Why Google Wants to Put Data Centers in Space | EP 162

Summary

This packed episode covers three major stories. First, Kevin digs into Google’s Project Suncatcher, an ambitious plan to build data centers in low Earth orbit where solar panels can be up to eight times more productive than on Earth. With terrestrial energy grids struggling to power the AI boom, space-based computing is moving from science fiction to serious corporate research, with Google planning prototype satellite launches in 2027.

Second, the hosts interview Dean Ball, a former senior policy adviser for AI at the Trump White House, who provides a rare insider view of how the administration’s AI Action Plan came together. Ball describes the competing factions on the right — from David Sacks’ anti-doomer accelerationism to Steve Bannon’s existential risk concerns — and argues that many AI safety problems are “eminently tractable” without slowing development.

Finally, in a “history mystery” segment, Professor Mark Humphries describes how a mystery Gemini model (likely an unreleased Gemini 3) performed what appears to be symbolic reasoning when transcribing 18th-century ledger documents, correctly converting between historical currency systems in a way that models “shouldn’t be able to do.” This suggests scaling laws may still be producing emergent capabilities.

Highlights

”We Are Going to Build Freaking Data Centers in Space”

yt-dlp --download-sections "*2:23-3:30" "https://www.youtube.com/watch?v=UBe4_V04Okg" --force-keyframes-at-cuts --merge-output-format mp4 -o "UBe4_V04Okg-2m23s.mp4"

“We are at the stage of this bubble where people have come to feel like we cannot provide enough electricity for the future we want to build on the planet that we live on. We actually have to get off the planet to realize our ambitions.” — Casey Newton, 2:23

”When Are the Pitchforks Going to Be Out?”

yt-dlp --download-sections "*30:00-31:00" "https://www.youtube.com/watch?v=UBe4_V04Okg" --force-keyframes-at-cuts --merge-output-format mp4 -o "UBe4_V04Okg-30m00s.mp4"

“The main question that people talk about is when are the pitchforks going to be out for this technology and what is going to cause the pitchforks to come out? I have come to the conclusion that rather than it being a singular issue, it’s going to be this kind of miasma of issues.” — Dean Ball, 30:00

”The Model Did Something Models Shouldn’t Be Able to Do”

yt-dlp --download-sections "*55:00-56:10" "https://www.youtube.com/watch?v=UBe4_V04Okg" --force-keyframes-at-cuts --merge-output-format mp4 -o "UBe4_V04Okg-55m00s.mp4"

“The model had to work backwards from a different currency system with a different base. The thing that makes that important is that models shouldn’t be able to do that. These models are basically trained in pattern recognition.” — Mark Humphries, 55:00

”NIMBYs to KNOPs: Not On My Planet”

yt-dlp --download-sections "*15:00-15:50" "https://www.youtube.com/watch?v=UBe4_V04Okg" --force-keyframes-at-cuts --merge-output-format mp4 -o "UBe4_V04Okg-15m00s.mp4"

“We’ll have gone from the NIMBYs saying not in my backyard to this new group of people that I’m calling the KNOPs that are saying not on my planet, and they want all the data centers just built up in the sky.” — Casey Newton, 15:00

”You Posted Your Way to a Job in the Federal Government”

yt-dlp --download-sections "*20:00-20:50" "https://www.youtube.com/watch?v=UBe4_V04Okg" --force-keyframes-at-cuts --merge-output-format mp4 -o "UBe4_V04Okg-20m00s.mp4"

“My Substack is called Hyperdimensional. I was recruited to the White House on the basis of primarily my writing on Substack. Probably a big chunk of it was my posts on X, which is maybe even more scary.” — Dean Ball, 20:00

Key Points

• Project Suncatcher (2:21) - Google’s plan to build data centers in low Earth orbit; solar panels up to 8x more productive in space
• 2027 Test Launch (10:00) - Google planning to launch two prototype satellites in partnership with Planet
• Sun Emits 100 Trillion Times (6:00) - The sun emits about 100 trillion times as much energy as the entire output of humanity
• Dawn-Dusk Orbit Path (6:30) - Special orbit that provides near-constant sunlight to space-based solar panels
• TPU Radiation Testing (9:00) - Google tested TPUs with proton beams; they withstood radiation beyond 5-year mission expectations
• StarCloud (10:30) - Y Combinator startup funded by Nvidia also building space data centers
• AI Policy Factions (18:00) - Ball identifies spectrum from David Sacks’ anti-doomer view to Steve Bannon’s existential risk concerns
• Kid Safety Bipartisan Issue (22:00) - Ball says Republicans concerned about LLM psychosis and teen suicidality with chatbots
• California as Default AI Regulator (35:00) - Ball argues California has become the de facto federal regulator of AI by default
• Woke AI Executive Order (28:00) - Applies only to government procurement, not consumer models; requires disclosure of system prompts
• Mystery Gemini Model (48:53) - Historian Mark Humphries found a model in Google AI Studio with radically better performance
• 1% Word Error Rate (52:00) - Mystery model achieved human-expert-level transcription accuracy on old handwritten documents
• Symbolic Reasoning in LLMs (55:00) - Model correctly converted 18th-century pounds/shillings/pence by working backwards from totals
• Scaling Laws Not Exhausted (58:00) - The historian’s findings suggest continued emergent properties from increased scaling
• Tail Risk Tractability (40:00) - Ball argues catastrophic AI risks are “eminently tractable” without slowing AI development

Mentions

Companies

• Google (2:21) - Project Suncatcher for space-based data centers; partnering with Planet for 2027 prototype
• StarCloud (10:30) - Y Combinator startup funded by Nvidia building space data centers
• Planet (10:00) - Satellite company partnering with Google on Suncatcher prototypes
• Axiom Space (11:00) - Another company exploring space-based data centers
• Nvidia (10:30) - Funded StarCloud’s space data center startup
• Anthropic (35:00) - Announced $50 billion commitment to build own data centers; Casey’s boyfriend works there
• OpenAI (35:00) - Referenced via Stargate data center project
• TSMC (35:00) - Chip fabrication company all hyperscalers compete for space at

Products & Technologies

• Project Suncatcher (2:21) - Google’s space-based AI infrastructure research project
• TPUs (9:00) - Google’s AI training chips tested for radiation resistance
• Starlink (8:00) - Referenced as proof that satellite data transmission already works
• Google AI Studio (48:53) - Google’s developer playground where mystery model was tested
• Gemini 3 (48:53) - Likely identity of the mystery model in AI Studio; expected December release
• SB53 (35:00) - California’s AI transparency bill that Ball found “rather reasonable”

People

• Dean Ball (17:48) - Former White House senior policy adviser for AI; author of Hyperdimensional newsletter
• Mark Humphries (48:53) - History professor at Wilfrid Laurier University who discovered mystery Gemini model’s capabilities
• Leanne Ley (50:00) - Humphries’ research partner working on AI-assisted historical transcription
• David Sacks (18:00) - President’s AI czar, anti-doomer faction
• Steve Bannon (18:00) - Represents the existential risk concerned faction on the right
• Jeff Bezos (11:00) - Interested in space-based data centers; sent Lauren Sanchez to space
• Eric Schmidt (11:00) - Also interested in space-based computing
• Alex Heath (49:00) - Reported on expected December release of Gemini 3

Surprising Quotes

“We are at the stage of this bubble where people have come to feel like we cannot provide enough electricity for the future we want to build on the planet that we live on. We actually have to get off the planet to realize our ambitions.” — Casey Newton, 2:23

“30-year veterans of thinking about tail risks — to them, you bring up AI bio-risk or AI cyber risk and they’re like, ‘Yeah, sounds like a serious risk. Okay, there’s a hurricane that’s tracking toward Florida. Let me go deal with that.’” — Dean Ball, 40:00

“Models shouldn’t be able to do that. These models are basically trained in pattern recognition. Predicting numbers is actually very difficult for models to do. The model had to work backwards from a different currency system with a different base.” — Mark Humphries, 55:00

“If we can’t deal with catastrophic tail risk, then we do not have a legitimate government. The whole point of government is to deal with this issue and we should just throw the thing out and return the money to the shareholders.” — Dean Ball, 42:00

“If you’re using group chats to conduct war or be racist, it could come back to bite you, which is a great reason not to conduct war or be racist.” — Casey Newton (from the Signal group chat discussion), 2:21

Transcript

0:00 Suncatcher, which sounds like a lost Led Zeppelin single, but is somehow a project to build data centers in space. The main question that people talk about is when are the pitchforks going to be out for this technology? So what I’m seeing turning on here for historians is that up to this point the models, we’ve been getting this sense that they’re starting to get good enough where you can feel like, yeah, I think I can trust the outputs on this.

