There is a classic result in industrial economics (Sutton, 1991): in industries where the cost of keeping up scales with the market itself, market concentration has a floor, and growth ends up entrenching the leader. Launch could become a textbook case. Reuse and cadence are an escalating bar, and Starship is many years, potentially a decade at scale, ahead… Putting out a launch market model this week 🧐

Falcon 9 has been priced at ~$70M for many years now, leading many to believe market prices have stayed flat. Heres what it looks like adjusted for inflation. F9 customer launch margin has increased from ~breakeven in 2013... ... to todays prices exceeding internal cost by, what we estimate to be 3.3x. Thats ~70% margin. (green line is the internal cost of Starlink, which is lower because they get far closer to filling the payload capacity!) Will be interesting to map these dynamics onto market prices with Starship around the corner....

Just released our AI Compute 'Keystone' Model: a market-clearing engine for AI compute, 2026 to 2040, with uncertainty built in - simulated across thousands of assumption permutations. It brings together nine months of our terrestrial and orbital compute work into one framework, built to test any assumption against a very uncertain AI future. Whilst building this, three takeaways have stuck with me particularly: 1. Tools, financial modeling and other, are changing fast. Elon replied to a post last week by an engineer who recalled AI seems akin to when they stopped reading assembly once they trusted the compiler.... "source code is on the verge of becoming like assembly". This describes exactly what I've felt is happening to financial modeling too. Traditional Excel simplified by necessity, because complexity buried in cells is hard to audit practically. LLM + code inverts that: richer architecture, live links to external data, tests built into every iteration. Our 46-tab Excel workbook and our Python engine compute everything twice; across 513 tracked outputs they agree to 2.11×10⁻¹⁴. Complexity used to cost audit-ability and Im starting to believe that paradigm is flipping... interaction with complex systems/models is being revolutionized by LLMs. This is the first model we're encouraging our clients to engage with via LLM, because the learning velocity is so much larger. 2. Its ironic that today orbital datacenter compute is considered a niche...because the model suggests it will be a necessary condition for AI continuing to scale in the 2030s. Our demand assumptions sit below the leading labs' claims, and orbital still ends up carrying the market. Terrestrial capacity peaks at 111 GW in 2031. By 2040 the fleet (excluding training) runs ~488 GW....~438 GW of it orbital (~90%). People treat orbital data centers as a bonus to supply. In our model, they are what keeps it scaling at attractive economics once the terrestrial grid can't. 3. Demand and supply are one system. Every one of the 5,000 futures is the same ebb and flow: demand multiplying against technology deflating. Cheaper compute recruits new demand....new demand funds capacity and faster chips...faster chips make compute cheaper still. If chips stop improving, prices stop deflating, and realized demand falls. The deflation and adoption equilibrium is a growth engine. The deflation and adoption equilibrium is a growth engine. This is why compute supply can grow 180x through 2040, whilst compute market revenue only grows ~7x to ~$3.2T a year... 96% of the volume growth is given back in price. No one has a narrow projection for a technology this unprecedented, us included. What Im proud to say we have is a framework built to evolve that will absorb history as it arrives and will narrow the space of futures that remain. We've published a public brief on the high-level takeaways from the model here 🧐: