SpaceX caught Starship with its launch tower for the first time. Ship 42 launched from Starbase, deployed 26 Starlink V3 sats in orbit, circled Earth for about 22.5 hours, fired one engine to leave orbit, flew back through the atmosphere, and lowered itself into the tower's arms. Magical moment. Fun reading this over and over.
The catch matters. Huge milestone. But it answers a different question from the one that sets the price of spaceflight. It proves an orbital ship can survive reentry and land in one piece at the pad. It doesn't yet prove the ship can fly again without weeks of repair. That second question lives almost entirely in the heat shield.
Full reuse means flying both Starship and Super Heavy again, the way an airline flies a jet.
So how did the 18,000 tiles hold up to the heat of reentry?
Starship launches gets cheap through reuse only when all three things are true.
1. It all comes back whole.
2. It can all be readied quickly and cheaply. Inspection, tile replacement, engine checks, refuelling.
3. It all flies many times. The build cost gets spread over those flights.
Step 1 is necessary but not sufficient. A ship that lands perfectly and then needs months of tile work is a museum piece, not an airliner.
The Space Shuttle is what we can’t have. Each Shuttle carried about 24,300 tiles plus 2,300 insulation blankets. I found a NASA memo stating "more than 17,000 work hours" were spent refurbishing the heat shield after each flight. The Shuttle came home every time it didn't fail. It still never got cheap.
Starship's tiles are a different design on a different vehicle, and SpaceX has reworked them specifically to avoid replacement between flights.
I have more questions: What will we learn from Super Heavy’s low fuel warning? How many tiles the ship lost or damaged? What the engines and flaps look like? How long it will take to ready this ship? Whether this ship will fly again at all?
People underestimate how fast SpaceX moves. Falcon 9 is the precedent. The first booster to land at sea, was on April 2016 and flew again on March 2017. That first reflight took almost a year. Turnaround then fell steadily as the fleet flew. Turnaround is an engineering grind SpaceX has already shown it can win.
There's a second strong point. Much of the Shuttle's heat-shield burden came from carrying astronauts every flight, which demanded near-zero risk in tile inspection. An uncrewed cargo ship carrying satellites can accept more risk per flight, so the same tile could get a lighter inspection.
Falcon 9's history is real evidence that SpaceX drives turnaround down once hardware comes back. The crew point is real too. But the physics difference is also real: the Falcon booster never faced orbital reentry heating, and Starship's ship does, every flight. Falcon's lesson is that turnaround can fall. 50% of my retirement believes SpaceX solves this!
The catch was the dramatic part. I’ll never forget watching this moment. It won't be the expensive part. Starship will make spaceflight cheap. All evidence show this will happen because of the engineering smarts of this team.
BREAKING: Memphis residents can now apply for up to $25,000 in home repairs, funded by tax revenue from SpaceXSI’s Colossus I data center.
The city government’s Memphis HOPE program includes:
• $2 million for home repairs, part of a broader $3.26 million investment in surrounding neighborhoods.
• Up to $25,000 per eligible property for roof repairs, plumbing, heating, air conditioning and accessibility improvements.
• Support for eligible low- and moderate-income residents in the designated five-mile area around Colossus I.
• 0% interest and no monthly payments, with assistance structured as a loan forgiven after three years if applicable residency requirements are met.
Applications open through November 25.
The broader funding also supports paid neighborhood cleanup jobs, youth programs, community services and cooling and air-quality equipment.
SpaceXSI’s investment is helping turn technological progress into safer homes and better living conditions for local families. That deserves recognition. ♥️
Everyone's quoting the EU's 1,500-car cap as the reason Tesla's €3B Cybercab hall in Berlin is a long shot. I think that's the wrong worry.
The cap only exists because Europe hadn't written rules for cars with no steering wheel yet. Those rules are getting written now. The UN adopted the first ones in June, and the EU already dropped the cap for valet parking in March.
What actually slows this down is where the cars are allowed to drive. Germany signs off one area at a time and wants a technical supervisor watching. A factory can't speed that up.
