We tracked 23 funded AI wrapper startups. Only two say they make money

Every week someone sends me a funding headline. Another AI startup, another few hundred million dollars, another valuation that sounds like a typo. And almost every time the same question follows: are these companies actually making money, or are they just very good at raising it?

Most of them are what people call AI wrappers. They did not build the model. They built a product on top of a model from OpenAI, Anthropic or Google, and they pay that model maker every time a customer uses it. Cursor began like that. So did Lovable, Harvey and Perplexity. People used "wrapper" as an insult for a while. Then some of these companies started growing faster than any software business before them.

So I made a list. Twenty-three AI application startups funded by Y Combinator or well-known VCs in the last two years, with what each one raised, what it earns, and whether it makes a profit or burns cash. Every number has a link next to it. Where a figure is an estimate I have said so, and where nobody knows, the table says "Unknown". I checked everything on 15 November 2026.

The short version

• Money in: these 23 companies raised about $16.5 billion between November 2024 and November 2026. At about ₹94 to the dollar (xe.com), that is roughly ₹1.55 lakh crore.
• Revenue: the 20 that have shared a number add up to about $7.3 billion a year in revenue, measured at different dates. Cursor alone is more than a quarter of it.
• Profit: only 2 say they are profitable, Gamma and Retell. Both are among the smaller companies here, and both raised very little.
• Burning: 2 are on record as losing money (Cognition and Rocket), and 2 more run thin margins that swing with model prices (Replit and Bolt).
• Exits: 2 were sold. Cursor went to SpaceX for $60 billion in shares and Manus went to Meta for over $2 billion.
• Dead: 1 shut down (Yupp), and the wider graveyard is much bigger than this list.
• Silent: the other 14 have never published a profit figure. That silence is the most honest answer to the question you came here with.

The tracker

Tap any heading to sort. The filters above the table group companies by money status, and you can switch the amounts to rupees or download the whole thing as a spreadsheet.

A few notes so you read it right:

• "Raised" counts only rounds announced between 1 November 2024 and 15 November 2026. Older money is left out, so Glean, for example, looks smaller than its total history.
• "Latest yearly revenue" is annualised revenue or ARR, the company's current monthly revenue multiplied by twelve. Each figure says whether it came from the company itself, from a press report, or from an analyst estimate (mostly Sacra). Treat estimates as rough.
• "Making money?" uses what is on the record. "Not disclosed" means the company has not said, which is different from "losing money". Most private startups never say.
• Who is on the list: application companies that mostly run on other people's models, funded by YC or known VCs in the window, with at least one public revenue or funding number. I left out the model makers themselves (OpenAI, Anthropic, Mistral) and companies like Bland that run their own models. A few here, such as Cursor, Lovable and OpenEvidence, now train some models of their own. They stay because they started as wrappers and still lean on frontier models. I broke my own rule once, for Retell, which raised its only round in August 2024. You will see why.

Who is actually making money

Two companies. That is the whole list.

Gamma makes AI presentations and simple websites. It crossed $100 million in yearly revenue with about 50 people, and says it has been profitable for over two years (Gamma, TechCrunch). Its only round in the window was $68 million at a $2.1 billion valuation. In August 2026 its CEO said it does not need to raise again (The Next Web).

Retell builds voice agents for call centres. It announced a $100 million annualised run rate this month and says it has been profitable since 2025, on a total of $4.6 million raised and fewer than 80 staff (Business Wire). Compare that with Cognition, which raised $3.4 billion in the same two years. Retell reached a ninth of Cognition's revenue on less than 0.2% of the money.

Both numbers come from the companies themselves, so read them as claims. Still, they show something useful. Both picked a narrow job that customers pay for every month, kept the team small, and charged enough to cover the model bill. Nothing fancy.

OpenEvidence deserves a mention. It is free for doctors and makes its money from drug and device ads, and Sacra estimates its gross margin at about 90% (Sacra). That is a very healthy number. But a good gross margin is still short of a profit, and the company has not published one, so it stays under "Not disclosed".

Who is burning, and why the model bill is the problem

Cognition, the company behind the Devin coding agent and Windsurf, is the clearest case. It said it hit $900 million in yearly revenue in September 2026 and raised more than $2 billion at a $48 billion valuation. In the same report, TechCrunch cited The Information saying its cash burn could reach $800 million this year, partly from a leased Nvidia cluster (TechCrunch).

Cursor is the most interesting one. It grew faster than almost any software company in history and still had a gross margin of about minus 23% in the quarter to January 2026, according to The Information. In plain words, it cost Cursor more to serve its customers than those customers paid. The same report says some investors turned down its next round because Cursor depended so heavily on Anthropic, a supplier that also competes with it. Margins have turned positive since, but you can see why selling made sense.

Replit's gross margin swung between 36% and minus 14% during 2025 as model costs moved (Sacra). Bolt.new sat at about 40% (Sacra). For comparison, a normal software company keeps 75 to 80% of every rupee of revenue after serving costs.

This is the bigger pattern, and the research firms agree on it:

• Bessemer's State of AI 2025 found that the fastest-growing AI startups, the ones it calls Supernovas, averaged about 25% gross margin, often negative. The steadier "Shooting Stars" averaged about 60%.
• ICONIQ's State of AI 2026 puts AI product gross margins at 45% in 2025, with companies expecting 53% this year.
• Closer to home, Rocket's COO Deepak Dhanak told the Financial Express that most AI coding platforms run below 50% gross margin, many at 20 to 30% (Financial Express, 6 March 2026).

