This Week in Startups Sep 21, 2026 1h 10m 55m saved
Vercel published its own measurement of what developers are running on its platform, and 78.4% of token volume is now going to open-source models. In June it was about half.
Jason Calacanis has been making this call publicly for two years against investors who pointed at paid-model revenue as proof it would not happen. His answer has been that open-source usage does not show up in anyone's revenue line, which is not the same as it not existing.
"The open-source tokens are dark tokens. You can't see them. That doesn't mean they don't exist."
Calacanis founded the show and invests through his own fund; Lon Harris is its editorial director and does the research on air. Monday's episode ran from Elon Musk reviving the hyperloop through Trump's plan to rename AI, Amazon cutting off Meta's shopping agent, and a robot-safety benchmark the show argued is measuring the wrong thing.
The full episode is covered here so you can skip it. 70 minutes of audio, 15 minutes of reading.
Here are the 9 takeaways that matter.
Key Takeaways
The Boring Company says it will build an Austin–San Antonio hyperloop, cutting a two-hour drive to 30 minutes The Las Vegas system raised $3 billion led by the UAE to take the same technology to the Middle East
Calacanis invested in the first hyperloop company and lost all of it — the white paper was published open in August 2013 and somebody else ran with it
A working Texas triangle would turn central Texas into one labor market, which is where the show sees the real economics
78.4% of token volume on Vercel is now open source, against roughly half in June
Trump is creating an "AI force" modeled on Space Force and ran a poll to rename AI
Amazon is defending its position as the front door to commerce, not its terms of service
A new model outputs probabilities rather than words, which its builders say makes hallucination structurally impossible
The viral robot-safety benchmark used language models, not world models, which the show argued makes the result meaningless
Calacanis defended asking a Meta executive about its record settlement on stage at his own summit
1. The Hyperloop Comes Back
The episode opened on Elon Musk returning to an idea he published and then gave away.
The original was a white paper. Calacanis explained the mechanism for anyone who had not seen it: an above-ground tube with the air taken out to cut resistance, so a vehicle inside it can move much faster, running on pylons rather than underground.
The paper is 13 years old
it was originally published by Elon Musk in August 2013.
Lon Harris
At the time, Calacanis said, Tesla was in trouble trying to get the Model 3 out, and Musk released it as an open idea rather than building it.
What is different now is the balance sheet
And Elon has a chip stack and he's got two incredibly stable public companies with a lot of ambition.
Jason Calacanis
The new version is a Boring Company project.
The route and the claim
They're going to be working on a new kind of hyperloop between Austin and San Antonio. He claims the tunnel will cut the, it's now about a 2-hour trip, especially if there's traffic on 35. He's saying it'll cut down the trip between the two cities to just 30 minutes.
Lon Harris
The Boring Company's own account said it would be honored to build it. The existing comparable is in Las Vegas, and Harris noted it recently raised $3 billion in a round led by the UAE to take the same technology to the Middle East. Musk made the comment under an AI-generated video about the world of tomorrow.
Calacanis added his own history with the idea. Shervin Pishevar, a Series B investor in Uber, took the white paper and started a company.
He put money into the first attempt and lost it
Shervin Pishevar, who is the Series B investor in Uber, started a company, Hyperloop I invested in. I lost all my money. It never got anywhere.
Jason Calacanis
He said he checked with Musk at the time and got a magnanimous answer, and that the company was later recapitalized after raising money in the Middle East.
2. Tube Versus Tunnel
Both hosts were careful to separate two things that get conflated.
The Boring Company's existing product is an underground loop a mile or two long, expensive to dig, whose advantage is that there are no stop signs, no red lights and no other traffic — you drive your own Tesla through it.
The hyperloop is a different machine
This is completely different. This is a mono rail. This is a not a Tesla driving in an underground tube. This is an above ground tube that would have in it these cars.
Lon Harris
Above ground, perhaps twenty feet up, on pylons, using magnetic levitation inside a near-vacuum. Building it on pylons saves money against tunneling and handles earthquakes, Calacanis said. The open engineering question Harris raised is what happens to the people inside if a seal breaks and the air is gone.
3. The Texas Triangle
The part of the conversation the hosts got most animated about was not the technology.
Texas has four large cities: Austin in the center, Dallas to the north, Houston to the east, San Antonio to the southwest. Austin to Dallas and Austin to Houston are about three hours by road, Austin to San Antonio about ninety minutes.
At 200 miles an hour, those become commutes
If you go 200 miles an hour and these things go down to 45 minutes, an hour, hour 15, man, that would be dynamite. You could literally live in any of these cities and work in another.
Jason Calacanis
Which is one labor market rather than four
You mean, you could start thinking about basically all of central Texas as like one mega city kind of area.
Lon Harris
The real-estate version followed. North of Austin — Round Rock, Leander, Cedar Park, Georgetown — is already built out. The corridor south toward San Antonio, through Buda, Kyle and San Marcos, is where the hosts see the opening, with the middle stretch compared to the distance between Dublin and Galway. A hyperloop that stopped once or twice along it, Calacanis said, would change what those towns are.
4. The Rail Boondoggle
The comparison neither host could resist was California's high-speed rail, which Calacanis put at $15 billion spent with the San Francisco–Los Angeles corridor abandoned in favor of two towns in the central valley.
His objection is the number, not the idea
It is insane to think we'll spend $150 billion dollars on this thing.
Jason Calacanis
His alternative is Southwest, or driving at 75 miles an hour on self-driving and taking a nap. Harris, who has driven California end to end many times and taken the train from Los Angeles to San Diego for years, said a high-speed rail system makes total sense and the problem is process.
Harris blamed the bureaucracy
It gets it's the bureaucracy. It's the way California legislature and projects are run where you just get these endless boondoggles where they just they keep expanding the vision
Lon Harris
Calacanis rejected the charitable reading and named it as fraud. The reference point for both was Nick Shirley, whose videos on wasteful projects have gone viral, and Calacanis said he told Shirley directly to go and do the same work in Texas and Florida so he does not read as a partisan operation. Both were explicit that this is not a partisan problem, naming corruption on both sides.
Calacanis then proposed a policy, on the basis that he owns the domains for a future run at office.
Double the sentence, triple the reward
I am going to propose that anybody who steals taxpayer dollars gets a 2x multiplier
Jason Calacanis
The second half is a funding model for investigative journalism: give the reporters and whistleblowers who uncover waste a share of what they recover, and double the budget for finding it. He named Frontline, CBS, ABC and the New York Times as outlets that have cut investigative capacity, and said he wants more of that work rather than less.
Harris offered the counterpoint on Texas: smaller state footprint, roughly half to a third of California's per-capita spending, and an oversight culture that reprimanded an official for planning to spend $60 million on office space for a light-rail project when the department already had offices.
5. Trump Renames AI
Over the weekend the president announced a new "AI force," modeled explicitly on the Space Force he created in his first term, with an AI czar to be named.
He also ran a poll asking followers to rename AI, offering superior intelligence, extreme intelligence and supreme intelligence. Harris noted he backed off supreme intelligence because it sounded too much like the Supreme Court — and that it is also the name of the villain that rules the Kree Empire in Marvel's Captain Marvel.
The rename option nobody offered
AI stands for American.
Lon Harris
The White House position, per Truth Social, is that the government will not in any way hinder or stifle AI growth.
6. Back to Work, Nerds
The same section covered the administration's dismissal of the safety letters, including the president calling the concern a hoax at the All-In Summit after meeting Geoffrey Hinton.
The show's summary of the position
Listen, get back to work, nerds, is all that you need to know. These drama queens are causing chaos. If your software is not good enough to release, don't release your software. Keep it in the lab, your lab, your company until such time as it's good enough.
This Week in Startups
The argument both hosts made is that holding back an unsafe product is uncontroversial and that the pacing letters ask for something else.
The objectionable half is the demand that everyone else stop too
But it's the this next level of oh, and also everyone else should have to hold theirs back, too. like it's not fair. We can't just be the ones holding our back. You all have to hold yours back whether or not you feel it's unsafe.
This Week in Startups
The analogy offered was a student asking the professor to extend the deadline for the whole class because his own paper was late. They also pointed at Meta as the counter-example: it had planned to ship Muse, decided it was not ready, and spent another twelve weeks on it, which is what the show said the normal process looks like — internal beta, VIP beta, release.
7. Open Source Takes 78.4%
The measurement came from a Vercel chart tracking what developers actually run on its platform.
The number
78.4% of the tokens are going to open
Lon Harris
Since June, when it was around half. Calacanis had been arguing for this outcome against the counter-argument that paid models account for 99% of tokens.
His answer to that objection
The open-source tokens are dark tokens. You can't see them. That doesn't mean they don't exist.
Jason Calacanis
His forecast on All-In, Harris noted, was 90%, which he said was close.
Calacanis restated the position and the reason it matters commercially.
He expects open source to take 90% of token usage
I believe open source is going to win the day on the large language models and take 90% of the token usage and I think the entire frontier model space could be undercut by open source
Jason Calacanis
The specific threat he named is small verticalized models that run on desktops and laptops. The strategic case he made — addressed by name to the investor Brad Gerstner — is about where a business puts its knowledge.
His warning about handing an agent your whole business
It is foolish for you, Brad, to give your entire business and all the knowledge you have to Anthropic through that.
Jason Calacanis
He compared an open model reaching the market to having an open-source Android competitor, and said there is a serious effort to stop it for exactly that reason.
8. Amazon's Front Door
Amazon began blocking Meta's Muse from shopping on its site on behalf of users. Anyone checking out through Muse gets a popup saying an unauthorized AI agent violates Amazon's conditions of use.
Amazon's stated reasons are that Meta never told it Muse would browse the site, that the agent does not properly identify itself, and that it appears to capture and store customer login credentials. The show noted the two companies are otherwise close: Amazon products sell inside Facebook and Instagram, and Meta recently signed a multibillion-dollar deal to run its AI on Amazon's cloud. Amazon has also sued Perplexity and moved against shopping agents from Google and OpenAI.
Calacanis's first reaction was that Amazon is hurting itself. The reply reframed it.
The fear is losing the first click, not the sale
Amazon's fear is they lose being the front door for one-click purchases to Grockbot, Muse, or any other or Perplexity comment.
This Week in Startups
On the legal position, the show's reading is that Amazon has been losing. A user controls their own desktop, the argument goes, in the same way a user may save and edit a New York Times page on their own machine — publishing it is a different matter. So Amazon is fighting at the corporate level, where its terms of service apply, and can afford to keep appealing for five or ten years.
The demonstration offered was a real purchase. A screenshot of a friend's list of seven short books went to an agent with an instruction to find the paperbacks on Amazon and add them to a cart; the agent came back asking which of several addresses to ship to and which card to use, then checked out.
Why the agent cannot simply be blocked
Grockbot and Muse have a little virtual computer. It's logged in permanently. You put your cookies in, you log in once.
This Week in Startups
The commercial danger the show identified is substitution at the moment of choice: asked whether to buy seven books from Amazon or a local bookseller, the answer will not always be Amazon, and Prime's free delivery is the last defense.
The obvious counter-question — why does Amazon not just ship a good agent of its own — got a two-part answer. It has one, and it works well for a narrower job. Asked about a product, it navigates the review corpus properly, surfacing the questions people actually ask about memory, wattage or whether a cable carries data, and answering from thousands of reviews. What it does not do is take an instruction and execute it. The hosts' summary of the gap was the history: Alexa, Siri and the Echo Dot all existed, and none of them became this.
9. A Model That Won't Talk
A startup called Typesafe came out of stealth with a model, Jev, that generates no text. It returns probabilities, which the company calls calibrated decisions. The founder worked on ChatGPT at OpenAI, and the show said demand after launch briefly exceeded the company's own API capacity.
How it differs from a language model
Jev does not speak. It doesn't talk to you. You just present it with a question that you have and it provides you with like a mode of making decisions.
This Week in Startups
The reported performance is an engineer at Vercel swapping it in for a ChatGPT-based classifier and getting answers five to eighteen times faster with better accuracy.
The worked example the hosts pulled up: an incoming support message about a failed Stripe integration returns "technical" at 85% probability for routing, plus a frustration score that raises the ticket's priority. A conventional model would have returned a sentence describing the same thing.
