Helsing's AI took full control of a Saab Gripen over the Baltic Sea and flew three air combat scenarios against a second Gripen flown by a human. Torsten Reil (@torstenreil), co-founder and CEO of the European defense company Helsing, says the system is called Centaur. It is trained with reinforcement learning inside a simulator that runs up to 25,000 times real time. The progression was their own simulation first, then an official simulator from one of the big defense primes, then a real jet last year. They now fly regularly against Eurofighter, Rafale, and Gripen pilots. Centaur already works in swarms, not just 2v2. Reil says that is why crewed fighter jets get hard to deploy, because you end up facing swarms of autonomous ones. Key takeaways already in your email via @PodWireHQ Source: David Senra (@FoundersPodcast)

Rocket Lab $RKLB put its first rocket into orbit on under $100M and 80 people, while Virgin Orbit took $1.2B from Richard Branson and went bankrupt. Sir Peter Beck (@Peter_J_Beck), the founder and CEO of Rocket Lab, says the startups he has watched fail most spectacularly are the ones that were funded the most. He grades his own company on what it achieves against the capital it spends, not on what it raised. Even as a public company Rocket Lab has never been the best financed name in its market, and Beck says everything still gets ground right down to the very last bit. Key takeaways already in your email via @PodWireHQ Source: Sourcery with @MollySOShea

Anish Acharya (@illscience) says Thomson Reuters $TRI and other big legal names traded down over an Anthropic legal plugin that was really just a zip of prompts. The a16z General Partner, who invests in consumer, points out that plugins are collections of skill files and skill files are just long prompts. The live debate was whether the frontier labs would vertically integrate up into the application layer. They have done the opposite and integrated down into inference and compute. Inference workloads are homogeneous, so a lab can build enormous scale in one part of the value chain. The application layer is thousands of combinations of pricing, packaging, and how each market wants to buy. Moving up is the opex heavy direction. Key takeaways already in your email via @PodWireHQ Source: The a16z Show with @jkhamehl

AI inference grows about 7x a year, but the content licensing deals publishers signed with the labs are fixed fee, so 50x more usage pays them exactly the same. Parag Agrawal (@paraga), the former Twitter CEO now founder and CEO of Parallel Web Systems, says even the publishers who landed a deal are on a broken model. Inference grows roughly 7x this year and 7x again next year. The deal size does not move with it. "None of them after signing a 2-year deal believes that their share isn't going to decline materially at renewal." And that was the only option on the table. Sitting in the head of the distribution was the one business model available, and most content on the web never got the offer at all. Key takeaways already in your email via @PodWireHQ Source: Training Data with @sonyatweetybird and Andrew Reed

A data center running at 95% uptime has basically zero buyers, and Neil Movva (@neilmovva) says he is that first buyer. The ex-Nvidia engineer, who co-founded Sail Research to sell the cheapest tokens on the market, gets there by cutting what everyone else pays for. No backup diesel generators. No redundant networking. One line of fiber instead of three, and no SLAs. He can live with the outages because his customers run agents in the background for hours. When a request dies, the control plane moves it to another GPU and that agent loses a minute or ten. The customer is asleep. Average throughput stays competitive, 99th percentile latency does not, and in return he gives them economics nobody else will match. Key takeaways already in your email via @PodWireHQ Source: Invest Like the Best with @patrick_oshag

Martin Casado (@martin_casado) says 20 people can now deploy $1B usefully, which turns software from an engineering-bound business into a capital-bound one. The a16z General Partner ran the same thought experiment across three eras. Give a 10-person startup $1B 20 years ago and it went into buying your own computers. Give it to them 10 years ago and you hired engineers, where the mythical man-month is very real and headcount stops turning into output. Engineering complexity was always the quiet limiter on how much money a small team could absorb. Casado says that limiter is gone and that we have never been here before. Steven Sinofsky (@stevesi), who ran Windows at Microsoft and is now a Board Partner at a16z, supplied the history. Computing was capital-bound for its first 30 to 40 years, then engineering-bound, and Casado's read is that we are capital-bound again. Key takeaways already in your email via @PodWireHQ Source: The a16z Show with @martin_casado and @eriktorenberg

Anthropic is now pulling as much as $50M of revenue per megawatt of compute that costs it $10M to $15M. Dylan Patel (@dylan522p), the founder of SemiAnalysis, says that spread is what flipped the labs out of venture-funded losses. Anthropic turned a profit in Q2. OpenAI could get there in Q3. A year ago the math ran backwards. Serving GPT-4 on Nvidia Hopper GPUs generated negative gross margin for OpenAI. Serving GPT-5.6, Opus 5, and Fable 5 clears the incremental cost several times over. "If I spend 10 bucks on inference capacity, I actually generate 50 bucks of revenue, and then I can turn around and incrementally spend all of that profit on training." Key takeaways already in your email via @PodWireHQ Source: Dwarkesh Podcast with @dwarkesh_sp

Recursive self-improvement is already running at OpenAI, just not where people keep looking for it. Tibo (@thsottiaux) says the models are rewriting the stack that serves the models. He leads Codex there. The version everyone argues about, models designing other models, is the research framing. What is actually working is narrower. They point the frontier models at the infrastructure on the critical path of running those same models. The inference stack, the hardware, the CUDA kernels. In his words it is all one big system, and you take that and point it back at itself. Asked if OpenAI is doing this on purpose, he said if we were not doing that, that would be pretty silly. Key takeaways already in your email via @PodWireHQ Source: Matthew Berman (@MatthewBerman)

Rocket Lab's $800M defense contract turned into more than $1B because the two rival primes that won alongside it had to buy Rocket Lab's components. Sir Peter Beck, the founder and CEO of Rocket Lab $RKLB, says national security has been an underserved market because the typical contract goes to a traditional prime on cost plus, which blows out on time and on money. Rocket Lab bids firm fixed price, and it is prime on two national security programs right now. The structural edge is that it supplies the companies it bids against. Reaction wheels, solar, and the rest of the parts inside a satellite. Lose the bid and the purchase orders land the next day anyway. "Even if we lose, we still win. And when we win, we win twice." Key takeaways already in your email via @PodWireHQ Source: Sourcery with @MollySOShea

Baiju Bhatt (@BaijuBhatt) is building data centers in orbit with no satellite, because the upper stage of the rocket itself becomes the data center once it gets to space. The Robinhood co-founder says the hardest part of running compute in orbit is not making power, it is dumping heat. The usual design carts mass and volume up just to unroll large sheets of radiator. So he stopped treating the rocket as something you throw away or fly back. It is already a big piece of metal, so it becomes the heat sink. "There's no satellite. The rocket becomes the data center satellite when it's in orbit." He priced a GPU hour from space from the bottom up, panel by panel and radiator by radiator, and says the integrated design gives Cowboy Space one of the lowest cost ways to do compute from orbit. Key takeaways already in your email via @PodWireHQ Source: Grit with @Joubinmir