The "hardware" here refers to two-dollar chips, twenty-dollar dev boards, and PCB prototyping orders, as well as the layer above—phones, glasses, toys—and the layer below: design tools, 3D printing, component data. Not data-center GPUs. This piece lays out the texts that large model companies and these hardware companies have written to and about each other, chronologically from 2014 to November 2026: who said "not officially supported," who rushed to claim "official," which side of the door was left open, and where the disputes ended up.
Between large model companies and maker hardware companies, no contract can be found in public texts—the relationship is sustained by text alone. Reading these texts reveals four patterns: the party that pulls harder says "official" more; the door opens on the "read" side first, closes on the "write, pay, train" side; labs don't build their own bodies, and every time they borrow someone else's; which path a dispute takes depends on whether there are actionable rights.
When reading this kind of relationship, don't look for contracts—look for the qualifiers in READMEs: who wrote "not supported," who wrote "official," and how the two sentences often describe the same project. Without a contract, the easiest party to walk away is the lab.
In April 2026, Anthropic posted a repository called claude-desktop-buddy on GitHub: a reference implementation of a Bluetooth protocol, plus firmware for an M5StickC Plus small screen, letting the Claude desktop app push permission prompts and recent messages to a small device on the desk. The README positions it in one sentence: this is for makers and developers, "isn't an officially supported product feature." (README) The contribution guidelines are blunter: "The best contribution is a fork." This repo is a reference implementation, "not an actively maintained project"; new features, new screens, ports to other boards—all declined. (CONTRIBUTING)
The same project received different treatment on the hardware side. On April 22, Espressif set up its own repo, esp-desktop-buddy, under Apache-2.0, turning this protocol into an ESP-IDF SDK; the repo's README doesn't describe its relationship with Anthropic. (GitHub) On May 14, Espressif issued a press release calling the M5StickC Plus the "first reference hardware" for this Claude hardware interface, and Anthropic recommended the M5Stack Cardputer as the development hardware at an event. (Espressif) On June 3, M5Stack's blog wrote: the StickC Plus Buddy is "Anthropic's official public demo." (M5Stack)
The README says the opposite. And there's another layer to the relationship: in April 2024, Espressif acquired a majority stake in M5Stack. (Espressif) The hardware praised in the May press release and the "official" mentioned in the June blog come from the same family.
Between these texts, no public contract can be found. Anthropic has never publicly written about purchasing or licensing anything from Espressif or M5Stack; Espressif and M5Stack have never written about receiving any authorization. The relationship exists only in what each side says about the other—and the accounts differ.
That is how this piece should be read. Between large model companies and maker hardware companies, there are no procurement contracts, equity stakes, or warrants like those for computing power—only READMEs, press releases, terms of service, changelogs, and llms.txt. This article treats "pushing away" as qualification, disclaimer, closing doors, and handing off; and "pulling closer" as releasing SDKs, writing adapters, opening interfaces, and announcing partnerships. Laid out chronologically, these texts reveal how the two sides pull toward and push away from each other.
The ESP8266 launched in May 2014, turning Wi-Fi into a two-dollar-class "attachment," and the maker community grew from there. September 2016 brought the ESP32; July 22, 2019, Espressif listed on the STAR Market; December 2020 saw the ESP32-S3 announced, targeting "AIoT"; the ESP32-S3-BOX in November 2021 was called an AI voice development kit for the smart speaker market; by September 2023, cumulative shipments exceeded 1 billion units. (Espressif milestones, 1 billion units)
JLCPCB took a different path, with a similar posture. Founded in 2006 for PCB prototyping, it added a components mall in 2012, launched the overseas EasyEDA in 2014, the domestic LCEDA in 2017, and merged three brands into the JLCPCB Group in 2021. (JLCPCB Group) Its narrative is about keeping the barrier as low as possible: EDA is free, the navigation page advertises "free prototyping" for 1-to-6-layer boards, and the IPO announcement page says "making it possible to produce even a single board" and "not defining the final product for the customer." (JLCPCB)
Both companies make money through openness. Espressif's annual report writes this as its business model: open-source projects, documentation, and communities influence technology selection during product design; developer adoption converts into long-term enterprise customers; and after mass production begins, sales can continue for five to ten years. (Espressif 2025 annual report summary) So these hardware companies welcome any platform that can drive shipments—they're not picky about partners.
Large model labs were not in this world. The closest thing findable is voice assistants: in August 2017, Espressif became a partner of Baidu's DuerOS conversational AI platform; in August 2019, it released the ESP-Skainet voice assistant framework; in July 2021, its audio front-end algorithm passed Amazon Alexa device certification. (Espressif and Baidu, milestones) The first generation of "AI partners" for maker hardware companies was cloud-based voice assistant platforms.
On the lab side, the one visible public move was an exit. OpenAI had conducted robotics research for years, decided to stop in October 2020, and announced it publicly in July 2021; co-founder Wojciech Zaremba explained that data was insufficient, "ultimately that was holding us back in terms of robotics," and a spokesperson also mentioned that other approaches were progressing faster. (VentureBeat) This was a "push away," and the reason was data, not cost.
This is the starting point: hardware companies courted all platforms that could drive shipments, large model labs retreated to software, and the two sides hadn't yet noticed each other.
After ChatGPT launched, hardware companies moved first, and fast. The chatgpt_demo in Espressif's ESP-BOX repo has a first-commit author date of March 24, 2023, and was merged on June 20. (GitHub commit history) On July 24, the Espressif developer blog published the OpenAI Component, an ESP-IDF component covering most of the OpenAI API except file operations and fine-tuning, with examples storing Wi-Fi and OpenAI keys in the device's NVS. (Espressif developer blog) The August 25 ESP32-S3-BOX-3 launch release wrote that the BOX-3 could become an online AI chatbot through platforms like OpenAI. (Espressif) The blog called this "a secure method" without explaining how it was protected.
On the OpenAI side, I found no official examples for microcontrollers in the same period. As we'll see later, its official reference page on short-term tokens only lists web and mobile as clients.
In 2024, several threads appeared in sequence.
On April 22, Espressif acquired a majority stake in M5Stack. (Espressif) On August 31, an individual account "78" created the repository xiaozhi-esp32; the author is Huang Guan, chairman of Shifang Ronghai, known online as "Shrimp Bro." According to GeekPark, they were a software team that chose ESP32 because it was beginner-friendly; when they open-sourced it, they "didn't know the Espressif team." (GeekPark, GitHub) On November 5, M5Stack released the Module LLM, a module that can run Qwen2.5-0.5B offline. (CNX Software)
On November 14, OpenAI created openai-realtime-embedded under its own official organization, authored by OpenAI employee Sean DuBois: an SDK for running the Realtime API on the ESP32-S3, adding an MIT license on December 17, followed by a Show HN post on Hacker News under the account Sean-Der. (GitHub, Hacker News) This was a large model lab, under its official banner, releasing an SDK for a maker chip.
