we just published a deep dive reaction to the SpaceX earnings from yesterday. Specifically comparing results against street expectations and our mach33 internal model. you can see the full analysis here. A couple core things stand out on the financial side. First, wall street expectations beat across the board. I believe the cope is that consensus made it too easy for them to beat, but the fact remains. Second, our model, which is widely considered among the most bullish outside of spacex, was dramatically conservative on a few core things. AI Compute for sure, and enterprise mix for starlink as well. note: our Starlink numbers are the most bullish in our models, due to our conviction on the Starlink V3 inflection. This earnings call confirmed our annual revenue numbers but our enterprise/consumer mix was off, which suggests ARPU decline will be slower than we model. On the technical side which was my personal favorite, i found these to be the most compelling... V3 Sats going to operational orbit on Flight 14 - I was publicly wrong on this. assuming they'd do at least one more test. They are going for it sooner than expected. 1,000 Sats by Q2 2027 - this implies about a 1.7 flight per month average by that time. Our global bandwidth suggests they'd be aiming for 4,000 sats by EoY 2027 or 45 more flight in H2 2027. This might be on the bullish side. Better enterprise mix would compensate for a miss on this cadence. Mini-towers/Femtocell hardware - this is a big reveal and not enough people fully grok the implication. We've long hypothesized that their terminal/sat/gateway manufacturing machine would be well positioned to cover urban density gaps with a creative solution. Classic SpaceX first principles. This aligns with our modeling thesis particularly in the 2028-2030 growth curves. Starmind Co-designed NVL72 rack (both earth and space) aligns closely with our findings on co-designing models with hardware and potentially leveraging wideEP designs (Nvidia published work on this) to maximize token efficiency. This would be extreme co-design for inference. Eager to see more on this. This also pulls our orbital data center thesis forward, as a significant portion of our modeling suggested needing co-designed chips. This is co-designing a rack with off the shelf chips which will likely increase the revenue generation capability of the starmind rack, which is an ROI lever. So racks may cost more token for token than ground but the ROI will be break-even sooner, and learning rates and nimbyism will do the rest for the crossover. makes Starmind sats economically viable sooner. Grok 5 getting engineering data is going to play a role in model quality and ultimately the data bottleneck which we've modeled as something to be more concerned about in the 10T-20T parameter model era. All in all, i'm biased but this call was incredibly insight dense. I'm looking forward to the next one.