Paris VivaTech take-aways + the real reason Builder.ai imploded — and how to dodge the same fate.
TL;DR for busy friends
- VivaTech 2025 felt like a candy store of “AI-powered” demos — yet one former unicorn, Builder.ai, has just crumpled under $85 m in unpaid AWS bills and a $37 m cash grab by creditors.
- The wipe-out proves that “fake-it-till-you-make-it” doesn’t fly once due-diligence kicks in — even amid peak hype.
- Meanwhile, real (but incremental) progress keeps rolling: Claude 4 (200 k-token context), Google Gemini 2.5 Pro preview, bargain-priced DeepSeek R1-0528 at ≈$1/M tokens, and Sakana AI’s self-rewriting Darwin-Gödel Machine.
- Take-aways:
- Don’t build your “billion-dollar startup” on a spreadsheet of lies.
- The AI gold-rush is real, but competition is brutal; either solve a hairy problem fast, master distribution, or pivot.
- Please build or fund a “Trust AI” startup
Dispatches from VivaTech 2025
Paris felt like CES on espresso this week: a record 150 k+ attendees milled through the halls.
- NVIDIA drew wall-to-wall crowds as Jensen Huang evangelised “sovereign AI” for every nation-state.
- AWS dangled $200 k in cloud credits for climate-tech founders hunting GPU time.
Great energy, great people — everyone proudly showing their tech babies.
(Sorry I took no pictures…)
Yet after dozens of booth chats with French-Tech friends I kept thinking:
“Your demo rocks… but will I still see you here next year?”
Too many stalls hid a thin wrapper: Claude API call, MCP servers, wired to a prompt-bitten UI, or an ultra-niche agent that wows but (likely) won’t scale.
Meanwhile almost no one tackled the elephants of AI trust, transparency, and reliability — a gaping market white-space. To be honest — I'm surprised very few players are tackling this problem.
I love the passion. But my gut says a chunk (most) of these products won't survive the next twelve months.
Cool demos—sure. Survivability? Less clear. The memory of one booth that wasn’t there this year made that contrast painfully obvious.
Case Study: The Rise-and-Rubble of Builder.ai
Builder.ai positioned itself as the no-code dream.
“We make building an app so easy even a caveman can do it.”
Their product, Builder Studio, claimed to let anyone build software just by prompting it—like ChatGPT but for apps.
No tech skills? No problem.
That was the pitch.
Except… it was a lie.
AI on the Outside, Humans on the Inside
Behind the sleek UI and “AI magic”…
Was a room.
A room filled with underpaid programmers in India, quietly rewriting broken AI outputs into actual working code.
AI wasn’t coding your app.
Humans were. Quietly. Painfully.
It was human labor hidden behind a chatbot.
Vibe coding—that was their term. A term so vague, it meant everything and nothing.
But That’s Not What Killed Them
Yes, the AI didn’t work.
Yes, the humans were doing the heavy lifting.
But the real reason for the collapse?
Fraud.
Builder.ai allegedly faked revenue through a roundtrip billing scheme.
In simple terms:
- They billed partners for work they never did
- The partners sent the money back
- On paper: huge revenue
- In reality: smoke and mirrors
Until one of their creditors called the bluff.
And seized $37 million from their accounts.
That was the end.
From $1.5B to $0: A Timeline
| Year | Milestone |
|---|---|
| 2023 | Valuation peaks at $1.5 B |
| Feb 2025 | New CEO Manpreet Ratia parachuted in [source] |
| May 20 2025 | Staff call: “We’re filing insolvency.” [source] |
| May 24 | Bankruptcy protection filed in 5 jurisdictions [source] |
| June 6 | Media autopsy headlines “largest AI scandal to date” [source] |
Meanwhile in AI-Land
| Model | What’s new | Worth caring? |
|---|---|---|
| Claude 4 (Opus & Sonnet) | 72% SWE-bench, 200k-token context, steady pricing | Better long-haul reasoning; not a quantum leap. |
| DeepSeek R1-0528 | Open-source, $1 per M tokens, single-GPU variant | Cheap tinkering for indie hackers. |
| Darwin-Gödel Machine | Self-modifying coding agent that rewrites itself via evolutionary search | Early days, but hints at AGI’s “recursive self-improvement” chapter. |
Business Lessons from the Collapse
What Builder.ai teaches us:
- AI hype sells, even when the tech doesn’t work.
- The AI landscape is hyper competitive; be the best or die.
- Manual work doesn’t scale—especially if you’re pretending it’s AI.
- Fraud is not a growth strategy.
If your AI startup can’t scale without hiding humans behind the scenes, you’re not building tech. You’re building a trap.
So What Should You Actually Do With AI?
Here’s the truth:
The best use of AI isn’t to “build your startup.”
It’s to LEARN.
Understand how these models and systems work under the hood—and learn how to do what you're already good at ×100 better. That's your edge.
Because whether the bubble pops like in 2001 or inflates into a singularity…
You win by learning what’s real and what’s fluff.
Final Note
VivaTech reminded me: the line between a breakthrough and a booth-sized mirage is thinner than ever. Ask every founder (or yourself):
“Is this a real AI innovation—just a couple of workflows, or a roomful of devs on coffee IVs?”
Stay curious, stack real value, and never let your cap table write cheques your ledger can’t cash.
Until next time,
— Charafeddine