In December, Volcengine launched the "AI + Hardware Zhiyue Plan" at its Force conference, partnering with Espressif and others: the ESP32-S3 integrates Volcengine's real-time audio and video capabilities, directly accessing Doubao. A Tencent news release described the division of labor: hardware manufacturers build the "body," chip design companies like Espressif build the "brain and nerves," and large models provide the "soul." (Tencent News)
The market moved faster than the companies' wording. On December 10 and 11, Espressif hit consecutive 20% daily-limit gains, with media attributing the surge to ByteDance's Doubao AI toy theme—this preceded OpenAI's repo license on December 17 and the subsequent Show HN. Rumors then circulated that ESP32 modules, chips, and dev boards were sold out; an Espressif representative responded: "the rumor is fake." (Xinshixiang, Sina reprint)
The same year, the AI hardware imagined by labs and startups was something else entirely: whole devices. The Rabbit R1 went on pre-sale at CES in January for $199, no subscription; Humane's AI Pin began shipping in April at $699, plus $24/month. (Rabbit, Humane, Wikipedia) What maker hardware companies imagined were modules and boards, sold to anyone who came. Both visions hit reality in 2025.
On March 25, 2025, OpenAI pushed three commits in a row to openai-realtime-embedded: "Direct users to Espressif implementation," "Remove GitHub workflows," with the body stating "No more code in this repo." The main branch was left with a single-page README; the original SDK code remained on the esp32 branch, with the README reading: "Espressif maintains a offical example" (spelling as in original). (GitHub) From creation to clearing, four months; the repo now has about 1.6k stars.
Espressif had actually already taken over: the openai_demo in esp-webrtc-solution was merged in early 2025 and continued tracking the Realtime API versions, migrating to the GA version's signaling flow on May 27, 2026. (GitHub) This example's README says the API Key is provided via environment variable, then exchanged by the device for a short-term token. OpenAI's official reference page on short-term client secrets aims to avoid leaking the master key, and the clients it names are "a web frontend or mobile client"—embedded systems, IoT, or microcontrollers don't appear. (OpenAI docs)
A company releases an SDK, hands it off to someone else four months later, and nothing in the public text constrains it from doing so. That's the benefit of having no contract, and also its cost.
The middle layer grew the same year. Xiaozhi's firmware added an MCP server on May 22, 2025, set as the default protocol on May 27. (GitHub) Its numbers are large and messy: GitHub stars as of November 10, 2026 are 32,970, MIT-licensed; in posts by Shifang Ronghai itself, connected device counts reached "over 1 million" in March 2026, "1.2 million+" in a June official site post, and "2 million" in an April FAQ; GeekPark's "100,000 units" in April 2025 referred to new connected devices over two months, of which only about a thousand were voice boxes sold officially. (Shifang Ronghai FAQ, GeekPark) Device count is neither sales nor active usage.
What's more worth examining is how each side writes about the other.
An Espressif announcement dated August 30, 2025, states that the open-source project "Xiaozhi" is a phenomenon-level application "led by the developer community and co-created by market users," that "Espressif and 'Xiaozhi' mutually reinforce each other," and cites it as an example of "new drivers for the company's future revenue growth." (Espressif announcement) The wording is community and users, not partner. The annual report summary disclosed the following March doesn't contain the word "Xiaozhi." Espressif's esp_xiaozhi documentation page only says "connects to xiaozhi.me service," with no statement of relationship or endorsement. (Espressif docs)
The Xiaozhi side is quieter. Its FAQ and two March 2026 official site posts don't contain the word "Espressif"; the FAQ only says it is "centered on the ESP32 microcontroller." (Shifang Ronghai) In the code, it's a different story: Xiaozhi's boards directory has a dedicated espressif/ subdirectory for Espressif boards, totaling 16; on July 27, 2026, the directory was reorganized by manufacturer; the ESP32-S31 went into mass production on July 27, and the repo already had its dev board support the day before, with Korvo-1 added the next day; ESP-Mosaico's board support was added on September 14, 36 days before Espressif's own release. (GitHub) The closest connection is in the commit history; the farthest is in the brand. At the Espressif Developer Conference Shanghai in November 2026, Shifang Ronghai and JLCPCB appeared together as ecosystem partners in a sharing session. (Espressif)
Espressif management cooled things down at the November 2025 shareholder meeting: two AI toys using Espressif chips were on the table—one using Doubao, one using Baidu's large model; the board secretary said, "current AI toy-related product shipments are still very small"; the chairman said that while the outside world expected sales to increase tenfold from this year to next, "at most it will grow slowly from 3x." (Sina Finance)
Also in November 2025, an OpenAI spokesperson told PIRG: "I can confirm we've suspended this developer for violating our policies." The suspended party was toy maker FoloToy, whose AI teddy bear Kumma had the weakest protections in PIRG's testing. (Futurism) FoloToy appeared in four ecosystems: Volcengine's Zhiyue Plan partner list, Espressif's developer conference ecosystem partner list, Xiaozhi's repo boards directory, and OpenAI's developer account. One account suspension severed only one of those links; according to Futurism, access appeared to have been restored about a week later, though neither company confirmed it. (Futurism)
Hardware companies were also tightening up. On October 7, 2025, Qualcomm announced the acquisition of Arduino; on November 21, Arduino posted an explanation for its updated terms: "As we introduce optional AI-powered features," the terms needed updating; it also assured that "Anything that was open, stays open." Adafruit reportedly led the criticism. (CNBC, Arduino) When AI features enter a maker platform, the platform starts to look more like a service with terms.
The whole-device route collapsed that same year. On February 18, HP bought most of Humane's assets for $116 million, and the Pin went offline on February 28; on December 5, Meta acquired Limitless, and the pendant was discontinued. (TechCrunch, TechCrunch) The Rabbit R1, per a Fast Company interview from September 2024, had at least 5,000 daily active users, while it had reportedly sold about 100,000 units. (9to5Google) These cases are the backdrop for labs subsequently releasing only reference designs, not whole devices.
In 2026, hardware companies opened doors quickly, and in consistent positions.
