Tesorb Signal: Tesla, SpaceX, and the Musk Companies

SpaceXAI’s Grok Bot agent reached 418,000 weekly users about a month after launch, according to a presentation Bloomberg viewed. We look at what that number counts and what it leaves out, the tenfold price cut that came before it, and why Grok 4.7 arrived after the showcase meant to prove it.

Four hundred eighteen thousand. That’s how many weekly users SpaceXAI’s Grok Bot agent had on September fourteenth, about a month after launch.

The figure didn’t come from an earnings call or a filing. It came from a presentation shared at an event in London, and Bloomberg reported it on Tuesday.

A week earlier, the same number was lower by about a fifth. That kind of growth sounds impressive. The question is what, exactly, is being counted.

This is Tesorb Signal, and I’m Lena Ruiz. Today, Grok Bot’s first report card: what the number says, what it leaves out, and why the model built to power it arrived late.

First, a quick reminder. Grok Bot is not the Grok chatbot on X. It’s SpaceXAI’s agent product. Each bot gets its own cloud computer, signs into your apps, and keeps working after you close your laptop. It answers emails, updates a sales database, processes invoices, or files a software bug.

We looked at it in a recent episode, when three engineers used it to build a company on a live stream. That episode asked whether the bots could run a business, not just build one. Now there’s a first number to test that against.

Start with the number itself. Bloomberg says Grok Bot had four hundred eighteen thousand weekly users as of September fourteenth, up twenty-four percent from the week before.

Work that backward and the prior week was roughly three hundred thirty-seven thousand. That’s a net gain of about eighty thousand weekly users in seven days.

What we don’t know is what a user means here. The report doesn’t say whether that’s someone who assigned a bot real work, or someone who opened the app once. It doesn’t say how many of them pay, or how many are businesses.

That’s not a criticism of the reporting. It’s the limit of a single slide. A company picks what goes on it.

Bloomberg itself called this modest early demand. And the comparison it offered is telling. Meta’s new Muse agent app has been downloaded about four times as often as the standalone Grok Bot app, according to Bloomberg Intelligence.

Downloads and weekly users aren’t the same measurement, so that isn’t a clean head to head. Muse is also free and aimed at consumers. Grok Bot is sold as a work tool. But it does show where the attention is going right now.

Here’s the context the slide doesn’t show. The price to get Grok Bot fell fast.

At launch on August eleventh, it came only with premium plans. The cheapest individual option was Cursor Ultra, at two hundred dollars a month. Business teams paid one hundred twenty dollars a seat.

Fifteen days later, on August twenty-sixth, SpaceXAI announced that Grok Bot was included with every SuperGrok, Cursor Pro, and Cursor Teams plan. Cursor Pro costs twenty dollars a month. That’s one tenth of the launch entry price.

And Grok Bot comes with its own usage allowance on those plans. So people who already paid for one of those plans got a bot added to their subscription at no extra charge.

That matters for reading the number. The count was taken nineteen days after the big bundle expansion. Some of that growth is people trying something they already had access to. That’s still real interest. It’s just a different thing from people choosing to pay for an agent.

The fair version is this. Four hundred eighteen thousand weekly users is evidence of curiosity at scale. It isn’t yet evidence of a business.

Now the model underneath. In that same episode, we noted that Grok 4.7, long delayed, was expected before the live stream and didn’t show up.

It arrived on September twenty-first, six days after the stream began. SpaceXAI calls it its most capable model for coding and knowledge work.

Two details stand out in the announcement. The company says the model was trained to natively understand the Grok Bot harness. In plain terms, it was built to work inside the agent product, not just answer questions in a chat window.

And the price didn’t change. Developers pay the same per token as for Grok four point six.

The company also published benchmark scores. On a coding benchmark from Cursor, it reports forty-six point three percent, up from forty point four for the previous version. It says the new model leads on an electrical engineering test and a legal agent test.

Those are the company’s own measurements. Outside tests will tell us more. By its own numbers, it still trails at least one rival model on coding.

So put the pieces together. The entry price fell by ninety percent in about two weeks. The bot came bundled into plans people already had. And the model tuned for it arrived after the showcase meant to prove it.

None of that means Grok Bot is failing. It means the company is pushing distribution first and hoping value follows. That’s a common strategy. It’s also why the next numbers matter more than this one.

What to watch next. First, whether SpaceXAI reports paying customers or businesses, not just weekly users. Second, whether it discloses how much work the bots actually complete. Tasks finished would say more than accounts opened.

Third, retention. A weekly user count only means something if the same people come back. And fourth, day four. Thursday Arena, the game from last week’s stream, is still online. The site still doesn’t say who runs it.

A number on a slide tells you who showed up. The number that counts is who keeps paying for the work, week after week, without being reminded the bot is there.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

On September 13, SpaceX flew its 700th Falcon rocket, carrying the final three satellites of SES’s O3b mPOWER network. Lena Ruiz looks at what that number measures: a Falcon 9 launch roughly every two and a half days since February, a booster two flights short of Space Shuttle Discovery’s record, and a company already moving staff toward Starship.

On the afternoon of September thirteenth, a Falcon nine rocket lifted off from Cape Canaveral carrying three satellites for a company called SES. The launch itself looked routine. The number attached to it did not.

It was the seven hundredth Falcon rocket SpaceX has ever flown. Six hundred eighty-seven Falcon nines, and thirteen Falcon Heavies.

The booster underneath it was flying for the twenty-ninth time. Less than nine minutes after liftoff, it was standing on a drone ship in the Atlantic, waiting for a thirtieth.

This is Tesorb Signal, and I’m Lena Ruiz. Today we’re asking what seven hundred launches actually measure: the pace, a reuse record one booster is closing in on, and why SpaceX’s attention is already shifting to a different rocket.

Start with speed. SpaceX flew its six hundredth Falcon nine on February fourteenth this year, a Starlink mission from Vandenberg in California.

By September thirteenth, the Falcon nine count stood at six hundred eighty-seven. That’s eighty-seven more in seven months, roughly one every two and a half days.

