AI Giants Lose Money on Subscriptions: This Isn't a Price War, It's a War for Users
Deep thoughts on AI and aspirations —— ByteThink Circle
Research firm SemiAnalysis recently did something simple but effective: actually worked under OpenAI and Anthropic’s various subscription tiers, then reverse-calculated the token value of these tasks using public API pricing.
The conclusion is stark: the more expensive the plan, the higher the subsidy multiple. For a two-hundred-dollar plan, the actual compute value consumed by users could be several times the subscription fee. In other words, these premium tiers were never designed to make money—they’re deliberately sold at a loss.
This violates pricing common sense, but aligns with battlefield logic. What’s truly being fought over was never the subscription fee itself.
Selling at a Loss to Whom
To understand this subsidy war, first see clearly who’s being subsidized.
Heavy users aren’t ordinary consumers. Those who can push their plan’s token allocation into subsidy territory are basically developers, enterprise decision-makers, and team leads. These people share one trait: their choices get amplified. When a technical lead becomes comfortable with workflows on a particular platform, their team, their product line, and their future employers will likely follow that choice.
So the real target of the subsidy is organization-level switching costs. Individual users can switch AI tools in five minutes; a team migrating their development stack takes months. Giants lose money retaining heavy users not for subscription renewals, but to become the default option for entire organizations. This math works within the CAC (customer acquisition cost) framework: if a decision-maker represents a fifty-seat enterprise contract, the subsidy lost on them is just a fraction of a large customer’s acquisition cost.
Three Types of Ammunition, Three Types of Strategy
In this battle, the structural differences in funding sources among participants matter more than subsidy intensity itself.
Google’s ammunition is advertising revenue—over three hundred billion dollars annually in cash flow from a mature business running for over twenty years. Using ad profits to subsidize tokens is like someone with an oil field entering a gas station price war; the cost comes from internal circulation, and they can sustain it indefinitely.
OpenAI and Anthropic’s ammunition is funding, cumulatively exceeding hundreds of billions of dollars. The difference between funding and profit lies in return expectations: investors want exits, and after exits, companies must face public markets’ ongoing scrutiny of revenue, profit margins, and cost structures. When Wall Street discovers a company loses several dollars for every dollar of subscription revenue, whether the growth narrative can sustain valuation becomes another question.
There’s another class of player rarely mentioned: cloud providers. Compute is their existing asset; for them, token subsidies are repricing of idle capacity, with marginal costs approaching zero.
Ammunition structure determines strategy: Google can fight a war of attrition, funding-driven companies must achieve positive cycles before running out of ammunition, and cloud providers are waiting to tax compute after everyone else burns out. The endgame of the subsidy war is largely determined not by who subsidizes more aggressively, but by who has lower capital costs.
Internet Playbook or Utility Playbook
Price war endgames typically reference two playbooks.
One is the internet playbook: subsidize for scale, monopolize, then raise prices to recover. Didi, Meituan, and Amazon all followed this path. This playbook works when lock-in effects exist—driver networks, merchant ecosystems, logistics infrastructure—where both users and supply sides can’t leave.
The other is the utility playbook: tokens become standardized commodities, like electricity and bandwidth, with undifferentiated products and zero switching costs. Competition crushes profit margins to zero, leaving a few infrastructure operators.
The trouble with the AI subscription market is that it’s stuck between these two playbooks. At the API level, it’s highly standardized. Multi-model compatible clients are proliferating, users can switch to cheaper models anytime, and lock-in effects are weak—this is the utility side. But at the workflow and habit level, stickiness is forming: prompt libraries, project context, team collaboration processes—these things accumulate on platforms, and migration involves real friction—this is the internet side.
My judgment is that the two layers will bifurcate. Base model capabilities will become utility-like, with price wars eliminating anyone’s ability to earn excess profits from models themselves; while the layer hosting workflows and context will become the new lock-in point, and where future profits truly reside. This explains why everyone is racing to build clients, memory, and ecosystems, not just APIs.
What This Means for Users
For individuals and teams, the strategy during this phase can be straightforward.
This is a buyer’s market for compute, and the window length is uncertain. Heavy users choosing plans can calculate directly based on “how many tokens will I consume”—during the subsidy period, real costs are far below list prices, so use full allocations when appropriate. Enterprises choosing platforms should recognize that being fought over is itself leverage; procurement negotiation space is much larger than list prices suggest.
At the same time, don’t budget assuming subsidies are permanent. All subsidy war histories point to the same outcome: players with the hardest ammunition structures remain, and pricing returns to rationality. Choices made today due to cheapness that involve deep lock-in warrant advance consideration of switching costs—data export, prompt migration, workflow portability—things that seem worthless today will suddenly become very valuable when prices rise. A crude but practical test: if your core assets exist only in a platform’s proprietary format, you’re already the locked-in party with zero bargaining power.
As for the giants, the question inverts: how to convert heavy users bought at a loss into profitable organization-level contracts before user habits solidify and funding runs out. The decisive factor in this battle isn’t subsidy intensity—it’s conversion speed.