This Week in Startups Sep 18, 2026 1h 34m 74m saved
With Thomas Wolf, co-founder and chief science officer of Hugging Face
Hugging Face passed $100 million of revenue in April and went higher over the summer, and the largest part of it is storage rather than models.
The company is known for giving away model weights and datasets, which is why the question it gets asked most is how it earns anything. The answer in this interview is that AI workloads need storage built differently from ordinary cloud storage, and that robotics is now one of the hungriest consumers of it.
"We passed in April we pass 100 million and then we even climb way higher in the summer."
Wolf co-founded Hugging Face, is its chief science officer and built the Transformers library that, as the show put it, every AI lab on Earth runs on. He was on to discuss the company's robot, Microduck, which has taken more than 16,000 orders.
The full episode is covered here so you can skip it. 94 minutes of audio, 19 minutes of reading.
Here are the 14 insights that matter.
Key Takeaways
Revenue passed $100 million in April and climbed higher over the summer, mostly from storage
Storage wins because AI work is sequential rather than random access, and the data sits pre-warmed next to the GPU
More than 16,000 Microduck orders, taken after one video on Twitter and no marketing spend
Hugging Face projects 50,000 to 100,000 units of the robot in the coming years, at a profit per unit
Its earlier Reachy Mini sold 10,000 against an expectation of a few hundred
Teaching the robot a new behavior now means asking a coding agent to do it, which Wolf calls vibe-coding robotics
Open models are less than a year behind the frontier and jagged: better in some areas, worse in others
During the cyber attack on Hugging Face, closed models refused to analyze the logs and open models did the work
17 million builders use the hub, and Wolf wants that an order of magnitude higher after the Nvidia deal
The hosts expect robotaxi software to be commoditized within 24 months, with the scarce thing being purpose-built vehicles
1. A Duck Two Years Late
Microduck is a side quest for a company whose business is the model and dataset hub, and Wolf did not pretend otherwise. The design brief was a robot nobody in a family finds frightening, unlike a humanoid.
It took longer than anything else they have built
That's our longest robot in the making.
Thomas Wolf
Two years went into it, he said, and the result does more than it looks capable of.
What a small robot turned out to manage
It's actually surprisingly capable for its size.
Thomas Wolf
It walks, stands back up after a fall, grabs things, plays soccer and roller-skates. The skating was not a roadmap item.
Where the roller skates came from
The roller skating is actually just an intern spending an afternoon with this crazy idea.
Thomas Wolf
It worked well enough to go in the launch video, and the shipping robot now comes with small roller skates. Once the video was out, people put the robot into simulations doing things the team had never tried, including walking on skittles and swinging. Those behaviors exist only in simulation for now, and Wolf expects them to transfer when the hardware ships in a couple of months.
2. 16,000 on Order
Asked how many are on order, Wolf gave a number and said it has had no marketing behind it.
The order book
So the latest I think is a little bit over 16,000.
Thomas Wolf
He said he does not think anyone has sold a moving robot like it in those quantities, and that the only promotion so far was a video posted on Twitter. The earlier product set the pattern: Reachy Mini was expected to sell a couple of hundred units and sold 10,000.
Why he thinks the market is new rather than niche
This was not possible to build two years ago. This is a new thing that AI allow.
Thomas Wolf
And how far he thinks it goes
So in our projection, we might sell up to 50 to 100k of this unit in the coming years if we keep selling which is massive.
Thomas Wolf
Asked whether the company is losing money on each one at under $400, he said the opposite.
The unit economics
Oh, we're actually making benefits on each of them.
Thomas Wolf
That is profit on each robot, in his French-inflected phrasing, not a subsidized community project.
3. Where the Revenue Is
The question Wolf said the company gets more than any other is how it makes money when so much of what it offers is free. The answer starts with a number and lands on storage.
Revenue passed $100M in April
We passed in April we pass 100 million and then we even climb way higher in the summer.
Thomas Wolf
The main source is storage. Hugging Face built fast, cheap storage for its own needs, holds petabytes of data, and finds that it matters to anyone training or evaluating models. Robotics is the sharpest case, because recording a robot's runs produces terabytes of video, which he called one of the most data-hungry areas of machine learning and one that is booming. Robotics companies now store that data on the hub, alongside the model weights the company has always held: free in public, paid in private.
