SPCX Q3 BEAT, 13 GW, DUAL CATCH | ep25 xitter.cf/i/broadcasts/1wxWjlNPP…

out of our state of connectivity deep dive, i've been consistently coming up against the question, What does explosive agentic growth do to connectivity infrastructure? to answer this i built a scenario explorer for when agentic upload would potentially overwhelm global bandwidth region by region between now and 2035. i've open sourced this tool here for anyone to explore: mach33-agent-bandwidth.verce… First screenshot is overwhelmed cells (~250km) 2030, second is 2035 at stress test case, which isn't all that crazy. what becomes obvious is that without substantial upload bandwidth certain regions particularly those without robust fixed broadband (regions overly reliant upon mobile upload) will likely start hitting limits during peak demand. potentially over 1B households by 2035. The default settings are a stress test. the mid range settings are more anchored to what the data supports today. worth pointing out that history suggest exponential forecasts undershoot, so I picked a number that forces the question but it's definitely NOT a ceiling. This also doesn't take Starlink (which is a massive upload bandwidth upgrade) into account. This is designed to show you how existing infrastructure will fare. However I do think it shows you what teams at SpaceX, Amazon and Blue Origin are seeing and building for and what a lot of traditional telcos are not building toward. I think it's worth noting that while the Mach33 team scraped, built and cleaned the underlying data set (for global household density, fixed and mobile infrastrtucture by cell) off of 3,125 quoted facts from 667 sources and 210 publishers over a month... the vercel app above was built on top of that dataset completely with HTML and deployed on vercel within a few hours completely with Grok Bot. you just need the right data, the right question and grok bot to unlock incredible alpha.

had a lot of fun (screenshot proof) chatting SpaceX with Adam Jonas a bit ago on a special version of rowdy robot. Its worth noting that I've been critical of wall street analysis in the past around SpaceX, but Adam and team are one of the rare groups who've shifted a lot of due focus onto the the technology and physics behind the business, and it shows. here are a couple fun pieces of the transcript if you weren't able to get the link directly. ON ORBITAL DATA CENTERS (STARMIND)... "Adam: So let's shift to data centers in space now, orbital data centers… You seem pretty geeked up on it. Tell us why and what kind of pushback are you getting? What do you think the market's missing? […] Aaron: People started coming out with this like, oh, physics doesn't work… Adam: You can't cool things in space, so why are we talking about this? Aaron: We had to kind of have some kind of independent voice on that that said, here, look at the math… Starlink… the satellite connectivity, represents an important thing because it's one of the big things where you can move bits into space and move bits back… 80, 90% of all the mass up there is basically something to do with this now… What I view as the orbital data center is now you're moving a bit up into space, and you're making it more valuable and moving that bit back. Adam: You're not just moving it from one place to the other, you're improving it. Aaron: You're improving it." ON SPACEX VS TESLA MANUFACTURING CHOPS... "Aaron: Their manufacturing capability at SpaceX is quite big. It was actually the one thing on your recent note that I brought up… SpaceX, you could argue, may have… Adam: Totally fair play. A hundred percent agree. We were describing Tesla as like the manufacturing layer… but there is a manufacturing layer there. And it's pretty gnarly. Aaron: Millions of terminals, thousands of satellites, hundreds of rockets soon. Adam: It's gnarly and it's scalable… it's fair play." ON MARS... "Aaron: I personally plan on going to Mars someday, like, that's my thing… Adam: You would do that? Do you know how long the journey is? Aaron: Yeah… I'm also investing in things I think that can help to reduce that. Adam: Like propulsion or life support or what? Aaron: A variety of things. Propulsion is one of them… But realistic, like, I legitimately want a Martian ranch… Adam: Well, I've seen your screensaver or your background is insane. Aaron: I want to be optimistic about it. It sounds a little nuts. But like I think that optimism starting to creep back in. Adam: Look… some people will laugh hearing you say that… I like to tell clients, buy the laughter, sell the silence."

solo live here on X in 10 mins. discussing orbital compute news, and little analysis on the state of global connectivity.

a very small group of people were paying attention to this 250kw VR72 starmind comment. Its incredibly interesting for those following the orbital compute and NVL72 rack operating conditions. we did a deep dive engineering analysis to try to explore what that target could mean for the future of starmind. this weeks public brief discuss some of those learnings. research.33fg.com/analysis/t… we had to build gpu renders to help explain the potential implications of running VR72 starminds at 250kw average. they are pretty cool. i highly suggest staying up to date on the starmind engineering reveals. this is one of the more consequential infrastructural engineering challenges of our lifetimes. understanding this stuff (even at a cursory level) gives a lot of insight into orbital infrastructure decisions across the board.

Live again in 20 mins. come join us here or @mach33 account.

speaking of 4.8, after learning about the spacex and starlink team's progress on pre-training and post training custom software rewrite, we revisited our Grok training model in an attempt to get a rough timeline on how long we think it will take 4.8 release. in today's brief we took a stab at estimating that timeline and the model suggested a Grok 4.8 releases in November. This is longer than our original estimates due to third party RL software stack, and larger model size. this is part of our team diving deeper into the engineering signal around AI, frontier models, and the inference hardware behind it. All to help us better understand the starmind and space inference. please don't crucify us in the likely scenario we're wrong but please do heap on the praise if we're right. Full logic open to the public here: research.33fg.com/analysis/h…

Great Moonshots podcast discussing Starship Flight 14 and impending Starlink bandwidth explosion yesterday. Particularly this thought provoking clip. For context the 150 launches is over 2 years not one, with the assumption that it ramps smoothly from once every 1.5 months today to launching about once every 3 days by the end of 2028. What makes that conservative is Elon says internal targets are closer to 1-2 launches per week by end of 2027. and our models suggest ~2.5 launches a week by end of 2028. In addition on the referenced chart we actually held about 10-20% of those launches for other uses, which leaves about ~1,000 Tbps (1,000 sats) on the table as well. Beyond that Alex makes the case that we need more payloads for all the upmass coming, and suggest Orbital compute may be filling that hole. Its clearly a bit of a chicken and egg debate, but our analysis has suggested there at least 8 markets unlocked by the dropping launch costs coming from Starship. Orbital compute is just the first most economically compelling (link to full analysis in comments- making this one public for a bit) Thanks for the shoutout Peter! and link to full podcast in comment.

To put V3 Starlink significance into context for investors, here is a starlink bandwidth estimate based on our conservative launch forecasts out to 2028. Our base case launch forecast suggests 8 Pbps of bandwidth (8,000 Terabits per second) in orbit by 2028, which only uses 151 Starship launches (1 launch every ~2 days, Elon and SpaceX are targeting 8k+ a year) 8 Pbps is enough bandwidth to more than 10x all of Starlink's bandwidth today, and put Starlink's global 2028 bandwidth supply about equal to ALL of the global internet traffic bandwith demand in 2025.

250kw Starlinks??? $/Tbps Starship | Episode #21 xitter.cf/i/broadcasts/1XxygwYbr…