FLIGHT 14 TARGETS END OF AUGUST First Starship to put operational Starlink V3 into orbit + attempt a ship catch. That single flight advances the entire stack: • Real payload delivery (not just tests) • Higher-load reusability data • The mass cadence required for 100 GW+ orbital solar AI compute From the earnings call: Capex is already directed at AI + Starship. Short-term numbers can absorb hits. The 5–10+ year objective remains Moon and Mars bases. This is how “imagine Mars” becomes engineering reality.

Yup, that’s one perspective we don't talk about enough. The heat shield isn’t just ”solved” because the tiles stayed on. The real signal, in my opinion, is that flight telemetry, high-resolution imagery, and post-splashdown visual inspection now line up so tightly that SpaceX can treat the design as a known quantity. That closes the experimental loop. From here the remaining work shifts from material physics to process control, robotic inspection, and high-rate manufacturing. That’s the quiet transition from ”can it survive one more flight?” to ”how fast can we build a fleet of them?” Hardware-rich iteration just paid off again😊 Pic from: @niccruzpatane 🤝

🎉 Today marks one full year of being officially on 𝕏! I’m so incredibly grateful for this platform and the community I’ve been lucky enough to build (and keep building every day). To every subscriber, every follower, every connection I’ve made, and every future connection still waiting to happen thank you. You’ve made this journey special. Thank you to 𝕏 itself for giving us this space, to everyone who’s helped make it better, and a big thank you to Elon Musk for creating something this powerful. Here’s to year two and everything still ahead. 🚀❤️

🚨 SPACEXAI UPDATES FROM Q2 EARNINGS GROK: >Grok 4.5 was a huge improvement >Grok 4.6 dropping ~next week >Grok 4.7 in three or four weeks >Grok 5 before end of the year “We will be incorporating the ENTIRE CORPUS of SpaceX data… ALL the data SpaceX has ever produced over the course of a quarter century into Grok training… This will make Grok by far the BEST ENGINEER” COLOSSUS COMPUTE: >building data centers faster than anyone else >efficiency of compute deployment also the highest >2GW+ online by end of 2026 >by the end of next year… SEVERAL times higher >closer to 10GW than 5GW NVIDIA x SPACEX: >we’ve decided to build EXCLUSIVELY on NVIDIA >Vera Rubin architecture is the best architecture >“we greatly value a close cooperation and partnership on many levels with NVIDIA” CURSOR: >acquisition should close quite soon >through almost all of the regulatory hurdles STARMIND AI1 SATELLITES: >optimized Vera Rubin NVL72 in orbit >"we expect to start launching these next year" >a radical simplification of the normal NVL72 rack >will cost less, and be more effective >so deploying them on the ground AND in space SPACEX CFO: >robust demand for cloud services >increasingly favorable economics with each agreement we sign >supply-demand imbalance in the compute market will continue >less than ONE YEAR payback on new compute capex >Q3 already contracted $6.7B more cloud revenue >ramping starting in October “We believe this puts us on a trajectory, including contribution from Cursor, to reach 100 BILLION of ARR by the end of this year.” $SPCX

AI twelve months ago predates Claude 4.5. That model was a genuine shock to the system. Two years ago feels like a museum exhibit. You look at it and say: you’re not going to believe this was state of the art. At the current rate of improvement, by the end of next year it is not clear there is anything digital that AI will be unable to do. Source I Elon Musk

SpaceX’s first public earnings call is live and the numbers are insane 🔥 Q2 2026: • Revenue: $7.81B (+92% YoY) • Adjusted EBITDA: $3.53B (+191% YoY) • Net loss narrowed to $541M • Starlink subscribers: 12 million (doubled YoY) • $14.1B in new Cloud Services Agreements • Capex $18.3B (mostly AI) • $100B cash + $47.5B backlog Key call notes so far: – Going exclusive on Nvidia GPUs (“best architecture”) – >2 GW compute by end of 2026 → closer to 10 GW by end of 2027 – Two successful Starship V3 tests in 90 days – Acquiring Cursor for $60B to accelerate AI – Elon: we’re playing the 5-10+ year game (Moon & Mars bases). Short-term earnings may take hits and that’s fine. First public report and they’re already operating at this scale. What do you think — undervalued or fully priced in? $SPCX