On April 8, Espressif launched its official documentation MCP service—a tool that only retrieves, doesn't execute, and can search both Chinese and English docs; the rationale is stated plainly: "Instead of guessing from training data or search results," let agents use real, up-to-date documentation instead of guessing. It also has limits: 40 calls per hour and 200 per day per user; for ESP-IDF versions no longer supported, it candidly admits it may return docs that don't match your SDK. (Espressif developer blog) On April 28, ESP-IDF 6.0 shipped with a built-in local MCP service, letting assistants like Cursor and Claude Code set targets, compile, and flash: "The assistant does not guess or simulate – it executes real idf.py commands." The safeguard mentioned in the article is that clients like VS Code and Cursor request confirmation before calling tools. (Espressif developer blog)
Espressif's own agent framework, ESP-Claw, had its repo created on April 17 and was released on April 21: the chip side receives messages or sensor events, calls a large model, then executes local Lua scripts. The release post names OpenClaw, Claude, and Codex as agents that can invoke devices via MCP; the README's model options are OpenAI, Alibaba Cloud's Qwen via Bailian, Anthropic, DeepSeek, and custom endpoints. The same README states: "ESP-Claw is still under active development." And "is not currently included in the Espressif Bug Bounty Program." (Espressif, GitHub) On October 20, the ESP-Mosaico was released—a whole device for "AI-assisted hardware creation"—with a measured line in the release: "AI-generated code still depends on a strong engineering foundation to run reliably on physical hardware." (Espressif)
Arduino is another sample. On August 12, App Lab 0.10's Agentic Mode went live, collaborating with agents via MCP and command line, with its own API Key; the exact words were "Claude is supported at first, with more providers coming soon." (Arduino) Its documentation site has an llms.txt. (Arduino)
What cloud providers do on these boards is something else: per-device licensing. Volcengine's hardware conversational agent documentation states that one License corresponds to one physical device, automatically bound and consuming a quota when the device successfully connects to the server, with 10 basic edition licenses included upon activation. (Volcengine) Alibaba Cloud's intelligent hardware documentation lists the Espressif ESP32-S3 and P4 as supported chips, while stating: "You need to apply for an intelligent hardware License," contacting a business manager; for the online version, "do not embed RTC_APP_KEY." (Alibaba Cloud) Contrast with Xiaozhi: individual users can use the Qwen real-time model for free after registration. (GitHub) The open-source world is free per person; the cloud world charges licenses per device.
On the Anthropic side, beyond the opening repo, three more things happened. Its Makers Station page revolves around the M5Stack Cardputer: download the repo, connect the device, point Claude Code at the folder, and "Claude will write the code and flash it over." (Anthropic) On August 27, Anthropic released a research preview of the Model Hardware Standard, a shared specification for AI agents to safely operate microscopes, liquid handling workstations, and robotic arms, offered only to an initial set of research labs and advanced manufacturers; Hugging Face and Raspberry Pi were named as partners integrating this standard, with the page saying the standard still needs work "before we open-source it." (Anthropic) Raspberry Pi is a partner integrating the standard here, not Anthropic's own hardware.
The third thing is terms. The "High-risk Physical Actions" section of Anthropic's Usage Policy, effective November 12, 2026: when model outputs are executed by hardware without human approval, a qualified person must be able to observe and stop it at all times; the device must stop or remain in a safe state when a human intervenes or it disconnects from Anthropic's services; and limits on speed, force, temperature, and the like must be enforced by the device itself or a controller independent of the model's output. The terms include exceptions in: consumer and household devices operating within manufacturer-built safety limits where the worst-case outcome is only discomfort or inconvenience; and code reviewed by a human before running, which also falls outside this section. (Anthropic Usage Policy) This section is limited to several categories: movement in shared spaces, forces that could injure, hazardous energy, the human body, safety systems, and industrial processes. The small screen on a desk doesn't appear in these categories.
The soft rejection of cloud-based large models by maker hardware companies appears elsewhere. On January 15, Raspberry Pi released the AI HAT+ 2, which can run large models locally; the official page pitches "locally and securely" and "reduces cloud API costs"; The Register quoted Raspberry Pi saying cloud models are 500 billion to 2 trillion parameters, while locally run ones are about 1.5 billion. (Raspberry Pi, The Register) In February, its official blog taught people to turn a Raspberry Pi into an OpenClaw agent, reasoning that running on a standalone device is a good way to mitigate security concerns. (Raspberry Pi) A cheap board becomes a cage for an agent. Seeed's SenseCAP Watcher product page is more blunt: continuously calling large models like ChatGPT is costly, so it uses an on-device model for detection first, with the large model only analyzing the results. (Seeed)
Finally, Meta. On October 2, Meta open-sourced the Muse Gadget SDK, which includes an ESP32 device-side SDK; the README requires each device to get a token from gadgets.muse.ai, read the SDK terms, and also states: "Side effects of tinkering may include bricked boards, voided warranties, brownouts, or bankruptcies." And "Proceed at your own risk!". The README doesn't mention Espressif. On October 19, Espressif published a release titled "ESP32 Powers the Muse Gadgets Ecosystem," stating that this ESP32 device SDK "is built on Espressif's ESP-IDF development framework." (GitHub, Espressif) The release doesn't say who made this SDK.
JLCPCB is the most significant player on this chain. On August 4, 2026, it listed on the Shenzhen Stock Exchange main board, code 001232; the offering price was 84.46 yuan, first-day close 208.01 yuan, total market cap approximately 115.6 billion yuan (per CNR reports); 2025 revenue was 10.232 billion yuan, and the first three quarters of 2026 saw revenue of 11.998 billion yuan, up 65.57% year-over-year, with net profit of 1.866 billion yuan, already exceeding full-year 2025. (CNR, Sina Tech) The IPO announcement page frames AI as demand: "As AI accelerates into the physical world." (JLCPCB)
Its relationship with AI splits into two doors.
The design-side door is open, and it's being opened officially. On August 5, 2025, LCEDA's GitHub organization released at least two official large model extensions: one that calls EDA extension APIs using natural language, and one that extracts pins from PDFs to generate symbols and footprints; the latter's README names Tongyi Qianwen and Zhipu, both requiring users to bring their own keys. On March 16, 2026, the official Skill easyeda-api-skill was created, with its README naming Claude Code, OpenCode, and QwenCode; it now has 1,074 stars. April 3 brought a WebSocket gateway extension covering Cursor, Cline, Windsurf, Trae, and others. On September 30, the official command-line tool easyeda-client-cli was created, with its README stating that the Pro client V4.1.70 onwards includes it, has an MCP mode, and comes with a document "for AI to read." On October 29, V4.1.93's changelog included three fixes for "AI automation" interfaces. (easyeda-api-skill, eext-run-api-gateway, easyeda-client-cli, changelog) A community project, easyeda-agent, created in June, already has 1,004 stars—nearly 50 times the official CLI's 21 stars. (GitHub) The official focus this year has been giving external agents a hand (Skill, gateway, CLI, MCP). That hand requires human consent first: the gateway extension requires manually checking "allow external interaction" in the EDA.
The ordering and data side door is in a different state. I placed real orders with JLCPCB in October and reported the test results: among 14 PCB APIs, there is no cancel and no validation; SMT has no order-creation API; order creation only accepts public URLs with no upload; there is no sandbox; the developer agreement doesn't contain "automation," "AI," or "Agent" at all; and llms.txt points to the web ordering form. ("Which Service Is Truly Agent Native") The international site JLCPCB's API page states that approval is required, based on the partner's previous orders with JLCPCB and company status, "not all applications will be approved," and no code review is provided. (JLCPCB) On May 28, its own AI customer service "Jia Xiao Zhi" went live, able to check order status, expedite orders, and assist with invoices; the page doesn't say whether it can create orders. (JLCPCB)
The terms walk the middle. The LCEDA project license agreement had no anti-scraping clause in its May 22, 2025 and April 14, 2026 versions; the August 18, 2026 version added §3.5: without permission, you may not use crawlers, robots, or scripts to batch or systematically scrape, download, copy, or mirror platform projects, nor use content obtained this way for data analysis, machine learning, "artificial intelligence model training," or similar purposes. (JLCPCB open-source hardware platform) This date falls between the April gateway extension and the September CLI: in the same year, the official side opened doors for its own users' agents while prohibiting others from using works from the open-source community to train models. The terms don't say why, and there's no evidence of causation between the two. What it rejects is bulk scraping of platform projects, not users' own agents operating their own EDA through the official gateway.
The data door had one real incident. The C-number (LCSC part number) in the component library has become a common part key for AI and code-based design in the community: jlcparts compiled JLCPCB's assembly library into a searchable database, tscircuit's jlcsearch syncs it daily, then feeds it to code-based design. These libraries are all community-built. In February 2022, the jlcparts maintainer found the service couldn't connect to LCSC and determined the IP was blocked; he contacted JLCPCB customer service for over a week with no response, and LCSC customer service's reply was brief, telling him to "stop immediately"; on March 7, the service was shut down, and after public attention, LCSC contacted him again with a proposal, restoring service on March 21. (Maintainer's account) This article found no text from JLCPCB authorizing, prohibiting, or commenting on jlcparts or jlcsearch.
The price notices don't mention AI. On January 8, 2026, double-sided and four-layer boards went up 20 yuan per square meter, citing "recent continuous increases in major raw material costs." The June 12 notice stated that copper-clad laminate had risen from about 100 yuan per sheet in previous years to around 230 yuan, with daily fluctuations; during the shortage period, orders exceeding 50 square meters would be subject to tiered surcharges, quotas would be activated when total orders exceeded capacity, splitting orders to circumvent this was strictly prohibited, and for temporary large orders, "even if the customer accepts high prices, the company will implement quantity controls based on capacity and may decline orders when necessary," with subsequent adjustments "no longer separately announced." On September 14, board material prices fell, with reductions up to 50 yuan per square meter; on September 21, single and double-sided boards each rose 40 yuan again; on October 27, orders for 6+ layers were "overloaded," with lead times extended by about 3 days, attributed to demand. (January, June, September 14, September 21, October) None of these notices mention AI or servers. The unit for quantity limits is area, regardless of whether the orderer is human or an agent.
PCBWay's approach magnified this into a single month. On October 20, it released the ProtoFlow plugin: users of this AI-assisted PCB design tool could send Gerber, BOM, and pick-and-place files to PCBWay's quote page with one click. (PCBWay) Eight days later, on October 28, it posted a "Notice for AI-Generated Design Files": AI-generated submissions may bring redundant files and conflicting requirements, and when materials are complex, "may not immediately begin evaluation"; "PCBWay reserves the right to decline inquiries and quotation requests based on AI-generated content or files." And for problems caused by AI content, it "shall not be responsible." (PCBWay) On October 30, it partnered with Tindie to launch the 2026 AI Hardware Design Contest, with a total prize of $2,500; "AI hardware" here refers to physical projects with AI capabilities, not AI design tools. (PCBWay) Welcoming hardware that is an AI product, the plugin going only as far as the quote page, and reserving the right to refuse AI-generated files: three things in ten days.
On the AI design tool side, nothing was agreed either. Flux's blog says its agent can quote and place orders, naming JLCPCB and PCBWay; JLCPCB's own Flux tutorial involves manually uploading files. (Flux, JLCPCB) tscircuit's documentation says in-app ordering currently supports only JLCPCB, redirecting to Stripe checkout, and doesn't say whether the two companies have a partnership. (tscircuit) This article found no partnership announcement between any AI-PCB startup and JLCPCB; the only one with an explicit partnership is ProtoFlow and PCBWay.
On November 9, Anthropic published a research paper saying Claude can design electronic prototypes end to end: Fable 5.2 designed three devices—a hands-free phone, a voice assistant hub, and a handheld virtual pet; three boards were produced by Diode from Claude's design files. The only manufacturer named is Diode, "a company that designs and manufactures circuit boards"; it doesn't mention the factory behind Diode, doesn't mention JLCPCB, and doesn't mention part numbers; a footnote states that Anthropic has a commercial relationship with Diode; Claude didn't connect to test equipment or assembly machines, and the boards were built and tested by Diode's engineers; the article states the boundaries: no electrical verification, antenna tuning, thermal testing, or certification was done, and it wasn't designed for mass production. (Anthropic) Diode describes itself as a U.S. PCB design and manufacturing company in the Brooklyn Navy Yard, taking clients from spec to mass production. (Diode) This time, the lab worked with a commercially related partner that produces its own boards; no public ordering portal appears in the article, and JLCPCB doesn't appear either.
The labs' attitude toward hardware isn't avoidance—it's borrowing.
OpenAI is the clearest case. After disbanding its robotics team in 2021, Caitlin Kalinowski joined on November 4, 2024, to lead robotics and consumer hardware. On May 21, 2025, OpenAI acquired Jony Ive's io, a $6.5 billion all-stock deal, with about 55 engineers, researchers, and product developers joining. (TechCrunch, TechCrunch) In February 2026, per court filings, the first device won't ship before late February 2027, and OpenAI also decided to stop using the "io" name. (MacRumors) On March 7, Kalinowski resigned over principled opposition to OpenAI's Pentagon deal; in August, per The Information, she joined Anthropic. (TechCrunch, The Information)
In the interim, OpenAI's first externally sold hardware was the Codex Micro, released and launched on July 15: made in collaboration with Montreal keyboard maker Work Louder, reportedly essentially a Work Louder Creator Micro 2 rebranded as OpenAI with preset Codex key mappings, $230, limited edition. TNW reported that manufacturing issues were already solved by Work Louder, and the device was made by them. (The Next Web, 9to5Mac, Axios) As of November 10, this article has not verified any news of OpenAI's own device being unveiled or sold.
Anthropic's approach is experiments, not products. June 2025's Project Vend had Claude run a small shop in its San Francisco office; restocking was done by employees of partner Andon Labs, who also served as the wholesaler without Claude's knowledge. (Anthropic) Project Fetch had researchers with no robotics background program a ready-made robot dog in one day; in the June 2026 second phase, Opus 4.7 completed it independently, about 20x faster than last year's best human team assisted by Opus 4.1, but the step of having the robot dog autonomously retrieve a ball still wasn't accomplished; the authors wrote: "We are plausibly entering the early era of physical agentic AI." (Anthropic) In both experiments, the hardware was someone else's, and neither had product plans.
Other labs and big companies take similar shapes. Google DeepMind partnered with Boston Dynamics at CES in January 2026, putting Gemini Robotics models into the next-generation Atlas—the robot body is someone else's. (The Robot Report) ByteDance's October 2024 Ola Friend headphones are Doubao's first AI hardware, with the hardware made by the Oladance team ByteDance had recently acquired. (QbitAI) Alibaba's Quark AI glasses, released November 27, 2025, are per Xinhua, Qwen's first step beyond the screen into the physical world. (Xinhua) MiniMax takes the model-provider route: in February 2025, it established an "Intelligent Hardware Industry Innovation Alliance," and its voice models are used in glasses and toys. (CLS)
What labs are willing to own are protocols, brands, and models. Bodies are rented out or borrowed through someone else's shell. The Humane, Rabbit, and Limitless whole-device cases are the backdrop for labs not making whole devices; in my view, the risk of whole devices has been left to toy makers and maker hardware companies.
The "hardware" above was two-dollar chips. One layer up are phones, glasses, and toys: the hardware companies here are bigger than Espressif, and their rights are more concrete. The same coyness has partly landed in court here.
Phones: the same companies are simultaneously partners and defendants. The Apple-Google relationship looks most like a partnership. On January 12, 2026, the two issued a joint statement that the next-generation Apple Foundation Models would be based on Gemini; the statement didn't disclose financial terms, with "about $1 billion per year" coming from prior media reports. (CNBC) Apple and OpenAI simultaneously occupy several positions. ChatGPT has been integrated into Siri and Writing Tools since 2024; in May 2026, per Bloomberg via TNW's retelling, OpenAI and its law firm studied options and might issue a breach notice, with the dissatisfaction being that the integration was buried behind friction (TNW's retelling language); TNW's article didn't carry Apple's response, and no lawsuit had been filed at the time. (TNW) On July 10, Apple separately sued OpenAI and others, alleging its hardware team misappropriated trade secrets; after an October 14 hearing, OpenAI agreed not to use Apple's alleged trade secrets during the proceedings, and per Daily Journal's headline, the two sides replaced a preliminary injunction with an agreement. (TechCrunch, 9to5Mac, Daily Journal)
OpenAI is also a defendant in two other lawsuits. X and xAI (later changed to SpaceXAI) sued Apple and OpenAI in the Northern District of Texas in August 2025, alleging the two made ChatGPT the only integrated AI assistant in iOS; both denied it; on September 14, 2026, the plaintiffs withdrew all claims against Apple, saying they had "resolved" the matter, with terms undisclosed, while the claims against OpenAI continue, with the trial reportedly set for January 2027. (9to5Mac, Unite.AI) §06 mentioned OpenAI no longer using the "io" name, which is connected to the iyO trademark lawsuit: filed in June 2025, a temporary restraining order was granted, the Ninth Circuit affirmed it in December of that year, OpenAI told the court in February 2026 it would stop using the name, and on July 27, both sides notified the court of a settlement in principle. (IPWatchdog, 9to5Mac) The three lawsuits concern trademark, trade secrets, and competition law respectively—all things with clearly actionable rights.
Chinese phones: barely reached court, went through risk control and public notices. On December 1, 2025, ByteDance's Doubao phone assistant and ZTE/Nubia's collaborative M153 engineering device were unveiled, with the assistant using system-level permissions to read the screen and simulate clicks to work across apps. (Nan Cai Compliance Weekly) From the next day, users reported WeChat popping up security alerts, and Taobao and some banking apps had login issues; on December 3, Doubao disabled WeChat operation features, and on December 5, announced it would regulate three categories: score/incentive farming, financial apps, and games. Tencent issued no announcement; per Business Standard's reporting, WeChat's statement to media was that there were no specific actions against Doubao, and it was more like routine risk-control auto-detection, so "WeChat blocked Doubao" was stronger than what the company itself said. (Business Standard) On January 26, 2026, per media reports, Pony Ma expressed opposition to "plug-in"-style approaches at Tencent's employee meeting; various outlets quoted him differently, so the original words aren't quoted here. (Jiemian News)
On September 14, 2026, the Doubao phone assistant launched the Screen Automation Operation Declaration Protocol (SAEP) and began a 30-day public comment period; on the 16th, the second-generation NaviX Ultra went on sale: the rule is protocol-first (MCP/A2A), with simulated clicks as fallback; app providers can declare refusal during the comment period, and those that don't will be gradually opened by risk tier after the period ends. The period ended October 14, with over 150 apps explicitly refusing, with counts ranging from 150 to 155 across sources; per PANews's list, Tencent-affiliated apps accounted for 66 and Alibaba-affiliated for 23, totaling about 57% of the ~155 (the percentage is my calculation). (Tencent News, PANews) This set of rules was introduced by the Doubao phone assistant itself; those that didn't declare refusal default into gradual opening, and it's not an industry standard. On the WeChat side, per 21st Century Business Herald, since June, phone makers' assistants have been operating WeChat via A2A: currently only messaging and audio/video calls, not yet Moments or money-related red packets, and in the June 5 test, only Honor's assistant worked. (21st Century Business Herald) From these texts, what's rejected is unauthorized screen reading plus simulated clicks, not AI itself.
Phone and glasses makers don't bet on a single lab either. The Samsung Galaxy S26 launched in February 2026, with Gemini handling agentic tasks, Perplexity for web queries, and in-house Bixby on-device; (CNBC) per Perplexity's official blog, Bixby's search and reasoning backend also uses Perplexity's API. (Perplexity) Rokid said in a March press release that its international-version glasses can switch between Gemini, ChatGPT, DeepSeek, and Qwen, contrasting with Meta's Llama-only approach; this is a one-sided manufacturer press release, with no independent confirmation from Google or OpenAI. (Fortune press release) This is the mirror of §06: labs don't bet their bodies on a single hardware company, and phone and glasses makers don't bet their "brains" on a single lab.
Toys: where risk is most concentrated, and the only layer with legislation. Mattel announced a partnership with OpenAI in June 2025, and in December told Axios there were "no plans" for the holiday season, with a spokesperson saying they hadn't abandoned product and experience work with OpenAI; as of November 10, this article has not verified any Mattel × OpenAI product launch. (Futurism) In late January 2026, WIRED reported that the plush toy Bondu's web console let any Gmail account read about 50,000 children's conversations; per Malwarebytes, Bondu was quickly taken offline and authentication was added after researchers alerted the company. (Malwarebytes) FoloToy, whose account was suspended by OpenAI in §03: per PIRG's March 2026 report, as of February 24, Kumma could still select GPT-5.1 in FoloToy's backend portal; OpenAI reiterated that FoloToy remains banned but didn't respond to the specific evidence PIRG provided. (PIRG) The account-level "rejection" doesn't match the model actually connected to the product; PIRG didn't provide a definitive explanation.
Legislation followed. China's "Interim Measures for the Administration of Anthropomorphic Interactive AI Services" was published on April 10 and took effect on July 15, prohibiting the provision of virtual relative and virtual companion services to minors, and requiring guardian consent for other anthropomorphic interactive services provided to minors under 14; the text doesn't specifically name AI toys. (Tencent News, SecRSS full text reprint) California SB 867 was signed on September 10, imposing a four-year moratorium on the manufacture and sale of toys with chatbot functionality, protecting children under 16; Padilla's office press release elsewhere wrote "under 18," an inconsistency, and didn't specify an effective date, with the bill stating it expires on January 1, 2031. (California Senate District 18, CalMatters)
Placing this layer's disputes side by side reveals that they take different paths.
Those with clearly actionable rights went to court; those with only terms and platform power stopped at terms, handled unilaterally by one side; those involving children, where neither of the first two suffices, went to legislation.
One layer further down, back to chips, models, and factories. The coyness here is mostly not arguments, but licenses, terms, and one quiet cessation of updates.
Shipping a model in firmware means carrying its license. Gemma 1 through 3n used Google's own terms of use, under which Google reserves the right to remotely or otherwise restrict uses it reasonably believes violate the agreement, and redistribution must carry the usage restrictions into downstream agreements; on April 2, 2026, Gemma 4 switched to Apache 2.0, and older versions already in devices won't automatically change licenses. (Google, terms page) When Mistral released Ministral 3B/8B in October 2024, self-deployment required contacting Mistral for commercial use, and the 8B weights were released for research use only; the launch post's user-facing line was "From independent hobbyists to global manufacturing teams"—generic marketing language. (Mistral) Per VentureBeat, the Mistral 3 family released in December 2025 switched to Apache 2.0. (VentureBeat) Llama 4's license requires prominently displaying "Built with Llama," and new models trained with it must start their names with "Llama," with products exceeding 700 million monthly active users needing to apply to Meta; the terms contain no device or hardware-specific clauses. (Meta)
DeepSeek's weights are MIT, but distilled versions inherit the base model's license: DeepSeek-R1-Distill-Llama-8B's page states it's based on Llama 3.1, "originally licensed under llama3.1 license," while the Qwen-based versions were originally Apache 2.0. (Hugging Face) During the February 2025 "adapt to DeepSeek" wave, per a Zhidxu tally (reprinted by Tencent News), 16 domestic AI chip companies announced adaptation or availability within 7 days, with the list comprising cloud, server, and edge accelerator cards—no Rockchip, Allwinner, or Espressif, though during the same period, there were many developer tutorials for running distilled versions locally on Rockchip's RK3588, just not on that announcement list; (Tencent News) A-share company Anka Micro subsequently clarified that its voice recorder solution accesses DeepSeek via API and "has not engaged in any substantive cooperation directly with DeepSeek or its affiliates." (Tencent News) "Adaptation" means at least three things: calling an API, running a distilled version locally, or official partnership. Open weights don't require a contract, and this article found no record of DeepSeek signing with any chip company.
Google wound down its maker-class accelerator sticks bit by bit, while "welcoming" chip companies. Coral's software repositories were set to read-only archive in sequence: pycoral on July 3, 2025, libedgetpu and libcoral on October 14, 2025, and finally edgetpu on April 19, 2026; this repo's README had long stated it was primarily an issue tracker, with legacy code "no longer maintained." This article found no official discontinuation notice. (pycoral, libedgetpu, libcoral, edgetpu) The day after archiving, October 15, 2025, Google Research released the open-source Coral NPU, a RISC-V design whose roadmap lists Gemma 3 270M as one of its reference architectures. (Google) On May 19, 2026, Synaptics exhibited the Coralboard at Google I/O's Gemma area, with limited editions given away as on-site experiences; on October 20, the Coralboard SL went on general sale through DigiKey, priced at $99.79 per VentureBeat, with a second model SR still unshipped. (Synaptics, VentureBeat) The old products were proprietary TPU-architecture USB/M.2 accelerator sticks, and the new ones are open-source designs; whether the two are in a successor relationship isn't mentioned on Google's roadmap page; in my view, the hardware has been yielded to chip companies.
Raspberry Pi is on both ends. On August 11, 2026, it co-posted with Google AI Edge that LiteRT can now run Gemma on the Pi 5, while in the same post, Hailo accelerator support remains "coming soon." (Raspberry Pi) On the other end, in December 2025, Eben Upton told The Register that AI infrastructure's memory pressure "is painful" but "ultimately temporary"; on November 2, 2026, an official post raised the 16GB Pi 5 by $80 to $385, citing memory costs and supply, with no mention of AI. (The Register, Raspberry Pi)
Autodesk Fusion is one of the connectors added to Anthropic's "Claude for Creative Work" on April 28, 2026, letting subscribed designers and engineers create and modify 3D models through conversation with Claude. (Anthropic) Autodesk's usage policy contains a broad clause that, on its face, prohibits using its products' outputs to train its own machine learning systems; it had been in contracts since May 2018, only drew wide discussion in October 2025, and was revised on December 8, clarifying that customers own their content and may use it for AI training, but may not use Autodesk products to replicate its functionality. (AEC Magazine) On the chip design side, on September 30, 2026, Synopsys and OpenAI announced a multi-year agreement to develop GPT-Synopsys, a chip-design-specific model: the two call themselves "preferred partners," OpenAI is authorized to use Synopsys's EDA tools, revenue is shared, and customer design data is protected and not used for training; the press release doesn't mention exclusivity. (Synopsys)
The domestic PCB manufacturer HQ NextPCB has built AI into its own KiCad distribution. Version 10.0.0 pre-installs an MCP for 72 tools, where users fill in their own API address, key, and model name, compatible with OpenAI's interface; (HQ NextPCB) the September 25 version 10.0.6-1 adds the DSH (DeepSeek Harness) agent platform, with documentation listing scenarios for connecting GLM, Kimi, Doubao, etc., depending on the configured service, while business functions like having the agent query HQ's component mall require first logging into the HQ agent platform. (HQ NextPCB) The protocol is open; the pricing query entry is in its own service.
