Unit economics of generative AI products - why most lose money

OpenAI is losing billions this year.

Anthropic projects similar losses. These aren’t startups fumbling toward product-market fit. These are the companies that defined generative AI, with millions of users and real revenue coming in. And they’re hemorrhaging money on every single query.

The unit economics of generative AI products are broken. Not broken in a “we’ll figure it out eventually” way. Broken in a “the fundamental business model doesn’t work yet” way.

Why AI economics work backwards from SaaS

Traditional SaaS has beautiful economics. You build software once, host it cheaply, serve unlimited users for basically nothing. An extra user costs you almost nothing. Gross margins reach extraordinary levels.

AI products flip this entirely.

Every user interaction costs you real money. Every prompt burns compute. Every response requires expensive GPU time. Inference costs almost always surpass training costs over a model’s lifespan - training happens once, but inference scales with every new user. And most companies report that AI costs are actively eroding their gross margins. That isn’t a margin problem. That’s a business model problem.

Building Tallyfy has given me a front-row seat to this contrast. Our traditional SaaS features cost almost nothing to serve at scale. Our AI features cost us every single time someone uses them. The difference isn’t subtle - it’s structural.

Total AI spending keeps growing at 44% year-over-year, with infrastructure alone accounting for a large share. GenAI has entered the Trough of Disillusionment - spending is skyrocketing while fewer than 30% of AI leaders say their CEOs are satisfied with the returns.

The cost squeeze that breaks businesses

Here’s the trap. Costs are dropping fast. Really fast. The cost of LLM inference has dropped by a factor of 1,000 in 3 years. For equivalent performance, costs decrease significantly every year.

Sounds promising, right?

Except usage grows faster than costs decline. ChatGPT reportedly cost OpenAI tens of millions just to process prompts in January 2023 alone. One month. Microsoft’s Bing AI chatbot needs billions in infrastructure to serve all Bing users - not to develop it, just to run it.

The trap works like this: you price your product based on current costs. Users love it. Adoption explodes. Your costs scale linearly with every new query. Yes, costs per query drop over time, but if your user base grows 10x and query volume per user doubles, you’re still losing money faster than before. The math doesn’t close.

Enterprise generative AI spending tripled in a single year, and the industry keeps pouring money in despite negative returns. OpenAI and Anthropic post massive revenue growth while burning through cash faster than they earn it.

That’s not a scaling problem. That’s a fundamental economic problem.

Three mistakes that destroy AI product economics

Most companies building AI products make the same errors. I’ve probably made some version of all three myself.

The first: treating AI features like traditional features. Companies bolt on AI capabilities to existing products without rethinking their pricing model. You can’t charge traditional SaaS prices when your cost structure looks like a utility company.

The second: underestimating how users will actually use AI. If you charge a flat rate and costs scale with usage, the distribution of queries per user becomes life or death for your margins. A small group of power users can make your entire product unprofitable. Companies are responding by raising prices substantially, but that limits adoption precisely when you need scale.

The third mistake is ignoring hidden costs beyond inference. A staggering 85% of organizations misestimate AI project costs by more than 10%. Data preparation, infrastructure, and ongoing maintenance add up to the majority of total project costs. Add cloud compute, software integration, developer time, data storage, MLOps, and continuous model retraining, and that vendor quote balloons by half again before you ever hit production.

The failure rate tells the story. The overwhelming majority of enterprise AI solutions fail. The average enterprise saw more than 80% of AI projects fail before they reached production. Only 11% of organizations have AI agents actually running in production - the rest are stuck in pilot purgatory, quietly shelved after cost overruns.

Every executive surveyed reported canceling or postponing at least one generative AI initiative due to cost concerns. Not some executives. Every single one.

How Midjourney broke the pattern

One company figured it out. Midjourney.

Midjourney generates hundreds of millions in annual revenue with a remarkably small team. Profitable in August 2022, six months after launch. No venture capital. No massive infrastructure spend. No desperate search for a business model that works.

They did something different. They focused on one specific use case - AI image generation - and charged appropriately for it. Tiered subscriptions from basic plans to professional tiers, with usage limits that match their cost structure. When users need more, they pay modest rates for extra GPU time. Simple. Transparent. Profitable.

The lesson isn’t “build an image generator.” The lesson is that specialized tools focused on specific high-value use cases can achieve sustainable unit economics. Midjourney doesn’t try to be everything. They picked a problem where users clearly understand the value and are willing to pay enough to cover real costs.

Compare this to companies trying to be AI platforms for everything. They’re trapped between needing scale to justify infrastructure investment and bleeding money on every query. Midjourney prints money with a team smaller than most startup engineering departments. I find that genuinely remarkable.

What actually makes AI economics work

Start with the cost structure, not the features. Know exactly what each user interaction costs you. Build monitoring so you can track cost per user, cost per query, and cost per value delivered. Most companies treat AI costs as “infrastructure” and lose visibility into what’s actually burning money.

Price for your actual costs, not for what competitors charge. If your AI feature costs you real money every time someone uses it, your pricing has to reflect that. Usage-based pricing isn’t just trendy - it’s the only way to align revenue with costs. Flat-rate pricing works when costs are fixed. AI costs are never fixed.

Gate expensive features carefully. Not every feature needs AI. Not every AI feature needs the most expensive model. Using appropriately-sized models and caching responses can cut costs dramatically without hurting user experience.

Explore multi-model routing. Routing tasks to cost-efficient models can cut inference costs by up to 85%. Simple questions go to small, cheap models. Only when quality checks fail does the request escalate to something more powerful. Some teams report cost reductions above 99% on specific workloads by routing intelligently instead of sending everything to a frontier model.

Focus on high-value use cases where users will pay enough to cover costs and a real margin. General-purpose AI tools have terrible unit economics because they try to be useful for everything at a price point that works for nothing. Specialized tools that solve expensive problems can charge appropriately.

Consider hybrid models. The biggest ROI consistently comes from automating knowledge work across customer operations and sales, not generic back-office tasks. Use AI internally to cut costs before you use it externally where costs scale with users.

Most AI products today are subsidized by venture capital. The business models don’t work. The unit economics are negative. Companies are betting that costs will drop fast enough, or that they’ll find pricing models that work before they run out of money. The market is responding: spending concentrates on fewer vendors, and the great AI consolidation is thinning out companies that can’t figure out the math.

Some will figure it out. Costs are dropping fast enough that if you can control usage patterns and price appropriately, sustainable AI products are possible. Midjourney proves it. So do specialized AI tools in vertical markets where value justifies premium pricing.

But most won’t. The companies trying to build general-purpose AI products at consumer prices are playing a game where the economics get worse as they grow. More users means more losses. Better features means higher costs. Growth becomes the problem, not the answer.

For mid-size companies, this matters because the vendors selling you AI products aren’t telling you about their unit economics problems. They’re selling you features and capabilities without admitting they lose money on every query you run. That affects their long-term viability, their pricing stability, and whether they’ll even exist in three years.

If you’re buying AI products, ask about their cost structure and pricing model. If you’re building AI products, fix the unit economics before you scale. And if you’re investing in AI products, understand that revenue growth without margin improvement is just buying your way to a bigger loss.

The technology works. The unit economics, for most products, still don’t.