The Great Pivot: Why the AI Race is Shifting from Training to Inference
For years, the AI narrative has been dominated by the massive compute power required to train large language models. This era belonged to the GPU, with Nvidia establishing a near-monopoly. However, a fundamental shift is occurring: the industry is moving toward AI inference—the process of actually running those models to generate answers and execute tasks.
The problem? Existing GPU architectures weren’t originally built for inference at a massive scale. As the demand for agentic AI workloads grows, the industry is hitting a wall regarding power consumption and heat. This has opened a window for a new generation of “inference-first” hardware designed to be leaner, faster, and significantly more energy-efficient.
Euclyd: Reimagining Silicon from the Ground Up
At the forefront of this European challenge is the Eindhoven-based startup Euclyd. Rather than simply iterating on existing designs, Euclyd is pursuing a “moon-shot” approach, building its architecture from scratch without relying on ARM or other standard architectures.
Their flagship solution, CRAFTWERK, is a system-in-package (SiP) designed specifically for large-scale AI inference. The technical specifications are staggering: it integrates 16,384 custom SIMD processors and 1TB of custom ultra-bandwidth memory (UBM). This UBM is claimed to deliver 8,000 terabytes per second of bandwidth, potentially outperforming Nvidia’s HBM.
The goal is efficiency. Euclyd claims its system can deliver 100x higher power efficiency for inference compared to Nvidia’s latest generation Vera Rubin chips. By processing data in multiple places rather than constantly moving it through a memory stack, Euclyd aims to slash the cost and energy requirements of AI infrastructure.
The Next Frontier: Photonics and the Conclude of Electronic Limits
While Euclyd optimizes electronic architecture, other European players are betting that electrons themselves are the problem. The U.K. Startup Olix is developing photonics-based processors that use light instead of electricity to move data and perform computations.
The industry is reaching a physical limit on how small electronic components can be made. As chips shrink, the heat they generate becomes a critical failure point. Photonics offers a potential paradigm shift, promising to bypass these thermal limits and provide a more scalable path for hyperscalers and governments requiring massive inference services.
This represents a battle Nvidia is watching closely. The chip giant has already invested $4 billion in photonics technology and acquired assets from inference startup Groq for $20 billion to protect its lead.
The Geopolitical Push for Sovereign Compute
The rise of these startups isn’t just about technical specs; it’s about geopolitical necessity. With U.S. Export controls and a heavy concentration of chip production at TSMC, Europe is facing a “sovereign compute imperative.”
Investment is flowing into homegrown silicon to reduce dependency on foreign tech. Companies like Fractile (U.K.), Arago (France), and Axelera (Netherlands) are eyeing nine-figure funding rounds to scale their operations. However, the gap remains wide: in 2026, European AI chip startups raised $800 million, compared to $4.7 billion for their U.S. Counterparts.
Structural Hurdles for European Silicon
Despite the talent, European startups face systemic challenges that their U.S. Rivals do not:
- Funding Gaps: A lack of massive, early-stage capital compared to the U.S.
- Ecosystem Maturity: A foundry ecosystem that still needs to mature to support volume deployment.
- Government Conservatism: A lack of a DARPA-equivalent agency to aggressively fund high-risk, high-reward tech projects.
- Labor Laws: Fragmented regulations across borders that complicate the recruitment of top-tier talent.
Frequently Asked Questions
What is AI inference?
Inference is the phase where a trained AI model is used to process new data and provide a result (e.g., a chatbot answering a question), as opposed to the training phase where the model learns from a dataset.
How does Euclyd differ from Nvidia?
While Nvidia uses GPUs (originally for gaming), Euclyd uses a custom architecture with its own processors and ultra-bandwidth memory (UBM) specifically optimized for inference efficiency.
What are photonic processors?
These are chips that use light (photons) instead of electricity (electrons) to move and process data, aiming to solve the heat and size limitations of traditional silicon.
Join the Conversation: Do you think Europe can successfully build a “Dutch Nvidia” to achieve tech sovereignty, or is the U.S. Funding lead insurmountable? Let us know in the comments below or subscribe to our newsletter for more deep dives into the future of AI hardware.
