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

Flatkey Compute's launch signals less a new entrant than a structural shift: AI compute consumption is migrating from shared, pooled infrastructure to dedicated, domain-isolated clusters. Technically, delivering a B300 cluster is a systems-integration problem, not a reselling one. Blackwell's HBM3E bandwidth demands, liquid-cooling density, and full NVLink topology mean Flatkey is effectively a hardware integrator wedged between NVIDIA and the end user, with supply-chain exposure running through TSMC's advanced-node fabs in Taiwan, China and SK Hynix's HBM allocation. The "dedicated" qualifier is the real differentiator: customers receive an unfragmented NVLink domain rather than virtualized slices, a qualitative leap for thousand-GPU training runs. On compliance, B300 sits at the frontier of US export controls. The compute-as-a-service delivery model is generating a regulatory gray zone: does provisioning dedicated B300 compute to a foreign entity carry the same legal weight as shipping the silicon? No 2025 enforcement precedent exists. Flatkey's compliance exposure dwarfs what its visible revenue line suggests. Competitively, CoreWeave has already locked in large-scale GB200/B200 deployments while Lambda Labs and Crusoe are positioning on the inference side. Flatkey is exploiting a narrow timing window during B300 cluster tuning. But NVIDIA's allocation priority among NeoCloud partners remains the true kill switch. Within 18 months, expect a first round of NeoCloud consolidation. Compute arbitrage—locking clusters at fixed cost, reselling by the hour—will become the dominant margin model, converging logically with data-center REITs. Players lacking power-infrastructure reserves and long-term NVIDIA supply agreements will be systematically squeezed out.