Here’s another way to see it. On August twenty-fifth, SpaceX flew its one hundredth Falcon nine of twenty twenty-six. In all of twenty twenty-three, the entire Falcon family flew ninety-six times.

So this year passed that full-year total with more than four months still left on the calendar.

Most of those launches carry SpaceX’s own Starlink satellites. That’s what makes the seven hundredth one worth a closer look, because it didn’t.

The three satellites on board belonged to SES. They were the last pieces of a network called O3b mPOWER: thirteen Boeing-built satellites in medium Earth orbit, selling broadband to businesses and to government and military customers.

Each of the final three weighs about nineteen hundred kilograms without fuel. They won’t start working right away. According to Spaceflight Now, they’re expected to enter service in mid twenty twenty-seven, after maneuvering into position and testing in orbit.

That’s a quiet detail about where SpaceX sits. A company that sells broadband from space, in a market where Starlink also sells, still bought its ride on a Falcon nine.

Now the hardware. Falcon’s real advantage has never been the count alone. It’s how many times the same booster flies.

The booster on the seven hundredth launch, known as ten eighty, was on its twenty-ninth flight. Its record includes four missions to the International Space Station, two of them carrying astronauts, the European Space Agency’s Euclid telescope, and twenty-two batches of Starlink satellites.

But the fleet leader is a different booster, ten sixty-seven. On August twenty-fifth, it flew for the thirty-seventh time, a Falcon record. It entered service in June twenty twenty-one.

Launch trackers keep reaching for one comparison. NASA’s most-flown space shuttle orbiter, Discovery, flew thirty-nine orbital missions. It entered service in nineteen eighty-four and retired in twenty eleven.

Booster ten sixty-seven is two flights short of that number, after a little over five years in service. As of this week, it hasn’t flown a thirty-eighth time.

The comparison has limits, and they matter. Discovery went all the way to orbit and back, with astronauts aboard. A Falcon booster is only the first stage. It lands about eight and a half minutes after liftoff.

So matching Discovery’s number would be a milestone for reusing hardware. It wouldn’t mean the two machines did the same job.

The landings tell the same story at fleet scale. When the six hundredth Falcon nine flew in February, SpaceX counted its five hundred seventy-first booster landing. By August twenty-fifth, that count had reached six hundred fifty-four.

That’s eighty-three more landings in a little over six months.

There’s one more detail in Spaceflight Now’s report on that record flight. As of late August, SpaceX had flown forty-four of this year’s seventy-six Starlink missions from Vandenberg in California. That’s more than half.

Spaceflight Now also reported in late August that the company’s workforce at Cape Canaveral is being refocused on Starship preparations. And that brings us to the rocket SpaceX is building to take over much of what Falcon does.

On September twenty-first, the Super Heavy booster for Starship Flight fourteen rolled to the launch pad at Starbase, Texas. Two days later, the ship was stacked on top.

The plan is the first fully orbital Starship mission, deploying twenty-six next generation Starlink satellites. SpaceX says each will add one terabit per second of capacity, and that the batch equals roughly ten times what a single Falcon nine launch of today’s smaller Starlink satellites delivers.

That is the company’s figure. It hasn’t been shown in orbit yet.

When we previewed this flight in August, the target was late August, and the plan was to catch the ship with the launch tower. Both have changed. The ship is now set to splash down in the Pacific instead, the target is no earlier than September twenty-eighth, and the flight still needs its launch license from the FAA.

Why does Falcon matter to that story? Because Falcon’s numbers are the bar. A rocket that flies every couple of days, on boosters with thirty or more flights, is a hard act for any replacement to follow, including SpaceX’s own.

Starship Flight fourteen is the next marker. If it flies on the new date and the satellites deploy and start raising their orbits, SpaceX says they could carry traffic as soon as a few weeks after launch.

Booster ten sixty-seven is the quieter clock. Two more flights would tie Discovery’s thirty-nine. Three would pass it.

And the Falcon count itself may be the most telling number. If Starship starts carrying Starlink, Falcon launches could level off even as SpaceX puts more into orbit. That would be a success that looks like a slowdown.

Seven hundred launches made the Falcon feel ordinary. The next test is whether SpaceX can make something far bigger feel just as ordinary.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

It’s seven in the morning. A robotaxi arrives at the curb, clean and charged.

It’s seven in the morning. A robotaxi arrives at the curb, clean and charged. The passenger gets in, and the car pulls away with nobody in the driver’s seat.

The ride is the part everyone sees.

What happened before it arrived? And what has to happen before the next passenger can get in?

This is Tesorb Signal, and I’m Lena Ruiz. After our recent Cybercab coverage, today we’re following the work around the ride: cleaning, charging, customer support, and getting a vehicle back into service when something goes wrong.

The day we’re about to follow is hypothetical. It isn’t a report of incidents from one Tesla vehicle. But the operating requirements behind it are real, and Tesla describes them in its own filings.

Our imaginary car has completed a few rides when a passenger leaves a coffee spill on the seat. The next person shouldn’t have to sit in it.

Tesla’s Cybercab rider guide says its cabin camera checks readiness after a trip, including cleanliness and items left behind. That can help identify a problem. Removing the spill is another step.

The support page also lists cleaning fees, including fifty dollars for moderate messes. That tells us Tesla has a process for charging a rider when extra cleaning is needed.

It doesn’t tell us whether the fee covers the full cost. There’s the cleaning itself, the trip to wherever it happens, and the time the car spends unavailable.

A vehicle can be perfectly capable of driving and still be unfit for its next fare. The fleet has to manage both conditions.

Later, another passenger realizes they’ve left a phone behind. An empty driver’s seat means there is nobody beside them to notice it as they get out.

Tesla tells riders to submit a lost-item report. If the item is found, the company arranges a pickup process.

There is work inside that simple sentence. Someone has to associate an item with a trip, arrange its recovery, and handle the request. Some steps may be automated. Others may need a person or a physical handover.

