The meter is replacing the subscription

In Gemini, one hard prompt can now be worth more than one easy prompt. That is the small product decision that gives the game away.

WIRED reported that Google has moved Gemini away from simple request limits towards a usage system where more complex AI tasks can consume more of a user’s allowance. The consumer experience changes from “I have this many goes left” to something fuzzier: the same subscription can feel generous or stingy depending on what you ask the model to do.

The obvious read is that Google is tightening the screws. Maybe. But I think the bigger shift is that AI is becoming harder to disguise as normal software.

The subscription was a useful fiction

Software subscriptions trained users to expect abundance. Pay the monthly fee, click as much as you like, complain when a feature moves tiers. That model works when the marginal cost of an extra action is close to invisible.

AI breaks that illusion. A long reasoning task, a multimodal request, or a giant context window is not the same economic object as a quick autocomplete. The product may present them as variants of “ask the assistant”, but the infrastructure sees different workloads with different costs.

That is why metered intelligence is creeping in. Google’s Gemini quota change is the consumer-facing version. Enterprise teams have already been living with token counts, rate limits, model routing, latency budgets, and cloud bills that make “unlimited AI” sound like a sales deck phrase from a more innocent time.

The parallel is electricity. A flat fee feels elegant until one customer is running a bedside lamp and another is running industrial machinery. At some point, the supplier needs a meter because pretending all usage is equal becomes economically absurd.

AI is reaching that point in public.

Open weights do not repeal scarcity

This is also the more interesting way to read Moonshot AI’s Kimi K3. TechCrunch framed the model as an open-weight Chinese challenger close enough to leading proprietary systems to unsettle the industry. The loud version of that debate is geopolitical: China, open source, frontier labs, strategic threat.

Fine. But the product economics are sharper.

If an open model can get close enough to proprietary quality, then the rent shifts. Owning the model matters less than operating it well. Distribution, inference cost, memory, latency, uptime, safety layers, and user experience become the battleground. The magic file may be open; the bill for running it is not.

That is uncomfortable for both sides of the AI market. Proprietary labs cannot rely forever on model access as the whole moat. Open-model builders cannot pretend weights alone create cheap products. Somebody still has to pay for the compute, and the user will eventually feel that cost through limits, delays, degradation, ads, higher prices, or some new pricing syntax designed to look friendlier than a meter.

Then the story leaves the data centre. TechCrunch reported that AI demand for memory is putting pressure on supply and affecting India’s smartphone market. That matters because it collapses the tidy separation between “AI infrastructure” and “consumer hardware”. The same boom that powers assistants and agents can feed into the availability and pricing of everyday devices.

For builders, the lesson is blunt: stop designing AI products as if capability is the only constraint. The constraint is capability per unit cost, under volatile supply conditions, with users who hate unpredictability.

That changes product work. Usage dashboards become part of the core experience, not billing plumbing. Model choice becomes a product decision, not an engineering afterthought. Teams need graceful fallbacks, cheaper default paths, and honest explanations when a task burns more allowance than expected.

The companies that win will not be the ones promising infinite intelligence for one tidy monthly fee. They will be the ones that make scarce intelligence feel legible, fair, and worth paying for.

The subscription made AI feel like an app. The meter will remind us it is a machine.

---