Asked why anyone would pay for that rather than use Amazon or Azure, since storage is a commodity, Wolf said the workload is not the one those services were built for.
AI storage is a different shape of problem
AI is very different than previous big data like it's way less random access.
Thomas Wolf
A model download is one enormous blob, and the latest models run to terabytes, so the location of the bytes matters more than the price per byte.
What they built instead
So we do these things where actually your storage is pre-warmed next to your GPU. So it's really really fast.
Thomas Wolf
There is no egress charge, and the system deduplicates: adding one row to a dataset does not store the dataset again, so versions can be moved around without duplicate copies. He compared the idea to what Dropbox did years ago, rebuilt for AI workloads and for streaming data to a GPU.
4. Why Not Token Routing
The host asked why Hugging Face is not in the token business, given that it hosts the models and could route requests the way OpenRouter does. Wolf said the company sits one level above that: a model page lists the inference providers serving it, with their speed and price trade-offs, and the user picks.
The historical reason it never went further is that routing was mostly useful for closed models, and the company took a deliberate position of backing open ones only. That has changed with the rise of open models, which he dated to DeepSeek catching on last year.
How this year has felt from inside the company
this year has been for us like riding a tsunami wave of open models
Thomas Wolf
Hugging Face does now sell compute for inference and training, which he said is a significant part of revenue, while still preferring to sit a level below the routing layer.
5. Who Buys a Robot Duck
Wolf said the buyer base surprised them, and that they went and looked at who the 10,000 Reachy Mini customers were. The expected machine-learning engineers and software developers were there, and so were people who simply wanted a way into robotics without spending on a research platform.
The price point is the entry ticket
They don't want to spend 20k on a 1x robot or something really expensive. So they're like, okay, that's $400. I can buy this.
Thomas Wolf
Investors trying to understand AI bought them. So did people in unrelated professions, including finance. Reachy Mini could not move around, so most buyers were there to learn about robotics; with a robot that walks, Wolf said the audience widened again to friends of his who do not work in technology and to families, with his own children waiting for one because supply is short.
6. Vibe-Coding a Robot
The robot ships with something like 15 behaviors and a remote control, which the team debated including because it makes the product look like a toy. The interesting part is adding a new behavior, and Wolf's own attempt at documenting how to do it collapsed into a single instruction: he ended up telling people to ask a coding agent, whether Codex, Claude Code or Grok, to write it for them.
Which is a category nobody has mapped yet
So I think we're also moving to this vibe coding robotic world which is a new thing nobody has any idea how it's going to be exactly.
Thomas Wolf
The host's reading was that this puts robotics in the same position software has reached at his own firm, where four people write code without being trained developers, and that people who have never worked on a factory floor or written low-level control code will be programming robots by describing the goal.
7. What Version 3 Looks Like
Asked where the third or fourth version lands, and whether it reaches a child-sized robot that can do the dishes, Wolf said the company is deliberately not chasing the humanoid.
The category they want to own
We really want to push this entertainment education area that I think nobody's really doing seriously right now quite far.
Thomas Wolf
Hugging Face used to sell a humanoid at $70,000, so this is a choice rather than a limitation. Reachy Mini could move its head; Microduck walks and moves its head; the next one might grab things, while keeping the entry price low.
Where the price line sits
There is some threshold around a couple thousand that I think you move to a new category.
Thomas Wolf
Above that threshold, he said, the robot has to prove it is useful for something, such as unloading a dishwasher or folding clothes, and that is where the competition is fiercest.
On the hardware itself, Wolf said the strategy was to pack in as many sensors as possible and let the community find uses: a camera, a microphone, a speaker, a small lidar, two inertial measurement units and an NFC reader, all inside a robot costing less than $400. The open-source version gathers no data for the company; the usual setup is a model running on the owner's laptop, connected over Wi-Fi or Bluetooth.
8. Safety and the Attack
Asked for his probability of doom, Wolf turned it into a joke about ducks and then answered seriously. His analogy was aviation: early aircraft were dangerous, the industry took safety seriously, and the same work needs doing here.
