GPT-5.2, Memory & space-GPU dramas, & robust robots

TLDR;

Welcome back! This week we have another week packed with insights and fresh news, from Gemini’s new deep research tool setting new records, OpenAI releasing GPT-5.2 literally seconds ago, the evolving drama around space GPUs, or the increasing supply shortage pressures in the memory chip market that could redefine AI in 2026.

At the end, we’ll also look at our portfolio, including a refreshing new look at one of the big stock winners so far in 2025 that could have an even better 2026, and a new aspect that most have missed: in AI agents, GPUs aren’t the bottleneck, meaning most people have AI hardware wrong.

Enjoy!

SPACE

Training Models in Space… Works?

One of the hottest debates in AI today is whether space-based data centers make sense. Everything started after NVIDIA announced that a startup they had invested in, Starcloud, was sending an H100 GPU into space.

Many, including NASA engineers, have since said this would never work. Yet, not only is NVIDIA excited about it, but Sundar Pichai, Google’s CEO, literally can’t stop talking about it.

In the meantime, Starcloud, the startup NVIDIA supports, has announced its first training and inference runs. They trained nanoGPT, a minute Large Language Model (LLM) trained on the works of Shakespeare, and also managed to run inference of a Google Gemma model. Quite the success for a project that was never intended to work.

TheWhiteBox’s takeaway:

There’s far from being a consensus on this matter, and the debates are heated. For Starcloud, this is the only way: let results do the talking. Whether this is just pure desperation from Hyperscalers running into power constraints on Earth or really a genius idea, time will tell (I honestly have no idea, I’m not a rocket scientist).

What I will say is that the debate is artificially created. We are artificially constraining power supply in the West due to restrictions on nuclear and solar power, two technologies that could scale to our needs (the former is essentially unlimited power, the latter has the shortest lead times among all power technologies).

And talking about power…

THEWHITEBOX

From Supersonic engines to Data Centers

One of the coolest recent stories is a supersonic jet company repurposing to build data center turbines. As the CEO explains in this post, it’s quite the story.

They initially aimed to deliver the supersonic product first (a jet that is faster than the speed of sound) and then the energy product (using the turbines for power generation). But as they saw the emerging importance of power in the US, they shifted and have since landed a 1.2 GW contract with Crusoe (one of the primary data center builders for OpenAI), based on their 42 MW turbines for a total value of $1.25 billion, which nets at around $1 billion/GW.

This has allowed them to land $300 million in new financing, a story that explains how it’s always about building something people want.

But why is this a big deal?

TheWhiteBox’s takeaway:

You…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.