Apple Just Paid Google for Frontier AI Instead of Building It. That's Your Market Signal.
On June 22, during WWDC, Apple announced that its new Siri and "Apple Intelligence" features would run on Google's Gemini models via a paid partnership. Not Apple's own foundation models. Google's, rented over an API.
Two days later, Apple's stock hit record territory. The market liked the choice to rent.
That's the signal. The company that owns the device, the OS, the silicon, and the distribution (the most vertically integrated platform on Earth) looked at frontier AI and decided: we're not building this in-house, and we're not waiting for our own models to catch up. We're paying Google.
Why Apple outsourced frontier AI
The math is brutal, and it looks like this:
In 2024–2025, the frontier AI labs (Anthropic, OpenAI, Google, Meta) committed between $5B and $50B each to training infrastructure, compute procurement, and research. Anthropic raised $5B at a $15B valuation. OpenAI is rumored to be at $110B. Google is spending in the tens of billions on Gemini and Grok-adjacent work.
Apple, with a $2.8T market cap and $115B in annual revenue, could absolutely afford the capex. But there's a second equation: moat erosion.
If Apple trains a frontier model, it competes with the same players spending more on compute. If Apple's model is 2% better than Gemini or Claude, it wins in a few verticals. If it's worse, it's a liability. Either way, Apple is now dependent on its own model quality for a feature that touches hundreds of millions of devices. The moat isn't the model anymore: everyone can license one. The moat is what you do with it.
So Apple paid Google.
This moves the competitive surface from "whose model is best" to "whose integration is best." Apple has 2 billion active devices. Google has a model that runs Gemini Advance. The deal is: Gemini runs on-device where it can (privacy, latency), and routes to Google's servers when you need the frontier stuff. Apple makes sure the experience is frictionless and branded as "Apple Intelligence." Google gets distribution and validation that their Gemini is good enough to power one of the most scrutinized product launches of the year.
For Apple, this is the same calculation they made about maps, mobile browsers, and cellular modems: buy commodity capability if the capex-to-ROI ratio on a custom solution is untenable.
What this means for your consulting practice
If Apple chooses to rent, the cascade effect is immediate.
A year ago, the pitch in a lot of enterprise AI shops was: "We'll build you a custom model fine-tuned on your proprietary data, and you own the moat." That's a $5M–$50M project. It requires hiring PhDs, renting GPUs for months, and managing model quality.
With Apple signaling that frontier AI is a commodity to rent, the defense for a custom model weakens. Your customer's first question is now: "Why wouldn't we just use Claude with fine-tuning, or Gemini with RAG, or Mistral with LoRA?" The answer has to be very specific. "Because your data is so proprietary and your accuracy requirements are so extreme that a rented model won't work" is a real answer. "Because it's cheaper to own it" is not anymore.
The new shape of enterprise AI consulting is narrower and deeper:
Domain-specific, non-model work. How do you integrate frontier AI into your hiring process (compliance), your customer support (accuracy), your sales forecasting (reliability)? The model is the commodity. The integration layer (the domain knowledge, the reliability guarantees, the audit trail) is the product.
The reliability layer. Rented models fail sometimes. They hallucinate. They cost more than you budgeted. The business that survives is the one that owns the monitoring, retraining, and fallback when the rented model doesn't perform. You can sell that as a consulting service: "We'll set up your alert system, your eval harness, and your escalation workflow so a model failure doesn't become a business failure."
Vertical depth, not model ownership. Build deep knowledge in one vertical: healthcare, legal, accounting, insurance. Rent the frontier model from OpenAI or Anthropic or Mistral. Combine the model with your domain knowledge, and charge based on the outcome, not the capability. "We'll reduce your medical coding error rate from 3% to 1%" is a defensible charge. "We'll give you access to Claude" is not.
The honest counter-take
Apple still invests billions in on-device machine learning and privacy-preserving inference. For use cases where you can't send data to a cloud API (medical imaging, call recording, sensitive financial data), Apple's on-device capability is a real differentiation.
And there are still verticals where a custom model makes economic sense: a financial services firm with billions in proprietary trading data, a pharmaceutical company with decades of experimental results, a defense contractor with classified information. The data is so sensitive and valuable that the ROI on owning the model (amortized over years, not quarters) is positive.
But the default, for most enterprises, is now: rent the frontier model, own the integration. Apple just told everyone that.
What I'd actually do
If you're currently selling "we'll build you a custom model," revisit your pipeline.
For the deals where the customer's value proposition is the model itself (e.g., "our private equity firm wants a model trained on our deal flow to predict returns"), keep those. They're still defensible, and they're still expensive.
For the deals where the customer wants "better AI for our process" (better customer support, better hiring, better forecasting), run a conversion conversation. Show them the math: Claude or Gemini off-the-shelf + your domain consulting + your reliability layer = same or better outcome for 1/10th the capex. If they want to own the model anyway, that's fine. But make sure they're choosing it for the right reason, not because it sounds more innovative.
And if you're building a tool or platform that wraps AI, watch the next 12 months closely. Rented models are going to get faster, cheaper, and better at domain adaptation (fine-tuning APIs, context windows, model distillation). Your tool's value is going to move further away from "we have access to the good model" and closer to "we've built a workflow that ensures the rented model works in your use case." That's sustainable. The access advantage isn't.
Author
Lukas
@lukcombinator