Etched Must Become Default Rack

This shift means Etched is not really selling a faster chip, it is selling a preassembled AI factory slice that has to win on total tokens per rack, power, memory, networking, and ease of deployment all at once. Nvidia and AMD now package compute, CPUs, interconnect, NICs, and software as rack products, so buyers compare full systems that drop into a data center rather than individual accelerators.

• Nvidia has made the rack the baseline buying unit with Vera Rubin NVL72, a system that bundles 72 GPUs, 36 CPUs, NVLink, SuperNICs, DPUs, and rack level networking into one reference platform. That turns the startup challenge from beating a GPU on specs to replacing an entire trusted deployment pattern.
  1 nvidia 3 nvidia 6 sacra
• AMD is following the same playbook with Helios, its first rack scale AI system now in production for 2026 deployments. That gives customers a second mainstream path to buy non Nvidia infrastructure without betting on a young chip company, which raises the bar for Etched from technical novelty to full system credibility.
  2 amd 7 amd 8 amd
• The startups converging on Etched's turf are also moving up the stack. Fractile sells complete servers or racks with runtime and fleet software, SambaNova sells dedicated racks plus cloud and managed deployment, and d-Matrix bought GigaIO's data center business to add rack scale systems engineering. The product boundary across the category is now the full serving system.
  4 sacra 5 sacra 9 dmatrix

The next phase of inference competition will be decided by who can become the default rack for decode heavy and agentic workloads. If Etched lands as a full rack standard, it can capture more value per deployment and lock in software, operations, and follow on capacity, instead of being treated as just another chip inside someone else's cluster.