Enterprise AI grows up: fewer bets, harder numbers, power contracts

Fewer bets

JPMorgan Chase is shifting from tracking 1,000 use cases to 50 strategic AI initiatives, according to Mayfield's Enterprise AI Playbook. The point is not that the bank is doing less AI. It is that counting use cases turned out to be a way of avoiding decisions, and 50 named initiatives with production timelines is a way of making them. Any large enterprise still reporting its AI progress as a count of pilots should read that as a warning. m:…8387

The rest of the enterprise picture points the same way. Nearly 6 in 10 enterprise executives plan to increase their use of open-weight AI models to manage costs, and New York requires major frontier AI developers to register by January 1, 2027. Cost control and compliance are now part of the buying decision, not an afterthought. The enterprises that survive the consolidation will be the ones that can say what each initiative costs, what it returns, and who owns it. m:…8387

Security is part of the same consolidation. United Airlines moved to block unauthorised third-party AI agents from booking flights, and Microsoft researchers demonstrated that editing fewer than 10 neurons in an AI model can create hidden backdoors. An enterprise that cannot say which agents are allowed to act on its behalf, and what its models will do under tampering, is not ready to run those initiatives in production. It is ready for an incident. m:…8387

The money moves to power and prepayments

CoreWeave expects approximately $9 billion in customer prepayments and deposits over the next three months. Prepayments are the strongest signal in this market: customers are paying upfront for capacity they have not used yet. That is demand you can count, and it is the number that matters more than any forecast about AI spending in general. m:…8387

The same week brought both ends of the infrastructure bet. OpenAI reportedly raised $15 billion at a $1.25 trillion valuation, while SoftBank cancelled 13 gigawatts of turbine reservations for AI data centres and took a $3.7 billion earnings hit. One side of the market is still raising enormous sums, and the other is cancelling power reservations it no longer believes in. The gap between those two moves is the real story: capital is getting pickier about which build-out it funds. m:…8387

Anthropic agreed to buy about $3.4 billion of Caterpillar power-generation equipment, securing its own electricity supply. Frontier developers are no longer just buying compute. They are buying the power to run it, years ahead. The enterprises watching this should notice what it implies: the constraint on AI is moving from chips to energy, and the companies that locked in power early will set the pace. m:…8387

What a return looks like

SurveyMonkey is the clearest example of a return measured properly. The company improved AI accuracy from 14% to 98% over nine months, by adding a semantic layer, business logic and clear definitions, in the account given by Larridin and SurveyMonkey. The starting point is the honest part: an agent with no understanding of the business is barely better than guessing, and the work that fixed it was unglamorous context engineering, not a bigger model. m:…3499

Onfleet reports a similar shape of return in delivery operations, with customers reporting up to 45% fuel savings and up to 55% more delivery capacity after moving to AI-powered routing. The routing engine is trained on more than 400 million deliveries, and the gains come from evaluating thousands of route combinations against real constraints, not from a single clever trick. These are the vendor's own customer numbers, so they should be read as the top of the range, not the average. Even discounted, they show what a measured return looks like: a percentage, a cost line, and a capacity line, not a count of use cases. m:…2094

Mayfield's Matt Huang identifies a new managerial class for this world, the swarm rider, a manager capable of orchestrating large agent swarms for productive ends. Whether that job title sticks or not, the underlying point is right: the scarce skill in enterprise AI is no longer building a demo. It is running a fleet of agents, measuring what each one returns, and cutting the ones that do not. That is what the JPMorgan consolidation, the CoreWeave prepayments and the SurveyMonkey accuracy curve all describe, from different angles. m:…8387