The Duck That Knew the Water Was Warm: What a $399 Open-Source Robot Reveals About EMS in China

The Duck That Knew the Water Was Warm: What a $399 Open-Source Robot Reveals About EMS in China In August 2026, a 25-centimeter robot duck sold more units in 24 hours than most hardware startups sell in a lifetime. MicroDuck — built by Pollen Robotics, the French company Hugging Face acquired in 2025 — opened preorders at $399. Within a day the order book passed $2.6 million. At peak, one duck sold every four seconds. Five days in, preorders crossed ten thousand units. New orders now ship in four to six months, and the company has set a sales target of 20,000 units, which would make it the best-selling robot in history by volume. Nine hundred and forty-one years ago, another duck made history. In 1085, the Song-dynasty poet Su Shi — better known in the West as Su Dongpo, one of the greatest writers China has ever produced — watched the river in early spring and wrote a short poem for a painting by the monk-artist Hui Chong: 竹外桃花三两枝，春江水暖鸭先知。 Beyond the bamboos, two or three peach blossoms; when spring has warmed the stream, ducks are the first to know. The line that survives in every Chinese schoolchild's memory is the second one: 春江水暖鸭先知 — "the duck knows the warm water first." Su Shi's point was not about ducks. It was about perception: the duck does not read a weather report. It swims in the river. Its body touches the water. So it feels the change before anyone standing on the bank does. MicroDuck is the 2026 version of that duck. It is the first widely available, affordable platform that lets an ordinary developer touch embodied AI — not read about it, not watch a demo video, but hold it, program it, and watch it learn. The duck felt the water warming. The question this article answers is how the water got warm in the first place — because that part of the story has almost nothing to do with AI, and everything to do with manufacturing. This is a supply-chain story wearing an AI costume. Read it that way and it will tell you three things: how a $399 robot is even possible (manufacturing economics), why it matters that it happened now (the Raspberry Pi moment for robotics), and what every hardware founder should steal from it (open source ≠ manufacturable). Part 1 — The $399 Question: How Is This Possible? For a sourcing engineer. Let's start with the object itself, because the spec sheet is the first surprise. Inside a body that weighs under 800 grams: 15 motors (Dynamixel XL330 servos) driving the legs, neck, and head One camera and a compact LiDAR (an 8×8 time-of-flight matrix) for perception Two IMUs — one in the body, one in the head — so the robot knows when it falls and when it's standing A microphone, a speaker, two NFC readers, plus Wi-Fi and Bluetooth A Rockchip RK3566 SoC — a quad-core Arm chip with a small neural accelerator, 1 GB of RAM, 32 GB of storage — running a 50 Hz policy loop that decides what the robot does next A removable NP-F550 camera battery (the same 2600 mAh cell used by video crews), good for about an hour of run time It ships with seven trained behaviors — walking, sitting and standing, kicking, grabbing, roller skating, and getting back up after a fall — and every one of them is a reinforcement-learning policy you can retrain on your own machine. The whole software stack is open source under Apache-2.0: the SDK, the physics simulator, and the full training pipeline, including the sim-to-real tricks that make a policy trained in MuJoCo actually work on real servos. Now the question that should stop every hardware founder cold: how can this cost $399? Pollen's previous robot, Reachy 2, sells for $70,000. That is not a typo. The gap between $70,000 and $399 is not a discount; it is a different economic universe. Yet MicroDuck is not a loss leader — the company has said the $399 price is not subsidized. So where did the money go? The answer is a supply chain, and it breaks down into five structural reasons. 1. There is not a single custom part in the machine The most expensive line item in any robot is usually the silicon or the servos. MicroDuck uses neither. The RK3566 is an off-the-shelf Chinese SoC already shipping in millions of consumer devices; the Dynamixel XL330 is a standard, catalog servomotor used across the robotics industry; the NP-F550 is a photographic-industry standard battery. Nothing in the BOM requires a non-recurring engineering (NRE) investment, no custom ASIC, no bespoke actuator development, no million-dollar tooling before the first unit exists. When every part is already in mass production, the bill of materials collapses. 2. The design reused a validated supply chain This is MicroDuck's second act with the same partners. Hugging Face's previous desktop robot, Reachy Mini — also manufactured by Shenzhen's Seeed Studio — had already worked through the painful questions: which suppliers, which tooling approach, which assembly line, which test process. By the time MicroDuck reached preorder, the supply chain had run the course once. Reusing it is like a startup hiring its own alumni: the mistakes are already paid for. 3. Shenzhen did what only Shenzhen can do MicroDuck is designed by Pollen Robotics in Bordeaux, France, and manufactured by Seeed Studio in Shenzhen — the open-hardware company that has been turning open-source projects into shipped products since 2008. Seeed handled the structural injection-molding tooling, the component procurement, and the full assembly. When the order surge hit, Seeed's logistics team shipped 3,000 units in three days. This is the part of the story that Western coverage routinely underweights. The Pearl River Delta — Shenzhen, Dongguan, and the surrounding corridor — is the only place on earth where you can source the PCB, the PCBA, the CNC-machined housing, the injection molds, the cable harness, and the battery pack within roughly a 30-kilometer radius, and have prototypes on your bench within days. That density is not a policy or a tax break. It is decades of accumulated infrastructure: tool shops, mold makers, board houses, component distributors, test labs, and — critically — the people who know how to coordinate them. [Estimate: no single authoritative density metric; this reflects industry consensus on the region's cluster structure.] The division of labor is the point. A French design team ships the product design and the full software stack. A Chinese supply chain does the tooling, the procurement, and the assembly. Neither side could do what the other does, and the combination is what turns a would-be $2,000 machine into a $399 machine. 4. The software pays for the hardware Open-sourcing the entire software stack is not charity; it is an economic strategy. The Apache-2.0 release means thousands of developers contribute behaviors, fixes, and documentation for free. Pollen publishes seven trained policies; the community will publish hundreds. The marginal cost of software is zero; the marginal cost of hardware is $399. When your product is a platform rather than a gadget, the software is the moat, and the hardware is the packaging. 5. The business model prices for scale, not margin A $399 introductory price on a product with a realistic path to tens of thousands of units is a land-grab price, and everyone involved knows it. Hugging Face's chief scientist set the target at 20,000 units because that is roughly the best any robot has ever sold. The first 20,000 units are not about profit; they are about establishing the platform, the developer base, and the ecosystem before anyone else can. What the same robot would cost elsewhere It is tempting to ask what MicroDuck would cost if assembled in the United States or Europe. The honest answer is that the question is almost unanswerable, because the ecosystem does not exist there at this scale — you would be paying for a supply chain to be built, not used. A rough order-of-magnitude estimate, accounting for US/EU labor rates, tooling amortization, and component sourcing at robot-industry volumes, puts the same machine at several multiples of $399 — likely four figures, possibly more. [Estimate: extrapolation from published labor and tooling benchmarks; not a verified quote.] The gap is not craftsmanship. It is infrastructure. The buyer takeaway: if you have ever wondered why a Chinese EMS can quote your electromechanical product at a price that looks like a typo, this is the mechanism. It is not "cheap labor" in the tired sense of the phrase. It is a fully amortized, hyper-dense supply chain that turns custom hardware into commodity hardware. (For a deeper look at how that machine works, our guide to robotics electronics manufacturing in China walks through the drivers, sensors, and control electronics side of the same story.) Part 2 — The Raspberry Pi Moment: Why This Happened Now For a program manager. The second thing MicroDuck tells you is about timing. To understand why the robotics community is calling this "the Raspberry Pi moment for embodied AI," you have to remember what the actual Raspberry Pi did. In 2012, the Raspberry Pi Foundation released a credit-card-sized computer for $35. It was not powerful. It was not beautiful. It was a bare green board with USB, HDMI, and a promise that a child could plug it into a TV and write their first program. Within years, millions of people — most of them kids who would otherwise never have touched a computer as something to create with — had used a Pi to learn programming. The Raspberry Pi did not invent the personal computer. It moved the personal computer from a desk in the family room to a million bedrooms, and in doing so it changed who could participate in computing. MicroDuck is the same move, one level up. Before August 2026, doing real research in embodied AI meant access to a university lab with a six-figure hardware budget, or a company with a serious balance sheet. Research-grade humanoids like Reachy 2 cost $70,000. A quadruped from the usual suspects runs five figures. The field was, in practice, gated by money. MicroDuck removes the gate. $399 is approximately the price of a mid-range CUDA GPU — and that GPU is all you need, because the training happens in simulation. Pollen publishes the complete reinforcement-learning stack: MuJoCo physics, the PPO algorithm, and a training harness that runs 4,096 parallel environments on a single CUDA GPU. Usable walking gaits emerge in roughly one to two hours of training. The repository even includes the actuator model — voltage control, back-EMF, friction terms — and domain randomization across battery voltage, motor delay, and terrain, so the policy that works in simulation works on the real duck. Export to ONNX, load it on the RK3566, and the robot runs your policy at 50 Hz. Do you see what this is? A single ML engineer with one GPU — no lab, no grant, no procurement department — can now train a robot to walk, deploy it on hardware the size of a coffee mug, and publish the policy to Hugging Face for the rest of the world to use. That is the Raspberry Pi moment: not a better robot, but a robot that changes who gets to do robotics. The flywheel The Raspberry Pi's magic was never the board; it was the community that formed around it. The same engine is visible here. Open policies accumulate on the Hugging Face Hub the way open models do. A developer in Berlin trains a gait; a student in São Paulo retrains it; a hobbyist in Osaka publishes a trick. Each contribution makes the platform more valuable, which attracts more contributors. Data accumulates, behaviors accumulate, and the platform compounds. This is the exact flywheel that made open-source software win — now running on hardware that costs less than a phone. The corporate footnote The context around the launch matters. Two days before MicroDuck opened preorders, The Information reported that Nvidia had agreed to acquire Hugging Face for $12.9 billion. [The deal is reported but not yet signed and could still fall through — at the time of writing, no final agreement has been confirmed.] Nvidia's calculus is straightforward: almost every open model runs on Nvidia hardware, and the more the open ecosystem grows, the more indispensable Nvidia's GPUs become. A $399 robot that puts physical AI in the hands of every ML engineer is, from Nvidia's perspective, a machine that prints GPU demand. You do not need a position on the acquisition to take the lesson. When the world's most valuable chip company pays 86× revenue to own the open-model community, the open-hardware side of the ecosystem gets a strong signal about where the platform war is going. Physical AI is becoming a software-defined market, and software-defined markets are won by whoever owns the developer ecosystem. What this means for buyers Here is the part that matters if you are a hardware founder, a sourcing engineer, or a COO evaluating robotics for your own product line. When robots become software-defined platforms, your procurement logic changes. You are no longer buying a robot. You are buying a mechanical platform that can run an open-source software stack — which means hardware modularity, upgradeability, and repairability matter more than turnkey polish. Iteration frequency goes up. When the community ships new behaviors weekly, your hardware must tolerate frequent firmware and policy updates — which means your electronics need headroom (processor, memory, connectivity), and your mechanical design needs to be revision-friendly. Small-batch, fast-turn production stops being a nice-to-have and becomes the default. Platforms iterate; iteration is small-batch by definition. This is precisely the operating model of a modern low-volume EMS partner: No MOQ, fast turns, and the discipline to build twenty units as carefully as two thousand. If your roadmap involves physical AI, this is the supply-chain profile you should be qualifying, not the one built for static consumer products. Part 3 — Open Source Doesn't Mean Manufacturable For a manufacturing engineer. Now the part that the headlines got wrong, and the part that matters most if you are building hardware yourself. Every major outlet called MicroDuck "open-source." Strictly, that is true for the software and false for the hardware. The SDK, the simulator, the RL training stack — all Apache-2.0, all on GitHub, all forkable today. But the mechanical and electronic design files are not open. Pollen has been explicit about this: its press materials tell journalists not to describe MicroDuck as open-source hardware, and the company confirmed to The Register that it has no plans to open-source the chassis. Think about how strange that is. A company whose entire brand is open source, run by the people who built the most successful open-source AI community in history, ships a robot whose body is closed. That is not an oversight. It is a deliberate, correct business decision, and it contains the most useful lesson in this entire story: Open source is a strategy for software. It was never a strategy for manufacturing. Here is why the boundary is drawn exactly where Pollen drew it. Software is the flywheel — the more open it is, the faster the ecosystem compounds. Hardware is the opposite: it is where the cost, the quality risk, and the brand risk live. Open the hardware and you give away the margin, you lose control of quality, and you hand your competitors a free path to your product. You also take on the burden of supporting a machine whose assembly you did not verify. Pollen keeps the software open to grow the ecosystem and keeps the hardware closed to protect the economics. That is not hypocrisy; it is clear-eyed platform strategy. For hardware founders, the lesson generalizes brutally. A GitHub repository with ten thousand stars is a specification, not a product. Between a great open-source design and a shippable device sit four gaps that no amount of community enthusiasm will close: Gap 1 — Design for manufacturability. Open-source hardware is usually designed by and for makers: traces sized for hand soldering, tolerances that make a mold designer wince, BOMs that assume infinite availability. DFM is not a review at the end; it is a redesign in the middle. The gap between "it works on my bench" and "it works in production" is where most open-source hardware projects die. Gap 2 — Supply chain reality. The components in a GitHub BOM may be obsolete, on allocation, or only available in reel quantities. A manufacturable design needs alternates, lead-time awareness, and a procurement path that survives the six months between first prototype and first production run. This is why component sourcing is a discipline, not a shopping trip. Gap 3 — Compliance. CE, FCC, UL, RoHS, and the various country-level certifications are invisible in the GitHub README and very visible on the customs dock. Certification is a manufacturing function: it has to be designed into the board and the test plan, not bolted on after. Gap 4 — Quality at volume. A hand-built prototype that works is one unit. The moment you are shipping a thousand, you need a test strategy, a burn-in process, FAI/CMM reports, yield data, and an EMS partner who will show you the numbers instead of the brochure. Quality is not a department; it is the production system. MicroDuck's manufacturing story is the proof that these gaps are crossable — but they were crossed by Seeed Studio, a company whose entire business is closing exactly these gaps for open-hardware projects. That is the real product that Seeed sells, and it is why "contract manufacturer" undersells what an EMS actually does. An EMS is not a factory you rent. It is the interface between a design and the discipline of volume production — which is exactly why the EMS vs. OEM/ODM distinction matters when you choose one. The buyer takeaway: before you commit to an open-source design as the basis of a product, or before you evaluate a "manufacturing-ready" open project, ask your would-be EMS the five questions that separate real manufacturing partners from order-takers: Will you do a DFM review before I commit to tooling — and will you show me the issues in writing? Which parts of my BOM are single-source, and what are the alternates with real lead times? What is your test strategy — and how do you know a unit is good before it ships? Which certifications are you equipped to support, and what data do you need from me? What quality documentation will I receive with a production run — FAI, CMM, yield, burn-in? If an EMS hesitates on any of these, that is the answer you needed. Part 4 — The Duck and the Poet: What Su Shi Knew in 1085 Let us go back to the poem, because it is not decoration. It is the clearest statement of the idea this entire article is about — and it gives us something no other manufacturing blog on the internet will give you: a thousand years of Chinese thinking about perception, as a gift from a poet who understood supply chains of the spirit. The poem is Two Poems on the Spring River by Hui Chong, written by Su Shi in 1085. Su Shi is not a minor figure. He is one of the "Eight Great Prose Masters" of the Tang and Song dynasties, a towering presence in Chinese literature, the subject of a thousand biographies and, in recent years, of a documentary that made him a folk hero again. The poem is a tihuashi — a poem inscribed on a painting — and the painting, long lost, was by Hui Chong (965–1017), a Buddhist monk from Fujian known for quiet water scenes so distinctive that they had their own name: "Hui Chong's little landscapes." The first poem, the famous one: 竹外桃花三两枝， 春江水暖鸭先知。 蒌蒿满地芦芽短， 正是河豚欲上时。 The standard English rendering, by the great translator Xu Yuanchong: Beyond the bamboos a few twigs of peach blossom blow; When spring has warmed the stream, ducks are the first to know. Now read the second line as a piece of engineering reasoning, because that is what it is. Why do the ducks know first? Not because they are clever. Because they are in the water. Their bodies touch the thing that is changing. The temperature signal reaches them directly, before it reaches the air, before it reaches the people on the bank, before any instrument someone could have built in 1085. Su Shi's line is a first-principles argument about where information lives: the people closest to a change feel it first. Chinese poetry has a name for this habit of mind — 以物观物, "seeing things through things" — the practice of using a concrete object to carry an abstract truth, instead of stating the abstraction. A duck, a peach branch, a warming river: four small objects, and inside them a complete theory of early signal detection. This is the poetic tradition that produced the line every Chinese schoolchild knows, and it is the same tradition that made Chinese manufacturing culture what it is: a culture that prizes being in the thing — in the workshop, in the supply chain, in the river — over reading about it from the bank. There is a second, quieter layer. Su Shi wrote this poem about a painting, and Chinese critics praised him for a specific gift: he wrote what the painting could not show. A painting can show you the ducks; it cannot show you the temperature. Su Shi added the warmth — the invisible signal — and in doing so turned a picture into an argument. "诗中有画，画中有诗" — "poetry within painting, painting within poetry" — the famous standard he articulated for Wang Wei. The poet is the painter's sensor. That is exactly the role MicroDuck plays in embodied AI, and it is why the duck is the right mascot for this moment. The robot industry has spent years producing demos — videos of humanoids walking, hands doing tricks, machines that look like they are about to change everything. Most of those demos are the painting: beautiful, staged, and missing the temperature. MicroDuck is the duck. It does not claim to do anything impressive; it simply puts you in the water. It lets a developer feel the state of the field — what reinforcement learning on real hardware actually costs, actually feels like, actually fails at — for the price of a mid-range phone. And if you are a hardware founder trying to read the temperature of the global supply chain in 2026 — tariffs, component lead times, shifting trade routes, the reshoring debate — take the poet's advice literally. The people who know what the water is doing are the people who are in it. You will not find the signal in a news feed or a webinar; you will find it by being in the river: talking to your EMS, visiting the factory floor, holding the first article in your hand. In a supply chain as dense and fast-moving as the Pearl River Delta, the duck is the only reliable instrument. Part 5 — What This Means for Your Next Hardware Build For a quality lead. MicroDuck is one product, one launch, one moment. It would be easy to file it under "cute robot news" and move on. That would be a mistake, because it is a legitimate leading indicator — the duck in the river — of three structural shifts you should be planning around: 1. Physical AI is about to have its Raspberry Pi moment. $399 hardware plus an open training stack changes who can participate in embodied AI. If your industry involves robotics, expect the talent pool, the tooling, and the expectations to change fast. 2. The hardware will be built in China, because that is where the economics live. MicroDuck's price is not a miracle; it is the Pearl River Delta supply chain doing what it does. The design was French; the manufacturing was Shenzhen. The same split is available to you — design where you think, build where the density is. 3. Open source is a strategy for software, not a strategy for manufacturing. The gap between a starred repository and a shipped product is real, and it is crossed by disciplined manufacturing: DFM, sourcing, compliance, and quality systems. Whoever closes that gap for you is worth more than the design itself. An action list, if you want to act like the duck If you are evaluating a robot or robot platform: ask about the software stack before you ask about the specs. A platform that runs an open, actively developed stack will age better than one with a proprietary SDK and a roadmap document. If you are building hardware on an open-source base: budget for the four gaps (DFM, sourcing, compliance, quality) as line items from day one. They are not overhead; they are the product. If you are qualifying an EMS: ask the five questions from Part 3, and add one more: "Show me a product you took from open-source design to volume production." Seeed built a company on that question. Your partner should have an answer. If you are planning a robot product: design for iteration. Small-batch, fast-turn production is the operating mode of this market, and the supplier who can do twenty units this week and two thousand next quarter is the one you want — the profile our own new product introduction practice is built around. At Nex-G, we manufacture in Dongguan, in the middle of the same corridor that built MicroDuck — EMS for electronics, CNC machining for the parts that have to be metal, and the No-MOQ, fast-turn discipline that software-defined hardware demands. If you are reading the temperature of your own supply chain and want to feel the water before you jump in, that is the kind of conversation we are set up for. On the floor the checks behind that promise are concrete rather than aspirational. Every BOM line is sourced through AS6081-aware channels with at least one cross-checked second source; the panel-house rule profile is used for DRC before the first build; and every first article gets an FAI per AS9102 plus an IPC-A-610 acceptance verdict from a certified trainer before the run scales; material compliance is documented to RoHS and REACH before parts ship. The duck does not need to know the river's name, its current, or its history. The duck needs to be in the water. Everything else follows. FAQ Is MicroDuck open source? The software is fully open source (Apache-2.0): the SDK, the simulator, and the reinforcement-learning training stack are all on GitHub. The hardware — mechanical and electronic design files — is not, and Pollen Robotics has explicitly asked media not to describe MicroDuck as "open-source hardware." It is an open stack on a closed chassis. Who manufactures MicroDuck? The robot is manufactured by Seeed Studio (矽递科技), an open-hardware company based in Shenzhen, China. Seeed also manufactured Hugging Face's previous desktop robot, Reachy Mini, and handled tooling, component procurement, and final assembly for MicroDuck. Why does MicroDuck cost only $399? Five structural reasons: (1) every component is an off-the-shelf mass-production part (RK3566 SoC, Dynamixel servos, NP-F550 battery) with no custom tooling; (2) the design reused a supply chain already validated by Reachy Mini; (3) Shenzhen's dense manufacturing ecosystem — tooling, sourcing, assembly, logistics in one corridor — keeps cost and lead time low; (4) the open-source software stack effectively distributes development cost to the community; (5) the $399 price is a platform land-grab, aimed at establishing the ecosystem before competitors. What is "the Raspberry Pi moment" for robotics? The Raspberry Pi (2012, $35) made personal computing accessible to millions of new learners and changed who could participate in programming. MicroDuck does the same for robotics: $399 hardware plus an open reinforcement-learning stack lets any developer with a CUDA GPU train and deploy robot behaviors, moving embodied-AI research out of well-funded labs and onto ordinary desks. Can I manufacture an open-source robot in China? Yes — and Seeed's work on MicroDuck is the proof. But "open source" covers the software; turning it into a shipped product requires closing four gaps: design for manufacturability, supply-chain reality (BOM availability, alternates), compliance (CE/FCC/UL/RoHS), and quality systems at volume (test strategy, FAI, yield data). An EMS with open-hardware experience is the bridge. What does "春江水暖鸭先知" (the duck knows the warm water first) mean? It is a line by the Song-dynasty poet Su Shi (1085), written for a painting of early-spring river scenes by the monk Hui Chong. It means: the duck swims in the water, so it feels the warming before anyone on the bank. The poem is a first-principles statement about perception — the people closest to a change feel it first — and is one of the most famous lines in Chinese literature. How much does it cost to manufacture a robot like MicroDuck? There is no public cost breakdown. What is public is the mechanism: a BOM built entirely from mass-production parts, tooling amortized across tens of thousands of units, and a Shenzhen supply chain that handles tooling, procurement, and assembly. The same design assembled at US/EU labor and tooling rates would cost several multiples of $399 — an order-of-magnitude estimate, not a verified quote. Is the Nvidia–Hugging Face acquisition confirmed? No. The Information reported on August 26, 2026, that Nvidia had agreed to acquire Hugging Face for $12.9 billion, but a final agreement had not been signed at the time of writing, and the deal could still fall through. MicroDuck's launch and its open-source positioning are significant in that context regardless of the outcome. Sources Pollen Robotics — MicroDuck official product page and tech specifications (accessed September 2026) Pollen Robotics GitHub — pollen-robotics/microduck (SDK/runtime) and pollen-robotics/microduck_rl (RL training stack), Apache-2.0 The Register — MicroDuck launch report including closed-hardware confirmation and NP-F550 battery detail (via TheWeblist syndication) Byteiota — MicroDuck technical breakdown (Dynamixel XL330, RK3566, MuJoCo/PPO training pipeline, 4,096 parallel environments) 与非网 (EEFocus) — MicroDuck supply chain report: Seeed Studio manufacturing, 3,000 units in three days, Nvidia/Hugging Face deal status 红星新闻 (Red Star News, via Baijiahao) — launch sales data and order timeline, September 1, 2026 新浪财经 / 爱范儿 (ifanr) — sales figures, Rockchip share move, LeRobot sourcing detail CGTN Radio — Xu Yuanchong's English rendering of Su Shi's Two Poems on the Spring River by Hui Chong Xu Yuanchong, translation of 惠崇春江晚景 (Su Shi) This article reflects the state of public information as of September 2, 2026. Sales figures, deal status, and shipping dates are subject to change. Figures marked as estimates are labeled as such and are not verified quotes. Photography from the Nex-G facility is not yet available for this article; product and factory imagery is being sourced. Build the next MicroDuck — with a partner who closes the manufacturing gapOpen-source software is only the start. The difference between a GitHub repo and a shipped product is DFM, sourcing, compliance, and quality systems. Nex-G builds electronics and precision CNC parts in Dongguan for hardware founders who iterate fast.Request a quoteMore articles Related articlesRobotics EMS in ChinaLow-Volume & Prototype EMSEMS vs OEM vs ODM vs JDMNPI: Prototype to Production