The $5B admission: The model is no longer the product.
Yesterday, Microsoft announced Microsoft Frontier Company, a standalone business backed by $2.5B and 6,000 engineering and change-management experts.
Two days before that, AWS committed $1B to a new Forward Deployed Engineering (FDE) unit. In May, OpenAI and Anthropic made nearly identical moves.
In total, four tech giants just dropped ~$5B in eight weeks to scale a brutal, face-to-face enterprise integration layer.
Why? Because the “software-only” model for enterprise AI is choking.
The data tells the story: MIT’s Project NANDA recently revealed that a staggering 95% of custom enterprise AI pilots return zero measurable P&L impact. The breakdown isn’t happening because the models lack intelligence. It’s happening because they hit a wall when meeting legacy data, rigid workflows, and organizational inertia.
Charlie Munger always advised studying historical bottlenecks to understand where value naturally clusters. We are watching a well-worn playbook repeat itself:
The 19th-Century Railroads: The economic boom wasn’t captured by the locomotive manufacturers. It was captured by the logistics networks and layout engineers who figured out how to physically cut tracks through messy, existing geography and cities.
The 1970s Mainframes: IBM built the breakthrough hardware, but the enduring enterprise moats were captured by EDS and Accenture. They did the unglamorous, on-site heavy lifting of systems integration.
Value always compounds where the friction is highest.
For the last three years, the tech world treated AI as a pure-margin software game. This week’s capital shift proves that the frontier has officially moved from model capability to enterprise absorption.
The winners of the next decade won’t necessarily be the labs with the highest benchmark scores. They will be the integration layers that own the trust, the context, and the local workflow.
Please note: this article is a quick reaction post, and I am still developing the ideas around this.