The Brutal Truth About AI Startups: Why 200 Companies Became 5 in 18 Months

Last December, I attended a rather unusual conference. Closed-door. About thirty AI founders crammed into a room. The organisers threw some numbers on screen that honestly made my stomach drop: at the start of 2023, there were over 200 startups in China building their own foundation models. By late September 2024, fewer than five teams still had their own models and were operating independently.

Fewer than five.

Let that sink in for a moment.

I remember it vividly — between March and July 2023, it felt like there was a new model launch every bloody week. Parameters went from 7B to 13B, then 70B, then 130B. The leaderboards became an arms race, pure and simple. We had this running joke in the community: if your GPU cluster could train a 70B model, you didn't even need a business plan. Investors would find you.

Fast forward to now. Those names that dominated the leaderboards? Nearly all gone. A few got acquired by big tech and absorbed into cloud divisions — their products dismantled into APIs. Some pivoted to vertical applications, quietly swapping their own models for someone else's. And two? Their websites return 404 errors. The founders don't even reply on messaging apps anymore.

Wait — I should clarify something. It's not literally fewer than five. It's fewer than five that are independently operating with their own models. If you count the ones still technically alive but completely off their original track, maybe a dozen or so. But that's not really the same kind of startup anymore, is it?

My own company tried training a legal document model in June 2023. Spent roughly £25,000 on compute alone — that's not counting the human cost or data preparation. When we finally tested it, the inference quality was worse than just calling a major provider's API, which had just slashed their prices. That moment genuinely made me question everything. When a tech giant's general-purpose model is good enough and priced at rock bottom, where exactly is your so-called technical moat?

I still haven't fully figured that one out.

It's... complicated.

The Price War: Who's Actually Bleeding?

Let me share some real numbers. Based on what we've been tracking, API pricing for mainstream models in China dropped over 90% between early 2023 and late 2024. Some lightweight models — think equivalents to GPT-3.5-turbo — now cost less than $0.001 per thousand tokens. What does that mean in practice? If you've got a product with 100,000 daily active users making hundreds of millions of calls per day, your inference costs might actually be less than your team's weekly coffee budget.

Sounds brilliant for developers, right?

Here's the catch.

I know someone who built an AI customer service product. Started in early 2023, fine-tuned on a popular open-source model, accumulated over 400,000 high-quality domain-specific conversation logs. Had paying customers and everything. Then this February, their biggest competitor became... the cloud provider behind that open-source model. The cloud provider bundled similar functionality into their own product suite and priced it 40% lower.

Why could they do that? Because they own the compute. They own the model. They own the distribution channel.

That founder and I grabbed drinks later. He said something I haven't been able to shake:

"We're not competing against other startups. We're fighting a landlord who's got nuclear weapons. And here's the kicker — we're renting our weapons from him."

That's the brutal reality of this price war. You think it's a technology competition, but actually, they're squeezing you on four dimensions simultaneously: compute costs, data flywheels, distribution lock-in, and ecosystem bundling. A startup might win on one dimension. The giants can apply pressure on all of them at once.

My biggest mistake last year? Underestimating ecosystem lock-in. We built a developer tool — code completion, automated refactoring, the works. Plugged into two different models for comparison. The product was genuinely good. Got solid traction on Reddit and Hacker News. But when a major cloud provider launched their own coding assistant, deeply integrated with their cloud services, IDE plugins, and DevOps pipelines... our growth curve went flat.

Not because users thought we were worse. Their feedback was: "The switching cost is too high. And anyway, both are free."

When free becomes the baseline, differentiation has to come from something irreplaceable. But what's actually irreplaceable? I've been wrestling with this for nearly a year.

Get Acquired or Get Obliterated? There's a Third Path

The conversation in startup circles these days is binary: either let yourself get acquired by a giant, take a decent payout, and become a feature in someone else's ecosystem — or fight to the death, betting you can find a niche too small for the giants to care about.

But here's what I've noticed: the smartest teams are carving out a third path.

What path? Becoming indispensable to the giant's ecosystem while remaining too niche — or too messy — for the giant to build themselves.

Let me give you an example. There's this three-person team in Hangzhou doing medical imaging AI. The founder came out of a university hospital. From day one, they never tried to build a general-purpose model. Instead, they took an open-source foundation model and poured their energy into medical imaging data — cleaning it, annotating it, ensuring regulatory compliance. Their core competency isn't the model at all. It's a data processing pipeline validated across three hospitals' radiology departments, plus deep relationships with over a dozen medical institutions.

Here's the thing about healthcare data: it's a compliance nightmare. Data can't leave the province. Can't leave the hospital. Sometimes can't even leave the department. When a tech giant wanted to enter this space late last year, they assessed the situation and realised that building this from scratch — the time, the compliance risks — would cost far more than just partnering.

The result? That tiny startup didn't get acquired. They became a technology supplier inside the giant's healthcare solution, retaining their own brand and client relationships.

This illustrates something crucial: the safe zones in the ecosystem aren't at the top of the tech stack. They're in the deep waters of specific business scenarios.

What are giants good at? General capabilities and scalable replication. What are they terrible at — and unwilling to do? Heavy customisation, long-cycle trust building, and the messy, labour-intensive work that requires deep industry know-how. That "messy work" is precisely what forms the strongest moat for startups.

I adjusted my own strategy along these lines. Shifted the team from "build a better model" to "solve a specific problem that no one else can solve in this particular context." It was painful. Had to kill features that looked cool but created zero defensibility. Had to resist the urge to do everything. Had to voluntarily give up parts of the general market.

But we survived. And honestly? We're doing alright.

From what I can tell, every team still standing has followed a similar path.

What Happens Next?

I reckon the next 12 months will bring three pretty clear trends:

First, the model-layer startup window is basically shut. I'm not saying there's zero opportunity, but "training your own foundation model" as a core competency is no longer a viable startup thesis. Compute costs, data acquisition difficulty, and the giants' pricing strategies have turned this into a pure capital game. The money you can raise? Nowhere near enough to compete.

Second, middleware and tooling will get re-evaluated. As models themselves become commoditised, what actually determines application quality is the engineering around them — data pipelines, evaluation frameworks, safety guardrails, multi-model orchestration. There might be new opportunities here, but only if you can establish standards faster than the giants can.

Third, the "last mile" in vertical industries becomes the main battleground. Legal, healthcare, finance, manufacturing — these sectors don't lack general intelligence. They lack people who can actually deploy that intelligence into specific business workflows. What's needed isn't stronger models. It's deeper industry understanding. You need to know how doctors actually work, how lawyers actually write documents. Reading a few papers won't cut it.

Had dinner with a friend in venture capital last week. He said something that's been rattling around my head ever since: "This wave of AI entrepreneurship isn't a race against technology. It's a race against the strategic patience of big tech. They can tolerate a business line losing money for three years. Your company might have 18 months of runway. So don't bet on big tech not doing something. Bet on them doing it badly."

That might be the clearest take I've heard all year.

So here's what I'm curious about: is your product going head-to-head with a giant's core strategy? Or are you quietly building a moat in some corner they've dismissed as not worth their time?

Drop a comment with your experience. I genuinely want to hear from teams still fighting independently — how are you thinking about this?

Tags: #AI #startups #machinelearning #techstrategy #pricing #ecosystem