The Bottleneck Nobody is Hiring For

Recently, I spoke with the CSO of one of Europe's most advanced photonics companies. They spent twelve years and $600 million building display technology. They are not abandoning it, but he is looking at the AI datacentre market and seeing something far larger. "The infrastructure buildout is the real opportunity," he told me, matter-of-factly.

Then he pointed me at something most people had skipped past. Jensen Huang does not excite easily. When he stood up at Computex and called Marvell the next trillion-dollar company, the room nodded and moved on, busy with frontier models and software agents. Hardly anyone clocked that the head of Nvidia, who has spent thirty years on what compute actually needs to work, was pointing straight at a connectivity chip business.

The CSO had noticed. The gap between those two reactions is the story.

My job is placing C-suite and board leaders across DeepTech and AI infrastructure in Europe, which means I hear where companies are heading before they announce it. Companies with deep hardware capability are repositioning around AI infrastructure, because the money pooling there now dwarfs the markets they were built for.

The harder problem it creates is a leadership one, and almost no one is hiring for it.

Connecting hundreds of thousands of GPUs into one coherent system is a physics problem, not a software one. And physics answers plainly: over the distances a modern datacentre spans, light carries more bandwidth than copper and burns less power doing it.

Push copper past about ten metres at these bandwidths and it gives out. The industry is now moving to replace copper with optics at every layer of the stack: rack to rack, then chip to chip, and eventually inside the package.

From Nvidia's AI Factory blueprint to Broadcom's co-packaged optics, now running a million hours without a dropout in Meta's labs and heading for commercial scale, the sector is solving one constraint. As clusters scale toward one million GPUs, the interconnect dictates the architecture. Marvell's sudden prominence is not a coincidence.

At Computex, Marvell unveiled its Teralynx T100, a 102.4 Tbps switch silicon purpose-built for AI infrastructure. It cuts power by 25% and slashes latency for intense training workloads. Huang wasn't being polite. He was pointing at where the constraint actually sits, in the connectivity layer that binds the GPUs into one system.

Marvell didn't build this portfolio on a hunch. They built it because every hyperscaler, sovereign compute cluster, and AI factory on earth is about to hit the same physical wall.

Marvell is selling the tools for the whole build, and most of the market is still watching the models instead.

Across France, Germany, and Switzerland, companies that spent a decade building hardware in photonics, semiconductors, and advanced materials are suddenly sitting in the path of the buildout. Some are pivoting deliberately; others are being pulled there by hyperscalers desperate to diversify supply chains away from Asia.

Europe is deep in this part of the stack, but its habit has been to invent in the lab and watch American companies commercialise it at speed.

That patience is now a liability. The technical edge is here; the speed to market isn't.

Everyone counts the engineer shortage. The harder gap is the one or two people at the top who can run these companies.

You cannot retrain a SaaS executive to negotiate a multi-year design cycle with Nvidia or run a semiconductor line. The profile is different. It takes real hardware intuition, the patience for long development cycles, and enough standing to sell into the procurement teams at the largest tech companies.

The CTO who can take an optical interconnect company from an academic lab to a hyperscaler contract is not on the job boards. Nor is the CPO who knows silicon photonics well enough to map a roadmap and still close a design win at Microsoft. People like this are already running a company, advising a sovereign fund, or spoken for, locked up by investors who spotted them three years ago.

Finding them is investigative work. An automated LinkedIn sequence will never surface them.

The $700 billion buildout is underway. The leadership race to support it has barely started.