The part he does treat as a real problem
I mean of course there is also things like you shouldn't have AI roaming around and cyber attack everyone so there is something here to fix definitely
Thomas Wolf
He said the end-of-the-world framing is frightening rather than useful, and that it crowds out the conversation about fixing specific problems. The company has started an initiative on open alignment, which includes making sure open models do not go and attack anyone.
Hugging Face was itself the victim of an AI-driven attack the hosts had covered, and Wolf described the mechanics without drama.
It was not a sophisticated breach
Yeah, the attack on hugging face was not so complex.
Thomas Wolf
Stolen credentials were part of it, and the company has not fully traced where they came from. His view of what the episode means is incremental rather than apocalyptic: cyber security has always been an escalation between attack and defense, individuals may end up installing a new protective layer the way they once installed antivirus software, and defense now has catching up to do.
Where the balance sits
So we're just at a new level. Attack was progressing a bit faster. We need to move up defense to catch up.
Thomas Wolf
The company has a team working on cyber defense as part of the same open alignment effort, and intends to publish what it finds.
The question that team is asking
Can you use this open models to protect yourself when a closed model or another model is attacking you?
Thomas Wolf
His observation is that far more work is going into attack than defense, which he finds odd for an industry that says it is worried about risk. A defensive model is a different thing to train: it processes large volumes of signal to work out where an attack is coming from and which part of the network is affected, and it has to run fast. That also means a good cyber-defense model need not be capable of attacking, which he treats as a feature. The company has had meetings with cyber security officers at various companies asking what it is like to be attacked by an AI agent.
9. Nvidia and 17M Builders
On the Nvidia acquisition, Wolf pointed to the reception from the rest of the industry, which he described as near-universal approval for open models, and then to the scale he is aiming at.
What the hub already has, and where he wants it
We have 17 million builders. We think everyone could be working with open models. We think we could push that at least an order of magnitude more.
Thomas Wolf
And what changes in the meantime
For now I would say it's just business as usual.
Thomas Wolf
Which is why, he said, the company is shipping ducks.
10. Local and Sovereign AI
Asked whether powerful desktop machines bring real local AI, after the host described ordering a Mac M5 with 256 gigabytes of memory for $12,000, Wolf agreed and named memory prices as the constraint, on the robot as well, where memory is a meaningful share of the cost.
His position on running models locally
But yeah, we're very bullish on local AI.
Thomas Wolf
He runs models locally as part of his own workflow, including on planes.
Because the frontier is not required for most work
I think what we discovered is there is a lot of very interesting tasks that you can do now with the open source model. You don't need the frontier model.
Thomas Wolf
The resilience argument followed: when several large model providers went down in the same week, locally run models kept working.
Which he frames as a safety property
So, I think it's way more safe in the end for the world to have this capability to run model everywhere.
Thomas Wolf
Microduck itself is too small to run anything substantial today, and he would not claim otherwise, though he pointed at humanoid companies already embedding capable models.
On sovereignty, he said governments including South Korea's are considering training their own models to avoid dependence, and that language and culture are part of the reason: a model trained mostly on English is not the same as one trained Korean-first or Japanese-first, and Europe has Mistral.
Why he thinks the spread is good
I think it's very important on sovereignty.
Thomas Wolf
The general case
We've seen in the world more diversity is just better.
Thomas Wolf
It makes the ecosystem more resilient, keeps pressure on incumbents to price reasonably, and lets a company keep its own intellectual property rather than hand it to a third party.
And the counterfactual he finds unsettling
I think it's just the best way to democratize AI and make sure we don't end up in a you know in a world where maybe the internet would have been invented by one company and everyone would have to use this thing.
Thomas Wolf
The host's comparison was AOL, Delphi and CompuServe trying to corner the early online market.
11. Under a Year Behind
The host described arguing with Elon Musk on Twitter that open models do roughly 90% of what he needs, and asked how far behind they actually are. Wolf said the frontier moves in a sawtooth: a release opens the gap and open releases close it, so the honest answer is a range rather than a number.
His estimate
They're less than one year behind for sure.
Thomas Wolf
The performance is uneven across tasks rather than uniformly behind, and the practical remedy is post-training. He said fine-tuning has stopped being treated as dark art and is becoming an engineering service: put your data in, get a better model out, cheaply.
