Google Put AI in Space. It Made Me Obsessed With My Own PC's Limits.
Google's 'Project Suncatcher' launched AI chips into orbit, which inspired me to create a personal LLM speed test to find out what my own hardware can actually
The short version
Google sent AI hardware into orbit on a SpaceX rocket. It’s a wild experiment to see if their chips can survive the harshness of space. This whole thing got me thinking about the physical limits of AI, not just in space, but right here on my desk. So I built a simple benchmark to test what my own computer is really capable of.
So Google really launched AI into space?
Yes, it really happened. On October 1st, Google quietly launched a refrigerator-sized satellite into orbit. The whole mission, dubbed ‘Project Suncatcher,’ rode into space on a SpaceX Falcon 9 rocket. This wasn’t for a new mapping service or satellite internet. The payload was a set of four of Google’s own AI chips. It’s a fascinating and bold move.
The recent launch of Google’s Project Suncatcher test satellite is a story that feels like it was pulled from a sci-fi novel. The satellite itself is relatively small, drawing about as much power as a standard hairdryer. That’s not a lot of juice, especially when you consider its mission. Its entire purpose is to be a testbed. A guinea pig. Google wants to see what happens to its sophisticated silicon when it’s bombarded with cosmic radiation and subjected to the brutal temperature swings of low Earth orbit. It’s one thing to run a tensor processing unit in a climate-controlled data center. It’s another thing entirely to have it flying through a vacuum at 17,000 miles per hour. This is hardware testing at its most extreme.
Why put AI chips in orbit anyway?
It sounds crazy at first. Why go through the immense trouble and expense of launching AI hardware when we have massive, powerful data centers right here on Earth? The answer has two parts: one for the near term, and one that hints at a much bigger, more ambitious future.
The immediate goal is research. Space is an incredibly hostile environment for electronics. Radiation can flip bits in memory, corrupting data and causing processors to crash. This is a known problem for all satellites, but it’s particularly tricky for the complex, dense circuitry of modern AI accelerators. Google needs to understand how its hardware degrades and fails in that environment. They need to develop ways to make their chips more resilient. If you want to use AI for things like on-board satellite data processing—analyzing Earth observation imagery in real-time without having to beam terabytes of raw data back to the ground—you need chips that can be trusted to work. This Suncatcher project is their way of collecting that crucial data. It’s a field test.
But the long-term vision is where things get really interesting. There are filings from companies like SpaceX that reportedly sketch out a future with millions of satellites forming a sort of distributed data center in space. The big bottleneck for AI on Earth is energy. Training and running large models consumes a staggering amount of electricity, which also generates a massive amount of heat that requires even more energy to cool. In space, you have a couple of advantages. You have near-unlimited solar power. And the vacuum of space is a pretty effective heat sink. The idea of shifting a significant portion of our global compute infrastructure into orbit, to escape terrestrial energy and cooling constraints, is no longer just a fantasy. It’s a strategic goal. Google’s little experiment is a tiny first step on that path.
How does this space stuff connect to my laptop?
Okay, a satellite bristling with AI chips in orbit is cool. But what does it have to do with you or me? For me, the connection is about understanding the physical reality of artificial intelligence. We often talk about AI as if it’s a cloud, a disembodied brain somewhere on the internet. We use APIs, get our results, and don’t think much about the hardware doing the work.
Project Suncatcher is a stark reminder that AI is deeply, fundamentally physical. It’s silicon, electricity, and heat. It’s hardware with limits. And seeing Google test those limits in the most extreme environment imaginable got me curious about the limits of my own hardware. Not in orbit, but right here in my office. What can my MacBook actually do? What about the gaming PC I built last year? We hear all these model names thrown around—Llama, Mistral, Phi—but how do they actually run on a machine I own?
This curiosity sent me down a rabbit hole. I wanted a way to measure my own hardware’s performance. A personal benchmark. Not just for bragging rights, but to get a real, tangible feel for what these models demand and what my computer can supply. I wanted to move beyond the abstract idea of “local AI” and turn it into a concrete number: tokens per second. How fast can my machine think?
How can I test my own AI hardware performance?
Building your own LLM speed test sounds intimidating, but it’s surprisingly straightforward. You don’t need a computer science degree. You just need a bit of patience and a willingness to use the command line. The goal is to get a large language model running entirely on your own machine and measure how quickly it can generate text. The key metric here is tokens per second (t/s). A “token” is roughly a word or a piece of a word, so this metric is a direct measure of the model’s text generation speed.
The easiest way I’ve found to do this is with a tool called Ollama. It’s an open-source project that simplifies downloading, managing, and running a huge variety of open-source LLMs. It works on Mac, Windows, and Linux, and it handles all the complex dependencies for you. Think of it like a simple app store for local language models.