Board manufacturers are most careful about the difference between "can produce a file" and "can go into production." In early September 2026, OpenAI released GPT-6 Astra, which in its demo routed traces in KiCad, and two board manufacturers subsequently wrote tests. JLCPCB's article wrote: with four lines of requirements, Astra designed a 44×34 mm dual-channel Class D amplifier board, with KiCad reporting "0 ERC and 0 DRC violations under the configured rules," but 9 categories of checks were turned off; the routing layout "came from an earlier amplifier task and was reused unchanged"; "No DFM upload, factory review, quotation, or order formed part of this project"; the conclusion was "a valid board file and a clean native project are not the same thing." (JLCPCB) NextPCB's test noted that the routing result, produced in about 12 minutes, was completed by Astra calling the open-source Freerouting. (NextPCB) Quilter (which sells its own AI PCB tool) said in an October 27 roundup that it found only one example of an LLM-designed, physically produced, and working board: a 4-layer RP2350 board with an e-ink screen, designed with Claude; JLCPCB manufactured and assembled 5 boards for €130; Freerouting left 49 unrouted connections, which Claude filled manually, and 65 DRC errors were sent back to the model for fixing; the author called it "not a complex PCB by any means." (Quilter, Author's blog)
Bambu Lab's MakerWorld sells AI generation while tightening AI slop. On July 30, 2025, a Tencent Cloud post said Hunyuan 3D would soon be fully integrated into MakerWorld; (Sina Finance reprinting Tencent Cloud post) on March 17, 2026, Meshy said its image-to-3D was built into MakerLab, with each export costing 2 credits; (PR Newswire) Aimia in late October put five third-party engines—Meshy, Tripo, Hi3D, Rodin, and Hunyuan—into one entry point, and per 4Additive (which self-identifies as AI-generated), Bambu Lab hasn't developed new algorithms and is currently on a waitlist. (4Additive) On the other side, since February 5, 2026, newly uploaded models must include at least one photo of an actual print; (Fabbaloo) the November 6 official blog wrote that MakerWorld won't simply divide content into "AI" and "non-AI," and the problem is that cover images promise quality beyond what the print delivers, with the CEO saying "AI itself is not the problem - the slop is." (Bambu)
The other two platforms take different approaches. MyMiniFactory acquired Thingiverse in February 2026; (Hackaday) in a 3D Printing Industry interview, Thingiverse's approach is creator self-labeling, visitor filtering, with AI-labeled content staying on the site but removed from the front page, and they say reliable classification at the mesh level remains difficult. (3D Printing Industry) Hackaday's August 29 column said its AI policy is the same old policy—write about clever hacks; red flags are circuits that don't make sense, violations of physics, and hardware projects with only renderings. (Hackaday)
On the manufacturing side's "write and pay," Slant 3D provides a counter-example to JLCPCB: a public MCP server requiring no key or account, where quote_checkout generates a payment link, and after the customer pays, the order goes directly to the print farm. (Slant 3D) It's not the only one: Make Parts Now's MCP requires no login to quote and generate a Stripe checkout link, with payment completed by the user; Harktech's requires no key or login and returns a pre-quoted checkout page; Fabworks requires login and similarly provides a checkout link after quoting. (Make Parts Now, Harktech, Fabworks) This article found no official public MCP from Xometry, Protolabs, Fictiv, PCBWay, or JLCPCB. Bambu Lab's printers are another case: the community Bambu MCP requires the printer to be in LAN Only mode with Developer Mode enabled, which disables cloud features, and the page contains no claim of Bambu Lab authorization or endorsement. (griches/bambu-mcp) The backdrop is that in January 2025, Bambu Lab added authorization authentication to its firmware, and in April 2026 sent a cease-and-desist to the maintainer of an OrcaSlicer fork, with the Software Freedom Conservancy confirming two AGPLv3 violations in May; this dispute itself is unrelated to AI. (SFC, Tom's Hardware)
Component data is essential for agents doing hardware. Digi-Key's API user agreement doesn't contain terms like AI, machine learning, LLM, or training ("agents" appears once in Section 10, referring to employees, contractors, and agents), but the restrictions are structural: no bulk downloading of data, no using API data to update or build your own database, may only display on your own site and internal applications, and may not disclose to third parties. (Digi-Key) The MCP servers around Digi-Key, Mouser, and Octopart found by this article are all community projects, with some pages disclaiming any affiliation or endorsement with JLCPCB, Mouser, Digi-Key, etc. (pcbparts.dev) Whether these terms constrain LLM agents from caching or summarizing Digi-Key's data is an inference, not a Digi-Key statement.
In this section, what's open is still reading and generation; what's left to humans is training terms, payment, and authorization.
Flattening twelve years of text yields this matrix.
In the first three columns, cells mostly have only one or two words: hardware companies pulled toward voice assistants, then toward ChatGPT; labs first retreated to software, then released an SDK. By 2025 and 2026, all four cells had text simultaneously, and often within the same project.
Pattern one: the party that pulls harder says "official" more.
Labs say less because "official" implies support, maintenance, and responsibility—a "not supported" is their firewall. Hardware companies say more because in sales, "official" is the word that turns a project into a category: "Claude hardware" sells better than "an open-source Bluetooth example." GeekPark noted that Sichip and Allwinner chips have already been adapted for Xiaozhi. Whether Xiaozhi doesn't mention Espressif because it doesn't want to be tied to a single chip vendor is my speculation; it hasn't said so itself.
Hardware companies aren't simply enthusiastic. Espressif's LLM solutions page states that devices don't need to run large language models directly, that model platform names on the page are only for illustrating connectivity or compatibility, with trademarks belonging to their respective owners; ESP-Claw is excluded from the bug bounty program; the ESP-Mosaico release acknowledges that AI-generated code can't replace engineering foundations. (Espressif) They pull close, but their disclaimers are clearly written.
Pattern two: the door opens on the "read" side first, closes on the "write, pay, train" side.