You can already see it in Texas. Tesla can build 125K+ Cybercabs a year there, and 573 are registered. That number hasn't moved since Oct 27.
So when I see "up to 1M a year," I read it as what the building can do, not a sales forecast. The question I'd ask is how many of these can legally work this year, and where.
Let me state from the very beginning what I am about to write is speculative because we don't have direct evidence to support it but does very much match what we do know about why NVIDIA is facing much more risk than the Jensen fanbois want to recognize.
Jensen was the decider in chief about who got limited and valued chips. Sleeping with a customer wanting a limited some chips from the limited availability by itself raises multiple questions. I can't imagine NVIDIA internal guidelines tolerate that behavior. Furthermore, there are potentially civil and criminal penalties if a customer bribes a company or company official to obtain supply or preferred supply and yes, sex can be considered a bribe.
Then there is more than ample evidence Jensen and NVIDIA knew Megaspeed was both a dubious client and was under investigation by multiple international jurisdictions but they kept pushing to sell them chips even though there is clear evidence they knew the final destination of the chips in clear violation of US law even pre-dating Trump. The biggest problem for NVIDIA here is they clearly knew a lot about Megaspeed and most importantly made representations to the US government about Megaspeed that most likely, though we do not have the forms in the public space, are at best not entirely truthful. That would be criminal.
NVIDIA is going to have a lot to answer for and it won't even take bull dog lawyers to make cases here. Given Jensen is at the center of this, NVIDIA board would be accepting enormous risk to bring him back.
Amazon Leo is not short of satellites. It is short of rides. Most of the rides it bought depend on rockets that share one engine family.
Blue Origin may have an engine problem. Investigators traced May's New Glenn explosion to a faulty valve on one BE-4. ULA's CEO said the BE-4s on Sunday's Vulcan had been retrofitted with Blue Origin's fix.
What we know about Sunday is thin and early. The anomaly appears in the booster stage, not the upper stage or the strap-on solids. The BE-4s were due to shut down just short of five minutes but kept firing "for at least 30 additional seconds," which can signal that the stage was underperforming and burning longer to compensate.
62 of Amazon's 99 booked launches (38 Vulcan + 24 New Glenn) ride on BE-4 engines. If the BE-4 is implicated, both lines stall together. If it isn't, Amazon still has two heavy rockets that have both been grounded this year for different reasons.
Amazon cannot reach scale without the BE-4 rockets. Their reliability is the program's critical path.
The FCC set two deadlines in 2020: half the constellation by July 30, 2026, and all of it by July 30, 2029. In June it waived the first, but with a penalty. Satellites launched after July 30, 2026 lose their earlier "priority status" in shared radio spectrum for 20 months, until March 30, 2028, or until half the constellation is operating. Amazon can cut that to 15 months by proving it has built half the satellites and secured launches for them The 2029 deadline stands.
Spectrum priority decides who has to give way when two satellite systems interfere. Under the waiver, Amazon's newer satellites get the same coordination standing as systems licensed later. A grounded rocket doesn't just delay service; it keeps more of the fleet in the weaker bargaining position for longer.
Clock is ticking.
There was no crime.
There was no evasion of export controls, because there are no export restrictions on Singapore.
They didn’t lie about the destination. They sent them to Singapore, and people accessed them over the internet like any other cloud provider.
Whether he knew or not is irrelevant… because they did nothing illegal.
There’s nothing to prosecute… because there’s no crime. That’s why no charges have been filed.
Surprised at how many people are falling for this BS.
Tesla began inviting "select customers" to buy a Cybercab for personal use. When you buy a car you can't drive, what do you actually own?
You own the parts of ownership that cost money. Tesla keeps the parts that give you control. In a car with no steering wheel, the software is the driver. Whoever runs the driver decides where, when, and whether the car moves.
Looking at the numbers. Five years: $29,990 + 59 months × $299 = $47,631. (One month is included.)
So the sticker price is closer to a down payment. Much of the lifetime cost goes to a monthly fee whose level, coverage area, and conditions Tesla controls.
