Why AI is the new Uber – and what it could cost your startup

I’ve been in a lot of conversations lately about AI transformation.

The pitch is usually the same: implement these tools, automate these workflows, reduce this headcount, improve that margin. The spreadsheet looks great. The board and your investors love it.

There’s just one problem. The economics underpinning those spreadsheets aren’t real.

Let’s be direct: AI token pricing is artificially low right now, and it is an intentional strategy — not a gift.

We’ve seen this movie before

Remember when Uber was cheaper than a taxi? For years, Uber priced rides well below cost, subsidised by billions in venture capital, with the explicit goal of making taxis economically unviable.

It worked. And once the competition was gone and the habit was locked in, fares corrected. The average price of an Uber ride has risen approximately 94% since those subsidy days ended.

The bargain was never real. It was a ruthless customer acquisition strategy at planetary scale.

Frontier AI labs — whether you’re calling the API directly, using ChatGPT, or buying software that wraps these models — are running the same play. The unit economics are, by any honest assessment, deeply unsustainable.

OpenAI’s more than $20 billion in annual revenue looks impressive until you consider that they’ve committed spending more than $600 billion on infrastructure in the next 4 years. Where, exactly, do you imagine that money is going to come from?

And they’re not alone. Frontier labs that build the models many products rely on are spending vastly more than they recoup. That gap is covered by venture capital, with an expectation of returns once your workflows are dependent enough that you can’t easily leave.

Anthropic’s recent price increases are not an anomaly. They’re a signal.

Why this is a startup problem specifically

Large enterprises have procurement teams, legal resources, and the leverage to renegotiate contracts. Startups, by and large, do not. What startups do have is a structural incentive to use current token pricing as a key assumption in their operating models, and that’s where real risk accumulates.

The AI cost-reduction pitch that’s most dangerous right now is this one: replace headcount (or don’t hire it in the first place) with AI tooling to growthmaxx before your next funding round.

On today’s token prices, that calculation can look compelling. But headcount reductions are structurally hard to reverse. Institutional knowledge leaves. Hiring pipelines atrophy. People you never hired don’t know how to run your processes. If token costs increase significantly (or even modestly compared to what Uber fares did), many of those businesses will find themselves without an ability to execute at a price they can afford to pay. That is not a theoretical risk. That is a predictable consequence of not doing the planning work.

Tokens have already hit cost parity with junior devs. Watch for the LinkedIn posts of tech founders inventing the concept of employees. Don’t worry, I’ll wait.

If you’re waiting on artificially subsidised pricing to validate your operating model, you are enabling your own future crisis.

What rigorous AI planning actually looks like

To find the answer, we need to think about this less as a technology adoption problem and more as a vendor risk management problem. Because that’s actually what it is. Here’s what that looks like in practice:

• Run your business case at 2×, 5×, and 10× current token costs. If it falls apart at 3×, your model has a problem, not the technology.
• Distinguish between high and low lock-in use cases. Well-defined, portable tasks (summarisation, classification) are low risk. Deeply integrated workflows built on proprietary model capabilities are high risk.
• Build portability into your architecture from day one, even if you intend never to use it. The optionality has value precisely because you may not need it.
• Separate “AI as augmentation” from “AI as replacement” in your planning. Augmentation use cases are resilient to price increases because the human, and their judgment, remains the asset. (This is also your reminder that people from underrepresented backgrounds are more likely to have their work seen as automatable – please consider the isms inherent in your AI adoption strategy.)

None of this requires you to reject AI. It requires you to be a good operator who sees through the product marketing hype currently driving the narrative.

The competitive advantage in this moment

Here’s the thing: most of your competitors are not doing this work. They’re moving fast, building on today’s pricing, and making structural commitments—including workforce ones—that will be difficult to unwind.

The startups that will be best positioned in three to five years are not necessarily the ones that adopted AI most aggressively. They’re the ones that adopted it most thoughtfully: that built resilient cost models, preserved human capability in the places that matter, and structured their AI dependencies with the same rigour they’d apply to any critical vendor.

I promise you: doing this planning work now will serve you so much better than whatever your competitors are building on a subsidy that won’t last.

The Uber story didn’t end when the fares went up. It ended when the people who’d restructured their lives around cheap rides had no good alternative. Don’t let that be your business.

• Aubrey Blanche is the Director of Ethical Advisory & Strategic Partnerships at The Ethics Centre and the founder of The Mathpath. She is currently a masters student in AI Ethics and Society at the University of Cambridge.