Shenzhen Foundry

@JIQQYJONES Let’s grab a couple of coffee ☕️

Tencent, Huawei, BYD, DJI. Headquarters within an hour of my office. Huaqiangbei, where you can source, prototype, and manufacture inside one square kilometer. Then the factory belt across Shenzhen and Dongguan behind all of it. Walked through DJI's new flagship this morning as one stop on that list. This is the whole thesis, compressed into a commute. #Shenzhen #AI #Robotics #GBA

VImag’s VMSM approach smartly shifts complexity from rare-earth materials to real-time control systems. In practice, success hinges on tight alignment between design, supplier processes for windings, and rigorous testing protocols. This is where mature manufacturing ecosystems deliver the biggest acceleration from prototype to reliable product.

🚨 BIG: An 🇮🇳 Indian startup may have achieved what global automakers have been working on for years. It could also help reduce the EV industry's dependence on 🇨🇳 China's rare earth magnet supply chain. Bengaluru-based Vimag Labs has secured a patent for a rare earth-free electric motor platform that replaces permanent magnets with software-controlled magnetic fields. If successfully commercialised, the technology could reduce the EV industry's dependence on Chinese rare earth supplies.

Magnet-free motor architectures like VImag’s address critical rare-earth supply risks through software-defined magnetic field control via power electronics and algorithms. The true differentiator lies in scaling: achieving consistent rotor winding precision, control integration, and volume production reliability. Shenzhen’s motor ecosystem has refined these exact processes over years of EV hardware execution.

VImag Labs blew my mind. They've built what they call a Virtual Magnet Synchronous Motor (VMSM). Traditional permanent magnet motors rely on rare earth magnets inside the rotor. These magnets are imported and expensive. Instead of using permanent magnets, VImag's rotor contains windings that are electronically excited and precisely controlled through software. As current is induced and managed in the rotor, it behaves like a "virtual magnet" allowing the rotating magnetic field from the stator to drive the motor just like a conventional permanent magnet synchronous motor. This means that the motor delivers the benefits of a permanent magnet design without actually needing rare earth magnets. Their current prototype is rated for 6 kW continuous power, with a peak output of 10 kW and 48 to 58 Nm of torque. The initial target market is EVs (everything from two wheelers, three wheelers, buses, and trucks). They may also use them in compressors and ceiling fans. The upsides are: 1. Lower cost. 2. Lower weight. 3. Smaller size. 4. Ability to control magnetic field. 5. Improved efficiency over PMSM. 6. Indigenous manufacturing and supply chain resilience. The company has been working on this tech since 2020.

Most stereo SLAM demos still fake the timing. They show a clean trajectory and a depth map, then quietly stamp everything with workstation arrival time. Later, when you try to align image + pose + depth for a real robotics dataset, you find out the three streams came from different physical moments. This is the live pipeline we just finished for a client: Head-mounted stereo → board-side H.264 → RTP → original camera frame ID and timestamp preserved all the way into visual SLAM and stereo depth. Camera moves, depth updates, features track — and every result stays locked to the exact same physical exposure. Clean validation: zero frame-ID mismatches, zero timestamp mismatches. This is the difference between a demo that looks good and data you can actually use. #PhysicalAI #EdgeAI #Robotics #StereoVision #WearableHardware #ORB_SLAM

@itzzme_pk Both show up, but the force/proximity stream usually degrades first under real hours of use. Sensor fatigue, cable micro-strain, and thermal drift hit harder than pure sync once the mechanical interface is solid.

@shenzhenfoundry Agree on the litmus :) second skin + matched sensor geometry is the whole game. Curious what breaks first at scale: sync drift across glove - ego - mocap, or the force/proximity stream staying usable after hours of real work?

Litmus test for tactile gloves in 2026: Can it wear like a second skin, run the same sensor geometry as a real robot hand, and spit out synchronized force+proximity data at millisecond latency? Most still fail the first two. This one passes. Same form factor as dexterous-hand sensors. Multi-device ready (Ego-Vision, optical mocap, data gloves). partners like Nvidia and BMW. Hardware doesn’t care about your pitch deck. It only respects physics and production constraints. If your embodied AI data pipeline still treats the hand as an afterthought, you’re not building a product. You’re collecting expensive anecdotes. #PhysicalAI #EmbodiedAI #TactileSensing #EdgeAI #RoboticsHardware #dexteroushands

Cameras are essential for physical AI. But what if the robot could also feel through its skin? 🤔 Electronic skin is already landing in concrete places: • 360° proximity + contact avoidance on collaborative arm joints and links • Full-body tactile coverage for humanoids • Real-time force feedback on exoskeletons and prosthetics • Local pressure arrays on soft grippers Vision covers the distance. Skin covers the last few centimeters and the moment of contact. Together they turn “safe in the lab” into something that can actually run next to people. #PhysicalAI #Robotics #HumanRobotCollaboration #TactileSensing #EdgeAI #Cobots #OEM

This compact camera from a Shenzhen team shows exactly why small form factor hardware is getting harder to fake. 📷 You can see how they handled the screen integration, physical controls, and overall thickness. Nothing feels overly thick or compromised just to fit the electronics. The real difficulty isn’t making it look clean in renders. It’s hitting usable battery life, thermal performance, and button feel while keeping it manufacturable and not ridiculously expensive. If you’re building any edge device or wearable that needs a screen and real-world usability, these kinds of public examples are worth studying right now. #HardwareDesign #PhysicalAI #EdgeDevices #camera #PrototypeToProduction #WearableTech #ShenzhenFoundry #shenzhen

Most printers force you to choose between fabric or rigid. This one doesn’t. xTool’s new O1 Omni Printer is a clear example. One desktop machine that prints on fabric, rigid parts, curved objects like bottles, and does transfers. The demos on t-shirts, caps, phone cases and skateboards are genuinely clean. For founders working on wearables, edge AI devices or robotics, this kind of tool coming out of Shenzhen lets you get real-world samples in days instead of weeks. We see the pattern constantly: Chinese engineering teams are getting very good at accessible fabrication hardware. What these tools don’t solve is turning good early samples into consistent, cost-effective production at volume. That part still requires deep experience inside the full China supply chain. #PhysicalAI #HardwareFounders #Shenzhen #Prototyping #EdgeAI #Robotics #3Dprinter

@soleio or like this one? x.com/shenzhenfoundr…

My pet theory is that the first breakout AR headset will use a monochromatic display

@vtchakarova Brutal. But give the Europe team credit — at least they picked a product with only one moving part. Try coordinating a few thousand components across a supply chain and the 'brutal' starts to look like Tuesday.😄

Brutal. Just brutal.

While the listed data market is concentrated among a few players, physical AI lets founders build proprietary data advantages directly from their deployed devices. Controlling the hardware iteration and production loop often creates a moat that's difficult to copy with off-the-shelf datasets.

Every single startup selling AI Training Data (July 2026) >50 cos sell data and RL environments to big AI labs and drive AI progress behind the scenes. They total ~$8.5B in rev and ~$100B in valuation, >75% of which are just 4 players: Scale, Surge, Mercor and Handshake.

Shenzhen’s density and speed are real advantages, especially for early Physical AI iteration. As these devices scale globally, advanced AI capabilities are increasingly delivered through subscription models running on the hardware itself. Hardware iteration speed directly sets the pace from prototype to scaled deployment — and accelerates how quickly the software layer penetrates markets and captures value worldwide. Over the next 3–5 years, this should further strengthen China’s position in the global software value chain. The culture that treats building physical products as high-status work sustains the full-stack talent pipeline.

Rebuilding industrial power is a long game. For founders building edge AI and physical products today, the real advantage is iteration speed. The teams that pull ahead embed manufacturing constraints into design from the start and keep close relationships with partners who can deliver prototype feedback in days, not months.

This is the playbook for how we rebuild America's industrial power.

🇺🇸1X NEO is coming from full robot integration and real task capability. 🇨🇳Wuji is betting on back-drivable, biomimetic joints with strong independent control. 🇨🇳LinkerBot is optimizing for speed, precision, and complex motions that can scale. 🇨🇳Xynova is going lighter with hybrid drive and cleaner movement. Three of them are Chinese teams pushing these different approaches in parallel right now. The mechanics are moving fast. The real constraint isn’t finger motion anymore, it’s turning that motion into something you can actually control, get useful touch data from, and manufacture at stable production volumes without everything falling apart. #DexterousHands #HumanoidRobotics #PhysicalAI #EmbodiedAI #RoboticHands

Many VLA and robotics teams are still duct-taping phones or building custom rigs for egocentric data collection. This is what proper production hardware looks like. Wide padded headband with easy rear adjustment for all-day comfort. Large multi-lens front sensor bar built for rich stereo vision, gesture tracking, and SLAM. 15 TOPS on-device compute running, Android 12 with full open interfaces. SenseTime vision stack. 370 g that actually feels wearable for hours. Already being used for industrial AR annotation, remote guidance, warehouse picking, and robotics perception training. If you’re building or training VLA models and need reliable egocentric capture + edge inference in one unit, what’s your current setup? #EmbodiedAI #VLA #PhysicalAI #EdgeAI #shenzhen

The front cameras are well integrated and protected. Headband looks comfortable for extended wear. Clean port placement on the side. It feels like practical hardware made for actual data collection rather than just another gadget. This Embodied AI Data Acquisition Headset comes from a factory we partner with in China. Lightweight (<150g), with binocular global-shutter cameras and a 1kHz IMU. It delivers clean, synchronized egocentric data through a fully open interface on Linux. It was built for teams training VLA and embodied AI models that need good first-person vision data without dealing with closed or bulky systems. Working directly with the factory gives us real speed on customization. #EmbodiedAI #EgocentricVision #VLA #RobotLearning #PhysicalAI

The “no boundary wire” claim is everywhere. What actually separates the machines that work from the ones that frustrate people is what happens when one sensor gets confused. LUBA 3 AWD uses real Tri-Fusion so the system can lean on LiDAR, NetRTK, or dual-camera AI vision depending on conditions — and still hold precision on steep slopes with AWD and a larger battery. Add flexible map editing and true zero-setup DropMow, and it stops feeling like a weekend project. A Shenzhen company pulling this off in their 2026 flagship while also covering pool robots shows how far the hardware stack here has moved. Most teams can make a demo look good. Few make the full integration reliable enough that customers don’t regret the purchase. #PhysicalAI #EdgeAI #SensorFusion #RoboticLawnMower #OutdoorRobotics

The "API to the physical world" framing is sharp. Once the hardware reaches this fidelity, the real constraint shifts upstream: generating enough high-quality, contact-rich interaction data in messy environments to make the policies reliable. Manufacturing excellence removes the hardware ceiling; the data loop determines whether it becomes infrastructure ：）

NEO’s Hands An API to the Physical World

@finkd 👀👀

(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.