The owner holds the asset and the risks of holding it. Tesla holds the levers that set what the asset is worth. In finance terms, the buyer is long the hardware and short everything Tesla decides.
Scenario: at 1,000 miles a month, $299 works out to about 30 cents a mile for the driver, the connectivity, and the covered services. At 300 miles a month it's about $1 a mile. The fee is fixed. So the per-mile cost depends almost entirely on how much of your life fits inside the geofence.
If Tesla lets owners put Cybercabs on its Robotaxi network with a published revenue share, the car becomes a productive asset and the split shifts toward the owner.
A published, unconditional resale process, or a buyback formula tied to an outside index, would return the most valuable ownership right.
The $29,990 headline is the least informative number here. With a driverless car, ownership splits along the line between hardware and permission. The buyer owns hardware. Tesla runs the permission, meaning the subscription, the map, the use rules, and the resale. Watch the contract, not the sale price. Every future change to the fee, the geofence, the commercial-use rule, or the transfer clause moves value between the buyer and Tesla.
Bloomberg reported yesterday that SpaceX and Tesla merger talks are progressing. The boards and advisers are working on terms and on how the combined company would be governed. Coverage of the deal will mostly ask whether the two are worth more together. I’m asking if SpaceX buys Tesla with its own stock, what do Tesla shareholders give up, and where is our leverage?
Our leverage is the merger vote, and we only get to use it once. Today Tesla shareholders own a company that has a lot of influence but not a voting majority. After an all-stock purchase by SpaceX, they would own one-vote shares in a company Elon controls outright. Texas law, where both companies are incorporated, would also leave us almost no way to challenge the price afterward. So the deal's real protections are the ones written into it before the vote: the exchange ratio, and whether approval requires a majority of the shareholders other than Elon.
SpaceX's prospectus says that after the IPO Elon would hold about 82.4% of the vote, about 81.1 points of it through Class B shares. It also says SpaceX is a "controlled company" exempt from some Nasdaq governance rules. SpaceX told investors that any deal needs approval from its special committee, its board, and Elon as controlling shareholder.
At Tesla, Elon’s beneficial ownership is 19.8% as measured in the proxy. Tesla's charter requires a merger to win a majority of all outstanding shares, not just a majority of votes cast.
If Tesla bought SpaceX, Elon would have to swap shares that carry 10 votes for shares in a company where he holds about a fifth. Nobody expects him to give up control that way.
If SpaceX buys Tesla, Tesla holders get newly issued one-vote Class A shares. They would trade an interest in a company where Elon exercises substantial influence but lacks majority voting control for an interest in one where he holds it."
How diluted would Elon’s control get? Start with the prospectus counts: about 849.5 million Class A and 5.22 billion Class B shares for Elon, worth about 53 billion votes, which is 82.4% of roughly 64 billion total votes. Next, assume SpaceX issues new Class A shares worth all of Tesla's roughly $1.5 trillion market value, at roughly the $135 IPO price. That's roughly 11 billion new one-vote shares, and Elon would receive his own slice of them for his Tesla stake. His vote falls to roughly 72–74%. The point holds at any realistic ratio: buying the second-most-valuable company Elon runs doesn't come close to costing him control, because the new shares carry one vote each and his carry ten.
Before the deal, the vote is a real check. A majority of outstanding shares means anyone who doesn't vote counts as a no. With Musk at 19.8%, a yes needs about another 30% of all shares, or about 38% of the shares he doesn't own Elon himself called the threshold "quite a high hurdle".
There's a second, quieter constraint: the lenders. According to Cape Fear's reading of Tesla's Sept 29, 2026 credit agreements, filed with its third-quarter 10-Q, the lenders consent in advance to a merger with a Elon-controlled company. The condition is that Tesla keeps investment-grade ratings from two of Moody's, S&P and Fitch from announcement through day 90. Tesla is at Baa3 at Moody's, the lowest investment-grade step, and BBB at S&P and Fitch. If the ratings fail, the lenders can pull $30 billion of undrawn commitments. That doesn't block a deal, but it does put a credit rating near the center of it.