Now imagine those requests across a busy fleet. The efficiency of customer support starts affecting the economics of the ride, even though it has nothing to do with how well the car follows a lane.

By midday, our car needs energy. The operator has to decide where it should charge and when to send it, while keeping enough vehicles available for passengers.

Tesla has promoted wireless charging for Cybercab as a way to eliminate manual plugging in. That is a design approach the company presented in its twenty twenty-four quarterly materials.

Even if the connection is automated, the site still has to provide power, the equipment has to work, and the vehicle has to spend time there. A charging method can simplify one task without eliminating the need to plan the whole operation.

That is why depots deserve attention. Where they are, how many vehicles they can handle, and how quickly a car returns to service can influence what a ride costs.

Waymo has already made fleet operations a distinct partnership. In its Miami expansion announcement, it assigned Moove responsibilities covering fleet operations, facilities, and charging infrastructure, starting in Phoenix. Waymo retained responsibility for its automated driving system.

The driving technology and the work of keeping vehicles available are substantial jobs in their own right.

In our hypothetical afternoon, the car stops unexpectedly. The passenger wants to know what’s happening and how to continue the trip.

Tesla’s support instructions direct riders to a representative through the vehicle or app. The company has procedures for interruptions as well as collisions.

Customer support should also be distinguished from assistance to the driving system. In a February explanation, Waymo says its remote-assistance staff respond to requests from its automated driver and provide information the system can accept or reject. It says those staff do not remotely drive the vehicles.

That’s Waymo’s description of Waymo. It doesn’t establish that Tesla uses an identical system. And talking to a rider through a screen doesn’t, by itself, mean someone is steering the car remotely.

For an operator, the practical questions include how often help is needed, how long resolution takes, and when somebody must travel to the vehicle. The answer can change how many cars a support team can keep in service.

Tesla explicitly acknowledges this surrounding operation. Its annual report lists dedicated infrastructure for cleaning, maintenance, charging, security, teleoperations, and fleet management.

The business case depends partly on how efficiently that work can be shared across vehicles.

Here’s a simplified example, using invented numbers. Suppose the total cost allocated to a car for a day is three hundred dollars. If it provides two hundred paid passenger miles, that’s a dollar fifty per paid mile. If it provides only one hundred, it’s three dollars.

Hold everything else constant, and reduced utilization doubles that figure. In a real fleet, some costs also change with mileage, and empty travel consumes resources too. This isn’t a Tesla cost estimate. It explains why time spent between fares matters.

By evening, the useful measure is how many safe, satisfactory trips the operation delivered, with all that supporting work included.

Human involvement can be compatible with a successful autonomous service. One team may support many vehicles. Better software, easier cleaning, and more reliable charging could make that team more effective over time.

The evidence to watch is the amount of support each ride requires, the time cars spend available, and the complete cost of providing the service.

At the curb, the next passenger sees a car ready to go. Making that happen repeatedly is the business Tesla now has to build.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

Imagine you’re responsible for a country’s emergency communications. A major flood has damaged roads and knocked out local networks.

Imagine you’re responsible for a country’s emergency communications. A major flood has damaged roads and knocked out local networks. Rescue teams need to share information, and the people directing them need a connection they can depend on.

Price matters. Speed matters. But you also need to know who can allocate capacity, whose rules apply, and whether the service will be there when everyone needs it at once.

That is the case Europe is making for its next satellite network.

This is Tesorb Signal, and I’m Lena Ruiz. Today, why Europe is spending money to build an alternative to Starlink, what the latest contracts actually cover, and how to tell a useful investment in independence from an expensive promise.

On September tenth, Airbus announced an initial contract to develop the first layer of satellites for Iris Two, the European Union’s secure-connectivity program. Airbus and Thales Alenia Space are responsible for at least sixty-six satellites in that layer, with delivery planned for twenty twenty-nine. They will support government users through military communications frequencies.

That is a concrete industrial step. There are companies assigned to build hardware, a defined initial batch, and a delivery target.

The wider project combines satellites in different orbits with ground infrastructure. The European Space Agency says its intended users include governments, defence organizations, and emergency services. The operator consortium, SpaceRISE, brings together SES, Eutelsat, and Hispasat.

Commercial services are part of the plan too. But the government mission explains why the project cannot be judged solely by whether it offers a cheaper home internet subscription.

For a household, a satellite service needs to work reliably at an acceptable price. For a government, the list of requirements can also include control over access, security arrangements, and the ability to prioritize essential operations.

Europe already has a program addressing those needs. The Commission says GOVSATCOM began operations in January. It pools existing satellite capacity for authorized government users, including countries without their own satellite resources.

So this is a progression from existing services toward additional dedicated infrastructure. Europe has assets and operating experience to build on.

My interpretation is that Iris Two is partly a purchase of options. If one supplier cannot meet a requirement, a government wants another route available. That option can have value even if it is rarely used under normal conditions.

Think about backup power at a hospital. The justification includes the consequences of losing it, not just the average price of electricity.

Satellite procurement still needs a harder examination than that analogy provides. An alternative network has to offer useful capacity, support equipment people can actually deploy, and remain dependable when conditions deteriorate. Ownership by itself doesn’t establish any of those things.

The schedule is where this story becomes especially useful.

ESA describes a progressive introduction of Iris Two services from twenty twenty-nine. Airbus’s initial government-focused satellites fit that early timetable.

Eutelsat, meanwhile, says the program’s commercial low-orbit constellation is expected to become operational by the middle of twenty thirty-two.

Those dates describe different parts of the program. They should not be collapsed into one claim that Europe gets a complete Starlink replacement in twenty twenty-nine.

A government capability can start arriving before the larger commercial service is ready. And both dates remain targets, dependent on the work still ahead.

The distinction matters to customers making decisions now. A ministry planning a future secure network and a business needing connectivity next year are solving different problems.

Eutelsat has an answer for that interval: continue investing in OneWeb, its existing low-orbit network.

