How DeepSeek V4 Connects to the US Power Grid

What It Covers

The White House invokes the Defense Production Act to expand US grid infrastructure as AI power demand threatens to double data center electricity consumption from 6% to 11% of US supply by 2030, while DeepSeek releases V4 at one-seventh the cost of comparable US frontier models, reshaping enterprise AI economics.

Key Questions Answered

• •AI Infrastructure Leverage: Cloud giants Amazon and Google are extracting significant strategic leverage from compute scarcity. Google's $40B Anthropic deal (structured as $10B upfront, $30B milestone-based) may give Google a 20%+ ownership stake at potentially 50% below Anthropic's $800B secondary market valuation, making infrastructure providers the structural winners regardless of which AI lab wins.
• •Power Grid as National Security Asset: The White House Defense Production Act memo authorizes the Secretary of Energy to make direct purchases and financial commitments to expand domestic grid manufacturing — transformers, transmission lines, high-voltage breakers, and electric core steel. Enterprises building long-term AI infrastructure plans should factor in multi-year grid expansion timelines as a binding constraint.
• •DeepSeek V4 Price Disruption: DeepSeek V4 Pro benchmarks near GPT-5.4 and Opus 4.6 at $1.74 per million input tokens — less than one-seventh the cost of Opus 4.6. V4 Flash undercuts Gemini Flash Lite by 80% at $0.14 per million inputs. Businesses running high-volume, non-frontier workloads should evaluate DeepSeek V4 as a cost-reduction lever immediately.
• •Open Source Dependency Risk: US enterprises adopting Chinese open-source models like DeepSeek create geopolitical supply chain exposure. If Chinese labs alter architecture or restrict access, dependent companies face sudden disruption. Procurement teams should treat open-source AI model provenance as a vendor risk factor alongside standard security and compliance assessments.
• •CPU Architecture for Agentic AI: Meta is renting Amazon's Graviton 5 CPUs — not GPUs — specifically for agentic workloads, signaling that CPU architecture may outperform GPU architecture for running AI agents at scale. Teams building agent infrastructure should evaluate CPU-optimized cloud instances alongside standard GPU clusters when designing cost and performance benchmarks.