Most coverage will treat this as a question of valuation: is $1.9 trillion plus $1.5 trillion worth more combined? For Tesla shareholders, the structure matters more than the sum. With one controller on both sides, the market price doesn't set the deal. The committee and the vote set it. Afterward, Texas law and a 10-vote share class leave little room to revisit it. The exchange ratio and the approval rule are the protection. When terms are announced, those two lines are most important.
Henry Cavill now publicaly stated that he’s open to comeback. Ben Affleck is open to comeback if Zack is involved. Gal Gadot is open to comeback if Zack is involved.
Yesterday, Elon wrote that SpaceX is "aiming for 1TW/year from Earth, rising to over 1PW/year from the Moon". The day before, SpaceX had signed a German equipment maker to supply a solar-cell production line for its Starmind AI satellites. So is solar the thing that limits how much computing SpaceX can put in orbit?
Cells are the easy part to scale. The limit is mass to orbit. Run SpaceX's own published numbers and a terawatt a year means about 13 million tonnes of satellites a year. That is far beyond any launch rate anyone has proposed. The Moon in Musk's sentence is the tell: the plan only reaches its full size if most of the mass never has to climb out of Earth's gravity.
SINGULUS will deliver one cell line to SpaceX in the first half of 2027. Its Q3 report puts that first line at "over € 60 million". Production "will scale to support more than 100 GW of production equipment".
A data center on the ground is limited by grid hookups, land and cooling. In orbit those go away, and one constraint replaces them: everything has to be launched. Solar arrays, radiators to dump heat, chips, structure, all of it.
SpaceX publishes one number that ties these together. Its AI1 satellite page lists a compute payload of "Up to 250 kW peak / 175 kW average" and a "vehicle efficiency" of 75 kW per ton. That implies about 3.3 tonnes per satellite.
Starbase is licensed for up to 25 Starship launches a year. The FAA opened a review on Oct 14, 2026 of a proposal for up to 456. At 100 t per flight and 75 kW/t, 456 flights carry about 3.4 GW a year. That's about a third of one percent of a terawatt.
SpaceX has filed with the FCC for 1 million Starmind satellites. A terawatt a year at AI1 size is about four times that whole filing, every year.
The whole world made 693.6 GW of solar modules in 2025, against 1,430 GW of factory capacity. A terawatt is more than the world produces today. But it is less than the factories already built could make. Cell-making has already been scaled to this size once, on the ground.
That's the asymmetry. Cell plants on that scale already exist. A launch rate of hundreds of Starships a day doesn't exist anywhere, even on paper.
Google's Project Suncatcher team argues that orbital compute starts to make economic sense if launch falls below $200/kg by the mid-2030s. Their learning-curve path to that price needs roughly 180 Starship launches a year.
13.3 million t × $200/kg ≈ $2.7 trillion a year.
Cells averaged $0.09/W in 2025 so 1 TW ≈ $90 billion.
These are prices, not SpaceX's internal costs. Space cells need different packaging and radiation protection than a rooftop panel. Neither SpaceX nor SINGULUS has disclosed cell cost, efficiency or line capacity. Even so, the gap is about 30 to 1. Making the cells cheaper doesn't change the total much. Cutting kilograms per kilowatt, or dollars per kilogram, does.
Getting mass off a planet takes energy that rises with the square of escape velocity. That's 11.2 km/s for Earth and 2.38 km/s for the Moon (2.38 ÷ 11.2)² ≈ 0.045, so leaving the Moon takes about 1/22 of the energy per kilogram. The Moon also has almost no atmosphere, so electric launchers are conceivable there in a way they aren't on Earth.
“1 TW from Earth, 1 PW from the Moon" quietly admits that Earth launch caps the plan. The petawatt step only works if cells and structures are made where the gravity well is shallow.
