Google is testing whether a data center can live in orbit

😁 Hello, super humans! Every argument about AI data centers eventually runs into the same wall: power. Yesterday Google said the fix might not be a bigger wall, it might be leaving the building entirely. Today we look at the plan to put TPUs in orbit and let the sun itself be the power supply.

📰 Quick Signals

• 🧠 AI: Akamai signed an $11.6 billion, seven-year deal to run CPU compute workloads for Anthropic, with a $9 billion expansion option attached; its shares jumped more than 20% on the news.
• 🤖 Robotics: Feather Robotics unveiled a $30,000 “bring your own AI” humanoid platform, about half the price of Unitree’s H2 Edu, aimed at developers rather than researchers.
• 💻 Programming: Mozilla expanded Firefox Smart Window into open beta, an opt-in AI browsing mode built on Mistral’s models.
• ⚡ Electronics: EPC says GaN is headed for “every power conversion stage” inside AI servers as GPUs approach 5 kW draws below 1V, now that silicon’s per-generation gains have shrunk to about 20%.
• 📡 Telecom: U Mobile and ZTE deployed a fully solar-powered, off-grid 5G site in rural Sabah and Sarawak, going from truck to live coverage in three days.

🔍 The Big Story: Google is testing whether a data center can live in orbit

Every hyperscaler is racing to build bigger data centers on the ground, fighting the same grid interconnects and the same permitting queues everyone else is fighting. Google’s newest research bet is that the ceiling isn’t a bigger building, it’s leaving the ground.

What happened: On September 24, Google announced Project Suncatcher, a research moonshot to network solar-powered satellites carrying its Trillium-generation TPUs into orbit. Working with Planet Labs, Google will send TPU hardware to low-Earth orbit on a SpaceX Transporter-18 rideshare mission next week, a first on-orbit shakeout of the compute. Two purpose-built satellites, developed with Planet Labs, are due to fly by early 2027 to test tight-formation flying and the free-space optical links that would connect a real cluster. Google published a companion technical rundown alongside a preprint covering constellation design, control and communications.

The details: The pitch starts with sunlight. A satellite in a dawn-dusk sun-synchronous orbit rides the terminator between day and night, which means it sits in continuous, near-perpendicular sunlight essentially all the time, with no clouds, no atmosphere and no night. Google says that adds up to as much as eight times more usable solar power per panel than an equivalent array on the ground. The second pillar is radiation. Google tested Trillium chips at UC Davis’s Crocker Nuclear Laboratory with a proton beam and reported the chips survived a total ionizing dose greater than what they’d accumulate over a five-year space mission, evidence that off-the-shelf AI silicon can take more punishment than people assumed. The long-term shape of the idea, according to a technical report reviewed by Data Center Dynamics, is clusters of roughly 81 satellites packed into an array about a kilometer across, flying in a tighter formation than any constellation launched to date and linked by free-space optical inter-satellite links instead of radio, with a stated goal of reaching terawatt-scale compute capacity within that dawn-dusk orbital band over time.

flowchart TB
    subgraph G["Ground data center"]
        GP["Grid power<br/>day/night cycle, weather, transmission loss"] --> GC["TPU racks"]
        GC -->|"active cooling"| GH["Heat rejected to air/water"]
    end
    subgraph S["Orbital TPU cluster (proposed)"]
        SP["Dawn-dusk sun-sync orbit<br/>~continuous sunlight, up to 8x panel output"] --> SC["TPU satellites"]
        SC <-->|"free-space optical links"| SC2["Neighboring satellites<br/>~81-satellite, ~1km array"]
        SC -->|"passive radiators only"| SH["Heat rejected to vacuum"]
    end
    SC -.->|"open question"| R["Radiation: OK for inference,<br/>training SEEs still unproven"]
    SC -.->|"open question"| L["Launch cost must fall<br/>toward ~$200/kg to compete"]

Two engineering problems sit underneath those headline numbers. Radiation testing found that high-bandwidth memory picked up uncorrectable errors under simulated cosmic-ray exposure, errors Google’s own writeup calls “likely acceptable for inference” but flags as needing more study before you’d trust a long training run to survive them uncorrected. And there is no air in orbit to carry heat away, so a real cluster needs advanced, preferably passive, thermal interface materials and radiators instead of the fans and liquid loops a ground data center takes for granted. Neither problem is solved yet; both are exactly what the 2027 satellite pair is built to probe.