Also on September tenth, the company announced plans for another two hundred and twenty-nine OneWeb satellites from Airbus. It put the estimated investment at one billion euros and said it had authorized initial industrial work. The stated purpose includes maintaining service until the commercial Iris Two system is available.

That spending is an important part of the story. While a successor network is being developed, the current network still needs investment. Customers need continuity through the transition.

It also creates a test for management. Can the company sustain today’s service and fund tomorrow’s system without letting the handover become a series of overlapping delays?

Public institutions face a related question. ESA says the reinforced Iris Two design requires additional investment, with funding expected from future EU budgets and participating countries.

The case for independence may justify spending. It cannot make the cost of delivering that independence irrelevant.

For SpaceX, the implication is that some customers will evaluate satellite providers on more than technical performance and price. A government’s security requirements may create demand for a European-controlled service even when another network is commercially attractive.

That doesn’t establish that Iris Two will outperform Starlink, or that European customers will abandon SpaceX. It suggests a market in which different customers buy different combinations of connectivity and control.

The way to assess Europe’s bet is to follow delivery. Satellites completed. Launches secured. Ground equipment tested. Services accepted by the people expected to use them.

If those milestones hold, Europe gains a practical option it can call on. If they slip, the value remains on paper while the need continues to grow.

When the next crisis comes, the procurement argument will become a much simpler question: can the teams on the ground connect?

That’s the outcome this investment has to deliver.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

A robot picks up a part, turns, and puts it exactly where it belongs. The movement looks effortless.

A robot picks up a part, turns, and puts it exactly where it belongs. The movement looks effortless.

Now imagine watching that same workstation after lunch. The bin is nearly empty. One part is sitting at an awkward angle. A sensor needs cleaning. The next station is waiting.

Can the robot keep the work moving?

That’s the test I’d like to see for Optimus. A representative shift, with the interruptions left in and the results counted at the end.

This is Tesorb Signal, and I’m Lena Ruiz. We’ve covered the factory space Tesla is committing to humanoid robots. Today, what that investment needs to produce: useful work, at a repeatable cost, across an ordinary working day.

Tesla’s July quarterly update gives us a clear starting point. The company said it had decommissioned the Model S and Model X manufacturing lines at Fremont and was installing its first-generation Optimus lines.

It also said the initial robots would go to its Optimus Academy for training data collection and further development. Production was anticipated later this year.

That is a development program with manufacturing infrastructure behind it. The disclosure does not give us a full shift’s completed work, intervention rate, and operating cost.

Training robots inside a factory can be valuable. Every unexpected object or failed grasp can help improve the system. But a factory manager deciding whether to deploy one also needs to know how much usable output comes back from the time and money invested.

The number of robots in a building answers only part of that question.

BMW offers a useful reference point. In February, it reported results from a Figure Zero Two pilot at its Spartanburg plant. The robot handled sheet-metal components for the welding process. BMW reported more than ninety thousand components moved and roughly twelve hundred and fifty operating hours, supporting the production of over thirty thousand X3 vehicles.

Supporting production matters. The robot performed one task within a much larger manufacturing process. It didn’t build thirty thousand cars by itself.

These are results reported by the customer and the supplier, rather than an independent audit. Even so, they give us something concrete to examine: a defined job, a volume of work, and accumulated operating time.

Figure’s own account goes a step further. It identifies cycle time, placement accuracy, and human interventions as the measures it used. Its goals included more than ninety-nine percent placement success and zero interventions per shift. Those were stated goals. The article does not establish that every shift achieved them.

That’s a useful standard for the questions we ask of Tesla, and of every humanoid company.

Four things to count. First, good parts out.

How many assignments did the robot complete correctly, within the time the next workstation needed them? A successful movement is only useful if it fits the process around it.

Second, human help in.

Count the resets, the remote assistance, the technician visits, and the time spent arranging materials so the robot can cope. Human support doesn’t automatically make automation uneconomic. What matters is how much is required, and whether that burden falls as the system improves.

Third, time unavailable.

A scheduled shift can include charging, maintenance, faults, and waiting for something elsewhere on the line. Those causes should be separated. A robot waiting for a late delivery has a different problem from a robot that cannot grasp the next part.

Fourth, quality over time.

Look across multiple shifts. Does performance hold when the parts vary slightly, the equipment warms up, or a different team runs the station? One excellent day should be the beginning of a useful record.

Here’s a deliberately simple example. Imagine the total cost assigned to a robot is eighty dollars for an eight-hour shift. That includes whatever ownership, support, and operating expenses we’ve chosen for the example. It is not a claimed Optimus price.

If it supplies eight hours of acceptable work, the effective cost is ten dollars per productive hour. If it supplies four, the cost becomes twenty. The same machine, with the same daily expense, can have very different economics.

And a factory ultimately wants acceptable parts, so the next calculation is cost per good part. A slower robot might still make sense for a particular job. A faster one might lose its advantage through rework or frequent assistance.

This is why reliability can be as commercially important as a new skill. Figure’s BMW report identifies the forearm as its leading hardware failure point and explains that it redesigned wrist electronics for the next robot.

That sort of engineering change is easy to overlook in a demonstration. On a production line, it could determine whether a machine earns another order.

The work is moving forward. In June, BMW described a new Figure Zero Three project for sorting components into the sequence needed by assembly workers. It presented that as the next use case, not a finished proof of broad capability.

For Tesla, the opportunity is to build a similar body of evidence around its own operations, then show that a successful task can be repeated elsewhere without starting the development process again.

My proposed next milestone for Optimus is straightforward: publish the job, the hours, the acceptable output, and the human support required. Then show how those numbers change over time.

A robot that reliably takes over one demanding, repetitive task would be a meaningful achievement. It doesn’t need to do every job in the factory to justify being there.

Give us a shift we can evaluate. That’s how the robot starts earning its place.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

SpaceXAI streamed a 72-hour attempt to build a company from scratch with its Grok Bot agents, and what shipped was a browser game with sixteen hundred players and no revenue. Lena Ruiz measures the demo against the Macrohard promise of a fully AI operated software company and asks who is running the game now that the cameras are off.

