FuriosaAI Raises 820 Billion Won, Valued at 3.8 Trillion, Challenges NVIDIA - 조선일보

The 820B KRW raise isn't a funding event—it's a repricing. At 3.8T KRW (~$2.7B), investors are betting on architectural substitution, not performance parity. FuriosaAI's WASP dataflow design bypasses CUDA's SIMT paradigm, striking at the 75% gross margin that NVIDIA's software lock-in generates. Technical ripple: If sparsity-native inference scales, HBM bandwidth curves diverge from NVIDIA's dense-matrix assumptions, forcing SK Hynix and Samsung to recalibrate roadmaps. EDA vendors face model-library rebuilds for non-standard architectures—upstream cost structures shift. Compliance ceiling: Korea's chip-sovereignty push mirrors the EU CHIPS Act, but BIS export controls on advanced packaging (CoWoS-class) impose a hard ceiling. FuriosaAI cannot access the most advanced nodes in Taiwan, China. That is the real cost of challenging NVIDIA and the structural cap on its valuation. NVIDIA's likely counter: not price cuts, but ecosystem deepening—accelerated CUDA on-premise licensing, tighter binding with Naver/Kakao, and a good-enough inference SKU (extending the A100→L40S→B200 trajectory) that undercuts FuriosaAI's cost argument. 12–24 month tail: Disruption won't come from one Korean startup. It will emerge from the convergence of FuriosaAI, Cerebras, Groq, and hyperscaler ASICs collectively compressing NVIDIA's inference margin from ~75% toward ~55%. Training remains NVIDIA's fortress; inference is where the war is actually being fought.