🤖 Can We Build Data Centers in Space?

🛰️ Can We Build Data Centers in Space?

💭 AI’s Dreaming up Millions of Materials. Are They Any Good?

📍AI Tier Tracker

🛠Trending Tools

🥪Brief Bites

💰Funding Frontlines

💼Who’s Hiring?

Today’s language models are getting bigger, faster, and hotter. In other words, the GPUs that run your favorite conversational chatbots like Claude, Gemini, and ChatGPT generate massive amounts of heat. This heat forces these GPUs to slow down, a process called “thermal throttling,” to prevent permanent damage. That’s why we want them cooler, so they can run faster and last longer.

We imagine space as a freezing cold void. In reality, it’s an airless heat prison. The only viable path to remove heat is through radiation, meaning every Watt of electricity used by GPUs must be converted into infrared light and beamed away into the void. It’s the least efficient form of heat transfer, requiring fragile and expensive radiator panels.

At a basic level, language models are essentially billions of numerical values called weights stored as “bits” (i.e., 0s or 1s) in a high-performance processor’s memory. Space is flooded with cosmic radiation and solar particles that constantly barrage electronics, flipping these “bits” into corrupted values.

The real dream killer of “space-based data centers”: the price tag. It’s brutally expensive. At a rate of roughly $1,000 per kg, launching a single standard server rack, which is easily at least 1,500 kg, would cost around $1.5 million.

For context, to replicate just 1% of Earth’s total computing capacity in orbit, you’d need to launch over twice the total mass humanity has ever sent to space. The capital required to achieve this would be in the trillions of dollars, just to save on electricity bills.

GenAI has been leveraged to generate enormous libraries of theoretical “super-materials.” For example, Microsoft’s “MatterGen” constantly designs atomic structures with specific magnetic, mechanical, and electronic properties. Now, the biggest bottleneck is actually figuring out how to make them.

“DiffSyn” relies on diffusion models, which gradually add random variation (i.e., “noise”) to more than 23,000 material synthesis recipes from over 50 years of scientific literature.

Then, “DiffSyn” systematically removes that “noise” to uncover hidden, high-quality patterns that enable it to learn the underlying statistical sequences for successful material synthesis planning. For instance, identifying the perfect combinations of raw ingredients and cooking conditions to actually construct a new material.

In 1980, the discovery of LiCoO2 by Professor John B. Goodenough enabled the creation of modern lithium-ion batteries, which power today’s smartphones, laptops, and computers. In the past few years, GenAI has assembled millions of “super-materials.” Yet, about 99.99% of them only exist on a computer screen.

It can take a synthetic chemist up to 10 years of trial-and-error to find the right recipe to make a “super-material.” The creation of “DiffSyn” aims to provide synthetic chemists with the right recipes immediately, replacing guessing with engineering.

Revena raised an $8M Seed Round to cut hospital revenue losses.

Linq raised a $20M Series A for AI-enabled assistants within iMessage.

Lotus Health AI landed a $35M Series A to provide world-class primary care for free.

Ebury (Paris, FR): Commercial Analyst Intern, Summer 2026

NVIDIA (Santa Clara, CA): Circuits Timing Engineer, Entry-Level

Figure AI (San Jose, CA): Helix AI Engineer, Robot Learning, Mid-Level

BitGo (Palo Alto, CA): Director of Engineering, Wallets, Senior-Level