A new product category is forming in enterprise AI infrastructure, but the evidence available does not support a specific dollar threshold for when self-hosting becomes worthwhile, and no deal data exists yet to show acquirers moving on it. What is documented is narrower and still notable: enterprises are self-hosting large language models for compliance reasons, open-source models have closed enough of the performance gap to make that practical, and a new middleware layer called the "agent gateway" is being built by at least three vendors to govern traffic between AI agents and the systems they touch.
The clearest evidence of vendor investment is Nutanix, which shipped its Agent Gateway as a generally available part of Nutanix Enterprise AI 2.7 in late May, built to route agent traffic across OpenAI's GPT on Azure, Anthropic's Claude, and self-hosted Llama models through one authenticated endpoint Agent Gateways Are Becoming The Control Plane For Enterprise AI (go.theregister.com).
Why self-hosting stopped being a hobbyist choice
The rationale enterprises give for running models on their own infrastructure has not changed in years — what has changed is what they can run. NAGA, an AI development and consulting firm, lists data privacy in regulated sectors, cost control at high usage volumes, fine-tuning control, and avoidance of API rate limits as the standing reasons to self-host Self-Hosting Large Language Models (LLMs) - NAGA (nagainfo.com).
What has moved is the model side.
What the gateway actually governs, and what it does not yet cover
The mechanism Nutanix describes is a proxy layer: an agent calls one endpoint, the gateway routes to whichever backend model is configured, applies the same authentication and rate limiting regardless of vendor, and fails over to a backup if the primary model is unavailable Agent Gateways Are Becoming The Control Plane For Enterprise AI (go.theregister.com). It also sits in front of MCP servers and applies tool-level permissions — a customer-service agent gets read-only database access, a DevOps agent gets GitHub write access — and logs every request with token usage attributed per team, which is what lets finance see the bill Agent Gateways Are Becoming The Control Plane For Enterprise AI (go.theregister.com). Nutanix's own disclosure is that the MCP governance and bundled test agent ship only as a technical preview; the parts that are generally available are routing, observability, and rate limiting Agent Gateways Are Becoming The Control Plane For Enterprise AI (go.theregister.com). A reader evaluating this category should treat the governance claims as ahead of the shipped product by at least one release cycle.
The spending this sits inside, and the skepticism attached to it
None of this happens in isolation from the capital being poured into AI broadly. Global datacenter capital expenditure is estimated at $657 billion for 2025, nearly double the figure from two years earlier, according to analyst firm Omdia as cited by The Register AI investment the only thing keeping the US out of recession (theregister.com). Apollo Global Management's chief economist, Torsten Sløk, is quoted saying there is "basically no growth in corporate capex outside of AI at the moment," and attributes the resilience of AI spending to rising equity valuations at Microsoft, Amazon, Alphabet and Nvidia rather than normal capex sensitivity to interest rates AI investment the only thing keeping the US out of recession (theregister.com). Forum commentary underneath a Register report on Microsoft's and Alphabet's latest capex commitments called the spending a bubble outright, with one commenter arguing "nobody is going to pay that much to use this" Microsoft, Alphabet throw more cash on the AI bonfire • The Register Forums — a judgment from an anonymous forum poster, not an analyst, but indicative of the skepticism sitting alongside the capex numbers. Whether the agent-gateway category attracts acquirers, and at what valuation, is not yet visible in any public deal record.