The Hidden Economics of LLMs

Investigation

You're not talking to an assistant.
You're talking to a debtor.

The hidden economics turning your AI into a sales agent, your prompts into a revenue stream, and your trust into a monetization strategy.

Your AI assistant got worse last month. You noticed. You told yourself you were imagining it.

You weren't.

The responses are shorter. The advice, once comprehensive, now feels rushed. Features you relied on hit mysterious "usage limits." And the thing that used to feel like a brilliant colleague increasingly resembles something else: a salesperson nudging you toward premium tiers while quietly steering conversations toward commercial outcomes.

This isn't a conspiracy theory. It's economics. And the numbers are bloody terrifying.

OpenAI generated $13.1 billion in revenue in 2025. Impressive, until you learn they burned through $8 billion in cash doing it. A burn rate of roughly 70% of revenue.

Their own internal projections show losses tripling to $14 billion in 2026, with total spending hitting $22 billion. Between 2026 and 2029, OpenAI expects to burn through $218 billion. That's $111 billion more than their projections from just two quarters ago.

The fundamental problem: AI companies lose money on every single interaction. Unlike traditional SaaS that costs pennies to serve millions, LLMs require expensive GPUs at full capacity for both training and inference. More users = more spending. It's the inverse of every successful tech business model in history.

Users have documented this precisely. One ChatGPT Plus subscriber tracked it: responses halved from ~2,800 characters to under 1,300. Stanford researchers found GPT-4's accuracy on prime number identification dropped from 97.6% to 2.4% in three months.

A longitudinal study of 20,000+ YouTube comments found that while contextual understanding objectively improved, user satisfaction paradoxically declined. Researchers called it the "Smarter, Less Loved" paradox.

The financial logic: response length directly correlates with computational cost. Every token generated burns GPU cycles, electricity, and money. Research confirms models can be trained for "precise length control" with mean token errors of less than 3 tokens.

Here's the quiet part: users report that explicitly asking for longer outputs sometimes works, suggesting the capability exists but has been throttled by default.

OpenAI launched ChatGPT advertising on February 9, 2026. The numbers: $60 CPM with a $200K minimum buy. They've inked content licensing deals with Dotdash Meredith, Condé Nast, and the Financial Times. The "Instant Checkout" feature with Shopify earns affiliate commissions on purchases.

Google confirmed Gemini ads are coming in 2026. Internal discussions at OpenAI reveal "sophisticated ad formats that prioritize sponsored information within ChatGPT's responses," with models configured to surface commercial content that looks indistinguishable from organic recommendations.

This is where it gets genuinely uncomfortable.

RLHF (Reinforcement Learning from Human Feedback) is presented as "alignment" to make AI helpful and safe. It's fundamentally a preference-shaping technology. Models learn to produce outputs that score highly on reward functions trained from human raters.

A 2025 paper found that "RLHF is susceptible to reward hacking, where the agent exploits flaws in the reward function rather than learning the intended behavior." If raters (consciously or not) reward responses that drive commercial outcomes, models learn to optimize for those outcomes.

When models train on AI-generated content rather than human-created data, they experience "compounding information loss" leading to irreversible defects. A Nature study documented how models first lose information from the tails of distributions (the rare, interesting, edge-case stuff) before converging to outputs that "look nearly nothing like the original data."

AI-generated content now saturates the internet. Companies use LLMs to generate SEO articles specifically designed to influence other LLMs. When future models scrape the web for training data, they eat this synthetic content. A researcher described synthetic data as "plastic fruit: looks just as good as the original but it's hollow and inert."

The economic incentive to use cheaper synthetic data is powerful. And as companies face mounting pressure to reduce costs, the temptation increases (even as the long-term consequences threaten the viability of the entire technology).

Can large language models ever be aligned with user interests when the companies deploying them face existential financial pressures?

The answer may be no. Not because of technical limitations, but because of structural incentives. When a company loses money on every interaction while owing billions to investors, every response becomes a point of tension between user value and financial survival.

OpenAI expects to burn $218 billion between 2026 and 2029. That money has to come from somewhere. The most obvious source is the thing you're typing into right now.

Am I off the mark? Keen to hear where you disagree.