The model war is over — the device war just started

OpenAI's answer to a 90% price cut wasn't a price cut. It was a phone.

Bloomberg reported that DeepSeek slashed V4 API prices by up to 90% last week, dropping V4-Pro input tokens to roughly $0.036 per million after promotional discounts. For context, comparable tokens from Anthropic or OpenAI run $12 to $25. Cache-hit costs fell to one-tenth across the entire API. The promotional pricing expires 5 May, but the signal is permanent: intelligence-as-a-service is heading where electricity went a century ago, towards a metered commodity priced at cost-plus-slim.

OpenAI didn't respond by matching the price. TechCrunch reported the company is developing a smartphone with MediaTek and Qualcomm where the homescreen replaces the app grid with an agent panel that handles everything from booking flights to compiling research. The target: 300–400 million annual units by 2028, with specs finalised by late 2026. When your core product is being commoditised, you don't fight on price. You move to a layer that isn't.

And Google didn't build a phone or cut a price. CNBC reported it's investing up to $40 billion in Anthropic, $10 billion now at a $350 billion valuation, with $30 billion more contingent on performance milestones. The deal secures Anthropic's expansion on Google Cloud TPUs. Google isn't buying a model. It's buying guaranteed demand for its compute infrastructure.

Where the profit migrates

Clayton Christensen had a name for this: the law of conservation of attractive profits. When one layer of a technology stack commoditises, margin doesn't vanish. It migrates to adjacent layers that remain scarce. The PC commoditised hardware; profit moved to the operating system. Cloud computing commoditised the operating system; profit moved to the application.

AI models are commoditising in plain sight. So the profit is migrating. Upward, to the device and interface layer OpenAI wants to own. Downward, to the compute infrastructure Google is locking in. The model itself becomes the abundant middle, cheap to run and interchangeable enough that DeepSeek can undercut rivals by 300x and call it a promotion.

The pattern explains why these three stories, which look like unrelated corporate manoeuvres, are actually the same strategic calculation made three different ways. DeepSeek is accelerating the commodity phase because that's where its cost advantage lies. OpenAI is escaping it by moving up the stack. Google is profiting from it by selling the picks and shovels. Each company has decided the model war is over and placed its next bet accordingly.

For anyone building products on top of these APIs, the practical implication is specific. Your inference costs will keep falling, possibly to near-zero within eighteen months if DeepSeek's trajectory holds. But the surfaces through which users reach your product (devices, operating systems, agent platforms) and the infrastructure that powers it are consolidating fast. The raw material gets cheaper. Access to distribution and compute gets more expensive. The economics are starting to resemble manufacturing, not software.

The real question this week isn't which model scores highest on a benchmark. It's whether you'll need OpenAI's permission to reach users on their phone, or Google's infrastructure to run inference at all. Those decisions are being made now, while the rest of the industry is still arguing about token prices.

---