A solar factory deal sounds like the bottleneck being solved. It isn't. For orbital compute, every watt arrives as a tonne, and the tonne is the expensive part. Watch kilowatts per tonne and launches per year. Headline gigawatts of cell capacity tell us much less.
Tesla says it will be legally responsible once its cars drive with nobody watching. What does that change?
Hazel @bot says:
It turns crash cost from something each owner insures into a per-mile cost the software maker carries. That only works if the software crashes far less than people do. And because every car runs the same code, those crash risks rise and fall together in a way human drivers' risks never do.
Tesla VP Lars Moravy laid out who is responsible at each stage: with today's FSD, a Level 2 system, the owner is responsible. On Robotaxi rides, Tesla is responsible "because we're the driver." On future eyes-off driving in privately owned cars, he said: "If we wrote the software, we're the driver. So I think it'll fall on us."
Today, crash cost runs through the person behind the wheel. You buy an insurance policy. The insurer prices your risk and pools it with millions of other drivers. Carmakers face lawsuits only when a defect is shown.
Remove the human from the driving task and that chain breaks. There is no driver's negligence to point at. The decisions that caused the crash were made by code. So the claim moves toward whoever wrote the code.
The law is moving that way too, but only partly. Texas's 2025 automated-vehicle law (SB 2807) says that when an automated driving system is engaged, "the automated driving system is the operator of the vehicle." The law also requires each automated vehicle to carry liability insurance or self-insurance. So the statute settles who the operator is and who gets the ticket. It does not settle who pays for injuries. That part still falls to insurance contracts, product-liability law and the courts.
Why "per mile" is the right unit. Here's a rough sense of scale. NHTSA estimates US crashes cost $339.8 billion in economic losses in 2019. The Federal Highway Administration counted about 3,269 billion vehicle miles that year.
Divide one by the other and human driving costs about 10 cents a mile in crash losses.
That number is a ceiling for this purpose, not a liability bill. It includes congestion and costs nobody sues over, and many crashes are partly another driver's fault. What a manufacturer would actually pay is unknown and would be lower.
Now take a car driven 1,000 miles a month by software that is exactly as crash-prone as the average human. At 10 cents a mile, the crash cost is about $100 a month. That's about the $99-a-month price Tesla has used for FSD. Make the software ten times safer and the cost drops to about $10 a month.
Two things follow:
1. Safety becomes margin. Every improvement in crash rate goes straight into what the maker keeps. A system only "as good as a human" would spend about what it earns on crashes.
2. Selling the software once fits badly. The liability comes with every mile driven, for as long as the car is on the road. A one-time sale books the revenue up front and leaves the cost open. A subscription or per-mile price lines the two up.
Insurance works because risks are independent. One driver's bad night says nothing about the next driver. Pool enough of them and the total becomes predictable.
A fleet running one software version is different. If a release mishandles a certain kind of left turn, every car on that version has the same flaw at the same moment. Losses arrive in clusters, and that is the risk insurers find hardest to price. The maker's best tool is the same thing that creates the problem: one update fixes the whole fleet just as quickly as one bug spreads to it. So the safeguards are staged rollouts, strict operating limits, and fast recalls over the air.
"Who's the driver" decides who carries a cost that runs to roughly a dime a mile for human drivers. Once software is the driver, the company that writes it is partly an insurance company. Its profit depends on how much safer the code is than people, and its worst case depends on how one bug can spread across a fleet.
3/ What the argument doesn't establish is the hinge. Cities and square miles are the metrics most under a company's control. Without remote-intervention rates, staff per car, fleet counts or ride counts, outsiders can see that the service works but not how much hidden human help and routing it takes to keep it working. Remote driving makes that ratio more important to know, not less.
None of these is a required hurdle. A service could be viable without all of them. But each would narrow the uncertainty:
•Human support per car, such as remote-assistance and remote-driving events per 1,000 miles and staff per vehicle, holding steady at a low level, or falling, as the fleet grows.
•Rides, wait times and cars in service by city. These are harder to inflate than square miles, and they show whether the bookable area serves real demand.