The Grimes County deal does not merely reduce taxes. It rewrites the marginal cost of scaling. A $20 million annual M&O cap on a planned $55 billion Terafab produces a $2.85 billion cumulative discount over 35 years 91.3 % savings against the $3.56 billion that would have been owed at a standard 0.45 % rate with 8 % depreciation. Under the $5 billion minimum commitment the same cap saves only $69.5 million. Every additional dollar of silicon, cleanroom, and tooling beyond the floor therefore carries an effective tax rate approaching zero. This is deliberate policy architecture: the county receives near-standard revenue on the minimum build while removing the tax penalty on maximal physical footprint. The incentive points in one direction build the largest possible compute and manufacturing base on Texas soil. @SpaceX @SpaceXAI

Grimes County SpaceX Terafab: 35-year fiscal model > $5B capital investment > 1,800 construction/tech jobs > $1B annual payroll County receives: > $10.035M upfront (60 days) > $20M/year PILOT (Yrs 1-10) > $20M/year capped M&O (Yrs 11-35) Municipal costs: $0 (Wastewater recycled on-site. Private fire/medical teams. Road repairs under separate SpaceX agreement.) Net county cash flow: > Year 1: $30.035M > Years 2-35: $20M each > Cumulative: $710.035M Zero local tax burden. Pure revenue. Local labor required. This is how industrial scale builds the multiplanetary base without extracting from the host.

TERAFAB IS STILL THE RATE LIMITER Current global advanced silicon meets ~2–3% of the sustained demand required for Optimus-scale production and radiation-hardened orbital compute. Development fab equipment ordered. Intel 18A veteran now leading as Director. SpaceX already paid Grimes County $10M ahead of schedule. Full location announcement incoming. One terawatt of annual output. Two purpose-built designs. Logic + memory + packaging under one roof. Without it the rest of the stack remains constrained.​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

STARMIND: THE CONSTELLATION THAT BECAME A MIND Starmind is the recently confirmed official name for Space𝕏 and 𝕏 AI's planned constellation of up to one million AI-enabled satellites, which will serve as an orbital computing network. The "D3" space-hardened chips being manufactured at the Terafab facility, Starmind is the actual satellite constellation where those chips will be deployed. Here is a breakdown of what Starmind is and how it will operate: Servers in Space, Not Just Pipes: While Space𝕏 's Starlink network functions like a fast communication pipe to relay internet data, Starmind satellites will function as actual servers in orbit. Equipped with onboard processors and massive solar arrays, these satellites are designed to run intensive machine learning inference and large language model workloads directly in the vacuum of space. Bypassing Earth's Data Centers: By computing AI queries in orbit, Starmind will be able to beam the generated outputs back down to users on Earth within milliseconds. This completely bypasses the need to route data through terrestrial data centers, which are currently facing severe power, cooling, and real estate constraints. If it scales as planned, Starmind could make SpaceX the "landlord of AI compute" in the same way Starlink dominates satellite internet. Deployment Timeline: The Starmind constellation builds upon an FCC application filed by SpaceX in early 2026 for an "Orbital Data Center system" that will operate at altitudes between 500 km and 2,000 km. SpaceX is targeting early 2027 to launch two "AI1" prototype satellites, with volume production expected by the end of 2027 at a new facility called Gigasat. Ultimately, Starmind represents the physical realization of the goal to move civilization-scale AI compute off the Earth, utilizing Starship's launch capabilities to build a space-based edge computing layer.

