The Big Three's Tuesday: 5 Gigawatts, MCP-Native Agents, and Thinking Images

Anthropic's Amazon compute deal, Google's MCP-native research agents, and OpenAI's thinking image model show the platform war has moved past benchmarks — it's about capacity and connected agents now.

Today's vendor moves aren't about model benchmarks — they're about the infrastructure and interfaces that determine who can actually ship production AI. Anthropic locked in a 5-gigawatt compute deal with Amazon, Google DeepMind shipped autonomous research agents with native MCP support, and OpenAI's new image model gained thinking capabilities. The pattern: the platform war is about agents with access, not parameters alone.

Anthropic announced an expanded Amazon collaboration on April 20, 2026, securing up to 5 gigawatts of compute for training and deploying Claude, with nearly 1 gigawatt expected online by the end of 2026. Amazon is investing an additional $5 billion now, with up to $20 billion more committed in the future, per Anthropic's announcement. For operational context on why the capacity matters: Simon Willison flagged this week that Claude Opus 4.7's adaptive thinking via the API no longer exposes a way to force reasoning — no budget_tokens, no effort: high parameter from prior models. The model decides for itself when to think, and that behavior only scales if the underlying compute can absorb the variable cost.

Google DeepMind's Deep Research and Deep Research Max, both powered by Gemini 3.1 Pro, launched on April 21, 2026, with arbitrary MCP support baked in — meaning the agents can securely connect to your own or third-party data through the same protocol Anthropic open-sourced last year. It is also the first Google research agent to natively generate presentation-ready visuals, with Deep Research Max tuned for exhaustive background tasks and Deep Research tuned for interactive latency. Separately, OpenAI shipped ChatGPT Images 2.0 the same day with a December 2025 knowledge cutoff and, notably, the company's first image model with thinking capabilities built in. It can search the web, generate multiple distinct images from one prompt, and double-check its own outputs before returning them.

For builders, MCP just became a competitive necessity. Google adopting it for Deep Research means developers targeting Gemini, Claude, and (via community clients) ChatGPT can write against one protocol for data-connected agents. For buyers, the Anthropic-Amazon deal signals compute availability is no longer the hard ceiling it was six months ago — which should eventually translate into higher rate limits and more aggressive enterprise pricing. For creators, ChatGPT Images 2.0's flexible aspect ratios (3:1 to 1:3) and improved text rendering make it production-viable, not just demo-viable.

The bottom line: MCP is table stakes for agentic products now — if you are building on a single vendor's proprietary tool-use protocol, you are already locked in and already behind.

Sources