Red River Insights - AI's Picks and Shovels top 10
AI's value pool is broadening - and the bottlenecks beneath the models are where it's going
Red River Insights - March 2026
Dear friends,
Just four companies, Alphabet, Amazon, Meta and Microsoft, are expected to spend about $650 billion on AI infrastructure in 2026, up from roughly $410 billion in 2025. AI’s first phase rewarded the model builders. The next phase is beginning to reward the bottlenecks beneath them. Scaling AI now forces capital into the physical layers underneath: power, interconnects, packaging, metrology, memory, storage, and cooling. Models create the demand. Infrastructure captures the choke points. And the choke points are where scarcity sits.
That shift is already visible. On March 2, NVIDIA committed $2 billion each to Lumentum and Coherent to deepen its optics supply chain, a reminder that even the biggest AI winners are now buying further down the stack. Cisco estimates that agentic AI queries can generate up to 25 times more traffic than a standard chatbot, which helps explain why optical networking - the use of light rather than electrical signals to move data between chips, because copper simply cannot keep up at AI scale (also known as photonics) - is shifting from technical detail to strategic bottleneck.
In our January edition, we covered the energy layer powering AI. This month, we focus on the layer that sits between energy and AI models: the enabling hardware and deep tech that translates raw power into usable compute. By AI picks and shovels, we mean the technologies that sell into many AI winners without having to predict which model, cloud, or application ultimately dominates. In Europe, those capabilities are concentrated in specialist parts of the stack: photonics, metrology (the science of ultra-precise measurement - in this context, the tools that inspect and verify chips, wafers, and optical components at nanometre scale), advanced packaging, alternative compute, cooling, and new storage architectures. Our ranking highlights the European startups rethinking how these critical AI infrastructure layers are built and scaled.
We did not assemble this ranking manually. It is the output of Ramp’s proprietary growth scoring system, an algorithmic framework refined over five years to identify the European companies whose trajectories are most leveraged to the AI infrastructure buildout. We excluded businesses whose primary exposure remains automotive, telecom, or other end-markets. The ranking expanded beyond ten: the European picks-and-shovels ecosystem proved deeper than we expected, spanning photonics, neuromorphic chips (processors inspired by the human brain that handle pattern recognition and sensor data at a fraction of the energy cost of traditional chips), DNA storage, and more. Cutting it at ten would have meant excluding companies that genuinely stand out.
RAMP’s AI Picks and Shovels top 12
We hope this sparks interesting conversations. If you have any comments or would like to suggest a startup that should be included, feel free to reach out to us. Joseph, Chloe and Olivier will be delighted to discuss these trends and rankings.
We’ve highlighted four key trends illustrated by these companies:
1. Light and copper: both winning, different layers
Five of our top 12 build photonic or optical technology: iPronics #7, NcodiN #8, PHIX #9, Akhetonics #10, and Scintil Photonics #12. The AI buildout is a massive tailwind for both copper and optics, but at different layers. Copper remains dominant for short-range connections inside servers and between boards, and the sheer volume of new data centers means copper demand is surging. But as AI clusters scale to tens of thousands of GPUs, optics are becoming critical where copper hits limits on power, reach, and bandwidth density: rack-to-rack, building-to-building, and campus-scale networking. NVIDIA signaled this in early 2026 by investing $4 billion in optical networking companies Coherent and Lumentum.
What do these companies actually do? Scintil Photonics (Grenoble, $58M Series B with NVIDIA participation) makes the laser chips that generate the light signals. iPronics (Valencia) builds programmable photonic chips that let operators reconfigure optical networks on the fly. NcodiN (Grenoble, another graduate of France’s powerhouse CEA-Leti ecosystem, not a coincidence that two of the five come from the same postcode, Grenoble is one of Europe’s deepest photonics clusters) moves data between chips using light instead of copper. PHIX (Enschede, Netherlands) packages all of these circuits into production-ready components. And outside the ranking, Salience Labs (UK) just launched a 32-port all-optical switch with backing from Applied Materials.
Europe’s strength in photonics is real, built on decades of public R&D in Eindhoven, Grenoble, and the UK. That said, each of these companies faces well-funded US competitors, this is not a position of dominance but of depth.
2. Not all AI chips look like GPUs
SynSense (#3) and SpiNNcloud (#6) build processors that work nothing like a conventional GPU. Instead of brute-force number crunching, they mimic how biological neurons fire, processing information in spikes rather than streams, consuming a fraction of the power. These are called neuromorphic chips: processors inspired by the human brain’s architecture, designed to handle pattern recognition and sensor data at radically lower energy cost than traditional silicon. SynSense (Switzerland, backed by Huawei, Samsung, and Baidu) targets devices where power is scarce: sensors, cameras, hearing aids. SpiNNcloud (Germany, spinout of the EU’s Human Brain Project) runs a neuromorphic supercomputer with 5 million processing cores.
These won’t replace NVIDIA in the data center. But AI is fragmenting across hardware: training stays on GPUs, while inference is splitting across specialized chips depending on the task. Optalysys (#4, UK) takes this further with photonic processors that compute using light, enabling encrypted AI workloads to be processed without ever being decrypted. Akhetonics (#10, Germany) is working on an all-optical general-purpose processor. European labs (imec, CEA-Leti, the Human Brain Project) invested heavily in these alternative architectures before they were commercially relevant, which partly explains why several of the most advanced startups in the space are European. Whether that head start translates into durable market positions is another question entirely.
3. The invisible factories behind every AI chip
Nearfield Instruments (#1) builds something most people have never heard of: metrology systems that inspect chips at the nanometre scale during fabrication. No advanced AI chip, from NVIDIA, AMD, or anyone else, ships without being measured by tools like these. The Netherlands-based company has raised over $218M across six rounds, including a $148M Series C from Walden Catalyst and Temasek, making it one of Europe’s best-funded deep tech companies.
PHIX (#9) plays the same enabling role for photonics: a packaging foundry without which many of the optical chip companies in this ranking couldn’t bring their designs to production. Also in the Netherlands, SMART Photonics runs Europe’s only pure-play Indium Phosphide foundry, a strategic asset that has attracted €190M. This is the ASML pattern at smaller scale: Europe builds the machines that build the machines. Foundry and metrology capacity is the scarcest resource in the semiconductor supply chain right now, and scarcity rewards those who own it.
4. DNA storage, satellite links, and where the stack goes next
Biomemory (#2) may be the most audacious company on the list: they store data in synthetic DNA molecules. One gram of DNA can theoretically hold 215 petabytes. The Paris-based company (backed by Bpifrance, $18M Series A) is targeting first commercial solutions in H2 2026 following the acquisition of Catalog’s DNA data storage assets. The promise is extraordinary density, orders of magnitude beyond anything else available, but the technology is not without hard constraints: reading DNA data is largely solved, but writing remains slow and expensive (thousands of dollars per megabyte today). The most realistic near-term use case is deep cold archival, where write speed matters less than density and durability. Data centers are running out of physical space, and that’s the wedge.
mBryonics (#5) builds laser links between satellites, exactly the technology SpaceX’s proposed orbital data centers would need. The Irish company bridges the AI infrastructure story with the space connectivity buildout, and its growth trajectory reflects it.
Submer (#11, Spain, €127M raised) tackles what may be the most immediate physical bottleneck in AI infrastructure: heat. As GPU clusters get denser, traditional air cooling can’t keep up. Submer builds immersion cooling systems, submerging servers in dielectric fluid, that cut energy use and let data centers pack more compute into less space.
Many issues are still up for debate
(Reach out if you have an opinion or to suggest a company we might have missed!)
Will European photonics companies build independent businesses or become acquisition targets? NVIDIA, Broadcom, and Cisco are all building out their optical supply chains. European photonics startups have the technology, but US hyperscalers have the purchasing power. The history of European deep tech is littered with companies that were bought too early. Can Scintil, Salience Labs, or iPronics become the next ASML rather than the next acquisition footnote?
Is neuromorphic computing a real market or a research project? SynSense has commercial customers (Samsung, Huawei), but the total addressable market for neuromorphic chips remains small relative to GPUs. Will edge AI create enough demand to sustain independent neuromorphic chip companies, or will conventional accelerators absorb the efficiency gains through architectural improvements?
When does the “memory wall” become a bigger bottleneck than compute? Some of the most sophisticated deep tech investors are betting that AI workloads are already memory-bound, not compute-bound. If true, the next wave of infrastructure value will shift from processors to memory and storage, a thesis that benefits companies like Biomemory and the emerging chiplet ecosystem more than GPU cloud providers.
Can Europe build a full AI infrastructure stack, or only pieces of one? The AI hardware ecosystem today is globally distributed: NVIDIA designs the chips in the US, TSMC manufactures them in Taiwan, ASML provides the lithography from the Netherlands, and hyperscalers deploy them worldwide. Europe has world-class positions in photonics, metrology, and alternative compute, but no equivalent vertical integration. Is a distributed European supply chain a feature (resilience) or a structural disadvantage?
More on RAMP's scoring method
The ranking of these startups is based on the estimated momentum of the company, but the algorithm does not assess the quality or reliability of the products/solutions developed by these companies!
In case you missed them, our latest top 10s are here: AI safety and governance (February26), Energy (January26), Nature & Biodiversity (December25) and Digital Biology (November25)
All the previous Top 10 are here.
Portfolio news:
Okeiro hit a major regulatory milestone: its iBox biomarker is the only surrogate endpoint with an accepted qualification plan in the FDA Biomarker Qualification Program, with the full package now under Agency review. The iBox predicts individual graft survival up to 10 years post-transplant and is already the endpoint of choice in active trials with Sanofi, argenx, and academic centers across three continents.
Otera announced a strategic partnership with Horn & Company to deploy agentic AI across complex, regulated enterprise processes at scale, combining Horn's transformation expertise with Otera's autonomous AI agents.
HyPrSpace landed a feature in Le Parisien profiling founders Alexandre Mangeot and Sylvain Bataillard as they prepare the first test flight of Baguette One. The Bordeaux-based launcher company also wrapped up its Lift-Off Challenge contest (3,000+ participants, only 30 solved it), with the grand prize winner earning the right to place an object in the rocket’s nosecone.
Worldia is expanding into Canada, with Canadian travel agents reporting they can build trips in under 10 minutes on the platform. The team also ran a rapid crisis response operation, reprotecting all affected travelers during the Middle East flight disruptions in early March.
You received this newsletter because we met or worked together and we thought you might be interested in an update, if we got it wrong, feel free to unsubscribe with the link at the end!