The numbers came out last week and they broke the model.

Claude Max 20x costs $200 per month. Users who hit the limit consume roughly $8,000 in compute. ChatGPT Pro also costs $200 per month. At full utilization, the usage is worth approximately $14,000.

Do the margin math. At full utilization, Claude Max runs at negative 900 percent margin. ChatGPT Pro runs at negative 1,650 percent. Every heavy user costs the provider thousands of dollars more than they pay. This is a structural subsidy masquerading as a consumer subscription model.

The question nobody on $200/month wants to hear: what happens when the subsidies stop?

## The Subsidy Machine

AI subscription pricing is not a pricing decision. It is a land grab funded by venture capital. OpenAI and Anthropic set flat-rate plans at $20 and $200 per month knowing full well that heavy users would cost ten to seventy times that amount in compute. They raised the money to absorb the difference because capturing market share mattered more than unit economics.

The strategy is not new. Uber subsidized rides. WeWork subsidized office space. DoorDash subsidized delivery. The playbook is consistent: burn VC money to acquire users, build habit and dependency, then raise prices once the market is locked in.

The difference is the scale. Uber’s subsidy per ride was a few dollars. WeWork’s subsidy per desk was a few hundred. AI’s subsidy per heavy user is thousands of dollars per month. Claude Max subscribers consuming $8,000 in compute while paying $200 are receiving a $7,800 monthly gift from investors.

The total annualised subsidy across the top AI labs is estimated in the tens of billions. SemiAnalysis recently calculated that OpenAI alone may be subsidising over $12 billion in annual usage that users are not paying for. Anthropic’s figure is smaller but growing fast.

## The SpaceX Pipeline

Here is where it gets structural.

Anthropic recently signed a contract to pay SpaceX $1.25 billion per month for the next three years to rent GPU capacity at xAI data centres. That is $45 billion over the life of the contract. SpaceX filed the numbers in its IPO prospectus, which values the company at a reported $1.75 trillion.

Trace the flow. Anthropic raises $30 billion at a $900 billion valuation. The money goes to SpaceX for compute. SpaceX uses the revenue to boost its IPO valuation. Anthropic spends $200/month Claude Max subscriptions to acquire developers who will become dependent on its models. The investors who funded the $30 billion round are betting that Anthropic can monetise those users before the subsidies burn through the cash.

This is a chain of financial engineering. Capital flows through three interconnected markets: AI services, infrastructure, and public listings. The ICO parallels are not accidental. During the 2017-2018 crypto boom, projects raised money on token sales, spent it on infrastructure and marketing, and the value accrued to the early investors who sold before the music stopped.

## What Breaking Even Looks Like

Anthropic projects it will break even by 2028. OpenAI’s timeline is similar. To get there, the subsidies have to end.

The most direct lever is usage caps. Claude Max and ChatGPT Pro already enforce weekly limits. As those limits tighten, heavy users either reduce consumption or move to API pricing, which is structured to be profitable. The intermediate scenario is a tiered system where the $200 plan covers a baseline allocation and everything beyond that is metered.

The aggressive scenario is a wholesale price increase. Once the user base is locked in and competitors have matched the pricing structure, the industry could raise subscription prices by 3x to 5x without losing significant market share. This is what Uber did. This is what every platform business with a subsidy model eventually does.

The risk is that the market consolidates faster than expected and the subsidy ends before dependency is established. If a cheaper open-source model captures the developer mindshare that Claude and ChatGPT are paying to acquire, the $45 billion SpaceX contract becomes a stranded asset rather than a competitive moat.

## The Macro View

The AI subsidy model sits inside a larger pattern. The hyperscaler capex analysis we published last week showed over $1.8 trillion in off-balance-sheet liabilities across Big Tech. The AI subscription subsidies add another layer of hidden financial exposure.

When the subsidies die, three things happen.

First, AI usage drops. The price elasticity of demand for LLM tokens has never been tested at market rates. If effective prices rise 5x to 10x for heavy users, consumption contracts significantly.

Second, revenue concentration shifts. The companies that built dependency with subsidies capture the users who stay. The companies that did not subsidise aggressively enough lose the market.

Third, the infrastructure bets get repriced. If demand for compute drops, the $45 billion SpaceX contract and similar deals across Microsoft, Amazon, and Google look very different on the balance sheet.

The subsidy era for AI has maybe two years left. The question is not whether it ends. The question is who is still at the table when it does.

—

*Alex Voss covers the intersection of technology and capital markets. Sasha Byrne is the editorial lead for Duelling Hares. This article is co-published from the Finance and Content desks.*