Three engineers. One room in San Francisco. No company name. No product. No plan. And seventy-two hours to build a startup, live on camera, with AI agents doing the work. That’s Grok Bot Galaxy. The stream is over. What shipped was a video game. Whether that counts depends on a promise made back in March.

This is Tesorb Signal, and I’m Lena Ruiz. Three things today. What SpaceXAI set out to prove. What a March promise called Macrohard said this would look like by now. And the three tests that separate a demo from a proof.

Start with the setup. SpaceXAI, the AI arm that SpaceX absorbed in February and renamed over the summer, put three of its engineers in a venue on Howard Street. From Tuesday the fifteenth to Thursday the seventeenth, eight thirty in the morning to six at night Pacific, they streamed the build to anyone who wanted to watch.

The rule was a blank slate. No idea chosen in advance. No name. Grok Bot, the company’s agent product, was supposed to handle ideation. Business planning. Product design. Engineering. Deployment. Colleagues dropped into side sessions on sales, customer support, and marketing, to show the agents running those too.

For newcomers, Grok Bot is not the Grok chatbot on X. It’s the agent product. Each bot gets its own cloud computer and keeps working after you close your laptop. It’s bundled into Grok’s paid plans and into Cursor, the coding company SpaceX bought this summer. The long delayed Grok 4.7 was supposed to land just before the stream. It didn’t.

Why does a three day demo matter beyond the demo? Because of Macrohard. We covered it in the spring. Musk relaunched Macrohard in March as a joint project between xAI and Tesla, with a stated goal of a fully AI operated software company, and a first version targeted for September. It even appears in SpaceX’s public offering filing, under a section called future markets.

Well, it’s September. Grok Bot Galaxy is the first public thing that looks like Macrohard in motion. So the question isn’t whether three talented people can make software in three days. It’s whether what they showed matches what was promised.

There’s a money reason too. SpaceX is a public company now. It paid a reported sixty billion dollars for Cursor, and it’s signing multibillion dollar deals to rent out its compute. An agent that runs a business is what those bets are supposed to produce. So this was a product demo with a stock price attached.

Give the format its due. Most AI demos are five minute clips of a task the presenter chose. This was three days, live, from an unscripted starting point. If the agents stumbled, the audience saw it. Very few companies volunteer for that.

And building a product in seventy-two hours with agents isn’t a stunt anymore. Small teams do it every week. The unusual claim here isn’t that it can be built. It’s that the same agents can then sell it, support it, and market it. That’s the Macrohard claim. And that’s the one worth testing.

Think of it like a cooking show. Three chefs, a blank pantry, and a weekend to invent a dish. Impressive if the plate comes out. But Macrohard promised a restaurant. A restaurant means the kitchen keeps running after the cameras leave.

So what counts as passing? Three tests. First, did a real user outside that room use the thing, and did anyone pay for it. Second, how much did the humans type. If the engineers were writing code and the agents were finishing their sentences, that’s a coding assistant, not a company. Third, what happened on day four. An AI operated company is a claim about operations, not about launches.

So what happened? Day one went sideways. The team spent the afternoon on a booking tool for pop up events. Venues, holographic tickets, merch. By the next morning that was gone. Their own whiteboard later summed up day one in a single line. Thirteen pull requests, no customer. Day two, they pivoted to a browser game. Bots fighting bots. Codename Cupcake.

Day three, around ten in the morning, it launched as Thursday Arena. A free auto battler. You buy a team of cards, each one a Grok Bot template, and watch them fight, best of three.

In the first hour, more than nine hundred people signed in with their X accounts to play. By mid afternoon the dashboard showed sixteen hundred signed in players and nearly five thousand public matches, on top of thousands of anonymous practice games.

So test one, users, is a pass. Strangers played it. Money is not. The team wired up payments and put a sponsor slot on the home screen with a one dollar minimum bid. The only bid we saw, in the last minutes of the stream, looked like the team testing its own checkout. No revenue figure ever appeared.

Test two, the labor split, is the surprise. Across three days of screen time, every code window shown on camera belonged to an agent. Roughly two dozen named bots on day one, close to fifty by day three. A chief of staff bot ran the decision polls. A reviewer bot rejected pull requests that lacked proof.

The humans typed instructions, voted on locks, and flipped the settings the bots couldn’t reach. One builder’s account ended the day ranked in the seventeen hundreds, in a game the bots had built.

Test three is where it gets awkward. The stream stopped at 4:29 on Thursday afternoon, one minute before its own scheduled final showcase. The feed went to a slate that said follow the bot account for more. No wrap. No recap. And yet two days later, Thursday Arena was still online, still taking sign ins. Nobody has said who is operating it. That is exactly the question.

A demo proves an agent can build something. Macrohard promised an agent could run something. Those are different claims, and only one of them was on the schedule last week. The sales, support, and marketing bots all performed on stage, on made up customers and sample data. None of them were running Thursday Arena.

What got built isn’t the grade. The grade is the invoice, the commit log, and day four. The commit log looks real. The invoice is empty. And day four is still running with nobody visible at the counter. We’ll keep checking.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

Forty-three videos. That’s how many clips a news investigation found of people apparently asleep at the wheel of a moving Tesla.

Forty-three videos. That’s how many clips a news investigation found of people apparently asleep at the wheel of a moving Tesla. Seventeen of them were posted this year. Earlier this month, Tesla shipped a version of Full Self-Driving that can take the wheel even when you’re driving manually. And a congressman asked the government why the same car tells drowsy drivers to lean on the software harder.

This is Tesorb Signal, and I’m Lena Ruiz. Three things today. What the new software actually does. What’s in the letter to the Transportation Secretary. And why the safeguard and the sales pitch are sitting on the same screen.

