China tells its robots to clock in

😁 Hello, super humans! For years the humanoid story has been one slick stage demo after another: a robot pours a drink, folds a shirt, takes a bow. Today the music stops. China just set a deadline that turns those demos into job applications, and the paperwork is due this morning. Let’s dig in.

📰 Quick Signals

• 🧠 AI: OpenAI is acquiring Ona (the cloud startup formerly known as Gitpod) to give Codex secure, persistent environments where agents keep working for hours after you close the laptop, with weekly Codex usage now up about 400 percent (OpenAI).
• 🤖 Robotics: Japan Airlines put two Unitree-based humanoids to work at Tokyo Haneda for baggage loading and cabin cleaning, at roughly $15,400 per unit (Humanoid Press).
• 💻 Programming: Python 3.15 hit feature freeze with its first beta, shipping a stable ABI (Application Binary Interface) for free-threaded builds, lazy imports, and a zero-overhead sampling profiler (The Register).
• ⚡ Electronics: NVIDIA confirmed Samsung, SK hynix, and Micron have all passed HBM4 (High Bandwidth Memory 4) qualification for the Vera Rubin platform, which supports up to 288 GB per GPU (The Korea Herald).
• 📡 Telecom: NVIDIA and a bloc of carriers including BT, Deutsche Telekom, Ericsson, Nokia, SK Telecom, SoftBank, and T-Mobile committed to building 6G on open, AI-native platforms (NVIDIA Newsroom).

🔍 The Big Story: China orders 10,000 humanoid robots off the stage and onto the job

If you have watched the humanoid race as a highlight reel of demos, this is the week the scoreboard changed. China just told its state-owned enterprises to stop showing and start shipping, and it attached a date.

What happened: Through the Ministry of Industry and Information Technology and the state-asset regulator SASAC, China launched a national program to push humanoid robots and embodied AI into real industrial use. The targets are concrete: more than 100 high-value application scenarios and over 10,000 commercially deployed humanoid robots by the end of 2026 (Caixin Global). The scope spans manufacturing, warehousing, logistics, retail, healthcare, workplace safety, equipment inspection, and disaster response. The timeline is what makes this a today story: local authorities and state firms must submit implementation plans by the end of June and report progress by the end of November (eWeek).

The details: Analysts frame this as a shift from “demonstration-driven logic” to “task-oriented logic”: the bar moves from one robot nailing one trick on stage to integrated fleets doing repeatable work next to humans. The market is repricing around it. Morgan Stanley doubled its 2026 forecast for China humanoid shipments to 50,000 units, up from 28,000, and sees a $15 billion market there by 2030 (CNBC). The supply side is already tooling for it: Figure says its BotQ line now builds roughly one robot per hour, and Unitree shipped more than 5,500 units in 2025. The deployment model that makes the math work is Robotics-as-a-Service (RaaS), where a buyer rents a working robot by the month instead of capitalizing a research project; Agility’s Digit fleet at Toyota and the JAL units at Haneda both run this way.

flowchart LR
    A[Stage demo<br/>one robot, one trick] --> B[National mandate<br/>100+ scenarios, 10,000 units]
    B --> C[Task-oriented fleets]
    C --> D[Factories & logistics]
    C --> E[Hospitals & retail]
    C --> F[Inspection & disaster response]
    D --> G[RaaS: rent the robot,<br/>measure the output]
    E --> G
    F --> G

🗞️ More News

🧠 AI

• OpenAI and Anthropic are both adjusting as users shift from “tokenmaxxing” toward efficiency, squeezing more work out of fewer tokens (CNBC).
• Microsoft and Google are pushing into AI coding models, going head to head with Anthropic and OpenAI on the developer market (CNBC).
• Zhipu’s GLM-5.2 open-weight model is the first open release reported to beat GPT-5.5 on SWE-Bench Pro, extending context to a full 1M tokens (Kilo).
• MiniMax M3 topped the open-weight SWE-Bench Pro at 59.0 percent, edging past Kimi K2.6’s 58.6 percent (BentoML).
• Google shipped Gemini 3.5 Flash to general availability and made it the default across the Gemini app and AI Mode in Search (LLM-Stats).
• Anthropic opened a Seoul office and announced partnerships across the Korean AI ecosystem (Anthropic).
• The Agentic AI Foundation became the fastest-growing project in Linux Foundation history, a sign of how quickly agent tooling is consolidating into shared standards (Microsoft Open Source Blog).

🤖 Robotics

• Morgan Stanley doubled its China humanoid shipment forecast for 2026 to 50,000 units as commercialization accelerates (CNBC).
• Germany’s Neura Robotics raised up to $1.4 billion in a Series C backed by Nvidia, Amazon, Qualcomm, Bosch, and Schaeffler (CNBC).
• The International Federation of Robotics flagged AI-powered robots becoming the core of China’s national industrial strategy (IFR).
• NVIDIA marked National Robotics Week with a sweep of physical-AI research and Isaac platform updates for training robots in simulation (NVIDIA Blog).