The reason is simple. A normal app costs almost nothing extra when one more customer uses it. An AI app pays the model maker for every request, and its heaviest users cost the most. I wrote about this trap for small businesses adding AI features in the AI feature margin trap. The startups above are the same story with more zeros.

The exits: sold before the answer came

Cursor was sold to SpaceX for $60 billion in shares. The deal closed on 14 August 2026, per SpaceX's own SEC filing, and SpaceX has since booked $608 million of Cursor revenue for the six weeks to 30 September (SpaceX 10-Q).

Manus, the general AI agent from Singapore, said it reached $100 million in yearly revenue eight months after launch and then sold to Meta in December 2025 (TechCrunch).

Both were great outcomes for their investors. Neither tells us whether the business could have stood on its own. A sale answers the investors' question. It leaves the profit question open.

The ones that died, and the ones that lied

TechCrunch keeps a running AI graveyard. A few names from it explain the risk better than any chart:

• Yupp raised a $33 million seed from a16z crypto to let people compare answers from hundreds of models. It closed in March 2026, less than a year after launch, saying it never found strong product-market fit (TechCrunch).
• Huxe, built by people behind Google's NotebookLM, shut down in May 2026, a day after Spotify shipped a similar feature. That is the platform risk every wrapper lives with.
• Relay.app closed this year after the big platforms built the same automations, and Ollie closed in November on a $7.5 million Khosla seed.

Then there are the numbers that turned out to be wrong.

• Builder.ai raised over $450 million, much of it before this window, with Microsoft among its backers. It told lenders it would make $220 million in 2024. The real figure was about $50 million, and it went into insolvency in May 2025 (Bloomberg).
• 11x, the AI sales rep startup backed by Benchmark and a16z, was reported to list customers that were not customers and to count short trial deals in its annual revenue (TechCrunch). Its CEO stepped down in 2025.
• Cluely's CEO admitted in March 2026 that the $7 million ARR he gave the press was false. His own Stripe numbers showed about $5.2 million (TechCrunch). That is why the tracker shows the lower figure.

So when you see "$100M ARR in eight months", read it as a claim. Sometimes it is true. Sometimes it is a trial deal multiplied by twelve.

The India angle

Two Indian names made the list, and they show both ends.

Emergent is a Y Combinator company with teams in Bengaluru and San Francisco. People describe the app they want in plain words and it builds it. It said it hit $100 million ARR eight months after launch (TechCrunch), then $120 million by July 2026, when it raised $130 million at a $1.5 billion valuation (TechCrunch). Only 8 to 9% of its revenue comes from India. That detail tells you who pays for these tools today.

Rocket, from Surat, does something similar. It raised a $15 million seed from Salesforce Ventures and Accel in September 2025, when it was at about $4.5 million ARR (TechCrunch). Its FY25 filings show about ₹13 crore net loss on about ₹2 crore operating income (Venture Intelligence). Early, and burning, which is normal at this stage.

The wider Indian picture is tougher. Tracxn counted 18 Indian AI startups that shut down in 2025, against 6 in 2024 (Analytics India Magazine).

What Y Combinator's own numbers say

YC has gone almost fully AI. Garry Tan said the Winter 2025 batch grew about 10% a week as a whole and that about 80% of the companies were building with AI (CNBC). For Winter 2026 he put average weekly growth at 14%, with about 88% of the batch AI-first (Lobster Capital recap).

Growth like that is real, and it is why the money keeps coming. But weekly growth measures revenue coming in. It says nothing about what that revenue costs to serve. Four companies on this list are YC-backed: Replit, Emergent, Vapi and Retell. One of them says it is profitable.

So, burning or earning?

My honest read: mostly burning, and mostly fine with it, for now.

Investors are paying for speed and betting that model prices will keep falling, which so far they have. If that holds, today's thin margins get fatter on their own. If it does not, a lot of these companies will need a buyer or a much higher price. The two that clearly earn money did it the old way: small team, narrow job, a price that covers the cost.

The fourteen "Not disclosed" rows are the real story. Many may be fine. But when a company shares its revenue every few months and never its margin, it is usually because the margin is the weaker number.

What this means if you run a business in India

You are probably not raising $400 million this year. But you are buying from these companies, or from smaller ones like them. A few things I would take from this:

1. Expect AI tool prices to move. When a vendor's cost per customer can swing from plus 36% to minus 14% in a year, its prices and limits will change too. Do not lock your whole operation into a tool on its launch pricing.
2. Keep your data portable. Yupp and Huxe were funded by serious people and still closed within months. Before you put leads, chats or customer records into any AI tool, check that you can export them in one click.
3. Check the claims your vendors make. "Ten thousand businesses use us" is the startup version of "$7M ARR". Ask for one customer you can call.
4. Small and profitable is a real option. Gamma and Retell got to $100 million without giant rounds. If you are adding AI to your own business, start with one job that customers already pay for, and price it so the model bill is covered.
5. Look at open alternatives where they fit. I compared a few in five open source AI tools that can replace what you pay for. Not every job needs the most expensive model.

Where this leaves your business

Most business owners I talk to do not care whether a tool is a wrapper. They care whether the phone gets answered and the lead gets a call back the same day.

That is what we build at Brand Bakery. An AI employee, such as an AI sales executive that calls new enquiries and books visits, runs on the same foundation models as the companies above. So we do the same maths they do: which model, what it costs per call, and what happens if the price changes next quarter. If you want to see what that would look like for your business, talk to us.

I will update this tracker as new numbers come out. If you spot a figure that has moved, or a company that belongs here, tell me and send the source.