Tesla redefines its accelerator roadmap with SI6 and Dojo D3 - Foro3D

Tesla's simultaneous launch of SI6 and Dojo D3 is not a product iteration—it's a sovereignty play in compute. The architectural signal is clear: inference and training are now decoupled into dedicated silicon. SI6 targets edge efficiency for vehicle-side workloads; Dojo D3 handles cluster-scale training throughput. The immediate upstream shock lands on advanced packaging and HBM supply. As compute demand migrates from NVIDIA GPUs to proprietary ASICs, the allocation logic for CoWoS-class packaging capacity gets rewritten, and the pricing power of Taiwan, China's OSAT players and Japanese memory makers faces repricing. On compliance, the in-house route shifts risk from procurement to fabrication. If process nodes dip below 7nm, foundry selection inevitably collides with export-control frameworks. Geographic concentration risk doesn't vanish—it just changes hands. In the competitive arena, NVIDIA's real threat isn't Tesla per se, but the 'de-NVIDIA' narrative spreading across automakers and robotics. AMD's MI300 and Intel's Gaudi 3 are losing their window. When anchor customers lock their software stacks to custom silicon, the premium logic of general-purpose GPUs erodes incrementally but irreversibly. The 18-month watch item: actual Dojo cluster deployment scale, and how deeply FSD's software stack adapts to SI6's ISA. If adaptation exceeds two quarters, hardware lead gets swallowed by software debt.