Flatkey Launches Flatkey Compute, a NeoCloud for Dedicated NVIDIA B300 Clusters - Your Valley

Flatkey Compute's launch marks a structural pivot in NeoCloud: from GPU rental to compute sovereignty. Technically, a dedicated B300 cluster delivers a physically isolated NVLink domain and InfiniBand fabric, eliminating cross-tenant bandwidth contention that cripples shared hyperscaler environments. For large-model training, this is a qualitative leap in all-reduce efficiency. Yet B300's liquid-cooling density—over 1,000W per GPU—imposes infrastructure demands far beyond H100-era facilities. Without scaled liquid-cooling deployment within twelve months, Flatkey's "dedicated cluster" remains a spec-sheet promise. Compliance-wise, B300's export-control status confines the addressable market to the US and a few exempt jurisdictions. TSMC (Taiwan, China) as sole foundry means any geopolitical friction severs supply. As a private entity, Flatkey lacks the governmental buffer hyperscalers enjoy, making it structurally more exposed than CoreWeave. Competitively, CoreWeave anchors frontier labs; Lambda owns inference. Flatkey's only wedge is mid-market dedicated clusters, but that segment is acutely TCO-sensitive. Once NVIDIA opens official B400 leasing in 2026, NeoCloud intermediary margins compress below 15%. Outlook: expect first-round NeoCloud consolidation within 12–24 months. Survivors need owned power assets or long-term PPAs. If Flatkey misses operational B300 delivery by Q4 2025, the market reclassifies it as proof-of-concept, not infrastructure.