🕶️ Air Force can only buy these balloons

Good morning, innovators. 🕶️
Hope your coffee is stronger than your Q4 pipeline.

The US Air Force just told a balloon company it's the only one able to do the job. Meta built its strongest model yet and kept the best version inside its own app. And in Europe, roughly 52 bodies can certify a medical device, but only one, by its own count, specializes in AI. It just raised $50M. Three different businesses, one shared idea: scarcity is the product. Models get copied, features get cloned, prices get matched. What does your company own that nobody can buy somewhere else?

In today's newsletter, we'll get into:

Every AI vendor reports a resolution rate. No two calculate it the same way, and none of them tell you whether the replies were actually good.

Running Agents in Customer Work is four 30-minute sessions for the leaders who have to make agentic AI work in support, success, and ops. Session 4 gets into how to score quality at volume without reading every transcript, and the questions to ask a vendor about what their resolution math leaves out.

Four Tuesdays, Oct. 13 through Nov. 3, 10 a.m. PT. Register today, one sign-up covers all four.

Register Now

In 2015, a few students in the Stanford Student Space Initiative had a problem: a standard weather balloon flies for about two hours, then pops. So they kept bolting things onto one.

Eleven years later, that club project is WindBorne Systems. It flies about 700 balloons around the planet, and the US Air Force just signed a contract saying nobody else can do the job. No bidding. CTO Andrey Sushko called it "a determination by the government" that "we were the only company able to provide it."

Run the tape. By spring 2019 the team had logged 26 test flights and four world records, per early backer Pear VC. That summer it found the market: by the founders' count, about 85% of Earth's surface isn't adequately monitored for weather. In 2023 came a $6M seed, after the CEO admitted the original cost edge "honestly wasn't" enough, so they made the platform 150x cheaper (his figure). August 2026: a $37M Series B and the first unpowered flight into a typhoon's eye. This summer: WeatherMesh 6, an AI model WindBorne says out-forecasts Europe's ECMWF.

Notice the order. Balloons first. The AI came later, trained on data only WindBorne collects.

Smaller than the headline. The contract is worth up to $49M, but the Air Force pays per order, and so far it has ordered $1.7M, about 3.5% of the ceiling. A ceiling is not revenue. Your investors know the difference, even if your press release doesn't.

Anyone can rent a transformer. Nobody can rent eleven years of balloon flights over the ocean. A moat built on a model lasts until the next open-weights release. A moat built on data you physically collect gets the words "only company able" written into a government contract.

Watch the order book. If the Air Force moves well past $1.7M, the sole-source bet is paying out.

In 2021, there were nearly 35,000 rounds of funding globally, with 20,000 of them being Angel or Seed rounds.

Despite a dip in angel investments during the pandemic, venture capital funding surged — not just in number, but in size. In fact, over $600 billion was raised in 2021, doubling the amount raised the previous year.

However, securing funding is still a rare achievement. In the US alone, more than 5 million new businesses were registered in 2021, but fewer than 4,500 early-stage companies managed to raise venture capital. That's just 0.08%.

This highlights how many companies aren't seen as "venture-backable" — which is one reason we pivoted our product to cater to all startups, not just those backed by venture capital.

Meta finally shipped a frontier model. Then it kept the better half for itself.

Muse Wonder Full, the bigger of two new models from Meta's Superintelligence Lab, won't be served through the API. The only way in is a Muse Code subscription, at $50 or $200 a month, per CNBC. Developers get the smaller Muse Wonder at $4 per million input tokens and $20 per million output.

Depends who's talking. Meta AI chief Alexandr Wang told CNBC the API would be priced "meaningfully below" OpenAI and Anthropic. TechCrunch read the same rate card as roughly equal to GPT-6 Sol or Opus 5.5. Meta's model card says Wonder scores comparably to Fable 5.1 and GPT-6 Astra, and beats them on MemoryBench, a memory test Meta chose to spotlight. Meta's numbers, Meta's tests.

Because Meta's money is ads, and CNBC frames this launch as a test of whether people will pay for AI at all. Sell your best model by the token and you're one row in someone's model-router dropdown. Keep it behind a subscription and the subscription is the business. Think of it as a movie theater: the ticket gets you in, the popcorn pays the rent. Wonder Full is the popcorn.

Two things. First, there's a new price-competitive model to benchmark this week, so if inference eats your margin, run your evals. Second, the pattern. When a lab's best model only lives inside its own app, everyone on the API is renting the second tier from a company whose own coding agent may compete with them. Build so you can swap models in an afternoon.

Watch whether Wonder Full ever reaches the API. If it doesn't, the subscription isn't a phase. It's the strategy.

Scale your growth: Join names like HubSpot, Ahrefs, and Deel. Put your product in front of 85,000+ startup founders. Let's talk.