IEEE Spectrum's inference-silicon survey frames the whole fight as memory bandwidth.

The September 15 feature catalogs GPUs idling 50-80% of the time waiting on memory, then maps the contenders: Nvidia's $20B Groq acquisition producing an LPU with 500MB on-chip SRAM and 7x GPU memory bandwidth, a Cerebras WSE-3 deployment pushing GPT-5.3-Codex-Spark past 1,000 tokens/second off 44GB of SRAM, Majestic Labs at 128TB DRAM per rack against GB300's ~20TB HBM3E, Etched's Sohu claiming 500,000 tokens/sec for Llama 70B. The cost driver behind the whole split is that HBM runs two to three times commodity DRAM.