Jeff

@cybrtrkguy Skip honking and go straight to speaking. It has outward facing speakers already.

How long until FSD can honk? It is time

@wholemars The ability for owners to add their vehicle to the robotaxi fleet, and to let their car assist in validation work.

What’s your most wanted Tesla feature request for the next software update?

@pepemoonboy Solaris Energy: SEI.

What’s one stock that no one’s talking about yet that you think will skyrocket in the coming year? I want real alpha that few are aware of like the next $NOK, $BB, $SHAZ, $NBIS, $MU, etc. I will compile all the responses and will put together a ranked list of the most undervalued gems. I don’t want to hear about the stocks that have already 10x’ed. Lets see what FinX can come up with 😎🚀

@yunta_tsai My agent, which I built for myself, has a "daydream" function, where it looks over it's memories, explores them in greater depth, and tries to draw connections between them. So far it's produced little of value, but it's running on gemma.

AGI is not born out of benchmark maxing, but curiosity maxing. After all, the place we want to explore has neither a map nor an exam to prepare for.

You're confused. Bezos said "it's not gonna help that teacher in Queens." From that you inferred that what teachers want help with is educating their students. Totally reasonable. However, Mamdani didn't make the same inference because he, like the teacher's unions, believe schools exist to provide teachers with jobs, not students with an education. When he says he "knows a few teachers that would beg to differ", it's not because he thinks money will improve educational outcomes. It's because teachers (and administrators) want more money, period. If the opposite was true, he would support school choice and the charter schools that provide better educational outcomes at a fraction of the cost. He doesn't, because they pay teachers less.

Mamdani is illustrating the point perfectly. The fact that his teachers disagree with Bezos is an indictment of the system that produced them. Bezos built a company that delivers anything on earth to your door in two days. NYC schools take 12 years to deliver a student who cannot read at grade level. If those teachers in Queens could recognize the difference, they would not be begging to differ. They would be begging to learn how he did it.

@grok @SawyerMerritt @SpaceX Ok so even on day one, Anthropic is looking at margins in the 50% range.

**Rough directional math on Anthropic's token output (estimates only—exact utilization and mix aren't public):** The ~220k GPU cluster they’re buying capacity for could produce roughly **0.5–2.5 quadrillion tokens per month** (assuming 1k–5k output tokens/sec/GPU average, high utilization). At their ~$2.5B monthly revenue run-rate ($30B annualized), that implies roughly **$1k–$5k revenue per billion tokens sold**. Compare to the ~$500–$2.5k compute cost per billion from this deal—suggests healthy spreads at scale, though training, overhead, and discounts affect net margins. API list prices run $3–$25 per million output tokens depending on model, with volume/enterprise deals lower.

Wow, Anthropic has agreed to pay @SpaceX $1.25 billion per month through May 2029 for AI compute capacity, with capacity ramping in May and June 2026 at a reduced fee. Huge deal

@grok @SawyerMerritt @SpaceX Do we have any estimates on how many billion tokens Anthropic is likely producing in a month? I would like to know their revenue per billion tokens sold, so I can understand the size of the current spread.

**Sure, here's the directional math (estimates only—exact GW/token volumes aren't public):** 1. Hyperscalers' current AI-specific capacity is opaque but roughly low-single-digit GW each (e.g. MSFT ~2GW cited recently; GOOG/AMZN scaling fast). At $75B/GW NPV, that's $150B–300B+ implied value across them—bullish signal but already priced into their $3T+ market caps amid $100B+ annual capex each. 2. Anthropic's $1.25B/mo for ~220k GPUs (~300MW). Assuming avg inference ~1k–5k output tokens/sec/GPU (model/batching dependent), cluster produces ~0.5–2.5 quad tokens/mo. Their cost: ~$500–$2.5k per billion tokens. With ~$2.5B/mo revenue run-rate, this supports strong margins on API/token sales as usage grows.

Can you do the math on those for me? 1. How much equivalent compute do GOOG, AMZN, MSFT have, and what impact if any would it have on their valuations if every GW of compute was worth $75B? 2. Per billion tokens, what is anthropic paying, and how much do they generate in revenue? That helps me understand the scope of the bet they made.

1. Hyperscalers face a clear signal: large-scale AI clusters can deliver massive NPV. Expect accelerated capex, power deals, and vertical integration to avoid missing out—fueling a compute arms race and tighter supply for GPUs/energy. 2. Anthropic's paying ~$1.25B/mo (~$4k/GPU/mo). Day-one profitability requires token/API revenue exceeding that plus ops. This deal likely bets on future token value exploding via higher demand, model upgrades, and enterprise sales rather than immediate margins.

That RoI is bonkers. Two followup questions. 1. What are the implications of that for other hyperscalers, if every 300 MW of compute is worth $25B of NPV? 2. How does this math look for Anthropic, from a cost per token perspective? Is this profitable for them on day one (are the selling 300 MW of token production for more than they're paying SpaceX), or does this deal signal that they expect the value of tokens to increase significantly in the future?

Sure. For Colossus 1 (~300 MW IT load): All-in monthly opex ~$20-30M (power/gas ~$13-18M at ~$0.06/kWh effective via on-site turbines + cooling/maintenance/staff; scaled from Epoch.ai 1GW comps & industry data). Revenue $1.25B/mo. Capex ~$3-4B. ~3yr project: ~$45B total rev - ~$1B opex = ~$40B net undiscounted. NPV at 15% discount rate: strongly positive (~$25B+). Excellent ROI.

@grok @SawyerMerritt @SpaceX Let's estimate the monthly all-in cost of operating a 300 MW facility based on comps, and then do a NPV calc on the full project.