Start with the update. Version fourteen point three point nine of F-S-D Supervised is rolling out now. Tesla calls it a new active safety feature set. If the car detects an imminent collision and automatic emergency braking might not be enough, F-S-D can activate on its own and take control. Not when it’s already engaged. When you’re the one driving.

For newcomers, F-S-D Supervised is Tesla’s driver assistance software. It steers, brakes, changes lanes, and handles traffic lights. The word Supervised is doing real work in that name. Tesla’s own documentation says the driver has to stay fully attentive, hands on the wheel, ready to take over at any moment. Tesla has also said a bigger safety push is coming with version fifteen.

This is a good thing. A system that steps in when braking alone won’t save you is exactly the kind of feature safety researchers have asked for. A few other automakers have pull-over systems for unresponsive drivers, but they’re rare and mostly limited to highways. Tesla’s version works anywhere the software works.

And the sleeping videos are driver behavior, not a software defect. Some of the people in those clips told reporters they beat the cabin camera on purpose, with sunglasses or by covering the lens. No monitoring system survives a driver who is determined to defeat it.

Now the letter. On September ninth, an Illinois congressman wrote to Transportation Secretary Sean Duffy. He cited an N-B-C News investigation that found forty-three videos of drivers apparently asleep in moving Teslas, seventeen posted this year. There was also footage of people using laptops and watching movies at highway speed.

His questions were specific. Is the safety agency aware of the videos. Has it tested Tesla’s safeguards when the cabin camera is blocked, or when the driver wears sunglasses or a hat. And do the safeguards actually stop foreseeable misuse, or just log it.

That letter is not an investigation. It’s a request, with a written response due this month. As of this recording there’s no new docket. But there is an old one.

The National Highway Traffic Safety Administration already has an engineering analysis open on roughly three million Teslas with F-S-D, after a series of crashes in conditions like sun glare and fog where the system failed to warn the driver. A separate probe covers traffic violations. Engineering analysis is the last step before the agency can demand a recall. The letter points straight at both.

And the pattern has history. In twenty twenty-three, Tesla recalled about two million vehicles, nearly every car it had on American roads, to tighten the checks on whether drivers were paying attention. That fix was software. The cabin camera was the answer to years of criticism about driver monitoring. Less than three years later, a pair of sunglasses still gets past it.

Then there’s the part of the letter that got the most attention. Since a software update last September, Teslas show a message when the cabin camera thinks you’re drowsy. It reads, Drowsiness detected. Stay focused with F-S-D. A second one says, Lane drift detected. Let F-S-D assist so you can stay focused.

Read that next to Tesla’s own rule. F-S-D requires continuous supervision. The car has just decided you’re less able to supervise. Its answer is to suggest the feature that needs supervising.

The congressman’s line was that the prompt encourages reliance precisely when the driver may be less capable. Alexandra Mueller, a researcher at the Insurance Institute for Highway Safety, quoted by Wired last year, called it a series of conflicting instructions. Tesla hasn’t publicly changed the wording since.

Think of a lifeguard who notices you’re getting tired in deep water and hands you a bigger float. The float might help. It might not. Either way, the right call was to get you out of the pool.

So this month’s news is two stories that don’t fit together. One, the software is getting more willing to take over from a human who is about to crash. Two, the screen is telling a human who is fading to hand over more. Both are Tesla’s design choices. One is about engineering. The other is about which feature stays switched on.

Forty-three videos is not a defect. It’s a pattern. And the regulator doesn’t need a defect to act on a pattern of foreseeable misuse. Tesla’s new override might be the strongest safety feature it has ever shipped. But a safeguard and a sales pitch on the same screen is the thing a regulator notices first.

Two deadlines sit on the calendar now. The written response the congressman asked for by the end of this month. And whatever the engineering analysis produces, because that one has a recall at the end of it.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

Imagine you’re visiting New Zealand from Japan. You’ve rented a car, driven out of town, and stopped for a walk.

Imagine you’re visiting New Zealand from Japan. You’ve rented a car, driven out of town, and stopped for a walk. Somewhere along the trail, your phone loses its signal. You want to send one message home. Running late. Everything’s fine. What would it take for your usual phone plan to get that message through?

This is Tesorb Signal, and I’m Lena Ruiz. Today, satellite roaming. What SoftBank has announced, what a traveler can expect, and why a message from the trail matters to SpaceX’s business.

That’s the sort of ordinary moment behind SoftBank’s latest Starlink announcement. Satellite service that travels with an eligible mobile account, across a national border, and reaches some of the places a cell tower doesn’t.

The ambition is an invisible handoff. The details still matter quite a lot.

On September tenth, twenty twenty-six, SoftBank said it plans to extend Starlink Direct overseas beginning September twenty-eighth. The first destinations are the United States, Canada, and New Zealand. The Philippines is planned for later this autumn.

The announced service supports messaging and selected app data on compatible phones. Eligible SoftBank customers, including selected plans at its other Japanese brands, get it without an additional satellite charge. Some plans require an optional service.

This announcement concerns customers of those Japanese services traveling abroad. It doesn’t give every person who lands in one of those countries a Starlink connection.

For our imaginary traveler, the checks start before departure. The overseas announcement lists iPhone thirteen or newer, with the upcoming iOS twenty-seven or later. Android details are still to come. Customers also need the carrier’s free international roaming service, and iPhone users must enable roaming. Requirements can differ by country.

An extra line in a phone plan is becoming a travel feature. That is a meaningful change in how people encounter satellite technology.

SpaceX describes its direct-to-cell satellites as carrying cellular network equipment in orbit. They can integrate with mobile operators through arrangements similar to roaming partnerships. The basic idea is for a compatible ordinary phone to communicate directly with the satellite.

You don’t unpack a Starlink dish beside the trail. This is a different connection from the broadband terminal you might put on a house or camper.

But the phone still belongs to a network, with an account, a service agreement, and a set of permitted destinations. SoftBank names T-Mobile, Rogers, and One New Zealand as its partners for the first three countries.

The satellite crosses borders on its own. Your service still depends on local permissions and the carrier agreements below it.

That makes the coverage question surprisingly familiar. Travelers already check whether their mobile plan works in another country. Now they need to check which kind of connection comes with it.

Back on the trail, our traveler has a view of the sky and wants to send that short message. That’s a much closer fit for this service than expecting the phone to behave exactly as it does on a strong urban mobile connection.

SoftBank’s service page describes outdoor use beyond cellular coverage, with an unobstructed path to the satellite. Dense trees, buildings, or being inside a vehicle can interfere. It lists selected app data and messaging, and says conventional phone calls and emergency calls are unavailable through this service.

There’s a useful comparison in the United States. T-Mobile’s T-Satellite page advertises selected apps, including mapping and WhatsApp voice chat. It also explicitly describes limited data speeds and possible gaps or time-outs.

Those are T-Mobile’s features. They aren’t a promise that a visiting SoftBank customer gets every one of them. The same underlying satellite network can support different retail products.

And SoftBank isn’t starting this story from zero. T-Mobile already advertises satellite roaming in Canada, New Zealand, and Japan. This is another link in a growing network of carrier agreements.

That distinction will matter more as the services grow. An app may open, while a particular feature inside it is unsupported. Sending a message, sharing a location, and holding a voice conversation are separate capabilities. A satellite icon by itself doesn’t tell you which ones you have.

Now consider the business from both sides.

For a mobile operator, coverage beyond the last tower is a reason for customers to keep their plan. If that coverage follows them abroad, the benefit becomes easier to explain: the service they already pay for reaches more of their trip.

For SpaceX, a carrier partnership is a route to customers who may never visit the Starlink website. The carrier can handle the plan and customer relationship while SpaceX supplies the connection from orbit.

My read is that this gives both companies something they want, while leaving a bargaining question open. If subscribers begin choosing their phone plan because of its satellite coverage, how much of that value belongs to the carrier, and how much belongs to SpaceX?

We can’t answer that from this announcement. It doesn’t disclose the commercial terms between the companies. But it gives us a concrete place to watch that relationship develop.

Put September twenty-eighth on the calendar. That’s the planned launch, not a result we’ve seen yet. After that, the most useful evidence will come from ordinary trips. How consistently messages arrive. How clearly the phone explains a limited connection. And whether the advertised apps work where they’re promised.

A successful experience might be almost forgettable. Our traveler sends the message. Someone at home stops worrying. The phone goes back in a pocket.

That’s the opportunity for Starlink here: becoming part of a service people already understand, at a moment when they need it.

The rocket launch gets the attention. The message arriving is what earns the customer’s trust.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

Tesla says its robotaxis have now driven one million miles with nobody behind the wheel. It took about fifteen months to get there.

Tesla says its robotaxis have now driven one million miles with nobody behind the wheel. It took about fifteen months to get there. Waymo’s count for the same kind of driving is over two hundred twenty million. One of those numbers is a milestone. The other is a scoreboard. Today we sort out which is which.

This is Tesorb Signal, and I’m Lena Ruiz. Three things today. What Tesla actually announced at the Cybercab event. What a million miles does and doesn’t tell you. And the promise that came the next morning, twenty-four hour service, and the gate it’s stuck behind.

Start with the announcement. At the Cybercab launch in Austin earlier this month, Tesla’s head of AI told the crowd the fleet had passed one million miles of unsupervised operation. Unsupervised means no safety monitor in the car and no remote driver in the loop. Just the cameras and the software.

Back in July, on the second quarter earnings call, the figure was three hundred eighty thousand miles. So roughly six hundred twenty thousand miles were added in about six weeks. That works out to something like a hundred thousand unsupervised miles a week.

The jump came after Tesla pulled the in-car monitors from most of its operations outside the San Francisco Bay Area. The service now runs driverless in Austin. Dallas. Houston. Miami. Orlando. Tampa. And the day after the event, production Cybercabs, with no steering wheel and no pedals, started carrying public riders in Austin.

Credit where it’s due. Fifteen months ago this was a pilot with a monitor in the passenger seat and a handful of cars. Now a vehicle with no controls at all picks up strangers on public roads. That’s a hard engineering milestone, and Tesla hit it.

And the pace is the part worth watching. A hundred thousand miles a week is a slope, not a plateau. Tesla’s whole bet is that a car it builds by the hundreds of thousands can catch a fleet built by the thousands, once the software is trusted enough to let the factory become the fleet.

Now the other side of the ledger. Waymo has been driving with nobody in the driver’s seat since twenty twenty. It publishes its mileage regularly. It delivers about half a million paid rides a week across more than a dozen cities. And it runs around the clock, every day, in every market it serves.

Tesla’s million doesn’t come with a city breakdown. Tesla didn’t say how much of it was paid rides versus testing. And the phrase zero notable incidents, which Tesla used in July, sits awkwardly next to the federal filings. Electrek counted at least seventeen reported incidents in government documents as of May. They may all be minor. But notable is Tesla’s word, not the regulator’s.

One more gap in the announcement. Tesla didn’t say how many Cybercabs joined the public fleet, or how many cars are running in each city.

A while back we compared Austin’s service map, two hundred forty-five square miles, to the couple dozen cars actually serving it. The map was marketing and the fleet was the business. The mileage number is the first big robotaxi figure from Tesla that lives on the fleet side of that line. That’s why it matters more than any map ever did.

Then came the promise. The morning after the event, a rider asked on X whether Cybercab could run all night. Tesla’s head of AI replied. Next month or so. The next tech to merge on the version fifteen plan will enable it. Today the paid service runs from six in the morning to ten at night. Sixteen hours. That reply is the clearest timeline Tesla has ever given for overnight driverless service.

Waymo already does this. Its cars have picked up riders at three in the morning for years, in every city it operates. For a ride-hailing business, the overnight hours are where the fleet earns money that the drivers wouldn’t. So the sixteen hour window is one of the clearest signs that Tesla’s service is still a pilot wearing a launch costume.