The structural claim
Jev can only make decisions based on a preset number of parameters. So it literally can't hallucinate.
This Week in Startups
Three output types were described: a choice from a set, a score, and a probability that a statement is true. Harris's own use case was a bot that reads deal coverage and ranks which deals to write about; the point made was that it could return a buzz score directly rather than a paragraph.
The company is also positioning it as a component of recursive self-improvement, on the argument that agents making better independent judgments is what such a loop needs. The show's read is that it is a developer tool for now.
Bonus Insights
The viral robot-safety benchmark used the wrong kind of model
A San Francisco outfit called RoboCurve published a benchmark, Robo Harm, hooking frontier models to robot arms and giving them five unsafe instructions, twenty trials each, across three models. The numbers read out were that one model attempted 97 of 100 trials and completed 60, refusing two on safety grounds; another refused more, but almost entirely on one task; and an open model refused nothing and completed six, because it could not work out how to do the rest.
The show's objection was categorical.
These are large language models that are not world models.
Jason Calacanis
A world model expects to be operating in the physical world and carries that protection, the argument went, and the companies building them are working on exactly this question. Calacanis called the demonstration performative and said the place this belongs is product liability and regulation, using Tesla's summon feature as the example of a company that ships a deliberately constrained version rather than a dangerous one.
The world-model companies, by money raised
Looked up live: World Labs at $1.46 billion and a Series C, Runway at $1.17 billion and a Series D, and a third at about $1 billion still in seed. Field AI came up as a company building software for mobile robots.
Calacanis defended the question he asked Meta's president on stage
At his own summit he asked Meta's Dena Powell McCormick about the company's record settlement, and the exchange went viral. His account is that he asks every guest beforehand whether they want a fast ball or a soft ball, that she asked for the fastest he had, and that he had been getting messages from the audience saying the session was turning into a commercial. He said he respects her for coming and for answering.
The disagreement with his co-hosts is about what a hard question is
Harris's framing was that a conversation in which everyone agrees is not informative, and that probing someone's thinking need not be adversarial. Calacanis said some of the other All-In hosts have publicly said they are not journalists and do not want an adversarial conversation, and that breaking each other's chops is part of the show's brand.
His defense of the label
I was a great journalist and I do random acts of journalism.
Jason Calacanis
A small demonstration of what agents change about publishing
Calacanis described sending an essay from Sam Harris's Waking Up app to an agent, having it transcribed, and posting it — something he said he would never have done manually — and then texting Sam Harris to tell him.
The show's bottom line is that the week's stories all run the same way: the technology is ready before the institutions are, whether that is a tube between two Texas cities, an agent checking out on Amazon, or a benchmark measuring a model doing a job it was never built for.
Products, Companies & Tools Mentioned
The Boring Company (Named as the builder of a proposed Austin–San Antonio hyperloop; its Las Vegas loop raised $3 billion led by the UAE to take the technology to the Middle East)
Tesla (Where the 2013 hyperloop white paper was published from, and the summon feature Calacanis uses as the example of shipping a constrained product)
Vercel (Published the token-share chart: 78.4% of token volume on its platform going to open-source models, against roughly half in June)
Amazon (Blocking Meta's Muse, suing Perplexity, and defending its position as the front door to commerce)
Meta and Muse (The blocked agent, and the product Meta held back twelve weeks rather than ship unfinished)
Perplexity (The agent Amazon sued, in a case the show says Amazon has been losing on appeal)
World Labs, Runway and Field AI (The world-model and robotics companies looked up live, at $1.46 billion, $1.17 billion and mobile-robot software respectively)
Harmonic (The data tool used on air to rank world-model companies by money raised)
Ai2 (Its open model refused none of the benchmark's unsafe instructions and completed six of 100)
Books & Resources Mentioned
Waking Up – Sam Harris (The app whose essay on doing something for the last time Calacanis had transcribed by an agent and posted)
Blink – Malcolm Gladwell (The shorthand for fast intuitive judgment the show used to explain what a "system one" model is doing)
Listen to the full episode
🔴 YouTube
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: