Some AI start-ups are in trouble, but not most of the value. A few frontier labs and infrastructure companies hold most of the money: OpenAI, Anthropic, xAI and Databricks. Below them sit hundreds of companies valued at $1–3 billion. The weakest of these resell access to someone else’s model, have no data of their own, and have gross margins that fall as usage grows. They are the ones exposed to down-rounds, acqui-hires and shutdowns. The count of AI unicorns depends on who is counting and how they define “AI,” so treat any single figure, including 372, as rough. The useful question is which companies have proprietary data, falling cost per query and customers who renew.
How many AI unicorns are there, really?
Counts run from a few hundred to more than 500. The number depends on whether a tracker counts only companies that build models or also software firms that call a third-party API. CB Insights’ unicorn list and Crunchbase use different tagging rules, so their totals differ. A software company that adds a chatbot to an existing product can be counted as an “AI unicorn” in one dataset and not in another.
The money behind these companies is real. The Stanford AI Index 2025 reports US private AI investment of $109.1 billion in 2024, and global generative-AI investment of $33.9 billion. Much of that went into a small number of very large rounds. A rising unicorn count therefore says little about how healthy the typical company is.
Which AI start-ups are most likely to survive?
| Type | What it looks like | Likely outcome |
|---|---|---|
| Frontier labs and platforms | Raise billions, own their models, sell to large enterprises, have cloud partnerships (OpenAI–Microsoft, Anthropic–Amazon/Google) | Survive. The risk is cost and concentration, not running out of money |
| Infrastructure (“picks and shovels”) | GPU clouds, data tooling, evaluation and security tools | Grow or get bought. CoreWeave borrowed heavily against its GPUs and went public in March 2025 |
| Applications with proprietary data | Products built for one industry, with workflows and data competitors cannot copy | Survive if gross margins hold as model prices fall |
| Thin wrappers | A user interface on top of someone else’s API | Most exposed. One feature added by the model provider can wipe out the product |
Failures since 2024 have mostly been quiet. Microsoft hired Inflection AI’s founders and licensed its technology in 2024. Google did a similar deal with Character.AI, and Amazon hired Adept’s leadership. Each deal left a company that no longer resembled what investors had funded. Builder.ai went into insolvency in 2025 after its revenue figures were questioned, despite having been valued at more than $1 billion.
What metrics flag a company in trouble?
- Burn multiple above 2: the company spends more than $2 of net cash for each $1 of new annual recurring revenue (ARR). Below 1 is strong.
- Net revenue retention under 100%: existing customers are spending less over time.
- Gross margin that won’t be disclosed or keeps falling: inference costs are rising faster than revenue.
- Secondary shares trading at a deep discount to the last round: insiders are selling below the headline price.
- Revenue from short pilots, not renewals: trial budgets that never turn into contracts.
Job seekers can ask about most of these directly. A company that won’t share retention or gross-margin trends with a finalist candidate is telling you something.
Is GPU scarcity still the main threat?
Not as it was in 2023. Then the bottleneck was TSMC’s advanced packaging capacity and the supply of high-bandwidth memory. Rental prices for H100-class GPUs have since dropped sharply as supply has grown. Model prices have collapsed too. The AI Index found that running a query at GPT-3.5 level became more than 280 times cheaper between November 2022 and October 2024.
Cheaper inference helps the companies that buy compute and hurts those whose business was selling it. The new constraint is electricity. The IEA projects that data-centre electricity demand worldwide will more than double by 2030. Lawrence Berkeley National Laboratory estimates US data centres went from 4.4% of national electricity in 2023 to a projected 6.7–12% by 2028. Waits for grid connections, not chip shipments, now decide where new capacity can be built.
How does regulation change the picture?
The EU AI Act (Regulation 2024/1689) came into force in August 2024 and is being phased in:
- Bans on certain uses took effect in February 2025.
- Obligations for general-purpose model providers took effect in August 2025.
- Most high-risk obligations follow later, and the Commission has proposed pushing some of them back. Check the Commission’s implementation page for the current dates.
Fines reach €35 million or 7% of global turnover for prohibited uses.
The NIST AI Risk Management Framework is voluntary guidance, not a US law. It matters mainly because enterprise and government buyers increasingly ask vendors to show they follow it. No reliable figure exists for compliance cost as a share of revenue. It depends on whether the product falls in a high-risk category such as hiring, credit or biometrics. For a general-purpose writing tool the cost is small. For a hiring tool it can mean a dedicated team. Fixed compliance costs fall harder on mid-sized companies than on the largest labs.
What this means if you buy AI tools for content work
As a video production company, we rely on AI vendors for transcription, rough cuts, upscaling, voice and generative B-roll. We have seen tools change pricing, cut credit allowances or change their terms partway through a project. Our rules:
- Never let a deliverable depend on a tool you can’t export from.
- Keep project files and raw media in your own storage.
- Check whether the vendor owns its model or rents it. Renters are more exposed to repricing upstream.
- Read the licensing and indemnity terms for commercial use before a client job, not after.
A tool that disappears during a campaign can cost days of rework. That cost is usually larger than whatever the tool saved.
If you are deciding whether to invest in, work for or build on an AI start-up, ask for four things: burn multiple, net revenue retention, gross-margin trend and who supplies the model. Then check its regulatory exposure against the EU AI Act timeline. A company that answers all four clearly is probably not the one in trouble.
