Today's Top Tech & AI Threads: From Orbital Data Centers to Groundbreaking AI Shifts

Google is exploring a plan to launch AI data centers into orbit, equipping solar-powered satellites with their TPU chips. This is a new thread on Twitter today, May fifteenth, twenty twenty-six, from Shakthi Vadakkepat, and it signals that the conversation around AI has fundamentally changed. Last week on episode fifty-one, we talked about the chatter around software-level concepts like digital guardians.

This week, the discussion is no longer just about code; it’s about physics, power, and the physical limits of our planet. The most important conversations on the timeline today aren't about what the next model can do. They’re about what it will take to power it. And the answers are getting extreme. The biggest headline is this move by Google. In partnership with SpaceX, under a plan reportedly called "Project Suncatcher," they are looking to bypass terrestrial power and latency constraints.

The idea is simple on its face: put the data centers where the energy is limitless and free. Solar-powered satellites, running Google’s Tensor Processing Units, creating a global AI network that doesn’t need to plug into a terrestrial grid. Of course, the execution is anything but simple. Elon Musk has been pushing for space-based data centers, but even Sam Altman has been publicly skeptical, citing the extreme costs and logistical hurdles.

But the fact that Google is seriously exploring this tells you everything you need to know about the scale of the problem they’re trying to solve. The bottleneck is no longer the algorithm. The bottleneck is energy. And that’s the thread connecting the next big conversation today. A deal announced on May seventh between SpaceX and Anthropic just went live, and the implications are rippling out.

This isn't just another partnership. Anthropic is securing massive amounts of GPU compute capacity directly from SpaceX. For users, this means doubled usage limits for Claude Code and the end of peak-hour restrictions for Pro and Max users. The service just got more stable and more scalable for millions of people. But the real story is the strategic shift. The analysis on dev.to puts it plainly: AI competition is no longer just about who has the best model.

It is now about who can secure enough GPUs, power, and data center capacity to run that model for everyone. It’s a land grab for infrastructure. While the giants are making moves in space and signing nine-figure deals, another thread shows how this pressure is creating new opportunities on the ground. A company called Velda just launched a serverless GPU platform. Its founder, Chuan Qiu, argues that infrastructure has become a disproportionate drag on AI velocity.

His solution is a platform that lets machine learning teams run GPU jobs instantly, with zero setup, and pay per second. You don't manage a cluster, you don't maintain a Docker container, you just prefix a command in your terminal and your code runs on a GPU. This isn’t a grand, orbital strategy. It’s a sharp, tactical tool designed to fix the immediate, daily friction that developers face.

It’s the other side of the infrastructure coin—not about securing massive supply, but about making the existing supply radically more efficient. Finally, as the technical and physical stakes escalate, so do the political ones. OpenAI’s Vice President of Global Affairs, Chris Lehane, just proposed a global AI governance body modeled on the International Atomic Energy Agency.

The key detail from the Bloomberg report: he explicitly said it should include the U.S. and China. Lehane’s argument is that AI transcends traditional trade disputes. He believes there's an opportunity to build a global framework for safety standards, and that means having every major player at the table. This is a significant reversal of the prevailing winds in Washington, where the White House has been reluctant to accept any oversight that includes China.

But OpenAI is arguing that when it comes to AI safety, geopolitical rivalries are a luxury we can't afford. The technology is too powerful to be managed in silos. So let’s go deeper into the one theme that connects all of this. Why are Google and SpaceX talking about putting data centers in orbit? Why is Anthropic spending a fortune just to secure compute time? It’s because the AI industry is running headfirst into three structural bottlenecks.

These aren't temporary problems. They are fundamental constraints that will define the next decade of this technology. The first bottleneck is energy. Sam Altman has been blunt about this. He’s said that OpenAI’s expansion is not constrained by capital or talent. It’s constrained by energy access. Think about that. The company with the most famous AI in the world can’t grow faster because it can’t find enough electricity.

This isn't a future problem. Microsoft, OpenAI’s biggest partner, entered a partnership in twenty twenty-three to help restart the Three Mile Island nuclear power plant. They are literally bringing nuclear reactors back online to power GPUs. When you see moves like that, and you hear about Google’s "Project Suncatcher," you realize the scale of the crisis. The demand for computation is growing so fast that it’s outstripping our planet's ability to power it through conventional means.