0:37 Casey’s photo op flop. So the other day I’m walking down Market Street and this very sweet young woman comes up to me and asks for a picture. She hands me her phone and they go and stand up against the street with their backs turned. And that’s when I realized these people have no idea who I am. They are just tourists and they want a picture of themselves in San Francisco.

1:43 I’m Kevin Roose, a tech columnist at the New York Times. I’m Casey Newton from Platformer, and this is Hard Fork. This week, Google’s crazy new plan to build data centers in space. Then, former Trump White House policy adviser Dean Ball tells us what Republicans really think about AI. And finally, it’s a history mystery with Professor Mark Humphries.

2:21 Well, Casey, today we are going to start by talking about space. We are going to build freaking data centers and put them in space. We are at the stage of this bubble where people have come to feel like we cannot provide enough electricity for the future we want to build on the planet that we live on.

4:00 Google has put out a paper on what it calls Project Suncatcher. They’re calling this a moonshot. The problem they’re trying to solve is twofold: it’s very hard to build stuff here on Earth, and the sun is a really good source of energy — it emits about 100 trillion times as much energy as the entire output of humanity. If you put solar panels in low Earth orbit on a dawn-dusk orbit path, they can be up to eight times as productive.

8:00 How does whatever is happening inside the giant terrifying bird get back to us on Earth? This is actually not that much different from Starlink. You’re sending data from satellites back to Earth. It might take a couple more milliseconds.

9:00 Google took a normal TPU and hit it with a proton beam simulating intense space radiation. Their newer TPUs actually withstood radiation much better than they thought — beyond what’s expected in a 5-year mission.

10:00 Google is planning to test this in 2027 by launching two prototype satellites in partnership with Planet. Other companies include StarCloud, a Y Combinator startup funded by Nvidia, and Axiom Space. Jeff Bezos and Eric Schmidt have also expressed interest.

15:00 We’ll have gone from the NIMBYs saying not in my backyard to the KNOPs saying not on my planet. Space debris is one concern, though experts say it shouldn’t be a big deal. Google positioned Suncatcher in the lineage of self-driving cars and quantum computing — willing to work on this for 8 to 15 years.

17:48 Interview with Dean Ball. Dean was a think tanker recruited to the White House primarily based on his Substack writing and X posts. He found coherent intuitions but not yet concretized policy positions. Intuition one: AI is the most important opportunity in decades. Two: there will be familiar and alien risks. Three: AI will play a huge role in American global leadership.

22:00 Ball describes the intra-right factions: David Sacks (anti-doomer), Steve Bannon (existential risk), national security people, and the kid safety concerned group worried about LLM psychosis and teen suicidality.

28:00 The “woke AI” executive order applies only to government procurement — not consumer models. It requires developers to disclose system prompts to government agencies. Ball argues using federal law to compel developers to change public models would be “unambiguously unconstitutional.”

35:00 Ball argues California has become the default central regulator of AI. He was broadly supportive of SB53, a transparency bill. Congress needs to act on AI policy, particularly on kid safety.

40:00 Ball says tail risks are “eminently tractable” — improved biosecurity, for example, was already needed before ChatGPT. He argues there’s no trade-off between mitigating tail risk and slowing AI development in the near term.

48:53 History mystery with Professor Mark Humphries from Wilfrid Laurier University. He and research partner Leanne Ley have been using AI to transcribe handwritten 18th-century fur trade documents. When testing a mystery model in Google AI Studio, he found one that achieved a 1% word error rate — matching human experts.

55:00 Most impressively, the model correctly interpreted tabular data from an 18th-century ledger written in pounds, shillings, and pence. It figured out that “145” meant 14 pounds 5 ounces of sugar by working backwards through a non-decimal currency system — a form of symbolic reasoning that models shouldn’t theoretically be able to do.

58:00 The implications: if this is Gemini 3, it suggests scaling laws are still producing emergent capabilities. What historians are seeing turning on is highly likely to start turning on in other knowledge work areas as well. Kevin confirmed Google does test unreleased models in AI Studio, making it likely this was an early Gemini 3 test.