to the compute engines themselves.
2. Power is photonics’ killer app, but manufacturing is the bottleneck
Van Hove says energy consumption is the first answer when data-center CTOs are asked what keeps them awake. Their CEOs demand growth, yet grids cannot arrive on the required timetable; he cites Google negotiating with Amsterdam for three years before moving its plans to Belgium.
The inflection came in November ’22, when ChatGPT went viral and made the need to switch to photonics apparent. NVIDIA progressively amplified its photonics push in ’24 and ’25, then demonstrated the intended scale through its ’26 investments in Lumentum and Coherent: scientists and designers had prepared the innovation, but “nobody was prepared for the speed of change.”
Traditional players such as TSMC, GlobalFoundries, and Tower have focused on the front end—making silicon-photonics wafers—but NVIDIA orders final chips. Lasers must scale dramatically; tool builders are not producing enough equipment; tools need better throughput, efficiency, and yield; and substrates remain “really, really cornered,” with only a handful of raw-material suppliers. Van Hove’s diagnosis is that the industry must solve every layer “all the way back to the beginning.”
3. Themaa is building the missing bridge from lab to fab
Van Hove saw a gap between large foundries handling the front end and university spinouts operating at research scale: no industrial player occupied the middle. Themaa’s answer is a large pure-play photonic foundry designed to industrialize decades of European research rather than let commercialization migrate elsewhere.
The team’s technical and operating background is part of that thesis. De Koninck entered silicon photonics 20 years ago, completed a PhD, worked at an early commercialization startup and imec, and joined NVIDIA five years ago to build its silicon-photonics technology to product-ready maturity. Van Hove, a longtime investor who became an operator late, says he saw the supply-chain gaps three or four years ago.
Serendipity supplied the wedge when a large European foundry went out of business during Themaa’s 2023–24 planning. Acquiring it sidestepped much of the three-to-five-year process of permitting, licensing, and constructing complex infrastructure, while placing the company 20 kilometers from imec amid industrial and specialized photonics talent.
De Koninck explains that imec began more than 30 years ago in electronics, became one of the birthplaces of silicon photonics, and grew into what he calls the world’s number-one research center for advanced electronics technologies. That nearby research engine underpins Themaa’s intended leap from proven science into repeatable production.
4. The roadmap is aggressive because customers are pulling capacity forward
Themaa plans to begin production in ’27 and work toward full ramp-up in ’28. Van Hove says the largest challenge is the timetable customers are imposing, while the encouraging signal is that requested volumes already exceed the company’s “wildly ambitious projections.”
Execution still depends on scarce tools, improved equipment, and substrates. The offset is a distributed European research pipeline across Greece, Germany, Spain, France, and Switzerland addressing individual technical constraints, with Van Hove seeing more genuine cross-border collaboration than before.
Hiring has not been the expected obstacle. Three of imec’s most senior people have joined Themaa’s board, and the company has attracted people from NVIDIA, AWS, TSMC, and Lam Research. Van Hove wants a “homegrown magnet for talent” that lets European specialists build locally instead of moving abroad or becoming bankers.
5. Europe has the science; Themaa must prove it can move at market speed
American tier-one potential customers challenged Themaa on four fronts: funding, talent, machines and substrates, and whether Europe could meet their timelines. Van Hove says Themaa is deliberately addressing the stereotype of Europe as slow and committee-driven by adopting a culture of failing fast, making decisions, and acting without “100% perfect information.”
Europe also lacks the same data-center network as the US and has stricter regulations. Van Hove sees energy-efficient infrastructure as an opportunity to build a homegrown European network that may look different from the US mega-hyperscalers, rather than reproducing them.
De Koninck learned the operating model at NVIDIA, where “the org chart is just a piece of paper” and anyone able to help a project contributes. Leaving what he calls the “golden palace” for the “pirate ship” was his chance to build world-class European manufacturing against one of AI’s central constraints.
Van Hove’s broader wager is that Europe can finally commercialize its own photonics leadership. He points to the University of Ghent’s claimed 35% share of 4,000 global peer-reviewed white papers on photonics from 2000 to 2015 and invokes ASML as proof that Europe has built a durable industrial champion before: “Europe is ready for more stories like that.”