What is true on Compute?

Jensen closed his GTC Berlin 2026 keynote with this Hermes Agent demo as the finale. He introduced it by saying the way to build an agent is to ask an agent to build it for you, so he installed and set up a customized Hermes in one prompt. Then he used Hermes, running on @nvidia RTX Spark, to create a simulated training environment for the @huggingface Microduck. Hermes planned an obstacle course with Nemotron and had GPT-6 Astra build it using FreeCAD, Omniverse libraries, Blender, Newton, and Isaac Lab. Hermes trained thousands of simulated ducks on that course until the validated policy was ready to run on the real robot in the physical world. At Nous, we believe open source is a practical tool as much as a philosophical ideal. It gives users the ability to build powerful, complex workflows chaining numerous local+cloud models, devices, and tools themselves instead of waiting for a company to enable them.

The AI compute shortage is getting WORSE and what's happening with NVIDIA's B200 GPUs right now is absolutely INSANE. According to Ornn Exchange, rental prices for NVIDIA's B200 GPUs just reached a new all time high of $8.54 per GPU hour, with prices climbing another 12.9% in just the past trading week. Normally, as technology gets older and more capacity comes online, you would expect rental prices to gradually decline. But we're seeing the exact opposite happening with B200s. Rental prices are moving higher, which shows that customers are willing to pay a premium just to secure access to compute. This is an incredibly BULLISH signal for the economics of AI infrastructure because higher utilization and stronger rental pricing can translate into better returns on deployed GPUs, especially for operators that have already secured capacity. And this is exactly why I remain so bullish on the long term opportunity for neoclouds. These companies are spending billions building out AI infrastructure and one of the biggest concerns investors have had is whether all this capacity will eventually lead to oversupply and collapsing rental prices but right now, the data is telling us a VERY different story. And demand for compute is already this strong TODAY, imagine what happens as these technologies begin scaling across the global economy.

$NBIS This is officially my new favorite website: B200 GPU rental price? +330% today. https://t.co/sRa2mWmwZD

American chattering classes love to hate our country’s financial services sector, but mark my words: the depth and versatility of U.S. capital markets will be our mightiest weapon in the singularity. Anti-finance sentiment, still fashionable on both left and right, is low IQ. https://t.co/UdGoe7G409

Nebius just secured a $3.5 billion financing facility from Citi to help fund its global AI infrastructure expansion, including data centers and GPU purchases. What makes this particularly interesting is the financing terms. The facility has a 364 day maturity and carries an initial interest rate of SOFR + 2.75%, potentially increasing to SOFR + 3.50%. SOFR is essentially a benchmark interest rate that banks use when lending money, meaning Nebius is borrowing at a relatively modest premium above that benchmark. Now, remember one of the biggest concerns surrounding neoclouds recently has been their ability to secure financing at attractive rates. Investors have been worried that companies like Nebius and CoreWeave would struggle to raise the billions of dollars needed to continue expanding, especially with borrowing costs remaining elevated and credit markets becoming increasingly cautious. Well, this announcement directly challenges that concern, at least for Nebius. Not only has Nebius secured another $3.5 billion in financing, but it has managed to do so at what I consider fairly attractive rates for a capital intensive AI infrastructure company. And here’s another important point. Nebius already has more than $40 billion in long term customer commitments from investment grade customers, including Microsoft and Meta. Those contracts give lenders greater confidence that the infrastructure being financed will eventually generate revenue. I’ve said this repeatedly, I don’t believe every neocloud will survive this AI infrastructure buildout. Access to capital is going to separate the winners from the losers, and Nebius continues to demonstrate why I believe it will be among the survivors. This is a very encouraging development for $NBIS shareholders, especially after all the concerns surrounding financing and the recent selloff.

