Google I/O bets that cheaper is the future

Google’s developer conference on 6 May was shorter on spectacle than its predecessors and longer on arithmetic, which is usually a sign that a company is thinking about its actual business. The centrepiece was a new Gemini variant, lighter than the flagship, with a price tag that undercuts last year’s equivalent by more than seventy percent, plus a set of agent-building tools that work across providers rather than locking developers into Google’s own stack. The demos were competent and the tone was disciplined, and the message underneath was the one the entire industry has been circling: the AI race has moved from capability to cost.

It is worth being clear about what that shift means. Two years ago the frontier labs competed on the benchmark table, and the table still matters for the top of the market. But the commercial centre of gravity has moved to inference, to the cost of running a model millions of times a day, and that cost now dominates every product decision from search to customer service to code completion. Google, which runs its own infrastructure and designs its own chips, is structurally better placed for that fight than most of its rivals, and the conference was an exercise in pressing the advantage. Every price cut announced on stage is a margin decision that someone else in the industry has to match.

I found myself more interested in the agent tools than in the model. The announcement that developers can now compose agents that mix Google, OpenAI and open-source models in a single workflow is a small surrender and a large strategic move at the same time. It concedes that no single provider will own the intelligence layer, and it bets that the integration layer, the plumbing that connects models to tools and memory and permissions, is where the durable value will sit. That is the same bet the model companies are making about their own platforms, and the fact that Google is making it publicly suggests the industry has passed the point where exclusivity was worth defending.

The conference also left the more difficult questions unasked, and I want to ask them here. The price war is being paid for by cheap energy and aggressive infrastructure spending, and the environmental and financial bills will arrive in some future quarter. The agent tools make it easier to build systems that act on the user’s behalf, which means the responsibility for those actions is about to become a product category of its own, one that none of the demos addressed. And the biggest unspoken item is the tension between the open-source posturing and the reality that every one of these platforms is a toll road with a different paint job.

The conference was more honest than its predecessors, and honesty in this industry is a competitive weapon. When a company stops selling you the future and starts selling you the unit economics of the present, it is telling you something real about where the sector is going. The AI business is becoming the cloud business, with the same virtues and the same vices: scale as a moat, price as a religion, and the customer’s data as the real product. Google is further down that road than its rivals, and on 6 May it said so, in the language of price per token, which is the only language that survives contact with an annual budget.