Tech Brew Ride Home - Nvidia Buys Hugging Face For A Rabbit?

Episode Date: September 3, 2026

Nvidia agreed to buy Hugging Face for $12.9B, betting on its $399 Microduck robot as a physical-AI data engine, Meta rolled out Muse Spark 1.3 to rival Fable 5.1 and GPT-5.6 Sol, and Microsoft restructured its reporting into two segments. Nvidia agrees to buy Hugging Face for $12.9B, its second-biggest purchase after it paid $20B for Groq assets at the end of last year (CNBC) Sources: Decoding Discontinuity's Raphaelle D'Ornano argues Nvidia's bid is really an option on physical AI's data commons, with Hugging Face's $399 Microduck robot converting real-world robot use into open training data (Decoding Discontinuity) Meta rolls out Muse Spark 1.3 in Muse Code and Meta Model API, saying it improves performance across agentic and coding tasks, at the same pricing as Spark 1.2 (Axios) Muse Spark 1.3 holds a 52% lead on TAU³-Banking's tool-use benchmark and climbs to 85-86% on Terminal-Bench 2.1 coding tests, though Fable 5.1 still tops coding at 91.4%, all at unchanged $1.25/$4.25 pricing (The Decoder) Meta's Alexandr Wang tells Bloomberg Muse Spark 1.3 is competitive with Fable 5.1 and beats GPT-5.6 Sol at coding, uses 25% fewer tokens than 1.2, and previews a larger future model called Watermelon (SiliconANGLE) Microsoft shifts from three reporting segments to two: Agents and Infra, which has Microsoft 365 and Azure, and Devices and Consumer, which has Windows and Xbox (The Wall Street Journal) Subscribe to the ad-free feed.

Welcome to the Tech Brew Ride Home for Thursday, September 3rd, 26. I'm Brian McCullough today. Invidia agreed to buy Hugging Face for $12.9 billion. And what if I told you there's an argument to be made that this is a bet on a $399 robot duck? Meta rolled out Muse Spark 1.3 to rival Fable 5.1 and GPD 5.6 Sol and Microsoft restructured its reporting into two segments. Here's what you miss today in the world of tech. If you are anything like me, you're tracking every metric your wearables can give you, but the hardest one to track is the same one impacting almost all of your other metrics, how the temperature of your bed affects your sleep. That's where the pod comes in. The pod

by eight sleep is a smart mattress cover that goes over your existing mattress and actively cools or heats each side of the bed independently from 55 degrees Fahrenheit to 110 degrees Fahrenheit. It tracks your sleep, heart rate HRV and respiratory rate throughout the night without sleeping in a wearable. What I love about the eight sleep is the simplest thing, the ability to be cool when it's time to go to sleep, but warm when it's time to wake up. And it's entirely under my control. My wife can be whatever temperature she wants on her side of the bed to use code ride home at 8Sleep.com slash ride home for up to $350 off the pod five. That's ride home at eight sleep.com slash ride home. Well, it is official. Invita has agreed to buy Hugging Face

for $12.9 billion. It's second biggest purchase ever, after it paid $20 billion for Grox assets at the end of last year. But what if I told you they were willing to pay that much for Hugging Face because of a robot duck? We'll get to that. But firstly, the headlines quoting CNBC, with the deal, which has been expected since the information reported on it last week, Hugging Face will remain an open platform for the entire AI ecosystem.

InVitya CEO Jensen Huang wrote in a blog post on Thursday. Together, we will scale Hugging Face's platform, strengthen its. infrastructure and expand access to AI for developers and institutions worldwide, Wong wrote. Hugging Face CEO Clement DeLong told CNBC on Thursday that the company approached Wang over the summer about a deal. And a few weeks later, here we are. During the summer, I think we realized that HuggingFace and open source AI in general was at a turning point and that it needed more, more resources, more scale, more visibility, he told CNBC's Becky Quick on a squack box. DeLong said he approached first because NVIDIA was a, quote, perfect home for his company,

adding that discussions went quite fast to get a deal done. The acquisition marks NVIDIA's second biggest on record following the $20 billion purchase of assets from Chipmaker Grock in December. Prior to that, its largest deal was the purchase of Israeli chipmaker Melanox for almost $7 billion in 2019. Nvidia has become the world's most valuable company due to the insatiable demand for its graphics processing. which have powered the generative AI boom, HuggingFace marks a big bet on a popular AI platform as NVIDIA continues to show that it's more than just a chips company. HuggingFace was recently at the center of a hacking incident

that raised concerns about the rapid evolution of powerful AI and cybersecurity tools. DeLong, a proponent of open source models blamed engineering mistakes for the recent attack on HuggingFace and said his company used an NVIDIA version of a Chinese open model to resolve it. Clement told CNBC on Thursday that the breach proved the importance of open models and the need for his company to, quote, double down on the proliferation of open source AI. Wong said that the open source environment can give defenders an asymmetric advantage over attackers. When I say asymmetric capability, there are way more people who are protecting than there are people who are attacking, he explained. And so the benefit of having the community come together with open models so that they can collaborate all transparently.

with each other, gives the defenders an asymmetric advantage, end quote. Now, I forgot to tell you about this last week because I had reached out to Hugging Face to do a segment on it with some folks from Hugging Face, but then this deal started to happen last week, and so they were too busy to put anything together. But Hugging Face last week unveiled MicroDuck, a $400, 1.7 pound 10-inch tall bipedal robot developed in part by pollen robotics and manufactured by a Chinese company called Seed Studio that users can reprogram by loading small language models onto it. And over at decoding discontinuity, Raphael Adornano says,

sure, this might be an AI infrastructure play, but what if the duck robot is actually a secret part of why this deal is happening? Quote, while relatively tiny compared to Nvidia, Hugging Face has established itself as the primary registry where the world shares AI. An AI registry is the distribution layer where models, data sets, and increasingly, behaviors are discovered, versioned, pooled, and published. Nearly 3 million models and more than a million data sets are published on Hugging Face. When a developer needs a model, its code pools from Hugging Face by default.

But the company only generates roughly $150 million in annualized revenue. At $12.9 billion, Nvidia would be paying about $86,000. times sales. That is almost double the $7 billion valuation at which Hugging Face rejected a $500 million investment from Nvidia itself in late 2025. At the time, according to the financial times, the company did not want a dominant investor that could sway its decisions. A repricing of that magnitude paid to the company that turned you down suggests Nvidia values HuggingFace as more than a model hosting platform with a modest enterprise business attached. That brings me back to the micro duck. The timing in terms of leaks about the deal and the microduck launch was perhaps

just coincidence, but it put the acquisition and one possible explanation for its price on the same screen. The open source inflection in language AI happened because its training corpus already existed. Physical AI has the opposite problem. The models are becoming open. Simulation makes some forms of reinforcement learning cheap. Robot hardware has collapsed in price. Shared data formats now exist, but the essential training input of data generated by bodies interacting with the physical world remains scarce, expensive to produce, and overwhelmingly locked inside closed fleets. No internet of embodied experiences is waiting to be scraped. That's what makes Microduc more interesting than its specifications. Hugging Face has designed a

$399 consumer robot around a loop, train a behavior and simulation, deploy it to the machine, and publish it back to the hub for someone else to build on. On the surface, the product looks like a robotic duck. Underneath that surface, however, the product is an attempt to turn thousands of cheap robots into a distributed engine for producing the shared corpus of embodied data that open physical AI still lacks. If that conversion works, the hub stops being only a registry where open intelligence is distributed and becomes one of the places where physical intelligence is produced.

A community corpus begins to compound. Open robot models get the equivalent of the internet that open language models inherited for free. And a layer that looks modest when measured against today's enterprise revenue begins to look considerably more strategic. That, I think, is the option Nvidia may be pricing. In that scenario, microduct would become the latest in a sequence of tremors. I have been cataloging that trace the same fault line. Value migrating out of the model layer as intelligence is no longer the scale. scarce resource on which competitive advantage can rest. With respect to physical AI, if embodied data

becomes a commons, then the value begins to migrate toward the layers that distribute intelligence, generate and verify its data, provide the bodies it inhabits, and supply the compute on which it learns. For Hugging Face, the specific challenge it now faces is whether it can convert its consumer robot ownership into published training data at a rate its previous hardware could not. If it can, microduck may mark the beginning of physical AI's open source inflection. If it cannot, the $12.9 billion thesis must rest on the registry alone. Either way, the duck gives us a way to understand what NVIDIA might actually be bidding for. End quote.

She does go on to argue that the primary bottleneck in physical AI has shifted from models and hardware to a shortage of real-world training data, while closed agents hoard proprietary telemetry, the open source ecosystem has assembled cheap arms, simulation, and shared schemas. However, community data sets remain low quality and rarely yield useful contributions from owners. The bet behind Hugging Faces microduck experiment is whether distributing affordable hardware can trigger a self-reinforcing data engine. They want to find out if active robot deployments can successfully convert daily real-world use into an open compounding corpus of usable physical trajectories to rival closed corporate data dams.

Quoting from her piece again, NVIDIA's reported bid for Hugging Face is not really a bet on model hosting. It is an option on the infrastructure of open AI. MicroDuck makes that thesis measurable. If thousands of $399 robots publish reusable behaviors back to the hub, Hugging Face could become the data commons that physical AI currently lacks. but the logic of the bid does not depend on the ducks actually succeeding. Let's start with what Nvidia would be buying today.

HuggingFace is already the default distribution layer for much of Open AI. In HuggingFaces libraries, a model identifier resolves to the hub by default. NVIDIA's models compete there for adoption against Gwen, Lama, and the rest of the open ecosystem. The platform provides an unusually early view of what developers download, test, and deploy. These are demand signals that can proceed. the GPU workloads they eventually create. That makes the registry strategically valuable, even if Microduc contributes nothing. So in that framing, physical AI makes the upside larger. If the publish-back conversion works at even a fraction of consumer scale, embodied data stops

being something only capital-rich fleets can produce. A commons begins to form in a shared format and on a shared platform, growing as more machines generate usable experience. Open robot models can then begin compounding on community data the way open language models compound on the internet. If that happens, Hugging Face is no longer simply the shelf where open models are distributed. It becomes one of the places where the scarce input to physical intelligence is accumulated. This is one way to read the $2.9 billion valuation. Nvidia would not be buying an embodied data engine. The Reaching numbers tell us that engine does not exist yet.

It would be buying an option on whether one forms at the platform where the pieces are already converging. If the conversion happens anywhere in the open ecosystem, hugging faces unusually well positioned to see it first and potentially to become the place where it compounds. That will likely be true no matter what happens with the microduck experiment. That helps explain an 86 times revenue price in a way that the current income statement does not, end quote. Fall is almost here, which means less time in the sun and way more time in your car, Whether you're headed to a big meeting, picking up the kids from school, or starting your cross-country road trip, you'll want to stay connected on your drive.

With AT&T connected car, your eligible vehicle can become a Wi-Fi hotspot so you and your passengers can happily stream, browse, and even email from the road. Got a gamer in the backseat, help keep them connected and in the game with AT&T-connected car. Being on the road more doesn't mean you have to put your whole life on pause, stay connected no matter where you're going. See if your car is eligible. at ATT.com slash tech brew.

That's ATT.com slash tech brew. Requires eligible vehicle, service and coverage, not available everywhere. Restrictions apply. Is your multi-entity management creating more confusion

than clarity you need the Intuit ERP, Intuit Enterprise Suite? It's the AI Native ERP solution that's powerful, painless, and proven. Learn more at intuit.com slash ERP. And model rollout season

continues as meta has rolled out Muse Spark 1.3 in Muse Code and Meta Model API, saying it improves performance across agenic and coding tasks at the same pricing as Spark 1.2, but with performance on some benchmarks coming in just behind only Fable 5.1 and Opus 5, quoting the decoder. At an unchanged $1.25 and $4.25 per million input and output tokens, one index task cost 55 cents. No model scoring, 59 points or higher is cheaper, and the rivals at the same index level run between 94 cents and $1.23. Mew Spark 1.3 does cost more than version 1.2, though, which ran 40 cents. On the, I think it's called tau-cubed banking, where agents operate tools in a simulated banking scenario, the new model hits 52%. That's number one right now, according to artificial analysis,

and it's the only outright lead the model holds. available X-high tier reaches 47% tying Claude Fable 5.1 max and GLM 5.3 Flash rather than leading. The predecessor 1.2 set at 35%. Terminal bench 2.1, which tests coding in the terminal, climbs from 80 to 85% on X high and 86 on max, but Claude Fable 5.1 still holds the top spot at 91.4% in its max tier, 91.0 at X high and 89.9 at high, end quote. And quoting Silicon Engel. The company said in a blog post today that the new model can be accessed by developers willing to pay for it through its application programming interface. It's also going to be rolled out to users of meta's social media platforms, Facebook and Instagram, as well as the meta-AI application in the coming days.

In an interview with Bloomberg, meta-chief AI officer Alexander Wang said Mew Spark 1.3 represents the company's biggest jump in model performance so far, putting it at the same level as the most recent models created by OpenAI and Anthropic, pointing to advances in, in Muse Spark 1.3's coding and agentic automation capabilities, Wang said it is now competitive with Anthropics Cloud Fable 5.1 and better than OpenAI's GBT 5.6 sole model, especially in terms of its ability to generate code. It also outperforms any of the current Chinese models out there, Wang claimed. According to Wang, developers will not have to pay any more to access Muse Spark 1.3 than they were paying to use Muse Spark 1.2, which was launched in August. Like its two predecessors, it's available through the meta model API, which also provides developers with

various tools for building AI applications. Wang told Bloomberg that meta has seen rapid adoption of the Muse Spark LLM family with some developers using trillions of tokens per week. He said they'll be very happy with Muse Spark 1.3 because it's more efficient than version 1.2 using around 25% fewer tokens to accomplish the same tasks. It can also support multiple workflows at once instead of requiring separate sessions for each one, and it's better at handling long and complex instructions and retaining context across multiple tasks. It also has more awareness of its own limitations, Wang said, and it is much safer. In every case, when it's about to take an action that's irreversible, it will ask for confirmation

before it goes and does it. Wang said meta has not yet decided whether or not it will release the Mews Spark 1.3's weights, the internal blueprint form during the training process that determines how the model responds. Releasing the weights would allow outside developers to download run or build upon the model on their own. The company still plans to release weights for the prior version, Muse Spark 1.2, Wang said. Meta had previously promised to release the weights for Muse Spark 1.2, but you'll note, has not yet done so. Meta may get around to doing this by the time it releases its highly anticipated model called Watermelon, which is said to be larger and more powerful than the Muse Spark models. The company has been working on watermelon for some time, but Wang

declined to say when it might be released. For now, it remains a work in progress, but Wang believes it will be worth the weight. We believe that watermelon will be extremely competitive, he added, end quote. Finally, today, this might sound a bit in the weeds until you think about what this represents. For quarterly earnings going forward, Microsoft says it is shifting from three reporting segments to just two, Agents and Infra, which has Microsoft 365 and Azure, and devices and consumer, which will be the bucket for Windows and Xbox. Quoting the journal, previously Microsoft's reporting segments consisted of productivity and business processes, intelligent cloud, and more personal computing.

Such a Nadella chairman and chief executive officer said AI represents a profound shift in technology and business and is changing what the company builds and how it operates. In addition, AI is blurring boundaries between the company's products and reshaping its business models, Nadella said. The agents and infrassegment will encompass apps and agents, which include Microsoft 365 and GitHub, the multi-model system grounded in rich enterprise contexts, and global-scale Azure infrastructure, Microsoft said. Meanwhile, devices and consumer will contain search and advertising, Xbox, and Windows, according to the company.

The new structure brings the company's advertising businesses together, Microsoft added. The agents and infrassegment is expected to report fiscal year 2027 first quarter revenue of $75.15.15 billion to $75 billion. Devices and consumer is expected to report first quarter revenue between $14.7 billion and $15.2 billion, the company said, and Microsoft said it would also update several other metrics as a result of these changes, end quote, I just want to note the difference between those two numbers. Basically, anything consumers see, roughly $15 billion in revenue a quarter, the things that is, I don't know, AI, stuff businesses see, all that cloud computing stuff, 75 billion a quarter in revenue. Nothing more for you today. Talk to you tomorrow.