"We either build the Terafab or we don't have the chips, and we need the chips, so we build the Terafab." - @elonmusk AND THIS IS THE STRATEGY BEHIND THE $20–$25 BILLION TERAFAB BET.... OUTPACING THE GLOBAL SUPPLY CHAIN Elon Musk has repeatedly warned that existing third-party foundries such as TSMC, Samsung, and Micron cannot scale manufacturing capacity fast enough to meet future demand. All current fabrication facilities on Earth produce only about 2% of the compute capacity his projects will require. Intense competing demand from Apple, Google, and Nvidia makes the choice binary: build the Terafab or accept a structural silicon bottleneck. MASSIVE DEMAND FOR OPTIMUS AND ORBITAL AI Tesla and SpaceX have established hardware scaling targets that require a dedicated custom silicon pipeline. Terrestrial AI (20% of output) will consume millions of custom AI5 and AI6 edge-inference chips for Full Self-Driving, the Cybercab robotaxi fleet, and Optimus humanoid robots. Scaling to 10 million Optimus units annually would require an estimated 20 million chips per year—roughly six times Tesla’s current automotive chip demand. Space-based AI (80% of output) will utilize radiation-hardened D3 chips for SpaceX’s Starmind constellation of orbital AI data centers. The objective is one terawatt of annual compute capacity powered by unfiltered solar energy in space. VERTICAL INTEGRATION AND THE RECURSIVE LOOP The traditional semiconductor supply chain is geographically fragmented, with design in North America, fabrication in East Asia, and packaging in Southeast Asia, creating a 12- to 18-month iteration cycle. Terafab consolidates chip design, lithography, fabrication, memory production, advanced packaging, and testing under a single roof. This vertical integration produces an extremely fast recursive loop in which engineers can design, test, identify flaws, revise the photomask, and manufacture a new batch in days or weeks rather than months. SILICON SOVEREIGNTY AND COST ECONOMICS By partnering with Intel to license its 14A (1.4-nanometer) process technology, Tesla is securing independence from foreign supply-chain vulnerabilities and geopolitical risk. Co-designing the software and the physical chips enables AI models to extract maximum performance from every circuit. The resulting control over the full stack drives down the long-term marginal cost of compute hardware to levels that make a multi-trillion-dollar interplanetary infrastructure financially viable.

TERAFAB IS THE RATE LIMITER Current global advanced silicon meets ~2–3% of the sustained demand required for Optimus-scale production and radiation-hardened orbital compute. One terawatt of annual output, two purpose-built designs, logic + memory + packaging under a single roof. Without it the rest of the stack remains constrained.

Elon Musk's Terafab & The Silicon Empire TerraFab: Tesla’s $119B semiconductor foundry in Texas is a masterclass in vertical integration. By owning 2nm AI silicon production (in JV with SpaceX, xAI & Intel), Tesla is securing AI5 (40x compute, 9x memory) for FSD/Robotics while building toward 200GW Earth + 1TW orbital compute capacity. In an era of severe chip scarcity, control over the physical hardware is becoming the ultimate competitive advantage. The future belongs to those who control the silicon, not just the software.

SpaceX’s actual mission ‘Make life multiplanetary.’ Not for profit. Not for prestige. To extend the probable lifespan of consciousness. One planet = single point of failure for everything we know. That’s the entire point.​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

100 GW/year of orbital solar-powered AI compute is an engineering equation, not a slogan. 1 Mt of payload to orbit annually at 100 kW/ton delivers it. Starship cadence of ~10 k flights/year closes the mass. By aircraft standards that cadence is modest; relative to today’s entire rocket industry it is transformative. This is the concrete path from power-constrained terrestrial data centers to continuous, radiation-hardened orbital compute at civilizational scale. Source I Elon Musk

Grok just held perfect vocal consistency + character tone across a full 5-minute monologue as Tom Ellis’s Lucifer. Zero drift. Zero breaks. The engineering is exceptional. I wrote the piece on free will the version that actually costs something. Monologues are my current focus. Writing comes first. First principles still apply: tools that can sustain a mind this long are the ones that scale.