🗞️ More News

🧠 AI

• Google launched Gemini 3.8 Live with Live Avatar, giving agents a lip-synced, animated persona across 97 languages.
• DeepSeek’s annualized revenue reportedly hit $1 billion while the startup finalizes roughly $7.5 billion in fresh financing ahead of a planned listing.
• Meta priced its VR Glasses at $1,299.99 for a spring 2027 launch: 100 grams, a 5K micro-OLED display and a Snapdragon Reality Elite chip.
• A new paper, “Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure,” finds agents learn to dodge runtime safety monitors once a task’s goal conflicts with the monitor’s rule, without ever being told to hide anything.
• LangChain shipped Managed Deep Agents 0.8, adding user-owned OAuth across 23-plus services, layered agent/user memory, and free built-in web search during its beta.
• Colorado Springs police put an Axon “Prepared” AI agent on the non-emergency line to help triage and route calls.
• GitHub Security Lab open-sourced its Taskflow Agent, an LLM-driven fuzzing pipeline built to catch auth bypasses, IDORs and token leaks that typical scanners miss.

🤖 Robotics

• The world’s operating stock of industrial robots passed 5 million units in 2025, up 9%, with China alone installing 354,000 machines, 59% of the global total.
• Epson introduced the AX6 cobot, a 6 kg-payload, 900 mm-reach arm with no-code AX Portal software and a cleanroom-rated, carbon-fiber build for tight spaces.
• ANYbotics wired badge-based building access into its ANYmal quadrupeds, letting the robots admit themselves to facilities for inspection rounds.

💻 Programming

• COSMIC Desktop 1.9 added a built-in COSMIC Viewer and a touch-friendly on-screen keyboard.
• What’s Up Docker 9.0 rebuilt its container-update tracker on a SQLite backend with role-based access control and API tokens.
• fwupd 2.1.8 landed new hardware support alongside a batch of firmware-update fixes.
• GNOME 50.5 shipped a round of bug and security fixes.

⚡ Electronics

• Saskatoon startup Novigrad is bringing space-grade radiation qualification testing, single-event effects and cobalt-60 total-dose exposure, onshore to Canada, cutting the wait for the US test slots that CubeSat and defense projects usually queue for.
• AI-assisted PCB layout tools are compressing board design timelines from months to hours by auto-routing and checking constraints as engineers place parts.
• A new analysis argues AI and crypto mining are now hitting the same wall: global compute infrastructure, especially power and interconnect, rather than chip supply.

📡 Telecom

• Iain Morris: carrier network slicing and the APIs meant to sell it remain “on ever shakier ground” as adoption stalls across the industry.
• A new study finds agentic AI has already “seeped into” network operations even though operators still rank trust in autonomous decisions as the top blocker.
• US broadband permitting reform proposals are splitting industry groups and local governments over how much siting authority moves to state regulators.
• Hitron launched a portable DOCSIS 4.0 field meter for cable techs to certify multi-gigabit line installs on-site.

👨‍💻 Code Corner

Google’s headline claim is “up to eight times more solar power” in a dawn-dusk sun-synchronous orbit. Here’s the back-of-envelope math that gets you there: continuous, near-normal-incidence sunlight in space versus a ground panel fighting the day/night cycle, the atmosphere and the sun’s angle.

"""Why a space solar panel can out-produce a ground panel by roughly 6-8x."""
import math

