🛰️ Google is Sending AI Data Center to Outer Space

Google is testing a new answer to AI’s infrastructure problem by putting compute in space. Project Suncatcher is heading into orbit with four AI chips and the results may shape how future data centers are built. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Google is taking Project Suncatcher into orbit to test whether AI infrastructure can actually work in space.

Google has already put its Trillium TPUs through launch vibration, extreme acceleration and proton radiation tests, with the chips surviving more radiation than they would receive during a five-year mission. The next challenge is proving they can keep working once they are actually in orbit. Cooling is especially difficult because space has no airflow, so the prototype uses heat pipes and radiators to move heat away from the chips. If those systems hold up, future satellites could carry dozens of TPUs and link together into larger AI computing clusters.

Google is testing whether AI’s physical infrastructure can eventually move beyond Earth. Today, scaling AI means securing advanced chips, huge amounts of power, cooling, land and grid capacity. Space could ease two of those pressures with near-continuous solar energy and far fewer land constraints, but only if Google can solve harder problems around heat, radiation, launch costs and networking. That makes this four-TPU mission more important than its size suggests. Google already controls much of the AI stack through Gemini, TPUs, cloud infrastructure and data centers. If Suncatcher works, it could extend that control into where compute physically lives and how it is powered. As AI companies increasingly try to own more of the infrastructure behind their models, orbital compute could give Google a very different way to scale.

Evolving AI: What The State of AI Engineering Report Reveals About AI in Production.

AI engineering has moved past experimentation, but are your observability practices keeping up? Datadog analyzed LLM telemetry from 1,000+ customers to reveal what's actually happening at scale: model adoption shifts, hidden token costs, and what the rise of agentic frameworks means for reliability.

Download the report to benchmark your AI stack against production reality.

Free Download →

Evolving AI: OpenAI upgraded ChatGPT Voice so users can now speak to connected apps, launch agent tasks and hand more complex work to ChatGPT Work.

OpenAI is effectively turning Voice into a front end for ChatGPT Work. Users can speak while Work handles multi-step tasks across connected apps and the browser, interrupt the agent to change direction, and follow the written results inside the same chat. If a task is still running when the voice call ends, Work can continue it in text without starting over. Existing plugin permissions still apply, and actions that require approval have to be reviewed on screen. The result is that Voice can now start, steer and hand off real work instead of ending with a spoken answer.

ChatGPT Voice can now save users from stopping what they are doing just to type out and manage a task. Someone can ask it to check email, review a calendar, gather information, start a document or launch a browser task, then keep steering the work by speaking. If the job takes longer, it can continue in text after the call ends. That makes Voice more useful for work that starts while someone is walking, commuting or away from a desk, and gives OpenAI another way to pull everyday tasks into ChatGPT Work.

Meta is preparing a 10-day global AI hackathon with $1 million in cash prizes. Builders will get access to Meta’s newest AI models, $150 in Model API credits, and hands-on workshops with engineers and researchers from Meta Superintelligence Labs. There will be separate tracks for independent developers and startup teams, and Meta says the event will be virtual, worldwide and free to enter. Applications are not open yet, so you can sign up now to be notified when they launch.

Get notified →

Evolving AI: C5R, an AI research company, built a research facility where AI can manage experiments across biology, chemistry and materials science.

C5R built Facility-0 to give AI models a way to work inside a real lab. Models can inspect available materials and equipment, design experiments as code, then turn those plans into instructions for machines or human operators. Once results come back, the model analyzes them, adjusts its plan and decides what to try next. Humans still handle some physical steps, so the research process is AI-directed rather than completely human-free. Early tests also show where models struggle, including trying to pipette frozen samples, contaminating DNA and failing to account for evaporating solvents.

C5R could matter most by increasing the speed and scale of scientific discovery. If AI can continuously plan experiments, run them, analyze results and choose the next test, researchers could explore far more possibilities than a human-led lab can handle today. That matters in fields like drug discovery and materials science, where progress often depends on testing huge numbers of molecules, formulations or conditions before finding something useful. The models are still unreliable and humans remain essential, but better performance could shorten discovery cycles and let small research teams pursue many more experiments in parallel.

QUICK HITS

👤 Google launched Gemini Live Avatar, giving enterprise AI agents near-real-time faces, voices and lip-sync across 97 languages.

💰 Anthropic signed an $11.6B cloud deal with Akamai for seven years, with the agreement potentially growing to $20B.

🎙️ Nvidia released Nemotron 3 Diarization, an open 100M-parameter model that identifies up to eight speakers in real time.

💻 Perplexity’s Portable Computer agent can now run locally on Windows PCs powered by AMD Ryzen AI Max processors.

 *partner link