The cloud giants bet the farm on AI infrastructure — now the crops are either coming in or they're not.

The Summary

The Signal

Microsoft, Google, and Amazon have been spending like they're building the Panama Canal. Data centers, AI chips, energy contracts — the scale of infrastructure investment over the past 18 months has been staggering. Now someone's actually asking: where's the revenue?

Gulati's framing is sharp. Azure isn't trying to be the model provider anymore. It's becoming the model router. Enterprises don't want to marry one foundation model. They want the flexibility to use GPT-4 for customer service, Claude for research synthesis, a small model for data classification, and their own fine-tuned model for proprietary workflows. Azure's bet is on becoming the platform that orchestrates all of that — and charges for the compute, the switching fabric, and the management layer.

"The central model orchestrator for enterprises — enabling tailored choices among foundation, small, and internal models."

This is the real architecture of Web4 emerging. Not one AI to rule them all. A mesh of specialized agents, each optimized for cost and performance, coordinated by infrastructure that makes the complexity invisible. The money isn't in the models anymore. It's in the orchestration.

But orchestration only works if customers see ROI. Gulati emphasized productivity gains in Office as the bellwether. That's the tell. If Copilot is actually making people faster at their jobs — measurably, not anecdotally — then enterprises will keep paying. If it's just a fancy autocomplete that people turn off after the free trial, the whole AI revenue story collapses.

The other companies are facing the same reckoning:

  • Google's Workspace AI features need to justify Gemini infrastructure costs
  • Amazon's Bedrock has to show enterprises it's worth switching from their current workflows
  • All three need to prove data center capex translates to customer product value, not just model benchmarks

The Implication

Watch the earnings calls this week for two numbers: AI-attributed revenue and customer retention rates for AI products. If Microsoft can show Office Copilot has high attachment rates and low churn, that's the signal that AI orchestration is real business, not just infrastructure gambling. If they dodge those questions with "strategic investment" language, the market will punish them.

For everyone building in this space, Gulati's framing is the future. Single-model companies are already obsolete. The winners will be platforms that let you compose models like Lego blocks — and charge you for the convenience of not having to be an ML engineer to do it.

Sources

Bloomberg Tech