The Money Card

· Jai An

In Where We Are in AI (The Flop) I argued that the big open questions of the last three years had all broken in AI's favor — and that one of the five cards on the table was customers are paying. For years the sharpest bear case was simple: the models are impressive, but where's the revenue? Who actually pulls out a credit card?

This post is that card turned face up and read closely. Not the labs — everyone knows OpenAI, Anthropic, and Google are enormous. The interesting question is the layer beneath them: the independent companies built on top of, and around, frontier AI. How big are they, really?

The answer, as of mid-2026, is: much bigger than the skeptics priced in, and accelerating.

The top tier: every AI startup above $500M run rate

Here's @deedydas's snapshot from July 13, 2026 — every pure-play AI startup at or above a $500M revenue run rate, excluding the three biggest labs. A = company-reported, E = estimate, R = rumored/in talks, † = gross marketplace volume.

Company — What it does — Run rate — Valuation
Cursor (Anysphere) — AI code editor · being acquired by SpaceX — $4B — $60B
Mercor — Expert marketplace for AI training data — $2B † — $20B
Scale AI — Data labeling & evals · Meta owns 49% — $2B — $29B
Surge AI — RLHF & frontier data · bootstrapped — $1.4B — $25B
Together AI — Open-source model inference cloud — $1B — $8.3B
Fireworks AI — Production inference platform — $800M — $15B
Lambda — AI training cloud · GPU clusters — $760M — $9B
Baseten — Model inference & deployment — $600M — $13B
Replit — Vibe-coding platform — $525M — $9B
Crusoe — AI factories & GPU cloud — $500M — $30B
Cognition — AI software engineer · Devin + Windsurf — $500M — $26B
Perplexity — AI answer engine & browser — $500M — $22.6B
Kling AI — AI video · majority-owned by Kuaishou — $500M — $18B
Lovable — Text-to-app vibe coding — $500M — $13.2B
ElevenLabs — AI voice & audio models — $500M — $11B
Higgsfield — AI video creation for social — $500M — $5B
Manus — General-purpose AI agent — $500M — $2B
Lightning AI — AI dev cloud · merged with Voltage Park — >$500M — $2.5B
Midjourney — AI image generation · bootstrapped — $500M — —

19 companies. ~$18.3B of combined run-rate revenue. ~$319B of paper value. And remember what this list leaves out: the three frontier labs (each a multiple of this entire table), the public "neoclouds" (CoreWeave, Nebius, Cerebras), the chipmakers (Groq), the AI-adjacent giants (Databricks at $5.4B, Anduril ~$3B), and ByteDance's Seedance (~$2B). DeepSeek is a wildcard — undisclosed, estimated anywhere from $200M to $1B.

In other words, $18.3B is the long tail beneath the giants — and the tail, by itself, is already a serious industry.

The tier right below: $100M–$500M

The band underneath is thick, and this is where the next wave of breakouts is forming. Starred figures are estimates; most of these companies don't report.

$300–450M — Mistral (~$450M), Fal (~$400M), Fluidstack (~$400M, projected), Suno (~$300M, confirmed), Harvey (~$300M, up from $195M six months earlier), Glean ($300M), Turing (~$300M), Modal (~$300M).

$150–300M — Cohere (~$240M, growing >50% QoQ), RunPod (~$240M), Sierra (~$150–200M), Synthesia (~$150–200M), Runway (~$150–300M), Clay (~$150M).

$100–150M — Hugging Face (~$130M), Abridge ($100M+), Gamma ($100M+).

Likely in the band but too opaque to place: HeyGen, Black Forest Labs, Decagon, Genspark, StackBlitz (Bolt), Vast.ai, Nscale.