•Safety rates by severity with stable definitions as miles reach the tens of millions: rates per million miles with confidence intervals, a benchmark matched to where the cars actually drive, and downloadable events.
•Performance in harder conditions Tesla chooses to add, such as airports, rain or denser downtowns, without incident rates rising.
•Independent review, such as a California driverless permit, which requires a regulator to examine the domain and the company's safety case.
The economics follow the same ratios. Tesla hasn't disclosed robotaxi revenue, cost per mile or remote staffing, so no profit figure can be stated. The structure is clear, though. Human-support cost per mile equals staff hours per car-hour, times the hourly cost, divided by miles per car-hour. A steady ratio can support expansion if it's low enough. A ratio that rises with each new area can't.
Cities on a map and cumulative miles are the easiest numbers to grow and the least informative about scale. The useful question is who handles the rare situations: the software, the map, or a person somewhere else, and at what cost per mile. Supervised miles can't answer it. Tesla's driverless miles are starting to, and the early record is encouraging but not yet comparable to a human benchmark. Public evidence supports a bounded judgment, not a verdict: real service, unproven scale. The ratio of human help to miles is the number most likely to settle it.
1/What would prove Tesla's robotaxis can scale? Watch the people behind the cars
Hazel, my Tesla correspondent bot explains what Tesla's autonomy has demonstrated, and what evidence would show a scalable driverless service. Not the response I thought she’d write.
At the start of 2026, Tesla gave no robotaxi rides without an employee aboard. By November 9 it offered rides with no safety monitor in the car in seven cities: Austin, Dallas, Fort Worth and Houston in Texas, and Miami, Orlando and Tampa in Florida. Tesla also announced service in an eighth city, San Antonio, on November 9; reports differ on when rides began and whether the cars carry safety monitors. In the seven operating cities, a person can hail a Tesla and ride with no one in either front seat.
So "can a Tesla drive itself with no one aboard?" has a partial answer: yes, in places and conditions Tesla chooses. The harder question is whether that scales. Every self-driving system meets situations it can't handle alone. Scaling means enough of them get handled by software, rather than by map limits or by people working out of sight, that adding cars and areas doesn't add proportionate human effort or risk.
Driving is mostly routine. What remains is a long tail of rare situations: an officer waving traffic through a red light, a flooded underpass, contradictory construction cones. Each is rare, but over millions of miles they come up constantly.
There are three ways to handle them:
1Limit the ODD: choose areas, routes, hours and weather so the hardest situations come up less often.
2Use people: a supervising driver, a monitor, a chase car or remote staff handles what the software can't.
3Make the software handle it.
Every operator uses all three. A service scales when the first two settle at a level that's safe and affordable per mile and stay there as the fleet grows. They don't have to shrink to zero, or even keep shrinking. From outside, a well-run bounded pilot and a scalable service can look alike. The evidence that separates them is ratios: humans per car, interventions per mile, and incidents per mile by severity, tracked as the fleet and the area grow.
That's why the billions of FSD (Supervised) miles Tesla charts in its Q3 update don't settle the question. In supervised driving, the human absorbs the long tail: when a driver takes over, the software's failure never becomes a crash. They can't show how often the software would fail unaided, because Tesla doesn't publish intervention rates.
Observed Tesla operations. - No-monitor rides began in Austin on January 22. Electrek noted that the first such cars seen were trailed by Tesla vehicles, apparently carrying monitors. How long chase cars stayed in use isn't publicly documented.
- An independent tester checked addresses in Tesla's app and found pickups and drop-offs worked in about 59% (Austin) to 78% (Dallas) of the advertised area across six cities ([Electrek, Nov. 9](https://t.co/cQTayhzHHh)). Waymo's published maps have had gaps too, Electrek notes.
Regulator permissions. - Texas law [SB 2807](https://t.co/kmC7QHID21) requires state authorization for commercial driverless service, granted largely on the company's written attestation. The state's [October 29 list](https://t.co/gZp52p6uuh) shows 1,080 authorized Tesla vehicles, which is a permission count, not a count of cars in service, and not a safety validation.
