Nvidia unveiled Vera, the successor to Grace CPUs, with NVLink-C2C designed to pair with Rubin GPUs and support multi-socket configurations. Developers running on-premises or hybrid infrastructure can now leverage Nvidia's latest CPU generation as a production-ready alternative to custom silicon for AI workloads.

DeepSeek ordered 160K Huawei Ascend chips to build a gigawatt data center, while China's domestic chip share approaches 90%. This signals accelerating decoupling of Chinese AI infrastructure from Nvidia dependencies and sustained capex intensity in frontier model training.

Power availability, not capital, is now the binding constraint on data center growth in 2026; geothermal offers 90%+ capacity factor with minimal grid dependency. Infrastructure teams must treat power procurement and grid interconnection as critical path items for any >100MW AI deployment.

XPUs now represent 73% of Nvidia's AI revenue; CEO Hock Tan stated custom accelerators achieve frontier model inference at half the cost of GPUs. Named deployments include Google TPU v8i at scale and one gigawatt of Ironwood for Anthropic, signaling major cloud providers are shifting margin toward custom silicon.

Far Labs is shipping software for distributed inference across residential GPUs using isolation, least-privilege resource limits, and encrypted auth. This approach addresses cloud inference cost pressure by monetizing idle consumer hardware while maintaining security and resource guarantees.