One footnote worth its own paragraph: gross vs. net

When I first drew up this band, someone asked why OpenRouter was missing. The answer is the single most important caveat in the whole exercise. OpenRouter is a marketplace, and which number you use throws it into a completely different tier:

Its two numbers land on either side of the $100–500M range, so it skips it entirely. That's the same trap Deedy flags with a † on Mercor, Surge, and Scale: count gross marketplace volume and you get a billion-dollar company; count the net take and it can be a tenth of that. Keep this in your back pocket for every AI revenue headline you read.

What the numbers actually tell us

Five patterns fall out of this data. They're the real payload of this post.

1. The demand question isn't just answered — it's accelerating. The bear case wasn't that AI revenue was zero; it was that it would plateau. It hasn't. Suno went $200M → $300M in about 90 days. Glean did the same in six months. Harvey went from $195M at the end of 2025 to ~$300M by May 2026. Cohere is compounding >50% per quarter. This is the "2026 is the year the doubt died" claim from The Flop, now with receipts: the second derivative is positive. Revenue isn't just growing, the growth rate is growing.

2. Today's cash is in the picks and shovels; tomorrow's bet is on the agents. Look at where the largest confirmed revenues sit: data (Mercor, Scale, Surge) and compute (Together, Fireworks, Lambda, Baseten, Crusoe, Lightning). Selling shovels to a gold rush pays first and pays reliably — that's an old pattern and it's playing out exactly. But now look at the multiples. The market is paying ~8x revenue for a compute business like Together ($1B → $8.3B) and ~52x for an agent business like Cognition ($500M → $26B). Investors are betting the durable, high-margin profits eventually land in the application and agent layer, even though the reliable cash today is in infrastructure. That tension — cash in the shovels, valuation on the agents — is the whole "who wins?" question from The Flop, priced in real time.

3. Code is AI's first killer app. Count the coding companies at or above half a billion: Cursor ($4B — the single largest on the entire list), Replit (~$525M), Cognition (~$500M), Lovable ($500M), plus Bolt/StackBlitz knocking on the door. No other single application category has this many breakout winners. When a new general-purpose technology arrives, the first place it pays off is usually the tool that builds more of the technology. Software eating software is not a someday story; it's the highest-revenue application of AI that exists right now.

4. A whole cohort hit escape velocity at the same time. Notice how many companies cluster at exactly "$500M" — Perplexity, Kling, Lovable, ElevenLabs, Higgsfield, Manus, Cognition, Crusoe, Midjourney. Some of that is rounding, but the signal underneath is real: a dozen companies crossed the half-billion threshold inside the same ~12-month window. This isn't one outlier running away from the pack. It's a generation of AI-native companies maturing in lockstep — which is what the early innings of a genuine platform shift look like, not a single-winner bubble.

5. The independents are already being bought. Cursor is being acquired by SpaceX. Meta owns 49% of Scale. Kuaishou controls Kling. Cognition absorbed Windsurf. Lightning merged with Voltage Park. Manus's Meta acquisition got unwound. The breakout winners aren't staying independent for long — the incumbents are spending their way into the layers that matter. Consolidation this early is a tell: the big players already believe these numbers, and they're not waiting for the multiples to get worse.

And one bonus, for the skeptics: you don't have to raise a fortune to win here. Surge is doing ~$1.4B and Midjourney ~$500M — both bootstrapped. In the right layer (data, image generation), AI economics can be good enough to fund themselves. The "AI is only a game for the mega-funded" story is already wrong.

Where that leaves us

The Flop argued that the foundational bets of the AI era had already paid off — the technology works, the capital showed up, and customers are paying. This is the third of those claims, audited. Nineteen independent companies over $500M, dozens more climbing behind them, growth rates that are themselves accelerating, and incumbents buying in before the window closes. The "will anyone pay?" card is not close. It's face up, and it's green.

What these numbers don't settle is the more interesting question — the turn-and-river question from The Flop: who ultimately keeps the profit? Right now the cash is in infrastructure and the valuations are on agents. Those two facts can't both stay true forever. The next few years are the market finding out which one was right.

The money card is on the table. The rest of the hand is about who gets to rake the pot.