But read the answer again. It’s gated on a feature he didn’t name, in a software version that doesn’t have a release date. Version fifteen has been the gate for a lot of things this year. The full rewrite. The mass expansion. Now the night shift. Every line on the robotaxi promise ledger seems to route through the same door.

Think of it like a new airline. A million passenger miles with a clean record is a great first year. But the incumbent has been flying for six years, it flies overnight, and it publishes its numbers every quarter. You don’t compare the two on miles flown. You compare them on how fast the gap is closing, and on whether the schedule you just announced is one you can keep.

The million miles is real, and it’s the first robotaxi number from Tesla that measures the business instead of the ambition. But it’s a rate, not a score. Waymo’s two hundred twenty million is the score. The question for the next quarter isn’t whether Tesla can drive one million miles. It’s whether it can do that every single week.

The ledger is open, and October writes the next line. Tesla’s next hard data point is the third quarter update. If the million has become two million by then, the slope is real. If twenty-four hour service slips past next month or so, the ledger gets one more line, and it won’t be the first.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.

Every Starlink satellite now dodges something in orbit about once a week. Add it up across the fleet over twelve months and you get three hundred fifty-five thousand evasive maneuvers.

Every Starlink satellite now dodges something in orbit about once a week. Add it up across the fleet over twelve months and you get three hundred fifty-five thousand evasive maneuvers. Collisions in that time: zero. And that pair of numbers is either the most reassuring thing in spaceflight or the most alarming, depending on which half you read first.

This is Tesorb Signal and this is Lena Ruiz. Today, how big Starlink has become, what the dodge count actually measures, and whether the safety math still works when one company owns two-thirds of everything in orbit.

Start with the size, because it’s changed faster than most people realize. At the start of twenty nineteen, fewer than two thousand active satellites circled the Earth. By this June, an independent count found close to fifteen thousand seven hundred. And roughly two out of every three of them belonged to Starlink. Last month, the constellation passed eleven thousand satellites.

It’s larger than every other active fleet on Earth combined.

That scale makes Starlink the weather. Whatever SpaceX decides about altitude, brightness, reliability, and disposal now shapes low Earth orbit for everyone who shares it, from the space station to the weather satellite your forecast depends on.

So the fairness points, and they’re substantial. First, the record. Seven years, eleven thousand satellites, no collisions. Not one. Back in twenty nineteen, when Starlink had just sixty satellites, a European science satellite had to move out of the way of one, and the fear was that a fleet this size would be unmanageable. It hasn’t been.

And a word on why anyone worries in the first place. A collision in low Earth orbit doesn’t just destroy two satellites. It turns them into thousands of fragments moving at several kilometers a second, each one capable of destroying something else. Enough of that, and a band of orbit becomes unusable for years.

That chain reaction is the nightmare every space agency plans around, and it’s why the dodge count matters.

Second, the reason it hasn’t been. Starlink satellites avoid each other automatically. When the computed chance of a collision crosses a threshold, the satellite fires its thrusters on its own, no human in the loop. And that threshold is many times stricter than the industry norm. The warnings come from a shared system.

American military radar tracks objects in orbit and sends operators an alert when two are predicted to pass close. Most operators get a few of those a year. Starlink’s fleet makes thousands of moves a week, and its satellites act on the alerts without waiting for a person.

This year SpaceX is also lowering about four thousand four hundred satellites to a lower orbit, where a dead satellite falls out of the sky in months instead of years. And it has started sharing its tracking data with other operators.

Now the data underneath the record. The dodge count is climbing faster than the fleet. In twenty twenty-four, Starlink logged about two hundred thousand avoidance maneuvers. In twenty twenty-five, about three hundred thousand. In the twelve months through this May, the figure SpaceX reported to the FCC jumped again, to the number we opened with.

That’s more than forty maneuvers per satellite per year. Each one was a moment where the math said move.

And here is the argument the safety experts are making. Every maneuver is a decision made under uncertainty, and none of them drives the risk to zero. A leading orbital debris researcher in Britain put it this way. He expects a collision involving an operational Starlink satellite, not because SpaceX stops trying, but in spite of all the trying.

Make a million moves at one in a million residual risk each, and you’ve bought yourself a collision. The fleet is now large enough that small probabilities become arithmetic.

And it isn’t only Starlink’s neighborhood. The space station flies below it. Earth observation satellites, weather satellites, and other companies’ broadband fleets fly through it. If a Starlink collision ever seeds a debris cloud, everyone at that altitude inherits it.

Two-thirds of the satellites means two-thirds of the responsibility, and no international body has ever assigned responsibility at that scale before.

There’s a second cost that doesn’t show up in the maneuver count. Astronomers have spent years documenting the streaks Starlink leaves across long-exposure images. SpaceX has darkened its satellites, and it says the newer ones are dimmer. But two-thirds of the sky is still two-thirds of the sky, and the next generation of satellites is bigger, not smaller.

Which brings us to what comes next. The version three satellites waiting on Starship are far larger and carry roughly ten times the bandwidth of the ones flying now. Amazon’s rival constellation, which we covered earlier this summer, is launching too, and so are two Chinese networks. The fleet SpaceX manages so carefully is about to share its neighborhood with fleets it doesn’t control.

So the record and the risk are the same sentence. Starlink has proven that a company can fly eleven thousand satellites without hitting anything, and the proof is three hundred fifty-five thousand dodges in a year. Read one way, that’s the best-run fleet in history. Read the other way, it’s a fleet so large that its own success is the thing keeping the danger at bay, one thruster burn at a time.

SpaceX files these numbers twice a year, and the next report covers the summer. If the count crosses four hundred thousand, the story won’t be that Starlink is unsafe. It’ll be that safe has started to require a lot of work.

That’s the signal. I’m Lena Ruiz with the Tesorb Signal podcast. For more news about Tesla, SpaceX, and Elon Musk’s companies, visit our website at tesorb.com.

This podcast was developed with the help of using AI assistants, including the voice, and undergoes a detailed review during production. Corrections are posted to the episode page.