AI's bottleneck has left the model.

Four unrelated stories broke this week: a vendor spending spree, a pricing war, a new standards coalition, and an enterprise adoption survey. Together, they say the constraint on enterprise AI has moved off the model and onto everything that has to happen after it is chosen. That matters for how a technology leader spends the next quarter, because most budgets and most attention are still pointed at the wrong layer.

The money already moved

Microsoft committed 2.5 billion dollars and 6,000 forward-deployed engineers to Frontier, a new subsidiary built only to embed AI inside client organisations, announced July 2. Amazon backed a similar effort with 1 billion dollars two days earlier, and OpenAI and Anthropic launched comparable deployment ventures in May. Vendors do not commit capital at that scale to a problem they consider solved. MIT's Project NANDA found 95 percent of enterprise generative AI pilots deliver no measurable profit impact. That is the gap this money is chasing, and it is a deployment gap, not a model gap. If your own AI budget still reads mostly as licensing and API spend, that is worth a second look against where the vendors themselves are now placing their bets.

The model itself is commoditising

Chinese-origin models now route more enterprise tokens on OpenRouter than US models, roughly 46 percent against 36 percent, and DeepSeek alone is the platform's largest single provider, according to a CNBC investigation published July 7. The gap is price, not capability. Open-weight Chinese models run 60 to 90 percent cheaper than the flagship US offerings. When cost decides model choice more than capability does, model choice has stopped being the strategic decision. The one that replaced it is data residency, export exposure, and whether the choice is defensible later to a client or a regulator, and that decision looks very different depending on which jurisdiction you operate from.

The lock-in risk moved up a layer

On June 17, Microsoft published the Agentic Resource Discovery specification, developed with Cisco, Databricks, GitHub, GoDaddy, Google, Hugging Face, Nvidia, Salesforce, ServiceNow, and Snowflake. It does not replace Anthropic's Model Context Protocol. MCP lets an agent use a tool once it has found one, ARD is the layer that helps it find and verify that tool in the first place. Two names are conspicuously absent from that partner list, OpenAI and Anthropic. The lesson for anyone running production systems is to treat discovery and execution as separate layers, and not build an entire agent stack on a single vendor's protocol, no matter how dominant that vendor looks this year.

Deployment still is not absorption

None of the above matters if the organisation cannot absorb what it bought. Publicis Sapient's 2026 Global Enterprise AI Report, based on 1,550 AI decision-makers surveyed across six markets and released June 17, found 73 percent of enterprises now use AI regularly or across most business processes, yet only 10 percent say AI is core to how their business actually operates. Only 38 percent say AI is fundamentally changing how their business runs, and 22 percent name their own organisational design, not the technology, as the primary barrier to success. A model installs in weeks. Trust, new habits, and ownership of what the model does take quarters, and that work sits with the buyer, not the vendor. No deployment number in a board deck captures that gap.

What this actually changes

On their own, these are four separate news items: a spending number, a pricing story, a standards announcement, a survey result. Together they describe one shift. Model selection is no longer where a technology leader's real work is. It moved to operating what was bought, protecting against protocol lock-in, defending the vendor choice to whoever asks later, and building a team that can own what the model does when it is wrong. Three things worth doing this quarter follow directly from that: check whether your last model-selection debate actually changed an outcome, map which parts of your stack depend on one vendor's protocol with no fallback, and name who owns absorption, not just deployment, for the last AI rollout your team shipped. Budget and attention should follow the constraint, not the model.