Morphotonics raised €40M because AI's bottleneck is the wire

Morphotonics, a Dutch deeptech company that has spent roughly twelve years building nanoimprint lithography machines, has raised €40 million to push that technology into data centres. TechCrunch reported the round, and I am commenting on their reporting here rather than repeating it.

The number is not the interesting part. The interesting part is that a company whose core skill is pressing a stamp into a light-sensitive layer is now being funded to fix a data centre problem. That tells you something specific about where AI hardware currently hurts.

---

🔍 What €40 million actually bought

Here is the shape of the round and the company, as reported:

Item	Detail
Amount	€40 million
Investors	3M Ventures, Innovation Industries, BOM, Invest-NL, the EIC Fund, Ernij Next, and the European Investment Bank
Structure	Extension of a round begun in 2024, taken in multiple installments
Base	Netherlands, with teams in China, Taiwan, South Korea and the US
Headcount	~30 in September 2024 → 60 now, expected to settle at 70–75
Revenue mix	~90% hardware, the rest services and licensing
Installed base	10–15 systems worldwide, targeting 50 within 2–3 years

Ten to fifteen machines. That is the entire global install base of a company that just raised €40 million. Each unit is expensive enough that sixty people can build a real business around it.

The process itself is easier to picture than it sounds. Conventional photolithography projects light through a mask to etch a pattern. Nanoimprint lithography does the mechanical thing instead: you make a mould, press it into a curable resist, and set it. It is a printing press for features far too small to see. Morphotonics has been using it for waveguides — the thin optical layers that carry an image across the lens of a pair of smart glasses, the category Meta's Ray-Ban Display sits in — and for vehicle privacy screens.

---

⚡ The bottleneck moved from the chip to the wire

The new market named in the story is co-packaged optics and photonic integrated circuits for data centre interconnects. That deserves a plain-English translation, because it is the whole point.

Inside a large AI cluster, the compute is not usually what is starving. The links between chips are. Today, signals leave a switch or accelerator as electricity, run across copper to a pluggable optical module at the faceplate, and only there turn into light. Every one of those electrical hops costs power and limits distance.

Co-packaged optics moves the optical engine into the same package as the silicon, so the signal becomes light almost immediately. Doing that at volume needs photonic waveguides patterned cheaply and repeatably — which is exactly what a stamping process is good at, and exactly what a company that has been patterning AR waveguides for a decade already knows how to do.

Key takeaway: The companies being funded right now are not the ones making transistors smaller. They are the ones making the connections between chips cheaper and cooler. When the money moves to the plumbing, the plumbing is the constraint.

One concrete number from the reporting: a next-generation machine currently under construction is expected to produce more than 6 million waveguides a year and to ship in early 2027. That is a manufacturing-capacity story, not a research story. The physics is considered settled enough to build a factory tool around it.

---

💰 What this does and does not do to your bill

I want to be honest about the timescale here, because the temptation is to draw a straight line from "cheaper optics" to "cheaper GPU hours for me." That line exists, but it is long and indirect.

Cost lever	Who controls it	How fast it moves
Model size, quantisation	You	Today
Batch size, context length, caching	You	Today
Choice of provider and region	You	This week
GPU rental market price	Cloud vendors	Months
Interconnect and optics cost	Suppliers like Morphotonics	Years

If you are renting GPUs from Sri Lanka and paying in USD, the top three rows are where your money is. The bottom row is real, and it is why cluster economics will look different in 2029, but it will not change your invoice this quarter.

So do the boring thing first. Before you touch architecture, put actual numbers on what you are spending: our GPU cloud cost calculator works out what an hourly rate becomes over a real training or inference run, and the inference speed calculator estimates tokens per second for a given model and card so you can tell whether you are compute-bound or memory-bound before you pay for a bigger instance.

---

🛠️ The twelve-year, sixty-person lesson

This is the part I would underline for anyone building a small technical company here.

Morphotonics operated, in the reporting's phrase, largely under the radar for about twelve years. It has around sixty people. It sells to a handful of customers. Its CEO, Hugo Da Silva, joined only in September 2024, by which point most of the hard engineering had already been done quietly.

Three things follow from that:

1. Selling to the boom beats joining it. Morphotonics does not train models or make chips. It makes a tool the people making the chips need. That position has fewer competitors and much longer contracts.
2. Narrow is defensible. Ninety percent hardware revenue from ten to fifteen deployed systems is not a broad market. It is a deep one. You do not need a large market to have a durable business; you need to be the only credible answer in a small one.
3. Boring compounds. Twelve years of process engineering is not a story anyone writes about until the money arrives. Then it reads like an overnight success.

None of this is a template you can copy directly. Semiconductor capital equipment is not a garage business anywhere, least of all here. The transferable part is the posture: pick a specific, unfashionable, physically hard problem that a larger industry has to solve, and stay on it long enough to be the obvious supplier.

---

💡 What this means for you

• If you work on AI infrastructure: read up on co-packaged optics now. The 6-million-waveguide machine ships in early 2027, which is roughly your window to learn this before it is assumed knowledge. The vocabulary — waveguides, photonic integrated circuits, optical engines in-package — is about to show up in job descriptions and vendor documents, and it currently has very few people who understand it.
• If you are a student: the demand here is for optics, materials science and process engineering, not another web framework. Those fields are badly under-supplied relative to how much capital is flowing into them.
• If you are costing an AI project: the supply-side story changes nothing you can act on this month. Measure your own usage, fix batching and model size, and re-check provider pricing. That is where your savings are.
• If you are building a company: you do not need to be in the loudest part of the market. You need to be necessary to it.

The headline says a display technology company is moving into data centres. What it really says is that the hard problem in AI hardware is no longer making the chip. It is getting data off the chip without setting the power budget on fire.