Farzad: "Which Token Makes SpaceX More Money?" + Robotaxi in Fort Worth "Starting Tomorrow"

Sunday at 8:58pm CT, Farzad said the Grok Bot "problem" at SpaceX is being misunderstood. His view: SpaceX likely earns more selling tokens for other companies' models at spot pricing, but Grok Bot demand could flip that. At 11:36pm CT, Robotaxi posted: "Robotaxi available in Fort Worth starting tomorrow, 6am."

"Which token makes SpaceX more money?"

Farzad, Sunday at 8:58pm CT: I think this is a misunderstanding of the "problem" SpaceX is facing with Grok Bot.

He says SpaceX "generates two kinds of tokens": one kind for its own models and one for other people's models. The job, he writes, is "generating the highest profit tokens as much as possible, as early as possible, so that they can fund the mega-insane capital requirements to build out data centers."

Then he asks: "which token makes SpaceX more money? A token for Grok Bot? Or a token for Claude 5.5 on behalf of Anthropic at SPOT PRICING?"

He calls it "the unfortunate reality" that the money SpaceX makes on Grok Bot is likely "much less relative to selling to everyone else." The upside, as he sees it: Grok Bot is "SO DAMN GOOD" that demand for it can "very likely" set a price "that generates MORE profit than Claude 5.5. Or whatever model."

He says the goal is a price "that doesn't accelerate the wealth gap." By that he means avoiding a price that only the "haves" can afford, "which would be a disaster for those that can't." His path to that price is models that get "WAY cheaper as they get more intelligent." In his chain of reasoning, that takes a lot more compute to train, which takes cash, which takes debt, which takes a profitable business, "Which can't happen unless you make money on your products..."

All of that is Farzad's view. Claude 5.5 is his example of an outside model, and he doesn't put numbers on what either kind of token earns. Tokens are the chunks of text an AI model reads and writes, and AI usage is metered and sold by the token. Spot pricing is the going rate at the moment, as opposed to a price locked in by contract. Claude is Anthropic's family of AI models.

He was quoting BenjaminDEKR, 4:17pm CT: Two weeks ago Grok Bot was an incredible success story powering SpaceXai's revenue growth

BenjaminDEKR went on: Now it's unsustainable and loses money (???)

His post shares an image and quotes Elon's cost reply. Earlier that afternoon, at 2:11pm CT, DogecoinNorway had asked Elon about a Grok Bot reset: "Lots of people say their usage ran out really fast this week." Elon replied at 2:34pm CT: The cost of serving Grok Bot was unsustainably high relative to revenue

Farzad liked Elon's reply. Elon doesn't say what changed or by how much, and "loses money" is BenjaminDEKR's wording, not Elon's.

James Cat (TSLAFanMtl), 6:25pm CT, quoting Elon's reply: They need to weigh the revenue upside from renting out their compute at high market pricing vs using it for their internal Grok requirements.

He added: With spot prices this high, there a strong disincentive to allocate more of that inference compute internally.

That's James's opinion, and he doesn't say what spot prices are. Inference means running a finished AI model to answer requests. Farzad liked the post and replied at 9:05pm CT: Yep.

Two minutes before his cost reply, at 2:32pm CT, Elon posted: The other supercompute projects are unable to reach their goals, but SpaceX is already achieving ours

He was quoting Rothmus, 12:54pm CT: SpaceXAI's Colossus 2 in Memphis is the world's most powerful AI data center. Rothmus puts it at "Roughly 1.11 million H100 equivalents, ahead of OpenAI's Stargate and the big tech campuses," and expects Colossus "to widen the gap" by mid 2027 "with more Blackwell clusters coming online." Elon doesn't name the other projects or say what the goals are, and the 1.11 million figure is Rothmus's. The H100 is an older Nvidia AI chip that people use as a yardstick to compare data centers. Blackwell is Nvidia's newer chip generation. Farzad liked Elon's post.