This is Kimi K3.1's voxel pagoda vs frontier models I also did GPT-6 Astra so you can see what I mean about it having similar design style, I attached carbon too from checkpoint 1 i'm redoing it now with the new checkpoint The actual tower is near identical in design to Astra, and you know what that means 🙏

How do you teach a humanoid robot to walk without programming every move? Reinforcement learning step by step 1. Build a simulation (MuJoCo or NVIDIA Isaac Lab) with joints motors sensors and a physics model. 2. Define the task: walk forward while keeping balance. 3. Write a reward function points for moving toward the target staying upright and matching the desired speed. Penalties for falls, wasted energy and unstable movements. 4. A neural network, the policy, reads the robot's state (joint positions, joint velocities, body orientation) and outputs target positions or torques for each joint. 5. An algorithm like PPO updates the policy over many episodes and millions of simulated time steps. 6. Close the sim-to-real gap with domain randomization, then test carefully in the real world.

How do you teach a humanoid robot to walk without programming every move? Reinforcement learning step by step 1. Build a simulation (MuJoCo or NVIDIA Isaac Lab) with joints motors sensors and a physics model. 2. Define the task: walk forward while keeping balance. 3. Write a reward function points for moving toward the target staying upright and matching the desired speed. Penalties for falls, wasted energy and unstable movements. 4. A neural network, the policy, reads the robot's state (joint positions, joint velocities, body orientation) and outputs target positions or torques for each joint. 5. An algorithm like PPO updates the policy over many episodes and millions of simulated time steps. 6. Close the sim-to-real gap with domain randomization, then test carefully in the real world.

Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.

🇺🇸 @ElonMusk’s two biggest companies, SpaceX and Tesla, report earnings this week. @SpaceX reports today, @Tesla tomorrow. Both are aggressively building for SI: last quarter alone, SpaceX put $15.8 billion of its $18.4 billion investment budget into it. That’s the kind of infrastructure spend that can actually scale superintelligence, robotics, and mass-to-space. Investors are also watching for any further alignment between the companies. Tesla’s new shareholder filing includes standard change-in-control provisions in Elon’s long-term stock awards (common in executive packages to protect incentives in a major deal). Any combination would still need the proper process, but the overlap in AI, chips, energy, and robotics keeps growing.

The memory shortage is now spreading to the parts UNDERNEATH the chips. According to Korean media, Samsung Electro Mechanics just signed a 360.5 billion won supply contract, about $267 million, for package substrates used in memory semiconductors.The customer is a large global company that was not named and the deal covers all of 2027. What makes this interesting is that it is the company's FIRST long term supply agreement for package substrates because until now its long term deals were for components like capacitors and inductors. It is also the seventh long term agreement Samsung Electro Mechanics has announced this year, bringing the total to roughly 4.3 trillion won. Customers do not lock in supply a full year ahead unless they are worried about not getting enough of it. I think this is another sign the AI buildout is creating bottlenecks across the ENTIRE supply chain, not just in HBM and DRAM.

You're not ready for this. One of the most human-like humanoid robots has just entered production. Tesla says Optimus production has started on its high volume manufacturing lines in Fremont with the first production quality units coming off the line. Tesla says the current process is suitable for a commercial product, although the hardware and manufacturing methods will continue to evolve. The company has also outlined plans for a Fremont line capable of producing up to 1 million robots per year. That is a target capacity, not its current output. How quickly can Tesla move from early production to manufacturing humanoid robots at scale? And what production numbers will we see over the next 12 months? Source @ SawyerMerritt

China wants to control exports of its most powerful AI models. A new draft adds 23 AI technologies to its export control list, including large models, AI agents and computing clusters. Frontier models above 2 trillion parameters, or trained with more than 10^26 FLOPs, would need a license before their weights or source code leave the country. Open source models that are already public are exempt.

Each new AI chip generation uses more power, but the amount of work it can do is growing much faster. The chips are getting more efficient. Our demand for AI is just growing even faster. $NVDA https://t.co/KhI9xEfmwT

Wells Fargo sees $NVDA shipping ~23 GW of AI infrastructure capacity in CY27 / FY28: $315 PT Vera Rubin supply chain is 2x that of Grace Blackwell and can assemble a VR rack in 5 minutes today vs two hours previously. https://t.co/Owd15LJsVF

Q3 2026 Earnings Call https://t.co/OHCTNvqSqK

First, Siemens (cuLITHO) and now Synopsys (CUDA-X) endorse Nvidia's GPU solutions for EDA. For computational lithography, it's imec using Nvidia CUDA-X libraries with the S-Litho modeling tool. https://t.co/oClLUhhDAK

We’ve spent the past year developing new primitives for GPUs. We believe this enables new algorithms to change how machines learn, open new markets, and dramatically reshape AI’s cost structure. We have raised a $1 billion Series A, co-led by Index and Lux. https://t.co/VUkQ68gL3U

A Python pipe buffer was holding camera frames until 64 KB had piled up. That quietly added hundreds of milliseconds to our π0.5 rollouts. New from @castorhat: how we found it, and what changed on the robot once it was gone. https://t.co/xTAueR4wOV

Tesla's Q3 earnings deck: "Optimus production began in Fremont on our high-volume lines. The bots coming off this line are the first production-quality bots that we have manufactured. While we will continue to iterate on the hardware and manufacturing process, we have proven that the current manufacturing process for Optimus is suitable for a commercial product."

Why is nobody talking about this? ToddlerBot is a 0.56 m, 3.4 kg humanoid robot built at Stanford by Haochen Shi, Weizhuo Wang, Shuran Song and C. Karen Liu. Its goal is to make robot learning cheap and easy to reproduce. The full hardware costs under 6,000 USD. Hardware - 30 active degrees of freedom: 7 per arm, 6 per leg, 2 in the neck, 2 in the waist - Dynamixel servos (XC330, XC430, 2XC430, 2XL430 and XM430 in the current open-source configuration) - 3D-printed structure: a full repair takes about 21 minutes of printing and 14 minutes of assembly - Jetson Orin NX 16 GB, two fisheye cameras, an IMU, two microphones and a speaker Sim-to-real - Zero-point calibration and motor system identification produce a digital twin of the robot - Walking: an RL policy trained in MuJoCo MJX and transferred to the real robot with no extra training, up to 0.25 m/s in any direction and 1 rad/s in place - Push-ups and pull-ups: keyframed motions, also transferred with no extra training Manipulation - RGB diffusion policies trained on 60 teleoperated demonstrations per task - Real-time onboard depth estimation at 10 Hz from the stereo fisheye cameras - A policy trained on one robot runs on another, and two robots tidied a room together Measured limits - Payload: 1,484 g, about 40% of its body weight - Endurance: 19 minutes of walking before motor heating affects stability - Durability: up to 7 falls before a part breaks - The cartwheel works, but its success rate is still low

$META on Muse: "I think one of the biggest constraints so far is that because this is a compute-intensive service to run, we have not launched it globally yet." Meta wants more compute, not less!

Q3 Shareholder Update → https://t.co/u0UcDOuMNa Highlights – FSD Supervised attach rate grew to over 75% of new vehicle deliveries in NA – Unveiled the next-gen Roadster – Record deliveries in Canada, Australia, Japan & Taiwan – Tesla Semi now in volume production. Top variant can go up to 500 miles at a full 82,000 lbs while achieving efficiency of 1.7 kWh/mile – about a third of the energy a comparable diesel truck burns for the same load – Ended the quarter with the largest backlog for our passenger vehicles in 4 years! FSD Supervised – 4 new approvals since Q2 (Croatia, Slovenia, Czech Republic, Slovakia) – FSD is now live in nine EU countries & used by over 100,000 customers logging more than 160 million kilometers since initial approval Robotaxi – Deployed Cybercab into our Austin fleet – Hit nearly 2 million cumulative unsupervised miles, 8x growth in the quarter – Austin service now runs 6 am to midnight, with wait times continuing to come down as we add cars & improve the orchestration algorithms – Added Grok to Cybercab with Grok Bot & Connectors, letting you interface with your inbox, calendar, files, chats and tasks, all by voice. Transport, office & living room in one Robotics – Started production of @Tesla_Optimus on our high-volume lines in Fremont. These are the first production-quality bots we've built, proving that the current manufacturing process for Optimus is suitable for a commercial product SI Compute – On track for 400 MW of training compute by year end, feeding both vehicle & humanoid autonomy – Austin Fab construction continues, equipment commissioning planned for 2027 Battery – Battery supply is still the limiter on how fast we ramp vehicles globally, so we're expanding on every front: pack capacity in Berlin, cathode and lithium refining in Texas, LFP cells in Nevada – Shipped the first Megapacks built entirely with our own LFP cells – Increasing 4680 output to support ramp of Cybercab, Semi & Model Y Energy – Megafactory Texas is up & running, with Megapack 3 production started – Highest-ever order rate & backlog for energy storage Grid constraints & power availability are making Megapack the answer for SI infrastructure demand: it excels at supplying large data center loads, smooths out demand spikes, helps new projects meet grid connection rules & makes the grid itself more stable All with less local fossil fuel generation Supporting Infrastructure – Added 2,100+ net new Supercharger stalls, up 14% year over year

BOOOM 🚨 $TSLA's Q3 2026 update is out. Non-GAAP EPS came in at $0.48, ahead of the roughly $0.43 to $0.45 analysts expected 🔥 But the most revealing parts of the deck have little to do with EPS. They show a company steadily shifting its weight from cars to robots, robotaxis, and AI 🔥 🧠 FSD is now on more than three out of four new cars in North America. Over 75% of new North American deliveries included FSD in Q3, up nearly 20 percentage points from Q2. Active FSD subscriptions reached 1.74 million, up from 1.48 million last quarter and 1.04 million a year ago. That count includes both upfront purchases and monthly subscriptions, and it leaves out free trials. Europe is opening up too. Croatia, Slovenia, the Czech Republic, and Slovakia have approved FSD (Supervised) since the end of Q2, bringing the total to nine EU countries. More than 100,000 European customers have already driven over 160 million kilometers on it, or nearly 100 million miles. Tesla says FSD subscriptions helped lift revenue and that FSD is driving demand wherever it's available. Put those together, and software is becoming a real revenue stream that can grow across the cars already on the road, not just with new sales. The outlook points the same way, with Tesla expecting hardware profits to be joined by faster growth in SI, software, and fleet-based profits. 🤖 Optimus has moved from prototypes to a real production line. Optimus production has started on Tesla's high-volume lines in Fremont. The bots coming off that line are the first production-quality Optimus units Tesla has ever built, and Tesla says it has proven the current process is ready for a commercial product. The capacity table lists Fremont as "Early Ramp" and shows a second Optimus site already under construction in Texas. Tesla hasn't shared a capacity number for either one yet. This is the jump from prototypes to manufactured units, and that's the step that matters most for bringing costs down. Building a second site before the first one has even ramped suggests Tesla expects demand well beyond what Fremont can handle. The missing capacity figure is a reminder that the ramp is still early. 🚕 Cybercab is now giving rides in Austin, and unsupervised miles are growing fast. Tesla started Robotaxi rides with Cybercab in Austin on September 3. By the end of the quarter, the Robotaxi service had logged nearly 2 million cumulative unsupervised miles, eight times the total from three months earlier. Austin service now runs from 6am to midnight. Texas now shows more than 125,000 units of installed annual Cybercab capacity with "Production" status. Tesla points out that installed capacity isn't the same as the current production rate, and it says battery supply is still the main constraint on vehicle volumes. It's ramping 4680 cells specifically to support Cybercab and Semi. There's also a telling line in the outlook. Tesla says deliveries will depend partly on "allocation decisions between sale to customers and use for our owned and operated fleet." In plain English, Tesla is openly weighing how many cars to keep for its own network instead of selling them. With the tooling listed as in production, batteries look like what will set the pace for Cybercab. Inside the car, riders get Grok with voice access to their inbox, calendar, files, and tasks, and Tesla pitches the cabin as "a second living room." Tesla is clearly thinking about what riders do during the trip, not just the trip itself. ⚡ Tesla plans to grow its AI training compute by about 50% in a single quarter. Cortex 1 and Cortex 2 in Texas are rated at more than 105 MW and 120 MW, and Tesla has another 40 MW of compute on top of that, for roughly 265 MW in total. The goal is 400 MW by the end of the year, and Tesla says this compute trains autonomy software for both its vehicles and its humanoid robots. Tesla is also building its own chip fab in Austin and plans to start commissioning equipment there in 2027. It hasn't said yet which chips the fab will make. All of this shows up in the spending. R&D rose to $2.76 billion, up 69% from a year ago, and capex hit $4.73 billion, more than double last year's $2.25 billion. Even so, free cash flow stayed positive at $1.5 billion, and cash grew to $45.4 billion. Because one compute pool trains both FSD and Optimus, progress on one can feed the other. A chip fab is one of the boldest vertical integration moves Tesla has made, and staying cash-positive while building it makes that bet much easier to fund. 🔭 Tesla is building every piece of its next chapter at the same time. Robots are coming off a production line, driverless Cybercabs are carrying riders, FSD is turning into recurring software revenue, and the compute behind all of it is about to grow by half. Each piece makes the others stronger, and together they point to a Tesla whose future depends less on how many cars it sells and more on what its AI can do.

Why hardware specialization is back

This might be one of the most minimalist snake robots ever built. • One motor. • No dozens of actuated joints. • Crawls, climbs between walls and swims. It's called SAW (Single Actuator Wave-like robot), developed by Dr. David Zarrouk's Bio-Inspired and Medical Robotics Lab at Ben-Gurion University of the Negev. How it works • A single motor rotates a helix inside the robot's body. • 3D-printed links follow the helix to create a continuous wave. • The links push against the surrounding surface to propel the robot. • Front wheels handle steering. The numbers • 57 cm/s on flat ground • Around 5× faster than similar wave robots • Around 8 cm/s in a tunnel, with contact against both walls • Around 6 cm/s while swimming in its waterproof version One gait, multiple environments • Gravel, sand and grass • Obstacles and narrow passages • Pipes, tunnels and water The same wave-like motion works across these environments without changing the gait. Why use one motor? Traditional snake robots often rely on 10, 20 or 30 motors. SAW uses a single actuator, reducing the number of moving parts and making the design lighter, simpler to manufacture and suitable for 3D printing. Possible applications • Search and rescue in collapsed buildings • Inspection of pipes and tunnels • Medical imaging and biopsies inside the digestive tract, with a miniaturized version around 1 cm wide or less The research was published in Bioinspiration & Biomimetics under the title Single actuator wavelike robot (SAW): design modeling, and experiments. What caught my attention is how the helical mechanism generates the motion with just one motor. The mechanical design does most of the work

$SPCX just reported ABSOLUTELY INSANE earnings, and I don't think people realize how quickly Elon Musk is scaling this company. SpaceX generated a staggering $14.6 BILLION in quarterly revenue, representing 161% year over year growth, while earnings per share surged 197% to $0.35. That's a 12% revenue beat and a 119% EPS beat against the estimates shown. And the numbers get even CRAZIER. SpaceX reportedly generated $4.4B in net income and $8.8B in adjusted EBITDA, with EBITDA growing an incredible 306% year over year. We're talking about a company that's not only more than doubling its revenue but potentially expanding its profitability at an even faster pace.

This is an interesting visual on just how big a splash CXMT made in Fab World. I just added them to my site's aggregator, and this treemap is the market cap for the fab operators in Asia. After adding CXMT, it's bigger than all the mid- to smaller fabs combined: SMIC, Nanya, UMC, ASE… Out of the gate, their market cap is about half the size of SK Hynix.

Superhuman Hardware Design Is The Easy Part

“Well, to be honest, Elon Musk is paying the most money in the country right now with this SpaceX little job that they got down there in Southaven, you know what I’m saying. It’s a killing,” said Ray Massey, who also traveled from Gulfport, Mississippi, in hopes of data center work.” One electrician had a message for Elon Musk about the work. “Tell Elon Musk to stay, we love this money you’re giving us man,” the electrician said.

