The most valuable position in AI is not the model. It is the trusted action surface, the place where an AI agent sees context, holds permission, and is allowed to act. On June 8, the largest company in the world reorganized its entire stack to keep that position.

At WWDC, Apple rebuilt Siri on a custom Google model and routed its heaviest reasoning to Google Cloud on Nvidia hardware. It conceded the model layer to one rival and the compute layer to another, and kept the layer it believes holds durable value: the surface where intelligence meets the user (Exhibit 1).

Apple conceded the model and the compute, and kept the surface

Capability is commoditizing, and Apple stopped paying to own a commodity. Requests now route through three tiers: simple tasks on the device, moderate ones in Apple's private cloud, and the hardest reasoning out to a 1.2-trillion-parameter Google model in the cloud (Exhibit 1). What stays constant across all three is the front door. The device decides what runs where, the operating system decides what context the agent sees, and the permission prompts are Apple's. If the surface owner can swap the model underneath without the user noticing, the model is a supplier, and suppliers face margin pressure.

EXHIBIT 1VFINTEL

Apple AI request routing, by tier, 2026

The model can be swapped; the surface stays Apple's

The trusted action surface
context · permission · the right to act, owned by Apple
On-device
simple tasks
Private cloud
moderate tasks
External cloud
heaviest reasoning, 1.2T model

Source: Apple, WWDC 2026; VFIntel analysis

Two positions in AI are genuinely scarce

AI has two scarce positions, and they are priced very differently. The first is raw compute, and the market has already given it an owner and a price. The second is the trusted action surface, and the contest for it is only beginning (Exhibit 2). Apple's roughly 2.35 billion active devices are the asset no supplier can provide, and they are what make its surface defensible.

EXHIBIT 2VFINTEL

The two genuinely scarce positions in AI

Compute is priced and owned; the action surface is contested

Raw compute
priced and owned
Trusted action surface
contested and open

Source: VFIntel analysis

Apps that are not legible to the agent layer get routed around

The least flashy announcement at WWDC may matter most. Apple expanded App Intents, the framework through which an app tells the operating system what data it holds and what actions it permits. For fifteen years the app was the unit of software; in an agent-driven environment the interface stops being the product. The apps that survive are the ones whose data and actions are structured, permissioned, and legible enough for the system to operate them on the user's behalf. Being legible to the agent layer becomes the requirement for staying in the transaction at all.

The rental economy's action surface is unclaimed

The lease event has no equivalent owner. About a third of American households carry the largest recurring obligation in their lives through it, yet there is no surface where an agent can see verified context about an applicant, hold permission from both sides, and act with legal standing. The reason it is unclaimed is that this surface cannot be assembled from software (Exhibit 3). Acting at the lease requires verifying identity to a housing standard, placing insurance inside the transaction, moving funds on regulated rails, reporting payment under federal credit rules, and preserving evidence both parties can rely on. Each is a regulatory relationship that takes years to establish, which is why neither an AI model nor a horizontal competitor can stand in for it.

EXHIBIT 3VFINTEL

What an agent needs to act at the lease, and the relationship behind it

The action surface cannot be assembled from software

REQUIREMENT TO ACT
REGULATED RELATIONSHIP
Verify identity to a housing standard
Bank sponsor and screening
Place insurance in the transaction
Carrier agreements
Move the money
Regulated payment rails
Report payment history
Bureau furnishment framework
Preserve evidence
Compliance architecture

Each is a relationship that takes years, not a feature.

Source: VFIntel analysis

What Apple proved

Winning the model war is not required to hold the most defensible position in AI. Apple conceded capability to the labs and compute to the chipmakers and still kept the layer where intelligence becomes action, because that layer is built from distribution, permission, and trust rather than parameters. The rental economy will source its models from the same labs as everyone else. What it does not yet have is its trusted action surface, and the lesson of June 8 is that positions like it do not stay open for long.

Robert Elensky

Founder & CEO, VFIntel

Robert built VFIntel on the premise that the rental economy's financial coordination failure is an infrastructure problem, not a product problem. He writes on regulated fintech, embedded insurance, and the structural risks accumulating across the enterprise software stack as AI agents become the primary actors operating within it.