What we took home from NVIDIA GTC Berlin

Robots, co-packaged optics and packed halls. Our team on three days at GTC Berlin.

October 22, 2026

GTC Berlin was busy from the first morning to the last. Three of us made the trip: Søren Sørensen, our chief growth officer, Daniel Tidselbak, our chief commercial officer, and James from marketing.

We didn't have a stand this time. We went to listen, walk the floor and meet as many customers and partners as we could. The halls were packed and the meetings ran back to back.

What we saw

Europe was front and centre. Jensen Huang's keynote spoke about European companies, sovereign AI and AI factories, along with ready-to-go systems and AI moving onto laptops, desktops and workstations. The floor matched it. There were distributors, integrators, neoclouds and a lot of very young startups.

Robots were everywhere. One prototype was built to install optical transceivers and cables in the field, because there aren't enough people to do the job. AI was turning up in audio and special effects too.

The speed of building stood out. NVIDIA showed a server that now goes together in under five minutes, down from about two hours. No cables, fully liquid cooled, and designed to be built and deployed fast.

For us, the network was the part that mattered most. Co-packaged optics is now in volume production on the Spectrum-X switch. It's a real step forward. But the whole industry is also chasing hardware that's faster to build and easier to service. Those two things pull in different directions.

Our view is simple. Pluggables, near-packaged and co-packaged optics will all have a place. Customers need the flexibility to choose the right one for each deployment. They also need a supplier who keeps pace as lane speeds move from 100G to 200G and beyond.

What does it mean for serviceability when the optical engine sits inside the switch package instead of in a module you can swap?"

Daniel Tidselbak, chief commercial officer

In conversation with Søren and Daniel

We asked them both five questions on the way home.

Who were you talking to, and what was on their minds?

Daniel: People right across the AI space. Two things stood out. The building keeps going, and the systems already out there keep running in all sorts of forms. And cost and time. Both are areas where we're relevant.

Søren: Things continue to move very fast. AI is going into every corner of our lives, from manufacturing to the rest of society. That pushes the need for even more GPUs and even more connectivity.

What stood out on the exhibition floor?

Daniel: Robots. Robots everywhere! I saw a prototype of a robot designed to install optical transceivers and cables in the field. It's there to solve the lack of manpower for those jobs.

Søren: The little robot duck Jensen took on stage. It's a small gadget and it's quite funny, but I think it's a major step. It makes AI playable and real for a lot of people. You train it, and it carries it out right in front of you.

It might be a small gadget, but it's a major step towards making AI real for many people.

Søren Sørensen, chief growth officer

Where's the biggest gap between what AI infrastructure needs and what the market can deliver?

Daniel: Damn good question. Power was a concern for some.

Søren: Yes, it's getting more efficient. But it's not getting as efficient as the need for more gear and more data centers. At the TD SYNNEX opening event we heard that Europe's data center capacity is set to grow several times over by 2030. It's mind-blowing. There's a control question too. One technologist I spoke to tests some of the most advanced robots around, and they still take the batteries out at night. People aren't comfortable with a robot running when no one is watching.

1.6T is the headline now. What does 3.2T really mean for the people building these networks?

Daniel: We're yet to see a reference architecture. The key is 200G-lane optics, so going from 100G lanes to 200G lanes is the big step. No doubt we'll see 400G lanes faster than we think. Right now, though, CPO is being pushed as the next frontier. And while the focus is on NVIDIA, AMD is starting to hit the market. That could change approaches and architectures for customers.

 

Søren: We saw the first CPO. It might be a thing, but most likely a limited one in a small ecosystem, because you lose flexibility. For the transceiver industry there'll be a big surge in the coming years, even at lower data rates. When everyone has an AI laptop, you need faster connections everywhere.

How did ATOP land with people this week?

Søren: We now have a brand that's well known among the neoclouds, and it's growing fast with the partners who work with high-speed and AI customers. They know us. They like to speak to us. We like to speak to them. That only comes with hard work. Being present, taking the meetings, taking the long days,

They know us. They like to speak to us. We like to speak to them."

Søren Sørensen, chief growth officer

Onto the next one

Three busy days, a lot of good conversations and plenty to think about. If we didn't get to meet you in Berlin, get in touch. We'd be glad to pick up the conversation.

What interconnects are used in GPU clusters?

Neither is better in general. Copper uses less power, adds less latency and costs less within its reach, which is around 10 m at 400G and 800G. Fibre is required beyond that and is also chosen inside copper reach where links must be serviceable without disturbing the cable plant or where cable bulk is affecting airflow. Beyond the rack, single-mode DR optics have become the common choice at scale.

How does interconnect choice affect GPU cluster performance?

Distributed training synchronises across nodes, so job completion time is set by the slowest link rather than the average. Interconnect choices that reduce link instability, keep FEC correction well inside its budget and avoid thermal throttling reduce tail latency, which is what actually determines training throughput.

What should a GPUaaS provider ask an interconnect supplier?

Ask whether the supplier manufactures or sources, whether they own their firmware and can code for the switch platforms in use, what the measured power figure is at realistic case temperature, what documentation and traceability comes with each batch, and whether the same specification can be guaranteed across build phases twelve to eighteen months apart.
arrow-right