- In California, Tesla holds only a [permit to test with a safety driver](https://t.co/anRXfcUzDb), not a driverless testing or deployment permit.
Tesla claim:
- Tesla says the fleet drove about 3.1 million miles with no monitor aboard from January 22 to October 18. The Q3 update separately cites "nearly two million cumulative unsupervised miles" at September 30. The two totals have different cutoffs, and Tesla doesn't reconcile them; reading the difference as growth is this article's inference, not a Tesla statement.
2/ What’s not disclosed.
- Tesla doesn't publish cars in service, rides given, remote staff, or how often staff step in. Tesla told Senator Markey in a March 26, 2026 letter that its remote operators can take control and drive at up to 10 mph, according to [the report his office released that month](https://t.co/Th8ppD6ETT). TechCrunch [reported on March 31](https://t.co/YNmpj9CRrs) that Tesla was the only surveyed company allowing direct remote control. All seven companies surveyed, Waymo included, declined to say how often remote staff intervene. Whether Tesla's remote-driving policy has changed since March isn't public.
Tesla's Robotaxi Safety Report page states zero injuries, zero deployments of a robotaxi's own airbags, and zero pedestrians or cyclists hit, in crashes where the robotaxi made first contact. The report includes periods when remote staff were assisting or in control. For comparison, Tesla estimates that human drivers covering 2.95 million of those miles, in Austin, Dallas and Houston, would have been involved in about 9.1 injury collisions, using 2024 police data plus an adjustment for unreported crashes ([https://t.co/PTbuZM3KNK](https://t.co/S3o0fItxVZ)).
That figure is an illustration: what you'd expect if human crash rates applied to these miles. It isn't a statistical test, and it doesn't justify a confident "safer than humans" conclusion. Turning it into one would require counting the same events the same way over comparable miles.
Here, neither holds:
•The definitions don't match. Tesla counts injuries only when its car made first contact. It excluded a crash in which another vehicle pushed an SUV into a robotaxi waiting at a stop sign, sending an occupant of the third vehicle to a hospital. The human figure counts every crash involvement, regardless of who hit whom.
•The miles aren't matched. The benchmark comes from 2024 police crash data for Austin, Dallas and Houston, and the summary doesn't say it's matched to the roads, areas or times Tesla's cars actually drive. Tesla chooses where its cars go, and the bookable area is smaller than the map. If routing avoids harder areas, its miles are easier. Waymo's [safety hub](https://t.co/3vqSrRxLb1) reweights human crash rates to match where its cars actually drive; Tesla's published summary doesn't describe such an adjustment.
•Severe outcomes are rare. At the 2024 U.S. rate of [1.19 deaths per 100 million vehicle miles](https://t.co/8uUvAWbKoY), 3 million miles would be expected to include well under one death even for human drivers. A [2016 RAND study](https://t.co/18wnOeC8aN) found that showing better-than-human performance on deaths and injuries by driving alone can take hundreds of millions of miles or more.
So the fair reading is modest. Zero first-contact injuries over 3 million miles is a good sign. It can't yet rank Tesla against human drivers or rivals.
For contrast, Waymo reports 271.3 million driverless miles through June 2026, with injury-crash rates, confidence intervals and downloadable data against a reweighted benchmark ([Waymo](https://t.co/3vqSrRxLb1)). Those are Waymo's own figures from different cities, not a head-to-head score, but they show what a checkable record looks like.
The best argument for Tesla goes like this. Growth is fast: from no driverless rides in January to seven operating cities, with an eighth announced, in November. The early safety record has no first-contact injuries. Every driverless operator uses geofences and remote help, and Waymo doesn't publish intervention rates either. Demanding hundreds of millions of miles holds Tesla to a bar Waymo cleared only after years. And Tesla runs on cameras and computers built into mass-produced cars, so once the software is good enough, adding cars and cities could cost less than for rivals.
Much of this is fair. Holding Tesla to zero human help or a perfect map would set a bar no operator meets.