Deloitte's AI infrastructure survey turns the enterprise AI factory into a budget meeting, framing adoption around model choices, hosting, budgets, and skills while saying more than 70 percent of surveyed respondents expect to operate AI factories at scale by 2028. [1]

That is a different story from the demo reel because a company does not become an AI factory when a model completes a task in a video; it becomes one when compute, data handling, hosting location, security, talent, and cost controls fit into an operating plan. [1]

OpenRouter's Series B announcement supplies one side of the buyer problem, saying enterprises are moving from single-model pilots to multi-model production systems with workspaces, spend management, guardrails, zero-data-retention policies, and provider failover. [2]

Anthropic supplies another side, with its Claude Opus 4.8 release describing effort controls and dynamic workflows for enterprise and team plans that turn agent behavior into something managers can dial up, dial down, and charge to a budget. [3]

The public argument will keep asking whether agents are magical or useless, but enterprises ask what the system costs when it runs every day, which data it may touch, who can approve a model, and where the workload lives, which is why Deloitte's frame matters: it strips the factory metaphor of romance and returns AI buyers to capital, labor, process, compliance, power, and throughput. [1] [2] [3]

-- THEO KAPLAN, San Francisco