NVIDIA MAY ALSO BE A BIDDER FOR AMBARELLA ( $AMBA ) - BETAVILLE

This is honestly insane. Marc Boroditsky, CRO of $NBIS, just revealed that the company has 1 GW of Vera Rubin GPU capacity coming next year, with significant customer demand already lined up. That alone could translate into tens of billions of dollars in ARR. 🤯 https://t.co/X2CNpaJaSH

NVIDIA has once again snapped up a massive amount of ABF substrate capacity... SEMCO!! https://t.co/5RRsvuWB3v

Memory buyers are locking in supply for the next DECADE. Google is reportedly in talks to secure memory supply all the way to 2035. HP and Dell have extended their deals with $MU through 2031. Micron now has 26 long term agreements covering about 35% of its revenue through 2030. Customers don't sign 9 year deals when they expect a glut but rather sign them when they are worried about running short. KeyBanc now sees DRAM contract prices rising 15% to 20% and NAND rising more than 20%. And $MU still trades at about 5x next year's earnings.

ELON MUSK TARGETS 13 GW OF AI COMPUTE BY 2027, WITH 440,000 NVIDIA VERA RUBIN GPUS IN A SINGLE TRAINING CLUSTER Musk says SpaceX is on track to bring roughly 200,000 $NVDA GB300 GPUs online per month for the remainder of 2026, with deployments continuing at a similar pace next year, primarily using Vera Rubin. The company expects to have 2.5-3 GW of compute online by the end of 2026, with another 10 GW potentially coming online in 2027. Musk also plans to have a 440,000-GPU Vera Rubin training cluster operational by the end of Q2 2027. To support the expansion, SpaceX is targeting an additional 20 GW of power and cooling infrastructure next year. Musk said there is a “non-zero possibility” of exceeding 20 GW of operational AI compute in 2027, although he acknowledged the probability remains small.

Latest CXMT update: Hefei & Beijing fabs are fully booked & preparing to add capacity thru 2028 (could increase to 600k wpm by 2028). Hefei-3, Beijing-2 + Shanghai? CXMT talks about G5 platform already matching industry leading mass produced nodes. LPDDR6 product hitting 12800 Mbps & lowering power consumption by 20% vs LPDDDR5X. Achieved this through SAQP. Company achieved 4F2 & other critical advancement in architecture. Again, this has hue application across mobile devices, edge compute, AIDC & smart cars.

Exclusive reporting by Korean media on Terafab: Terafab is excluding Chinese suppliers entirely. It wants to work with Korean suppliers because Japanese products are expensive. However, it is asking Korean companies to offer prices even lower than those of Chinese suppliers. Terafab decided in advance which companies to approach and which products to discuss at SEMICON West 2026, then visited their booths as soon as the show opened. Leading global semiconductor equipment suppliers Lam Research and Applied Materials have set up dedicated Terafab teams. The initial facility, with a capacity of 100,000 wafers per month, is expected to be completed before 2030. For logic chips, Terafab is working with Intel and is also in talks with TSMC. It is understood to be developing a separate manufacturing process for memory chips. A representative from an overseas equipment supplier said, “My understanding is that Musk initially explored acquiring production facilities from Intel, TSMC, and Micron, but shifted toward partnerships after being turned down. I have never seen this level of drive anywhere else in the semiconductor industry.” A representative from a Korean partner company said, “Even when we send inquiries in the early hours of the morning in the U.S., we get a response within 30 minutes to an hour. Their engineers have decision-making authority, so working with them is completely different from working with typical customers.” https://t.co/OD7AZhh0Ew

ELON MUSK SAYS SPACEX CAN BRING POWER ONLINE 12 TO 18 MONTHS FASTER THAN ANYONE ELSE And he has a plan to add 20 gigawatts next year On the earnings call, Musk said SpaceX $SPCX is turning on about 220,000 Nvidia $NVDA GB300 chips every month, and that pace will speed up next year. He expects to reach close to 3 gigawatts of compute in the next few months. He says he's highly confident SpaceX adds another 10 gigawatts next year, for about 13 total. The bigger plan would add 20 gigawatts, and he says the odds of getting there are "rising rapidly." He kept the details of that plan private, calling it "highly proprietary" because if he spelled it out, others would try to do the same thing. One advantage he did share comes from rocket engines. Turbine blades and vanes are one of the biggest bottlenecks in the industry, running 12 to 18 months behind the rest of the turbine. SpaceX is now casting its own, using what Musk calls the most advanced casting capability on Earth. "That means that we can bring power online 12 to 18 months faster than anyone else."