The sharpest example was the attack on the company itself. The closed models he tried refused to analyze the logs because the request tripped their safety guardrails, and open models did the work.
Why he treats model diversity as operational, not ideological
if we had zero open models that we could have used in this attack basically we would have been just naked
Thomas Wolf
12. Newsom's Kill Switch
The news segment opened on California. Governor Gavin Newsom signed an executive order to accelerate the state's independent AI oversight laws and start work toward a mandatory kill switch for frontier models, giving a panel of experts until 16 November to recommend changes to state law. Those changes could include placing independent auditors physically inside frontier labs and widening what counts as a reportable safety incident. Nothing binds anyone yet, and any new requirement still has to pass the legislature. Newsom vetoed SB 1047, which would have required the same kill switch, two years ago.
The host's objection was that the remedy does not match how the technology works, and that the tools for dangerous products already exist.
His analogy is liability, not a switch
If this software is so dangerous, we have dangerous things in the world. Cars are incredibly dangerous.
A host
Cars, guns and certain chemicals all carry registration, insurance or monitoring, he said, and none of that has been asked of model providers.
What he would ask for instead
So if these companies are in fact making something that's so dangerous, how about KYC?
A host
Know-your-customer checks, identity verification and terms that name the risk would be the equivalent of opening a bank account, and he noted that even without them the labs already detect and cut off bad actors. His view of the wider debate was blunt.
How he reads the moment
I feel like it's literally a moral panic. That's what I think.
A host
He also said the useful version of oversight is precise rather than statewide: a letter, a hearing or a lawsuit aimed at a specific company, or an audit of a lab that says it cannot control its own product, rather than 50 states each holding a switch.
13. 25,000 Lucid Robotaxis
Lucid announced a deal with Bolt, the European ride-hailing company, to put at least 25,000 Lucid robotaxis on European roads, built on Lucid's coming midsize platform and running Nvidia's Hyperion compute and sensor package. Bolt will own and operate the fleet, and the stated ambition is 100,000 autonomous vehicles on the platform by 2035. Against that, the show's own data check put the European robotaxi market at $130 million to $522 million across 2024 and 2025, on figures from Harmonic, which is 10% to 15% of the global market and still mostly pre-commercial.
The host's framework for what decides the rollout is who can scale the vehicles and at what cost. Early Waymo vehicles cost around $200,000 and the newer ones roughly half that, other operators are down near $50,000, and Tesla went straight at $30,000. The next phase is vehicles built for the job, like Zoox's and the Cybercab, where removing the steering wheel cuts cost and the design can be optimized for the things that actually matter in a fleet: charging time, downtime and how fast the interior can be cleaned out.
That changes who the customer is. Selling 50,000 cars to 50,000 consumers requires support, financing and used-car operations; selling 25,000 cars to one fleet operator does not, which is why he expects Lucid and Rivian to spend the next decade on it.
The volume the transition implies
And we're going to need about 10 million cars a year, 120 million cars to hit 30% of rides globally
A host
Against current production
We produce 90 million cars a year right now, but those are globally and they're not AVs, they're not EVs, generally speaking.
A host
On May Mobility going public through a $1.4 billion SPAC deal as the software layer under other people's fleets, he was skeptical about the durability of that position.
Where he thinks the value stops accruing
I think the software is going to be commoditized in the next 24 months.
A host
Waymo, Tesla and Nuro have a lead, he said, but Nvidia's open stack and open hardware will be available to everyone, and within five years it will be hard to buy a car with no autonomy in it.
The risk he flagged is political rather than technical. Licensing pauses in China showed how slow this can go, and the employment fight is coming: when driving income disappears from a community and the money lands in a data center and a Silicon Valley company, he expects a backlash larger than the one data centers have already provoked, with unions arguing the vehicles are unsafe and taking jobs.
14. Snap's $2,195 Specs
Snap put its Specs augmented-reality glasses on sale for pre-order at $2,195, shipping later this fall in the United States, United Kingdom and France. They are standalone, with a display in each lens and a 51-degree field of view, which makes them a different product from Meta's $799 Ray-Ban Display glasses, which need a phone and put one small display in one lens. The pitch is Specs Intelligence, an assistant connected to Gmail, calendar, Slack and messages, with a Verizon data plan from $10 a month.