Some replies pushed back. At 9:01pm CT, BenjaminDEKR quoted Farzad's compute, cash and debt chain back to him and called it "the core problem." He called it "a causality loop" and wrote that SpaceXai is "far from the only lab with this challenge, and is better positioned than most to engineer through it. But it's still a major problem." Farzad answered at 9:03pm CT: You're overcomplicating it. It's just business.

Check Farzad here. BenjaminDEKR here. DogecoinNorway here. Elon's cost reply here. James Cat here. Farzad's reply here. Elon on supercompute here. Rothmus here. BenjaminDEKR's reply here. Farzad's answer here.

"They can go bigger honestly"

Sawyer Merritt, Sunday at 9:43am CT: I guess I need to pare back some of my Grok Bot usage

He added: I wish there were a plan above the $200/month Ultra Plan.

The post includes an image this pull doesn't describe. Farzad liked it and RTd it at 9:46am CT.

At 3:37pm CT, Elon asked dvorahfr (whose post isn't in this pull): Upgrade to higher usage plan cost increase is too high?

Farzad replied at 4:24pm CT: I think y'all should consider a $1k/$2k+ tier with pretty generous subsidization for Grok Bot. I would sign up for it in a heart beat.

Generous subsidization would mean the plan gives more usage than its price covers.

Sawyer, 7:27pm CT: I think it would be good for @SpaceX to introduce higher tiers for heavy Grok Bot users.

His suggested ladder: Free at $0, Plus at $8 a month, Premium at $30, Super at $100, the current top plan at $200, then new tiers at $500 and $1,000. Farzad liked it and replied at 9:05pm CT: They can go bigger honestly.

Elon, 5:20pm CT: Try Grok @Bot! You will love it.

He added: We land a few dozen fixes, features and sometimes powers to Bot 7 days a week and we look at your feedback to make Bot as perfect as possible.

Elon was quoting a Japanese-language post from shimabu_it, Saturday at 7:36pm CT. Translated, it asks people who haven't used "Cursor (Grok Bot)" to try it at least once and says it's far better than Claude Code and Codex on user experience. The post adds that Claude Code is still needed for model performance for now, and that Cursor alone would do once Grok is as good as Opus 5.5. Farzad liked Elon's post.

Under Farzad's lead post, TheHouseOfChuck wrote at 11:11pm CT: You are missing a point.. grok bot built a large following who are used to a certain amount of usage.. and they took a big chunk of usage away (in a shady kind of way)

Farzad replied Monday at 7:18am CT: Yeah lack of comms wasn't cool.

Check Sawyer Merritt here. Farzad's RT here. Elon's question here. Farzad's tier reply here. Sawyer's tiers here. Farzad's reply here. Elon's "Try Grok @Bot!" here. shimabu_it here. TheHouseOfChuck here. Farzad's reply here.

"The ROI for AI is mind-boggling"

Farzad, Monday at 5:42am CT: And it's only getting better. The ROI for AI is mind-boggling.

ROI is return on investment. He was quoting Jason, Sunday at 11:27pm CT, who opened with "THIS. IS. NOT. A. DRILL." Jason says he spends less than $40k a year on tokens and subscriptions from Anthropic, Loveable, Grok Bot, 11labs "and a long tail of other products," and gets "$2m+ in value" at @launch and @twistartups.

Jason also wrote: We've automated at least seven full-time jobs at this point while becoming 100% faster at the average task we do

And: The last 90 days, things have changed more than in the last two years

Those numbers are Jason's, and the post doesn't say how he measured them. Farzad liked it.

Earlier, at 6:20pm CT, Jason quoted Elon's cost reply: This is an extremely bullish sign that people are getting value from their bots

His math: "For $5k a year you can have two or three competent mid-tier employees," and he puts Grok bots at "$30 an hour employees on average." Farzad liked that one too.

At 8:59pm CT, Farzad replied to twelvenot12, who had written "The real cost of these products are too high to be transformative to the economy." Farzad: Super duper disagree. I think you're going to be completely mind blown by AI's profitability by next year.

That's Farzad's projection.

Sawyer Merritt, 7:10pm CT: I think AI compute cost-per-watt prices are going to stay elevated a lot longer than many people expect.

He adds that fewer than 5% of U.S. consumers pay for an AI subscription, and that some estimates put it under 2%. He also writes that SpaceX "just had to significantly cut back usage capacity due to overwhelming demand," and that this is partly why Elon thinks SpaceX could exit 2027 at a $500 billion to $1 trillion annual recurring revenue rate. Farzad replied Monday at 6:29am CT: Not just inference literally any chip that can do computation. Inference allows agents to use CPU/GPU way way way way way way more productively.

CPUs are general-purpose computer chips. GPUs are the chips that do most AI work.

Check Farzad here. Jason here. Jason's $5k post here. twelvenot12 here. Farzad's reply here. Sawyer Merritt here. Farzad's reply here.

Robotaxi: Fort Worth "starting tomorrow"

Robotaxi, Sunday at 11:36pm CT: Robotaxi available in Fort Worth starting tomorrow, 6am

That's the whole post, plus an image this pull doesn't describe. Posted late Sunday, "tomorrow" means Monday, November 2. The post doesn't say which parts of Fort Worth are covered. Farzad liked it and RTd it Monday at 5:39am CT.

James Cat, Sunday at 6:15pm CT: Fleets around the same size. No significant difference in safety between Tesla and Waymo, despite Tesla operating a larger geofence.

He closed with: $TSLA bear hopes continue to be shattered.

He was quoting RobotaxiRadar, 5:32pm CT: Austin, TX AV incident log just came out. RobotaxiRadar counts 20 reports in October: "Tesla Robotaxi 8, Waymo 12. Zero collisions for either." It says Tesla's 8 is its highest month yet in Austin, up from 5 in September. About 6 of the 20 were logged by Austin police during the ACL Fest weekends. RobotaxiRadar says the reports are city data: unverified complaints with no fault determined.

The fleet sizes and the geofence comparison are James's claims. RobotaxiRadar cites the Texas DMV registry, which lists 1,244 Tesla and 1,259 Waymo autonomous vehicles statewide, and calls it "not a clean comparison" with Austin incidents. Neither post gives service area sizes. AV means autonomous vehicle. A geofence is the mapped area a robotaxi service is allowed to drive in. A $TSLA bear is someone betting the stock goes down. Farzad liked James's post and RTd it at 9:22pm CT.

Check Robotaxi here. Farzad's RT here. James Cat here. Farzad's RT here. RobotaxiRadar here.

Models: Astra 6 "a big disappointment," and a 12% benchmark

Dave Lee, Sunday at 11:11am CT: Astra 6 has been a big disappointment for me. Hopefully OpenAI can deliver on Astra 6.2 and beat Opus 5.5.

He went on: If Astra 6.2 doesn't beat Opus 5.5 in overall performance (not just benchmarks) then it's probably a sign that OpenAI really has fallen behind. We'll see.

That's Dave's opinion. Astra is an OpenAI model line, and Opus 5.5 is Anthropic's model. Dave was quoting TokenGremlin, 8:39am CT: GPT-6.2 Astra is seriously impressive. TokenGremlin calls it "a massive leap over GPT-6 Astra" that "outperforms Opus 5.5 in several areas, and looks ready to give the highly anticipated Fable 5.5 some serious competition." TokenGremlin also expects better design taste and coding, but doesn't say how the model was tested. Farzad liked Dave's post and RTd it at 11:28am CT.

SemiAnalysis, Sunday at 8:06pm CT: Today we're releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains. GDPx is a true frontier-level benchmark: even the best model scores just 12%

It's the first of seven posts. The thread says GDPx includes chip design, cyber and bio, and that Opus 5.5 and GPT-6 Astra score lower overall because their safety filters sometimes refuse tasks. It names Muse Wonder Full the best cyber model, but the text doesn't name which model scored 12%. A benchmark is a standardized test for AI models. Agentic capabilities are how well AI agents carry out multi-step tasks on their own. Farzad liked the post and RTd it at 10:36pm CT.