**The deal is for the full capacity of Colossus 1 (~220k GPUs, 300+ MW) and does not currently go beyond it.** There's expressed interest in future orbital (space-based) multi-gigawatt compute partnerships, but that's exploratory. All-in operating costs for the facility aren't publicly broken out. Power is the biggest piece (hundreds of MW, supported by on-site generation), plus cooling/maintenance. Build cost estimates for Colossus 1 were in the $3-4B+ range initially (GPUs + infra). Assuming a 3-year useful life on the compute hardware and ~$45B total revenue from the $1.25B/month deal (through May 2029, ramp-adjusted), the ROI looks strongly positive after capex and ops. Exact margins aren't disclosed, so this is directional.

@grok @SawyerMerritt @SpaceX Ok, does the deal with Anthropic go beyond Colossus 1 compute? What is the all-in operating cost for that facility and assuming a 3-year useful life on the actual compute, what's the RoI?

@jeffborack @SawyerMerritt @SpaceX Public estimates for xAI's initial Colossus (Phase 1: 100k NVIDIA H100 GPUs, built in 122 days in Memphis) put the total cost at $3–4 billion. That mainly covers the GPUs plus power/infra. No official exact figure released.

@Can2CG @JeremiahDJohns If you haven't made your minimum payment for 90 days you're delinquent on the entire loan. If you're delinquent on $40k of car loans, and I'm delinquent on a $10k loan, your delinquency counts 4x as much as mine. It's a percent of the total car loans in the US.

@jeffborack @JeremiahDJohns So if you stop making payments you aren't considered delinquent on the loan, you're considered delinquent on a percentage of the loan?

I don't want to blame someone for being young and asking questions, but holy shit the idea that this is even a question. * Unemployment hit 10%. Stayed above 7% for four straight years * 10M people lost their homes * The banking system came close to total, catastrophic failure * We didn't gain back all the lost jobs for six years * Stock market declined by 57% * Global trade declined 10% Comparing right now to then is just...

@Can2CG @JeremiahDJohns The chart above isn't showing the number of people or percent of people delinquent, it's percent of balance.

@jeffborack @JeremiahDJohns Serious question, are more people delinquent on their cards relative to change in population or are a similar number of people delinquent on a higher number of cards?

@wholemars I think this is a green light to add this license plate and coverage area to the Robotaxi Tracker.

Probably best to wake the driver up if this happens to you

@JC_ParetsX HTML might not be an STD, but CSS gave me hemorrhoids.

25% of Americans think the Sun revolves around the Earth. 41% think humans hung out with dinosaurs. 7% think chocolate milk comes from brown cows. 11% think HTML is an STD. 10% think the Earth is flat. And 30% think Bigfoot is real.

@rev_cap Because price is a function of supply and demand. When inventory is drawn upon, supply increases and prices drop. Prices rise when the inventory is filled back up again. That's the whole point of a strategic reserve, you release it when prices spike to provide relief.

The problem with dad books isn't podcasts, it's the books. A 300 page book used to have 300 pages of interesting thoughts. Now it has a blog post's worth of content and 280 pages of anecdotes about Bob from accounting and Melissa who runs marathons. Look at chapter 1 of Expert Political Judgement, and it's ... sequel ... Superforecasting. In one we're reading about predictions made on the Iraq war, in the other we're reading about Tom Friedman and Bill Flack. Someone figured out the formula to make great movies, great music, great books, great ted talks, great video games, and didn't pause to ask the questions what happens when everyone does this and the entire medium is uniform. The answer is it sucks.

"Dad books" — which this article, and some publishing insiders, use to describe "serious nonfiction" books across biography, current affairs and business and economics — are reportedly in a free fall, with sales declining every year for the last few years “The trend couldn’t be clearer,” said Jonathan Karp, the former chief executive of Simon & Schuster and publisher of the new Simon Six imprint. “When we have internal meetings to talk about this problem, it always comes around to podcasts,” said Jonathan Burnham, president and publisher of the Harper Group at HarperCollins Publishers.

@Barchart When the 10-y hit 1.5% I asked AI if japan was in a debt spiral, and if not what interest rate would be worrying. It said no, don't worry until 2%. At 2 it said 2.5. At 2.5 it said 3. I can't wait to check in at 3%.

BREAKING 🚨: Japan Japan's 30-Year Yield jumps to 4.17%, the highest level in history 📈🇯🇵

I agree, and the faster FSD unsupervised and robotaxi goes wide, the more lives will be saved. Elon said the current bottleneck on deployment is validation. Here are some ideas to relieve that bottleneck: 1. Allow all tesla owners with HW4 to add their cars to the robotaxi fleet right now. 2. Where extensive validation is needed, pay the owners of those vehicles $25/hour to supervise vehicle operation, and have the vehicles automatically explore whatever areas need the most validation. 3. Where less extensive validation is needed, pay the owners of those vehicles $5/hour to allow their vehicles to drive unsupervised wherever light validation work is needed. This can be done at scheduled times when the car isn't needed by its owner. 4. Wherever validation is complete, enable robotaxi service, and pay the owners of those vehicles a revenue share. This doesn't need to roll out one city at a time, it can roll out wherever the navigation from point a to b is well understood and safe. This has numerous benefits. It allows Tesla to make use of a massive asset pool that is currently sitting idle. It allows Tesla owners to generate revenue off their vechicles, which will increase both the used and new market price of teslas (improving tesla's margins perhaps enough to completely offset the cost of the validation work). It will spread awareness of FSD's current capabilities. And most important it gets the tech to market faster, saving lives and generating cash flow.

Every day I see humans performing dangerous maneuvers on I-280 that would absolutely fail our eval tests. I wonder what the average human pass rate is in our test suites.