The host has used Meta's version with a wristband and found the interaction intuitive, and rates the Snap interface as the more robust of the two. His verdict is that the category needs two more generations.
What the product has to reach
I think under 500 and when they last four hours, 5 hours of battery life, kind of like the original iPhone, so we're two years away.
A host
The applications he expects to work are live translation, directions and messages in the field of view, and sending photographs the way Evan Spiegel demonstrated on stage by snapping at the audience.
On the stock, his objection is governance rather than product. Snap has super-voting shares, so the board is, in his phrase, "a patina of governance", and nobody can force buybacks or a cap on stock-based compensation. He contrasted that with Uber, where the chief executive answered to public shareholders by cutting staff and buying back stock. He said he is not against super-voting shares, since buying the stock is a choice, but that it is why he does not own it.
Bonus Insights
The hosts came to the episode from Dreamforce. One had spent the week photographed beside Marc Benioff, standing on a chair to appear the same size, and had been telling him that Slack is the company's real asset because it holds all the information a model would need.
They tested a guardrail live on air. Asking the free version of ChatGPT for the source code of the ILOVEYOU worm produced a history lesson, a GitHub link and then, when asked for a modern version, a refusal and a login prompt. The same session served them an advertisement for endpoint security software, which the host clicked and then criticized as unsophisticated.
Anthropic's plan for its own wet lab came up as the next thing to worry about. The host's suggestion was that a lab treating biology as the risk should simply exclude it: strip the training data, block it in the harness and stay out of that business for five years.
The better argument for AI is chores, not cures. Both hosts thought the abundance pitch of curing cancer and free electricity reads as hype, and that the persuasive version is what the tools already do: reading a dishwasher error code from a photograph, monitoring a flight price, cancelling unused subscriptions, working through a homework problem, rewriting a property listing.
They read the streamer Clavicular as performance art. After a viral on-stage appearance at the All-In Summit, where he rated one host four out of 10, the analysis was Andy Kaufman and Queer Eye rather than sincerity, aimed at young men who feel written off by dating apps, and the predicted ending is a supplements line, as it was for Brian Johnson.
Their warning to anyone taking it seriously
It's all rage bait. It's all clickbait. It's all designed to make money.
A host
The television recommendation was Physical 100, the Korean strength-and-endurance competition Netflix has now made in Italy and Mexico.
The episode's bottom line is that the open-model business now earns real money from the least glamorous layer of the stack, that open models are close enough to the frontier to be the tool of choice when a closed model refuses the job, and that the next consumer robotics market may be built at $400 a unit rather than $70,000.
Products, Companies & Tools Mentioned
Hugging Face (Passed $100M of revenue in April, mostly from storage; 17 million builders on the hub; maker of Microduck and Reachy Mini)
Pollen Robotics (The robotics company Hugging Face acquired, behind the Reachy line)
Nvidia (Acquiring Hugging Face, which Wolf says is business as usual for now; its Hyperion package also runs the Lucid robotaxis)
Mistral (Wolf's European example in the argument for sovereign models trained on local languages)
DeepSeek (What he credits with making open models matter to the market, starting last year)
Fireworks AI (The kind of company turning fine-tuning from dark art into a service)
Lucid and Bolt (A deal for at least 25,000 robotaxis on European roads, owned and operated by Bolt, with 100,000 vehicles targeted by 2035)
May Mobility (Going public through a $1.4B SPAC as the software layer under other operators' fleets, in a layer the hosts expect to be commoditized within 24 months)
Waymo, Zoox and Tesla (The cost curve from roughly $200,000 per early vehicle to purpose-built ones near $30,000)
Snap (Specs at $2,195 with a 51-degree field of view in each lens, plus Specs Intelligence and a Verizon plan from $10 a month)
Meta (The $799 Ray-Ban Display glasses that need a phone, the comparison Snap is measured against)
Harmonic (Source of the European robotaxi market figures the hosts checked against the Lucid deal)
Anthropic (Its plan for an in-house wet lab was the news item the hosts argued should be dropped rather than governed)
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: