Samsung has co-led a 200 million euro, about $230 million, funding round for Euclyd, a two-year-old Dutch startup building AI inference chips designed to rival Nvidia’s dominance, the companies confirmed Monday. Somerset Capital Partners, the Scaleup Europe Fund, managed by EQT, and Innovation Industries joined the round, which was the Eindhoven-based company’s Series A. Samsung’s individual contribution was not disclosed.
Euclyd, founded in 2024, is developing a processor and memory architecture built specifically for AI inference, the stage when trained models generate answers, images or predictions. Unlike Nvidia’s general-purpose GPUs, the company says its design can reduce both the electricity consumption and the operating cost of running AI data centers. That pitch matters because inference, not training, is where most of the industry’s projected compute spending will land as models move from labs into everyday products.
A crowded field of Nvidia challengers
The round lands in a market that has turned hostile to complacency about Nvidia’s lead. In February, Reno-based Positron AI raised a $230 million Series B at just over $1 billion, claiming its Atlas chip delivers three times the compute per watt of Nvidia’s H100. Positron’s second-generation chip, Asimov, supports 2TB of memory per accelerator and 8TB per Titan system, bandwidth the company compares to Nvidia’s Rubin GPU, and is on track to tape out in October 2026 with production slated for early 2027. At rack scale, Positron says these figures translate to memory capacity totaling more than 100TB.
The giants are building their own silicon too. OpenAI has partnered with Broadcom on its first in-house processors, Google designs its own TPUs with Broadcom, Amazon builds Trainium chips, and Meta develops custom accelerators. Nvidia, meanwhile, agreed to license technology from startup Groq and hire its CEO Jonathan Ross, who helped start Google’s AI chip program, in a deal reported at $20 billion, its largest ever, as it works to defend the position all of these efforts are attacking.
| Company | Approach | Latest funding |
|---|---|---|
| Euclyd (Netherlands) | Purpose-built inference architecture | EUR 200M Series A, Sept 2026 |
| Positron AI (US) | Atlas/Asimov inference chips, 3x perf-per-watt claim | $230M Series B, Feb 2026 |
| OpenAI | First-party chips with Broadcom | Partnership, in production |
| Nvidia | GPUs plus Groq licensing deal | $20B Groq acquisition of assets |
Why Samsung is buying both sides
Samsung’s position in this market is double-edged. The company supplies critical memory and fabrication technologies to Nvidia itself, making it a key player in the incumbent’s supply chain, while simultaneously backing challengers through investments like this one. The hedge is rational. If hyperscalers shift meaningfully toward alternative architectures, Samsung wants to be positioned as their foundry and memory partner regardless of whose chips win.
It also reflects where the demand signal has moved. Hyperscalers are openly looking for cheaper inference solutions as their AI capital spending draws investor scrutiny. Several of the largest have told analysts they expect inference to become the majority of their compute workload within a few years, and inference economics, power draw and cost per token, now drive procurement decisions that GPUs alone used to settle. Data center power availability has become the binding constraint on AI expansion, which makes any architecture that promises lower consumption per operation a procurement conversation rather than a curiosity.
The round underscores growing investor appetite for alternatives to Nvidia’s dominant graphics processors, with chip giants joining AI and robotics rounds totaling over $250 billion in 2026 to date, according to Crunchbase.
The European angle
Euclyd’s location matters as much as its architecture. European deep-tech funding has strengthened over the past two years, and EQT’s Scaleup Europe Fund exists partly to keep chip talent on the continent rather than watching it migrate to US labs. The Netherlands has a genuine chip heritage through ASML, the company whose lithography machines every advanced fab depends on, and Eindhoven’s Brainport region hosts a dense supplier network that a hardware startup can actually use.
Samsung’s foundry capacity gives the startup a credible manufacturing path that most seed-stage hardware companies lack. Access to advanced nodes is the single hardest input for a chip startup, and a strategic investor who also runs fabs changes that equation. The companies did not announce a tape-out date for Euclyd’s first chip.
The hard part is still ahead
Euclyd is two years old and has not taped out a chip at the scale of its claims, and it is entering a market where Nvidia’s CUDA software ecosystem remains the default for most developers. Hardware alternatives have repeatedly found that matching raw specifications is easier than displacing the software stack. Positron, three years in, is only now shipping its first-generation chip and faces the same test. Cerebras, Groq and SambaNova, earlier generations of the same thesis, have spent years fighting for meaningful market share against the same headwind.
What distinguishes the current wave is the buyer. In 2022, alternative chip companies were selling into a market where Nvidia had no supply constraints. In 2026, cloud providers cannot get enough GPUs and are actively second-sourcing. A challenger no longer needs to beat Nvidia on merit; it needs to be good enough to relieve a shortage, which is a much lower bar and explains why funding has accelerated across the category.
For Nvidia, the sum of these efforts is still small relative to its revenue, but the direction is the story. Every major cloud provider, several national governments and now a growing list of specialized startups are funding the search for something cheaper to run models on. Most will fail. The ones that do not will reshape the economics of an industry currently spending hundreds of billions a year on compute.