What's open is documentation, Skills, MCPs, CLIs, tutorial blogs: Espressif's documentation MCP and ESP-IDF MCP, JLCPCB's official Skill, CLI, and MCP mode, Arduino's Agentic Mode, JLCPCB's tutorials for Flux and AI tool rankings. What's closed is ordering, approval, scraping, and training: JLCPCB's 14 PCB APIs have no cancel and no validation, the developer agreement lacks the word "AI," and the international site's API requires approval; the project agreement banned scraping and AI training on August 18; PCBWay reserved the right to refuse AI files; Arduino changed its terms for AI features; OpenAI's one-time suspension of FoloToy's account.
The manufacturing chain has the same shape: Autodesk opened Fusion's Claude connector, while the training terms are in a separate document; public MCPs like Slant 3D can quote and generate payment links, with payment completed by humans; Digi-Key's API agreement doesn't mention AI but prohibits bulk downloading and database building.
Reading has no cost and no responsibility; writing and paying involve money on one side and accountability on the other. It's reasonable for vendors to open the responsibility-free half first. But this leaves so-called "Agent native" stuck halfway: it can read manuals and produce boards, but ordering relies on web pages, and there's no API for cancellation.
Pattern three: labs don't build their own bodies, and there's no contract.
OpenAI's first externally sold hardware is Work Louder's keyboard, Anthropic's experiments use Andon Labs' store staff and a ready-made robot dog, DeepMind uses Boston Dynamics' robot, and Anthropic's boards are made by Diode. These are all borrowed. Borrowing means you can return: OpenAI pulled its SDK from the main branch and suspended a toy maker's account. Nothing in the public text constrains these actions. The hardware companies' actions look more like investments: Espressif continued maintaining openai_demo after OpenAI pulled the code, Xiaozhi's repo had ESP32-S31 board support the day before mass production, and Espressif published its own release when Meta's README didn't mention it. Without a contract, the easiest party to walk away is the lab.
Pattern four: which path a dispute takes depends on whether there are actionable rights. Figure 4 already divided this layer's disputes into three paths. Trademark (iyO), trade secrets (Apple), and antitrust (xAI)—those with clearly actionable rights—go to court, with the same lab often on the defendant's bench, two of them related to its hardware projects. Those with only platform power stop at terms, handled unilaterally by one side: WeChat's risk control, the Doubao phone assistant's SAEP public list, OpenAI's suspension of FoloToy, JLCPCB and PCBWay's terms; the other party's recourse is to disable features, or not respond and keep using. This doesn't mean there's no legal right to sue, just that this article found no lawsuits between Tencent, Alibaba, ByteDance, and Doubao. Those involving children, where neither of the first two suffices, go to legislation: China's measures and California's moratorium. Between maker-level hardware companies and labs, none of the frictions found in this article have reached court: not because there's no friction, but because friction happens in READMEs and terms, with no actionable rights.
There are two additional pieces of circumstantial evidence. Partnership and litigation aren't mutually exclusive: Apple and OpenAI are simultaneously partners and opponents. Hardware companies also hedge: Samsung, Rokid, and Apple don't bet their "brains" on a single lab, just as labs don't bet their bodies on a single hardware company—it's the same posture.
The following four are judgments about mechanisms, not date predictions; each includes its basis, how to test it, and what would prove me wrong.
Judgment one: terms will precede interfaces. Basis: JLCPCB added anti-scraping and AI training terms on August 18 first; PCBWay posted its AI file notice on October 28 first; Arduino changed its terms in late October 2025 first, and when responding on November 21, AI features were one of the reasons; while JLCPCB's ordering-side developer agreement still doesn't contain the word "AI," and JLCPCB's API page was last updated on September 9. Test: check whether the next revision of JLCPCB's open platform or JLCPCB API first adds "automation, AI, Agent" language, or first introduces interfaces with cancel, validation, or sandbox. Wrong if: by end of 2027, JLCPCB's ordering APIs add cancel, validation, or sandbox, while the developer agreement still doesn't contain these words.
Judgment two: labs will continue using "reference implementation + developer mode + usage policy" rather than products; hardware companies will continue calling them "official." Basis: OpenAI pulled its SDK from the main branch in March 2025; Anthropic's buddy is marked not officially supported, with the BLE API only available in developer mode; the Model Hardware Standard isn't open-sourced yet; OpenAI's first externally sold hardware is a keyboard made with Work Louder. Test: check whether the next official hardware interface is still labeled "developer mode" or "research preview," and whether hardware companies' press releases use stronger language than the lab's original text. Wrong if: by end of 2027, Anthropic or OpenAI moves an interface for maker boards out of developer mode and into official product support.
Judgment three: labs will continue partnering with commercially related parties that produce their own boards, rather than going through public ordering portals. Basis: Anthropic's November 9 article names Diode, a commercially related PCB design and manufacturing company, as the manufacturer, without mentioning JLCPCB or PCBWay; in design tools, JLCPCB and PCBWay are destinations in a dropdown; PCBWay's integration goes only to the quote page. Test: check who makes the boards the next time a lab publishes electronics design results, and whether a public ordering process is named. Wrong if: by mid-2027, a lab's official article names JLCPCB's or PCBWay's ordering process and states it was completed by an agent.
Judgment four: the next case to reach court will be one with clearly actionable rights, not a terms dispute. Basis: the three cases in Figure 4 that went to court all concern trademark, trade secrets, and competition law; the terms-based frictions of JLCPCB, PCBWay, Arduino, and ByteDance SAEP have so far produced only announcements, lists, and account actions. Test: check whether the next lawsuit involving a lab and a hardware company is based on trademark, trade secrets, patents, or antitrust, or on service terms or scraping. Wrong if: by end of 2027, a lawsuit appears with service terms, scraping, or AI training terms as the cause of action, where one party is an AI lab or maker-level hardware/manufacturing platform.
This piece has several points that can only go so far. The status of OpenAI's own device as of November 10: this article has not verified any news of a unveiling or sale. The claim that Anthropic sent Cardputers to creators comes only from secondary sources and isn't in the main text. JLCPCB's IPO and financial figures are quoted from media retellings of the prospectus. Shifang Ronghai's device counts, revenue, and other figures come from the company's own posts and are inconsistent with each other—they can only be taken as stated figures, not as sales. The phones, glasses, and toys section in §07 mostly comes from media retellings; first-party documents from Apple, Tencent, ByteDance, and OpenAI weren't individually verified; the Doubao phone assistant rejection list numbers range from 150 to 155 across sources; FoloToy's status is only as of PIRG's report dated February 24; California SB 867's effective date isn't clearly stated in the press release; in §08, Mistral 3's license comes from media without checking the model card, Coral's archiving has no official discontinuation notice, and Quilter's roundup is a competitor's retelling. The agreement between Espressif and Xiaozhi wasn't verified by this article, so "no contract" means: none found in public text, not proof of nonexistence.
First published 2026-11-10