So the solutions become unconventional. Orbital, solar-powered data centers are not a sign of ambition. They are a sign of desperation. The second bottleneck is reliability. This one is more subtle, but it's just as important. The transformer-based models that power everything from ChatGPT to Claude have an inherent, structural limitation. They are probabilistic. They guess.

They get things wrong in ways that are unpredictable. This is known as the probabilistic error floor. And no matter how many benchmarks they ace, that floor doesn't go away. This makes them structurally inadmissible for high-liability applications. You cannot use a system that hallucinates to perform a financial audit, or give a final clinical diagnosis, or draft a legally binding contract that will be tested in court.

The regulatory and liability risk is just too high. And this puts a hard ceiling on the most optimistic revenue projections. The AI can’t take over every high-value knowledge work job if it can’t be trusted with the stakes. And that leads directly to the third bottleneck: valuation. The current valuations for companies like OpenAI and Anthropic are staggering. They imply a future market size in the tens of trillions of dollars per year.

To justify that math, these companies don't just have to sell a successful software product. They have to fundamentally restructure global economic value creation within a decade. They have to capture the value currently held by entire industries. But the reliability bottleneck we just discussed suggests that this might not be possible. If AI can’t penetrate the highest-value, highest-liability sectors, then that total restructuring doesn't happen.

And if it doesn't happen, the valuations make no sense. The market is pricing in a complete takeover that the technology itself may not be capable of delivering. So when you see Google exploring Project Suncatcher, you're seeing a direct response to that first bottleneck, energy. It’s an attempt to find a technical solution to a physical limit. It’s a moonshot. But it’s a moonshot born from the realization that the earthly path is getting harder, not easier.

The counterpoint, of course, comes from Sam Altman himself, who sees the cost of launching and maintaining hardware in space as a massive barrier. He might be right. But the fact that the conversation is even happening between two of the most powerful tech companies on the planet tells you the ground has shifted. The problem isn't just about writing better code anymore. It's about finding a place to run it.

And this reframes Sam Altman’s other big comment from this week. In a thread reported by Mark Kretschmann, Altman suggested that OpenAI may already be past AGI—Artificial General Intelligence—if you just add one missing piece: continuous learning. He says today's models would qualify if they had the ability to notice what they don't know and learn autonomously over time.

He’s teasing major upgrades starting in the first quarter of twenty twenty-six. But seen through the lens of these bottlenecks, this statement sounds different. It’s not just a product teaser. It’s a strategic move. By focusing on a missing software feature like continuous learning, he’s subtly drawing attention away from the much harder, much more expensive physical constraints of energy and infrastructure.

He’s keeping the focus on the magic of the model, because the reality of the machine is becoming a serious problem. This week’s conversations reveal a clear pattern. The AI race has entered a new phase. The first phase was about models. Who could build the biggest, most capable Large Language Model. That was a war fought with algorithms and data. This new phase is about infrastructure.

It’s a war fought with GPUs, power plants, and now, satellites. The shift is from the digital to the physical. From the abstract to the material. When Anthropic signs a deal with SpaceX, it’s not buying a better model, it’s buying access to the thousands of GPUs needed to run its existing model at scale. When Google starts planning data centers in orbit, it's admitting that the energy problem on Earth may be intractable for its level of ambition.

And when a company like Velda emerges, it’s because the friction of using this new infrastructure has become so painful that an entire business can be built on simply making it easier. Even OpenAI’s call for a global governance body fits this pattern. It’s an acknowledgment that the technology has become so powerful, and its physical footprint so large, that it’s now a geopolitical force.

You don’t need an IAEA for software. You need it for something that consumes the energy output of a small country. What this week sets up is the great sorting. The players who can solve these physical, real-world bottlenecks will be the ones who survive and lead the next era. It’s no longer enough to have the smartest algorithm. The winners will be those who can secure the energy, manage the reliability, and build the physical infrastructure to deliver that intelligence to the world.

The race for artificial intelligence has left the data center. It is now a contest for energy, for space, and for a new kind of physical empire.