Apple-Google AI Deal Brings Competition Policy Questions Into Sharper Focus

Apple just bet its AI future on a partnership with one of its biggest competitors—Google—when it agreed to have Gemini power Siri and other Apple Intelligence products.

Reports put the deal at approximately $1 billion annually. Compare that to the $20 billion Google pays Apple each year to be Safari’s default search engine. The asymmetry in magnitude and direction is instructive. Google effectively pays Apple to not compete on search. Whereas with AI, Apple is paying Google for an essential service. One is a non-compete payment. The other is a service fee. And at that price, possibly a signal that the model layer is commoditizing faster than the infrastructure buildout suggests. If foundational models were scarce and differentiated, Apple would pay more. Instead, Bloomberg reported that Google won the contract partly on price. When the most advanced AI models compete primarily on cost, the moat may not be in the model layer.

The deal raises a question regulators will need to answer: is the model layer commoditizing, or is Apple simply outmatched?

Two worlds

The question is whether Apple concluded that the model layer is a commodity that was not worth contesting, or whether Apple concluded it simply cannot compete at the model layer regardless. Apple has made strategic sourcing decisions before, whether it is Samsung screens or Intel chips, and knows how to distinguish differentiating technology from commodity infrastructure. But both possibilities have policy implications. And these aren’t mutually exclusive. Models might be commoditizing while remaining inaccessible to anyone outside the hyperscaler coalitions between large tech firms; a commodity input controlled by an oligopoly.

If Apple is right that models are commodified, then the hyperscalers’ infrastructure buildout looks even more puzzling. Hundreds of billions are flowing into improving models where the most sophisticated buyer pays $1 billion annually for access to these precious weights. Now, a lot of the money flowing into generative AI goes toward building inference capabilities and Apple is going to run the Gemini model on its own Private Cloud Compute servers. But still the suggestion is that the hyperscaler coalitions would be fighting over a commodity business while real value accrues elsewhere—at the interface, at distribution, or at data integration points.

If that’s right, the current data center buildout looks increasingly precarious. As I’ve argued elsewhere, the infrastructure has limited shelf life as AI chips rapidly become obsolete. If model provision becomes a commodity margin business, the hyperscalers need application-layer revenue to justify the capital expenditure. That’s a hard road. It means competing with their own customers, extracting rents from developers building on their platforms, or hoping enterprise demand materializes at prices that cover the investment. If they bet wrong, these are stranded assets with no salvage value.

If Apple is surrendering because it can’t keep up, then the market is consolidating around Google, Microsoft, and Amazon not because of inherent economics but because of path dependency and resource concentration. Apple’s exit narrows the field. That’s a different policy problem: not misallocation, but concentration.

Indicators to watch

How do we know which world we’re in?

Model differentiation. If frontier models converge in capability over the next year or two, that supports the commodity read—Apple is paying for something any major provider could supply. The enterprise market offers a useful signal: watch whether large customers can switch providers based on price or capability. Price-driven switching suggests commodity dynamics. Capability-driven loyalty suggests durable differentiation at the model layer.

Apple’s capital allocation. Apple says it’s still interested in building advanced models internally. If that effort accelerates, this deal was a bridge. If the team keeps shrinking and timelines slip, it was a white flag. Where Apple invests instead–on-device processing, privacy infrastructure, interface innovation—reveals where the company thinks differentiation actually lives.

Hyperscaler pricing and integration. If Google, Microsoft, and Amazon compete aggressively on API pricing, that confirms commodity dynamics at the model layer. And commodity dynamics create pressure to capture value elsewhere—which means building applications that compete directly with developers on their platforms. The model layer becomes a loss leader for application-layer extraction. The rental market for inference capacity already shows price compression. Watch whether that extends to API access, and whether the hyperscalers simultaneously move up the stack into applications their own customers are trying to build.

The search parallel

In my ProMarket piece last October, I argued that the same barriers preventing search competition would operate in generative AI: scale, distribution, defaults, data advantages, and content access. Apple’s Gemini deal confirms the pattern.

Apple’s choice came down to three hyperscaler-backed options for access to a model: Google, Microsoft-backed OpenAI, or Amazon-backed Anthropic. The market structure I described—a small group of coalitions controlling the market—is consolidating as predicted.

It’s worth pausing on the observation that the company best positioned to introduce AI competition, with billions of active devices, the most valuable customers in technology, chose to license from the incumbent rather than build a challenger. If Apple, with those resources and that distribution, cannot mount an independent AI effort, what chance do actual startups have?

Policy stakes

The deal validates hyperscaler dominance and highlights the stakes of Apple’s strategic choices. But the implications for policy run deeper. If models are commoditizing, competition policy should focus on the layers where value concentrates: infrastructure economics, distribution control, and data integration. If models are not commoditizing but Apple simply couldn’t compete, the field is narrowing to three or four coalitions with no realistic entry path for challengers. That’s a concentration problem requiring structural remedies like interoperability standards or mandatory data-sharing. Finally, if models are commodities but the same three coalitions control production, the risk is parallel pricing power over an undifferentiated input—requiring scrutiny of vertical integration and potentially structural separation between model provision and cloud infrastructure.

Regulators should be watching this space closely. Commodity dynamics call for attention to infrastructure subsidies and distribution bottlenecks. Monopolization calls for structural separation or mandatory access. Apple’s choice doesn’t yet answer which world we’re in, but it sharpens the question considerably.