Musk raises Tesla AI5 memory to 96GB after cut

Musk's bump of AI5 memory from 72GB to 96GB LPDDR5 is not a spec tweak—it is an anchoring move in the automotive AI chip race. Technically, 96GB signals that FSD's on-device inference models have crossed the 7B-parameter threshold, demanding 80GB+ headroom for KV caches and intermediate activations. Choosing LPDDR5 over LPDDR5X is a deliberate power-thermal tradeoff: in automotive contexts, bandwidth-per-watt matters more than peak bandwidth. Memory controller design, PCB stack-up, and battery thermal management all get recalibrated around this figure. On the supply chain, 96GB adds roughly 20% to BOM cost versus 72GB, but the subtler risk is bargaining power. As Tesla becomes the largest single buyer of a given LPDDR5 generation, Samsung, SK Hynix, and Micron gain leverage on capacity allocation and pricing. Musk's insistence on not being the sole player at the lowest tier is really a benchmarking statement against NVIDIA Thor and Qualcomm Snapdragon Ride Elite roadmaps—a preemptive defense against a generational spec gap. Competitively, NVIDIA Thor targets 24GB GDDR6 while Qualcomm pursues heterogeneous memory. Tesla jumping to 96GB resets the industry baseline from sufficient to abundant, forcing rivals to either follow and inflate sector-wide costs, or differentiate architecturally. Within 18 months, automotive AI chip memory will collectively migrate from the 16-32GB band to 64-128GB. On-device large models become a hard requirement. Memory capacity and bandwidth are replacing TOPS as the new arms-race metric.