AI-pilled-then-killed

The economics of LLMs are nightmarish. And enterprises may be sleep-walking into a trap.

A $200/month subscription to tools like Claude Code can conceal an enormous amount of underlying compute usage. Under heavy use, the token consumption implied by such a subscription can translate into thousands of dollars at retail API prices. Analysts have circulated estimates showing that the usage allowance of the highest subscription tiers could correspond to roughly $5,000 worth of API-priced tokens in a month. [1]

Whether the exact number is $2,000, $5,000, or something else is almost beside the point. The underlying dynamic is what matters. The models produced by frontier labs are aggressively subsidized in the form of flat-price subscriptions, generous limits, or bundled features.

This is not generosity. It is market capture.

And if you are building a company or integrating these models deeply into your organization, the implications are uncomfortable.

Because the economics of large language models are structurally misaligned with the companies building on top of them.

Subsidies First, Pricing Power Later

The frontier model providers, most notably OpenAI and Anthropic, are currently operating in a phase that should look familiar to anyone who has watched the development of modern technology platforms.

The playbook is simple: subsidize usage, capture developers and enterprise integrations, consolidate market share, adjust pricing once dependency sets in.

None of this is especially unusual. Cloud computing went through a similar phase. So did ride-sharing. So did food delivery. When a platform sits at the base layer of an ecosystem, the priority is adoption, not margins.

But LLMs introduce a new complication.

The cost structure is not merely opaque. It is fundamentally unstable.

Running large models requires expensive GPUs, enormous data centers, and continual model retraining. Even as hardware improves, the scale of usage tends to expand faster. Every step forward in capability increases token consumption and computational intensity.

Agentic workflows make this worse. Instead of one prompt and one response, AI systems increasingly run dozens or hundreds of internal model calls to complete a task. A single user interaction can quietly generate large amounts of inference activity.

Which is why the economics of "unlimited" AI tools tend to break down quickly. Heavy users can generate backend compute costs that dwarf what they pay in subscriptions. [2]

For the companies providing the models, this creates a strange equilibrium. Subscriptions are attractive because they encourage adoption and experimentation. But the underlying economics push toward usage-based pricing.

That tension eventually resolves itself in one direction. Historically, it resolves toward higher prices.

The Structural Problem for AI Startups

Now consider what this means for startups building on top of frontier models.

If you build a product that depends on an external model provider, your gross margins are permanently exposed to the pricing decisions of that provider. Every API call is a cost center. Every improvement in model capability can also increase your cost structure.

And crucially, the model provider can always move up the stack.

If an application layer product becomes successful enough, the model provider has three options: increase API pricing, launch a competing feature, or restrict or revoke access.

None of these options are theoretical. Variations of this pattern have appeared repeatedly across technology platforms. Operating systems, social networks, cloud providers, and mobile app stores have all followed similar trajectories.

Platform risk is not new.

What is new is how concentrated the AI stack currently is.

There are only a handful of companies operating true frontier models. Training them requires vast amounts of capital, specialized hardware, and elite research talent. That creates a power asymmetry between the base layer and the application layer.

Startups and incumbents building AI products often behave as if they are building on stable infrastructure.

In reality, they are building on someone else's pricing policy.

The Flywheel That Really Matters

If the cost structure were the only issue, this would already be a difficult situation.

But there is a second dynamic that matters even more.

Data.

When organizations integrate tools from companies like Anthropic or OpenAI into their workflows, they often connect them to internal systems. Documents. Emails. Code repositories. CRM records. Knowledge bases.

All of this information becomes part of the interaction surface between the model and the organization.

The frontier labs are not merely selling inference.

They are learning.

Every workflow integration reveals how real organizations operate. Every use case exposes new vertical opportunities. Every pattern of usage suggests new product features.

This creates an enormous feedback loop.

Access to real operational data allows the model providers to build increasingly specialized tools across industries. Legal drafting. Financial analysis. Software engineering. Marketing operations. Customer support.

In effect, the AI labs gain a panoramic view of knowledge work across the economy.

And they can productize it.

This is where the strategy becomes both risky and brilliant.

Risky, because the upfront costs are enormous. Brilliant, because the long-term strategic position becomes extremely powerful.

The Disintermediation Problem

There is another aspect of the AI narrative that many organizations seem reluctant to confront.

The investment thesis behind the major AI labs is not simply about improving productivity.

It is about replacing labor. And replacing labor in the limit means disintermediating the firms that organize that labor.

If large language models become capable enough, many forms of knowledge work become software features. Legal research, market analysis, software debugging, contract drafting, financial modeling. Tasks that currently support entire industries.

This is not a subtle shift.

It is the explicit economic premise behind the valuations attached to the frontier AI companies.

You cannot simultaneously believe that AI will replace large segments of knowledge work and also believe that the firms performing that work will remain structurally intact.

One of those assumptions has to give. Which leads to a somewhat uncomfortable observation.

Organizations are currently feeding the systems that may eventually compete with them.

The Last Durable Moat

For most companies doing knowledge work, the real competitive advantage is not the technology stack. It is institutional knowledge.

The accumulated understanding of how a particular organization, industry, or process actually works. The tacit knowledge that develops through years of experience. The unwritten rules. The judgment calls. The context that never appears in public documentation.

Much of this knowledge does not exist on the internet. It lives inside organizations.

And that makes it one of the few remaining durable moats in a world where models have already consumed most publicly available information.

If AI is going to meaningfully replicate high-value knowledge work, it will need access to this kind of organizational context.

Which is exactly what many companies are now providing through deep integrations. Slack archives. Internal wikis. Engineering repositories. Operational playbooks. Customer histories.

In other words, the crown jewels.

There are perfectly rational reasons for doing this. The productivity gains from AI tools can be significant. Teams can automate routine tasks, accelerate research, and streamline workflows.

But it is worth recognizing the strategic tradeoff being made. Organizations are lowering the drawbridge over the one moat that might actually matter.

A More Cautious Approach

None of this is an argument against using AI.

The productivity benefits are real, and the technology will continue improving. But it does suggest a more cautious posture.

Three principles are worth considering.

First, treat frontier models as infrastructure with variable pricing. Assume costs will rise over time, not fall.

Second, avoid building core business logic that depends entirely on a single model provider. Build in ways that are platform or model agnostic. This way if the pricing underneath you shifts, you have options to move to without much stress or downtime.

Third, think carefully about what internal data should be integrated into external AI systems. Not all knowledge needs to be externalized. And not every workflow needs to be connected. There are ways to expose only minimally relevant context to frontier models and keeping all other knowledge local.

Better yet, there are also ways to "own your LLM stack", by serving inference locally. This is a fast growing area and both the software, models, and hardware are making this increasingly feasible.

Let me illustrate with just one example.

People think they need to subscribe to Anthropic because Claude Code is so good. Very few realize the fundamental difference between a model and a harness. Claude Code is a harness - which can be used with non-Anthropic models, including local ones! You can use the excellent Claude Code harness with something like a Qwen open-source model, served locally and all your data remains air-gapped.

The current moment in AI feels similar to the early cloud era. Adoption is rapid. Capabilities are improving weekly. The incentives all point toward deeper integration.

But infrastructure layers eventually exercise power. When they do, the companies that built their businesses entirely on top of them tend to discover the downside.

The economics of LLMs are already complicated. The strategic implications may be even more so.

Footnotes