South Korean chip startup FuriosaAI invades European datacenters

FuriosaAI began deploying its RNGD AI accelerator line at Equinix’s LS2 datacenter in Lisbon, signaling a push into Europe to capture rising sovereign AI compute demand. Each RNGD PCIe card delivers 512 teraFLOPS dense FP8 performance with 48GB HBM3 and 180W TDP, and eight cards run in a 3kW NXT server with up to 384GB HBM. The company also plans a third-generation RNGD accelerator with Broadcom, targeting higher performance and cluster scalability using faster HBM4/HBM4e, though availability is likely delayed by memory ramp timing.

Analysis

This is best read as a validation of the inference-optimization trade, not a headline-size earnings event. The market mechanism is rack economics: when a 3 kW, air-cooled system can be dropped into existing colo space, the bottleneck shifts from chip performance to available power and deployment speed. That structurally favors EQIX and other colocation operators with stranded capacity, while nibbling at the premium attached to very high-TDP GPU stacks in power-constrained European markets.

The more meaningful second-order winner is AVGO. Every additional custom accelerator program that needs Ethernet/PCIe scale-up, packaging, and memory integration reinforces Broadcom’s role as the common enabler across fragmented AI silicon efforts. TSM gets a quieter long-duration tailwind from more non-NVIDIA wafers and advanced packaging content, but the near-term financial effect is immaterial unless these programs move from pilot to meaningful volume.

Contrarian view: the consensus may overread this as a competitive threat to NVDA. The limiting factor is not chip awareness but HBM4 supply, software maturity, and deployment scale, so this is a 6-18 month share-erosion story at best, mostly in inference and sovereign-AI use cases. What would falsify the bull case for efficient inference hardware is a rapid decline in European power costs or a shift by buyers toward liquid-cooled, hyperscale-only procurement, which would reduce the premium on air-cooled density and slow colo monetization.

Key Decisions for Investors

• Add EQIX on any pullback over the next 1-3 months: this is a modest but durable positive for colo utilization and retrofit economics; expect the thesis to work only if European AI tenants continue choosing deployability over peak performance.
• Overweight AVGO versus NVDA on a 6-12 month horizon: AVGO is the cleaner way to express the custom-silicon and AI networking buildout, while NVDA faces incremental inference substitution risk; thesis breaks if Broadcom AI commentary slows or NVDA maintains overwhelming software lock-in.
• Small tactical long TSM on weakness for a 6-18 month view: the upside comes from broader custom-accelerator wafer demand and HBM4 packaging, but size should be limited because the article signal is too early to move fundamentals materially.
• Avoid chasing a standalone NVDA short here; use this as a watch item for sovereign-AI tender wins, power-price trends, and HBM4 availability before leaning into a larger competitive-trend trade.