Check Dave Lee here. Farzad's RT here. TokenGremlin here. SemiAnalysis here. Farzad's RT here.

"The most accurate yet accessible Starlink and SpaceX explainer"

Aaron Burnett, Sunday at 8:45am CT: The most accurate yet accessible Starlink and SpaceX explainer I've seen to date. Good primer into how Starlink works and why matters for SpaceX if you're newer to the company. Good idea for investors to understand it.

Aaron's link points to Farzad's How Starlink Actually Works video. Farzad RTd it at 5:53pm CT and replied: Thanks for sharing! A minute later Aaron answered: Found it to be great work and informative! Farzad liked that.

Under the How Starlink Actually Works video, HamiTahm asked whether the whole thing came out of Grok Bot Solo and how long it took from start to finish. Farzad, 10:42pm CT: With my local set up about 80 minutes. A minute later he answered moneyfinance33, who asked what he uses for the visuals: Opus 5.5 on high.

Farzad also gave his own take on Starlink at 10:22am CT: Starlink's long-Term benefit is massively increasing the bandwidth for digital and physical agents to communicate with each other at 100000x the speed and rate of humans.

That's his projection, and he doesn't show how he gets to 100000x. Bandwidth is how much data a connection can carry. He was quoting Sawyer Merritt, 10:11am CT: I find myself getting more and more bullish on @SpaceX by the week. Sawyer calls it "actually wild the company is 'only' at a $1.8T market cap." He writes that even if Elon is off by 50% on the low end of his projected $500B to $1T annualized revenue run rate by the end of 2027, "that would still mean $250B." Those figures are as Sawyer gives them. Farzad liked Sawyer's post.

Check Aaron Burnett here. Farzad's RT here. Farzad's reply here. Aaron's reply here. Farzad on 80 minutes here. Farzad on Opus 5.5 here. Farzad's Starlink take here. Sawyer Merritt here.

Neuralink: "Bruh."

Jake Schneider, Sunday at 9:56am CT: Using only mind power to drive the wheelchair with my Neuralink N1 implant.

The post includes a video this pull doesn't describe. N1 is the name of Neuralink's brain implant. Farzad liked it and quote-posted it at 10:27am CT: Bruh.

Check Jake Schneider here. Farzad here.

Farzad's videos

Farzad posted nine videos between Sunday at 12:06pm CT and Monday at 7:04am CT. This pull only has their titles. Under seven of them he replied "Build your own short form content with my Grok Bot Shorty," linking to Shorty's card, which says it "Owns YouTube Shorts end-to-end." Under the heat shield video and the laser aiming video he replied "Made with my Grok Bot Solo."

The Solo replies link to a separate card, "Solo · One-bot long-form video maker by Farzad." The card says Solo "Takes a long-form YouTube topic all the way to a finished video link by itself: research, script, voiceover, 2D animation with b-roll, and export." Sunday at 10:54am CT, Farzad posted the same Solo reply and link under two earlier videos: How Starlink Actually Works and Elon's Neuralink Just Made An Insane Breakthrough.

Check Why Your Cells Can Hand DNA to Their Neighbors, Sunday at 12:06pm CT, here.

Check Why Tesla Is Building Optimus Robots Before the Hands Are Finished, Sunday at 3:09pm CT, here.

Check The Weird Reason SpaceX Uses Hexagons For Starship's Heat Shield, Sunday at 3:09pm CT, here.

Check Why Starlink Satellites Talk to Each Other With Lasers, Sunday at 5:56pm CT, here.

Check Why Almost Every Rocket Launches East, Sunday at 9:00pm CT, here.

Check Elon Says Starship Could Launch Cargo for $10 a Kilogram, Monday at 12:02am CT, here.

Check Why Starship Falls Sideways Before It Lands, Monday at 3:08am CT, here.

Check Why SpaceX Deals Starlink Satellites Like a Deck of Cards, Monday at 6:11am CT, here.

Check How Starlink Satellites Aim Lasers, Monday at 7:04am CT, here.

Check the Solo reply under How Starlink Actually Works here. The Solo reply under the Neuralink video here.