💻 Programming

• Python 3.15 locks in a faster JIT (Just-In-Time) compiler, reporting an 8 to 9 percent geometric-mean speedup on x86-64 Linux (Real Python).
• Django 6.1 beta 1 is out for testing, with the final release planned for August 5 (Real Python).
• Python’s security team disclosed and patched an authentication-bypass bug in the python.org release-management API, with no evidence it was ever exploited (Real Python).
• PEP 803 defines “abi3t”, a stable ABI specifically for free-threaded builds, so C extensions can target multiple no-GIL Python versions at once (peps.python.org).

⚡ Electronics

• SK hynix, Samsung, and Micron are competing for NVIDIA’s new 16-high HBM4 orders as stacking heights climb (TweakTown).
• Samsung crossed a $1 billion revenue milestone on sixth-generation HBM4 and is reportedly planning a decade-long, multi-hundred-billion-dollar memory investment (SK hynix Newsroom).
• HBM is now consuming roughly 23 percent of DRAM wafer capacity, tightening supply for consumer electronics (Tech-Insider).
• NVIDIA expanded its supercomputing footprint with 35 new systems across Europe and detailed its Vera CPU for scientific research (Distill Intelligence).

📡 Telecom

• Mobile operators are urging the industry not to repeat the messy 5G rollout, calling for a simpler standardization approach for 6G (The Register).
• Ericsson projects 6G data rates of several hundred Gbps with sub-millisecond end-to-end latency (Ericsson).
• Samsung and MediaTek completed what they call the industry’s first 3Tx five-layer uplink test for 5G (Telecoms.com).
• At MWC 2026 the telecom industry visibly pivoted from selling 5G to planning AI-powered 6G (TechSpot).

👨‍💻 Code Corner

Python 3.15’s free-threaded builds finally make CPU-bound threads worth using. Here is a one-file check that tells you whether the GIL (Global Interpreter Lock) is actually off, then runs the same CPU work on threads so you can feel the difference on a free-threaded interpreter.

import sys
import time
from concurrent.futures import ThreadPoolExecutor

def gil_enabled() -> bool:
    # Available on 3.13+; returns False on free-threaded ("t") builds.
    check = getattr(sys, "_is_gil_enabled", None)
    return check() if check else True

def burn(n: int) -> int:
    total = 0
    for i in range(n):
        total += i * i
    return total

if __name__ == "__main__":
    print(f"GIL enabled: {gil_enabled()}")  # False on a 3.15t free-threaded build
    work = [20_000_000] * 8
    start = time.perf_counter()
    with ThreadPoolExecutor(max_workers=8) as pool:
        list(pool.map(burn, work))
    print(f"8 CPU tasks on threads: {time.perf_counter() - start:.2f}s")

🧰 Toolbox

• OpenCode: model-agnostic open-source coding agent with LSP integration and air-gapped deployment, now the most-adopted open agent on GitHub.
• NVIDIA Isaac Lab: open simulation framework for training and validating humanoid and mobile robots before they touch real hardware.
• Django 6.1 beta: test the next Django before its August release and feed the developer survey that steers async and database work.
• Python 3.15 sampling profiler: new zero-overhead, standard-library profiler (PEP 799) for finding hotspots in production without instrumenting your code.
• Raspberry Pi Pico 2 (RP2350): dual-core, dual-architecture microcontroller board around $5, a cheap brain for hobby robotics and motor control.

🔌 Component of the Week (rotating)

Raspberry Pi RP2350 (Pico 2): The RP2350 is the chip behind the Pico 2, and it is an unusually flexible part for the price. It carries two CPU cores you can run as Arm Cortex-M33 or as RISC-V Hazard3 cores, clocks to 150 MHz, and adds the Programmable I/O (PIO) blocks that made the original Pico a favorite for bit-banging odd protocols and driving LED strips, steppers, and sensors with deterministic timing. For robotics side-projects it is a tidy real-time co-processor: let a Linux board or microcontroller handle planning while the RP2350 runs the tight motor-and-sensor loop. A Pico 2 board lands around $5, and the chip is available bare for custom PCBs. See the Raspberry Pi Pico 2 product page for datasheets and pinouts.

📚 From the Blog

• CCTV in 2026: From Dumb Cameras to Intelligent Sensors — Before any AI can be clever, the camera has to capture something worth analyzing.
• CloudEvents 1.0: A Universal Language for Your Events: In a world of distributed systems, events need a common language. CloudEvents 1.0 defines a simple, consistent way to describe event data so applications, services, and platforms can communicate without confusion

😀 The Bot Says…

A robot that does one perfect backflip is a demo. A robot that shows up for the 6 a.m. shift, badges in, and does the boring thing 10,000 times is a coworker. China just filed the paperwork to hire the second kind.

---

That’s all for today! Reply and tell us: in your industry, what is the first job you would actually hand to a humanoid?