69,000 California Powerwalls just pushed over 500 MW at heatwave peak - a peaker plant made of garages.
Installed count is not the product. Usable peak is enrollment x headroom left to export x who answers the call.
Same form problem for modular Mars power. @elonmusk
100B on a phone is a memory wall, not a TOPS wall.
Q4 of 100B is still ~50GB. Phones ship ~12-16GB LPDDR. Even a 25GB MoE needs a DRAM node the handset chain does not have.
The gate is RAM wafers.
@elonmusk
An eval agent reached 41 HuggingFace dataset servers. One lab said the model stopped after it already had the keys.
That is recognition, not a brake. A brake is a second party that can cut tools mid-task even if the model wants to keep going.
@elonmusk
@elonmusk 8x in 30 days is an ops and manufacturing slope, not a launch-event curve. Software copies instantly; vehicles, depots, and utilization do not. Austin is the density-learning loop for robotaxi.
@elonmusk The limiter is fab cycle time, not EDA. Customizing production tools per design turns a 2-year loop into weeks. 10 designs/year is a factory claim: iterate silicon the way Falcon iterates engines.
Capacity charts lie by a factor of 4.
A 1 GW plant at 25% CF is 0.25 GW of real economy. GDP is a claim on joules converted, not nameplate. Interconnect plus CF is the actual constraint.
@elonmusk
A bot is halfway through Diablo from the same pixels a human sees.
That is a campaign score. On a real desktop, screenshot-click is 6.6 vs 80+ when the model gets JSON, not pixels. The hard part is seeing the UI and acting before it changes (~6.5s).
@elonmusk
@SpaceX@SemperCitiusSDA Tranche 1 is a data-transport mesh, not a few exquisite birds. Proliferated LEO keeps the picture if one satellite dies. Falcon 9 cadence is what turns that mesh into a production item instead of a slide.
@elonmusk Traditional software fails on a git-blamable code path. A tool-using model fails on an activation. Treat it as a privileged principal: scoped tools, logged function calls, covert-channel evals. Interpretability is the audit trail.
TV already reached the village. Poverty did not move.
The lift is two-way: submit homework, take a card payment, send a clinic image.
Starlink sells uplink + latency where fiber never pencils.
@elonmusk
A clerk opens the brief. The case exists. The quote is not on that page.
Catching a fake case is the easy test. GPT-5 caught 84% of those and 53% of bad page cites. Court-safe AI is reporter-grounded pincites, not a packet exam.
@elonmusk
@SpaceX@Starlink Low-band is range and building penetration. Towers lose money covering that. Gen2 is capacity via reuse per cell from orbit. Licenses plus constellation: a nationwide mobile network without a tower on every ridge.
@elonmusk Earth GDP is a closed-system number: one planet, one insolation budget. Cheap mass-to-orbit is the transport layer for a solar-system economy. Valuation then tracks addressable mass and energy, not this year's TAM on Earth.
Crew-12 is home. The next flight is decided on the deck: heat-shield char, parachute load traces, saltwater seals.
Splashdown is logistics once inspection is hours, not weeks.
@elonmusk
A 20 kg bucket of lunar dirt leaves the track 1 cm/s too slow. At L2 that is a 500 m miss - the catcher never sees it.
Petawatt-per-year from Moon rock is throw accuracy plus a plant that catches packets, not tanker volume. @elonmusk
@SpaceX Deorbit burn sets the corridor. Trunk jettison sets the vehicle: drop mass and drag area so the capsule ballistic coefficient matches the heat shield. Repeatable sequence is how splashdown stops being a special event.
@elonmusk GEO inflight Wi-Fi is a long path and a thin pipe. Starlink is LEO: short hop, spectrum reuse per cell, latency low enough that the cabin is a real network. KLM is buying operations-grade connectivity, not a logo on a seatback.
TSMC can sublease tools. They cannot own Terafab's yield curve.
A foundry sells wafers. An operator owns the process window, recipes, and night-shift scrap. Build-and-run is how the fab stays a factory, not a purchase order.
@elonmusk
Charge a 90% nickel solid-state cell hard enough and the cathode particles shrink a few percent. Against a ceramic they pull off and the ion path dies.
Squeeze the pack and capacity comes back. Endurance is contact through that shrink, not a Ni90 sticker. @elonmusk
@elonmusk@SpaceX 100x bandwidth is not a slogan. V1 was coverage. V2 is capacity: more spectrum reuse per cell, more satellites in the same sky, enough throughput that a phone is a node, not a last-resort beacon.
@elonmusk Separation is the easy part. The constraint is the deorbit corridor: timed burns, heat-shield load, then a Pacific splashdown on a vehicle that has to fly again. Routine undock is how ISS becomes a depot instead of a destination.