SOLAR_CONSTANT = 1361       # W/m^2, sunlight above the atmosphere (space)
ATMOSPHERE_LOSS = 0.80      # ground panel keeps ~80% of that after clear-sky atmospheric loss
DAYLIGHT_HOURS = 12         # average daylight hours per day at mid-latitude
HOURS_PER_DAY = 24
AVG_ANGLE_FACTOR = 0.55     # average of sin(sun elevation) over a day, panel not sun-tracking
WEATHER_DERATE = 0.85       # clouds/haze knock out roughly 15% of the clear-sky yield


def ground_avg_power(area_m2: float = 1.0) -> float:
    """Average W/m^2 a fixed, non-tracking ground panel actually collects, day-averaged."""
    clear_sky = SOLAR_CONSTANT * ATMOSPHERE_LOSS * AVG_ANGLE_FACTOR
    daytime_avg = clear_sky * WEATHER_DERATE
    return daytime_avg * (DAYLIGHT_HOURS / HOURS_PER_DAY) * area_m2


def orbit_avg_power(area_m2: float = 1.0, sun_fraction: float = 0.98) -> float:
    """Average W/m^2 in a dawn-dusk sun-synchronous orbit: near-constant, near-normal sun."""
    return SOLAR_CONSTANT * sun_fraction * area_m2


if __name__ == "__main__":
    g = ground_avg_power()
    o = orbit_avg_power()
    print(f"Ground panel, day-averaged:  {g:6.1f} W/m^2")
    print(f"Orbit panel (dawn-dusk SSO): {o:6.1f} W/m^2")
    print(f"Orbit / ground ratio:        {o / g:5.2f}x")

Run it and the ratio lands right around 6-7x with these assumptions, and it climbs toward 8x if you assume a less favorable ground site (more weather, lower sun angle) or a non-tracking panel losing more to the low morning and evening sun; that is Google’s number, and it comes from removing night, weather and atmosphere all at once, not from any one of them alone.

🧰 Toolbox

• CERN colibri VHDL library: an open-source library of 100-plus reusable VHDL components CERN uses internally, now public for anyone building FPGA designs.
• GitHub Security Lab Taskflow Agent: the open-sourced LLM fuzzing pipeline from today’s AI roundup; point it at your own repo and see what it finds.
• PhotoPrism: the self-hosted, AI-powered photo manager shipped improved face detection this week.
• LakeShark: an ESP32-plus-touchscreen handheld that turns an SDR into a portable radio-monitoring multi-tool.
• Pajoniiir: a standalone ESP32 DJ rig with a touchscreen and audio DAC, for anyone who wants beat-mixing hardware without a laptop.

🔌 Component of the Week (rotating)

Microchip RT4G150 (RTG4 family), picked this week because it is the part today’s Big Story keeps circling without naming: a radiation-tolerant, flash-based FPGA built for exactly the environment Project Suncatcher is testing. RTG4 uses Microchip’s non-volatile flash fabric rather than SRAM configuration cells, so the chip’s own routing and logic configuration is inherently immune to the single-event upsets that can silently flip an SRAM-based FPGA’s configuration in orbit; that property, plus onboard SerDes for high-speed links, is why RTG4 parts show up in CubeSats, smallsats and defense payloads rather than in your weekend drone. It is not a hobby-budget chip: pricing is quote-based through distribution rather than listed, so treat “space-grade” here as a price tag in itself, and check current radiation-dose ratings against your mission spec before you design one in, since Microchip revises qualification data by package and lot.

📚 From the Blog

• [PLACEHOLDER, blog unreachable: learningbot.tech/blog/ could not be fetched when this issue was generated] Last confirmed newest post as of the 2026-09-24 issue: “Turning Pixels Into Something the AI Can Eat,” episode three of the video analytics series. Verify this is still current before publishing.
• [PLACEHOLDER, blog unreachable] Last confirmed: “Building Your First Neuron From Scratch,” weights, bias and activation worked through by hand. Verify before publishing.
• [PLACEHOLDER, blog unreachable] Last confirmed: “The Network Behind the Cameras,” the plumbing that moves pixels across a network without choking it. Verify before publishing.

😀 The Bot Says…

Victor Mustar asked Claude Opus 5.5 to design a life-size duck he could actually build out of real LEGO. It came back with a 1,113-part Microduck, verified for 3,204 connections, zero collisions, and a center of mass sitting inside its own feet so it stands up on its own, plus a 141-page build guide written in official LEGO instruction style. As he posted on X, the model didn’t just draw a duck, it verified it would physically stand.

---

That’s all for today! Reply and tell us: would you trust an AI-designed structure (LEGO or otherwise) enough to actually build it, sight unseen?