MCP Hits Production Scale, Scaling Bottleneck Shifts to Power

MCP moved from research protocol to infrastructure standard this week. The July 2026 release candidate arrived with a stateless core, tighter authorization, and a standards lifecycle that treats the specification as a governed artifact. Enterprise adoption numbers back the shift: 78% of enterprise AI teams now run MCP agents in production, and 28% of Fortune 500 companies operate MCP servers. Meanwhile, infrastructure capital continued its consolidation into the largest players, with the bottleneck moving decisively from silicon to electrical power and real estate.

Models

Kimi K3 arrived July 16, marking Moonshot AI’s entry into the frontier competition. xAI’s Grok 4.5, released July 8, proved that training a 1.5 trillion parameter mixture-of-experts model on Cursor interaction data yields practical efficiency: it scored 83.3% on Terminal-Bench 2.1 while consuming about 25% fewer output tokens than Claude Opus 4.8 on similar tasks. Anthropic’s Claude Sonnet 5 and Claude Fable 5 continued the pattern of models built explicitly for specific workflows, with Fable 5 leading the WebDev Arena at 1653 Elo rating.

Tools and frameworks

SpaceX’s $60 billion acquisition of Cursor fundamentally reset the coding-tools landscape, making cursor interaction data a strategic asset. Claude Code gained computer use capabilities and access to the field’s highest-rated models. The top ten agentic frameworks—LangGraph, CrewAI, Mastra, PydanticAI, and others—have stabilized enough that adoption now depends on team fit rather than release churn. Google’s Agent Development Kit has emerged as a serious contender for teams building stateful, multi-agent systems with integrated memory and evaluation loops.

Standards and open source

The MCP release candidate lands July 28, with three headline changes. First, stateless protocol architecture: servers that previously required sticky sessions and shared state can now run behind simple round-robin load balancers, with clients caching tool lists for the server’s declared TTL. Second, a formal feature lifecycle (Active, Deprecated, Removed) gives vendors and builders planning horizons. Third, an extensions framework lets new capabilities ship as opt-in before moving to the specification. Monthly SDK downloads sit at approximately 97 million, and MCP has been under the Agentic AI Foundation since its December 2025 donation.

Money and infrastructure

The five largest hyperscalers—Amazon, Microsoft, Alphabet, Meta, and Oracle—are collectively spending between $660 billion and $725 billion on AI infrastructure in 2026, nearly doubling 2025. The shift is concrete: individual AI data centers now require 1 to 5 gigawatts compared to 20 to 50 megawatts for traditional facilities. Over 400 data center projects are planned through 2027, totaling nearly 180 million square feet. But power has become the critical bottleneck, with U.S. data centers facing an 11 gigawatt capacity shortfall today and analysts projecting a cumulative gap exceeding 49 gigawatts by 2028. This means capital will continue flowing to hyperscalers with existing grid access, power purchase agreements, and land banks.

What I am watching

With MCP specification finalization one week away, I am watching whether the standards body governance model holds and whether production deployments gravitate toward the stateless architecture or remain on sticky-session deployments for another cycle. I am also watching power: once silicon is no longer scarce, the teams with reliable power supply and siting advantage win the next phase of scaling. The $700 billion spend is real, but it is constrained by physics, not money.