When intelligence is free, what do you own?

Stop calling cheap inference a moat.

There's a tidy story in boardrooms right now: training is commoditizing, inference efficiency is the new battleground, and whoever drives cost-per-token lowest wins. It's comforting, because it's measurable. You can put cost-per-token on a slide and watch it fall.

It also gets the strategy backwards.

The cost of a unit of machine intelligence is collapsing faster than compute did in the PC era or bandwidth did in the dotcom boom: roughly 10x every year. Hitting GPT-3.5-level performance cost about $20 per million tokens in late 2022; 2 years later, it costs around 7 cents. Gartner expects trillion-parameter inference to cost over 90% less in 2030 than in 2025. The direction isn't in dispute.

If the cost curve is falling for everyone, on the same trajectory, using the same chips and the same tricks, how is being on it an advantage?

A moat is something competitors can't easily copy. The falling cost of inference is the single most copyable thing in the entire AI stack. Treating it as a moat is like a manufacturer believing cheap steel was its edge while every rival bought from the same mills at the same falling prices. Cheap inference is table stakes. It’s the price of playing, not the prize for winning.

The second trap is the word "commoditizing." Yes, a frontier model's capability now diffuses to open-weight alternatives in 3 to 4 months. Thus, you can rent near-frontier intelligence cheaply and almost immediately. But producing the frontier still costs hundreds of millions of dollars and gigawatts of power, resources that perhaps a dozen organizations can marshal. Even if you can rent the capability, you cannot cheaply become the lab. Betting your strategy on a model advantage that was never yours to hold is how companies get caught flat footed.

If efficiency is table stakes and the model is rentable, where does durable advantage live? It lives in the four things that don't deflate when the token does:

• Custom silicon at gigawatt scale: when tokens cost nearly nothing, margin survives only for those who own the means of production.
• The serving and orchestration stack: using smaller models for routine tasks and frontier models for complex work can cut costs. Over time, this routing system becomes deeply embedded and costly to replace.
• The workflow surface and proprietary data: software embedded in a system of record keeps retention near 99%. A model can copy a feature in thirty seconds. However, it cannot copy years of customer data that a company is authorized to use.
• Compliance, trust, and the edge: in a regulated industry, the cheapest token is worthless if the deployment fails legal review.

The common thread is that the cost of a token is collapsing toward zero, but the cost of completing a task, including retrieval, orchestration, compliance, and review, is not. As token costs shrink, these other elements account for a larger share of the total cost.

The winners of the next phase will not be those with the lowest cost per token. As inference becomes cheaper across the market, the real advantage will come from delivering a reliable, completed task at the lowest cost, within the required latency and compliance.

Here is the question for your next strategy review: when intelligence is free, what do we own that stays scarce? If the honest answer is “nothing,” it may be time to rethink your AI strategy.