Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Kulipa, a startup making infrastructure for crypto cards, has shut down — learn more below.

• What nobody tells you about writing agent skills

• AI has solved a lot of math conjectures in the last 2 weeks — learn why this matters below

This Week In Startups

🔗 Resources

Code Review (As We Know It) Must Die

Adding ambient AI: How to put AI to work where nobody is looking for it

The Endgame Of Vertical Integration

What nobody tells you about writing agent skills

📰 News

ChatGPT brings unlimited text chats to free users

Meta launches Muse Code, an AI agent for large code bases

Apple finally fixed Siri

OpenAI’s new AI smart speaker will reportedly sell for between $300 and $400

💸 Fundraising

Array Labs, a satellite radar startup, raises ⁠$21M strategic funding

ProphetX, a sports prediction market, raises ⁠$35M funding round

Mirae, an AI platform for continuous care of autoimmune diseases, raises ⁠$5.4M funding

Biota, an environmental diagnostics startup, raises ⁠$3M seed round

Fail(St)ory

Crypto Mastercard

Most crypto payments still hit the same wall: you can’t buy groceries, pay for dinner, or book a flight with a wallet full of stablecoins.

Kulipa was trying to solve that by building the infrastructure to let crypto wallets launch their own payment cards.

It worked fine until the cards stopped working.

What Was Kulipa:

Imagine you’re a crypto wallet. Your users already keep USDC in your app. They don’t want to cash out every time they buy coffee, and they don’t care about blockchain rails when they’re standing at a checkout. They just want to tap a card and move on.

That’s exactly what Kulipa made possible.

Instead of asking merchants to accept crypto, Kulipa went the other way. It connected stablecoins to the payment networks people already use. The customer paid with a normal Mastercard, the store received a normal card payment, and Kulipa handled everything happening in the middle.

For wallet companies, that removed an enormous amount of work. Launching a card program usually means finding licensed banking partners, passing compliance checks, integrating fraud systems, connecting to card processors, and figuring out how to move money between crypto and traditional finance. Kulipa wrapped all of that into a single API.

Its biggest selling point was that users didn’t have to preload a card account. Their USDC stayed in their own wallet until they actually bought something. At the moment of payment, Kulipa moved just enough funds to settle the transaction. The wallet stayed self-custodial, but spending felt like using any other debit card.

That was a pretty compelling pitch for crypto wallets. They could offer a feature people immediately understood without becoming payments companies themselves.

By April 2026, Kulipa said it had signed around 20 customers, issued more than 120,000 cards, and was seeing 70% month-over-month transaction growth. Partners ranged from crypto wallets to fintechs serving people who preferred saving in stablecoins instead of local currencies.

Then one of the licensed issuers behind the cards ran into trouble, and the whole setup unraveled fast. Within weeks, users lost access and partners were scrambling for a replacement.

The Numbers:

• 💰 Funding: ~$9.2M raised

• 💳 Cards issued: 120,000+

• 🤝 Customers: ~20 wallets and fintechs

• 📈 Reported growth: 70% month-over-month transaction volume

• 💵 Peak monthly volume: ~$9M (January 2026)

• 📉 Card services stopped: July 29, 2026

Reasons for Failure: 

• One issuer became a single point of failure: Kulipa depended on regulated issuers to connect its software to Mastercard. After one key issuer reportedly ran into regulatory trouble, Kulipa had to move programs to another partner with more limited coverage. Within weeks, cards started disappearing in some markets before the entire service stopped.

• The product relied on too many moving parts: Kulipa looked like a software company, but every payment depended on banks, card networks, compliance providers, processors, and regulators working together. Each layer added another dependency outside the company’s control. That makes the business much harder to operate than a typical API startup.

• Customers outsourced a critical feature: Wallets loved Kulipa because it let them launch cards without becoming payments companies. The tradeoff was that one infrastructure provider now sat between them and their users. When Kulipa went down, products from around 20 different companies went down with it.

• Growth came before resilience: Kulipa moved quickly, reaching more than 120,000 issued cards and around 20 customers in a short time. That’s impressive traction, but payment infrastructure earns trust through years of reliability, not months of growth. Once partners started looking for replacement providers, that trust was difficult to recover.

Why It Matters: 

• Map every regulated dependency before scaling; one weak partner can shut down the whole product.

• When customers build on your infrastructure, reliability matters more than feature velocity.

Trend

AI Math

AI has had a ridiculous two weeks in mathematics.

Anthropic helped disprove an 87-year-old conjecture. OpenAI answered by publishing ten new results on open problems in mathematics and theoretical computer science. These weren’t benchmark questions or Olympiad exercises. They were problems that mathematicians genuinely didn’t know how to solve. 

This isn’t the first time AI has contributed to original math. The difference is the tempo. One breakthrough felt exceptional. A string of them, from multiple labs, starts to look like something else.

Why it Matters

• Frontier AI is getting measured on original research. Benchmark scores don’t move the conversation like they used to. New discoveries do.

• The cost of exploring ideas is falling fast. AI can search through thousands of possible directions. Humans spend more time deciding which ones are actually right.

• Mathematics is probably just the first proving ground. It’s a field where new results are public, verifiable, and hard to fake.

What happened?

The biggest story came from Anthropic.

On July 20, mathematician Levent Alpöge casually posted on X that, while watching the World Cup final, Claude had helped find a counterexample to the Jacobian conjecture, an 87-year-old open problem. The entire result fit inside a tweet. Within hours, mathematicians started checking it. It held up.  

A few days later, OpenAI published ten new results across mathematics and theoretical computer science, generated by an internal model called Astra. Instead of presenting one headline result, it released a collection spanning graph theory, geometry, coding theory, complexity, and more. 

Neither company was working in isolation. AI had already been making steady progress on open math problems throughout 2026. What’s different is that the announcements are arriving much closer together, and they’re increasingly being used as public demonstrations of what frontier models can do.

It’s hard not to see some competition here. Once one lab shows its model can produce original mathematics, everyone else suddenly has an incentive to show they can too.

The Trend

Math is becoming the new benchmark.

Coding benchmarks are crowded. Leaderboards move by fractions of a point. Original research is much harder to dismiss. Either a result is new or it isn’t. Either other mathematicians can verify it or they can’t.

That’s why these announcements matter beyond mathematics.

If that capability keeps improving, mathematics won’t stay the main story for long. The same search process can be applied anywhere progress comes from exploring a huge space of possible ideas: chemistry, materials science, biology, economics, or algorithm design.

We’re still early. Human experts are doing most of the verification, and these results need time to be absorbed by the research community. But the last few weeks made one thing much clearer: frontier AI companies have started competing on scientific discovery, and mathematics is where they’re choosing to keep score. 

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico