Samsung backs Nvidia AI chip rival in $230 million funding round as GPU alternatives boom

Dutch AI-chip startup Euclyd raised €200 million ($231 million) in a Series A co-led by Samsung, Somerset Capital Partners, EQT-managed Scaleup Europe Fund and Innovation Industries. Euclyd is developing a non-GPU architecture for AI inference that it says could lower AI data-center energy use and costs, targeting hardware sales and IP licensing. The company plans to begin deploying physical systems in 2028 and serve thousands of enterprise customers by 2030, though its technology has not yet been validated in commercial deployments at scale.

Analysis

This is not a near-term revenue threat to NVDA: an unproven architecture with a multi-year commercialization path cannot displace the CUDA/software ecosystem or validated hyperscaler supply chain in the next 12-24 months. The relevant signal is that inference is becoming the primary battleground for cost-per-token, where memory bandwidth, power draw and rack-level integration matter more than peak training FLOPS. That favors memory suppliers and advanced packaging providers before it affects accelerator incumbents' revenue, particularly if customers diversify inference capacity while retaining NVDA for training.

Samsung's involvement raises the probability that memory-centric or tightly integrated logic-memory designs gain a credible manufacturing route. MU and SK Hynix (000660 KS) are cleaner public read-throughs than NVDA, because a broader set of custom inference ASICs increases high-bandwidth-memory and advanced-memory content even when GPU share falls. The offset is that custom silicon can ultimately pressure memory pricing if architectures use lower-cost commodity DRAM rather than HBM; disclosed memory configuration and performance-per-watt benchmarks are the key missing data.

Hyperscalers' internal-chip efforts remain more consequential than venture-backed challengers. AMZN, GOOG and META can use proprietary inference silicon to lower depreciation and power intensity, but the financial benefit appears only when utilization is high enough to amortize design, software-porting and deployment costs. Consensus may overstate the immediate NVDA disruption while understating a 6-18 month risk to inference accelerator pricing and gross margin if cloud customers demonstrate materially lower cost-per-token on non-NVIDIA systems.

There is no actionable implication for EQT (the U.S. natural-gas producer); the named investment manager is not a fundamental catalyst for EQT Corp. Treat any equity-price linkage from automated news classification as noise. The thesis for inference disaggregation is falsified if NVDA maintains a clear total-cost-of-ownership advantage after networking, software migration and utilization are included, or if custom-chip deployments remain confined to narrow internal workloads.

Key Decisions for Investors

• No directional NVDA trade solely on this development. Maintain existing exposure; reassess only after independently audited inference benchmarks show at least 30% lower all-in cost-per-token versus comparable NVIDIA systems and a named scaled customer commitment. Near-term downside thesis lacks a revenue catalyst.
• Watch-list long MU versus short NVDA over a 6-18 month horizon only if hyperscaler capex disclosures show custom-inference deployments increasing while HBM bit-demand guidance remains firm. The pair isolates a shift from accelerator value capture toward memory content; exit if MU pricing guidance weakens or NVDA inference revenue/gross-margin guidance accelerates.
• For AMZN, GOOG and META, monitor quarterly depreciation, AI infrastructure capex and management commentary on inference unit economics rather than treating proprietary silicon as immediate margin upside. A sustained reduction in capex intensity or improving operating margin alongside rising AI usage would support a 12-month long bias, particularly versus software peers with less infrastructure control.
• Set an alert for commercial deployment disclosures, foundry partner selection, tape-out timing and disclosed memory stack. Until those milestones are public, regard the private-company valuation as venture optionality rather than a tradable semiconductor supply-chain signal.