Getting started is simple. First, you install Ollama from their website. Then, you open your terminal (or Command Prompt on Windows) and pick a model to test. I’d recommend starting with a smaller one. A great choice is Microsoft’s phi-3-mini. It’s powerful for its size and runs well on a wide range of hardware, even without a monster GPU. To download it, you just type one command:
ollama pull phi-3-mini
The tool will download the model files, which might be a few gigabytes. Once it’s done, you’re ready to run your first test.
What’s a good setup for a local speed test?
With a model downloaded, you can now run it and measure its performance. The process is simple. You use the run command in Ollama and give it a prompt. For a consistent benchmark, you want to give it a task that requires a decent amount of text generation. A good prompt is something like, “Write a short story about a lighthouse keeper who discovers a mysterious object washed ashore.”
So, you’d type this into your terminal:
ollama run phi-3-mini "Write a short story about a lighthouse keeper who discovers a mysterious object washed ashore."
The model will start generating the story right there in your terminal. It might take a few seconds to get going (this is the “time to first token”). Then, the words will start streaming out. After the model finishes its response, Ollama does something beautiful. It prints a little summary of the run. It looks something like this:
total duration: 25.1s
prompt eval count: 23 token(s)
prompt eval duration: 1.1s
prompt eval rate: 20.91 t/s
eval count: 489 token(s)
eval duration: 24.0s
eval rate: 20.37 t/s
The number you care about is that last one: eval rate. This is your tokens-per-second score. In this example, the machine is generating about 20 tokens per second. That’s a very readable speed, like a fast typist.
Now you have a baseline. You can run the same test with different models. Try a bigger one, like llama3:8b. How does the score change? If you have a laptop, run it on battery versus plugged in. If you have a gaming PC with an NVIDIA or AMD graphics card, Ollama should use it automatically. You’ll see a massive jump in your eval rate. On my MacBook Air with an M2 chip, phi-3-mini hums along at around 30-40 t/s. On my desktop with an RTX 3070, that same model screams past 150 t/s. And a larger 8-billion-parameter model that is sluggish on the laptop becomes perfectly usable on the desktop. This is the kind of tangible, practical insight I was looking for. It’s no longer an abstract spec sheet; it’s a number that tells me exactly what to expect.
Is running local LLMs even worth it?
With all this talk of hardware and benchmarks, it’s a fair question to ask. Why bother, when I can just open a tab to ChatGPT or Claude and use a far more powerful model for free (or a small subscription fee)?
The answer comes down to a few key things: privacy, ownership, and offline capability. When you run a model locally, none of your data ever leaves your machine. Your prompts, your documents, your conversations—they all stay with you. For anyone working with sensitive information, or just generally creeped out by big tech’s data appetite, this is a huge win. You also own the process. There’s no API that can be shut down, no terms of service that can change, no waitlist to join. If you have the model on your hard drive, it’s yours to use however you want, whenever you want. Even on a plane with no Wi-Fi.
Of course, there are tradeoffs. The biggest models, like GPT-4, are still far too large to run on consumer hardware. The quality of local models is getting incredibly good, but for the most complex reasoning tasks, the frontier models in the cloud still have an edge. Speed can also be a factor if you don’t have a decent GPU.
But for me, that’s okay. It’s about having options. I use cloud models every day. But I also now use local models for summarizing meeting notes, rewriting emails, and brainstorming ideas, all with the confidence that my data is staying put. And thanks to my little benchmarking project—inspired by Google’s adventures in space—I have a much clearer picture of what works, how well it works, and what the physical reality of AI looks like on a machine I actually own. It makes the whole field feel a lot more grounded.
FAQ
What is Google’s Project Suncatcher? Project Suncatcher is a test satellite launched by Google that carries four of its AI chips. Its purpose is to study how the hardware performs and holds up in the harsh environment of space, specifically dealing with radiation and extreme temperatures.
When did the Project Suncatcher launch happen? According to the report, the satellite was launched on October 1st aboard a SpaceX Falcon 9 rocket.
What is an LLM speed test? An LLM speed test is a way to measure how quickly your own computer can generate text using a local large language model. The most common metric is tokens per second (t/s), which tells you how many words (or parts of words) the model can produce each second.
Do I need a powerful GPU to run a local LLM? Not always. Smaller, highly optimized models like Phi-3-mini can run quite well on modern CPUs or the integrated graphics found in laptops like Apple’s M-series MacBooks. However, for running larger models or for getting the best possible speed, a powerful, dedicated GPU from NVIDIA or AMD is essential.