Business Is Our Ability to Predict the Future

The average person has a huge misconception about how businesses are evaluated.

Most people assume business is about current earnings, profit margins, or revenue. That’s how we’re taught to think: If you want to start a company, how are you going to make money?

On the surface that framing is correct, but it’s overlooking the real question: what is going to change in the future that will allow this company to be ahead of the curve? In that sense, business is actually the monetary expression of predicting the future.

With that framing, let’s take a look at the hotly debated topic of orbital data centers. If you evaluate them based on what we know today – that the cost of building a data center on the planet is cheaper than launching one 17,000 mph into orbit – then they don’t make any sense. But that is the wrong frame because it doesn’t consider any future predictions.

Put simply, orbital data centers are not a bet on today’s world. On paper, they are extremely hard to justify as an engineering system. They even look a little insane. But they start to make sense if the following predictions are true:

• The effective cost of terrestrial data centers will continue to rise as power, permitting, and regulation become more difficult
• Most AI compute is increasingly used for inference, rather than massive training clusters
• The price of launch continues to fall such that the cost of putting compute in orbit becomes a relatively small share of the overall cost

If any of these bets prove false, orbital data centers become harder to justify. If all three are true, then the business succeeds.

The problems orbital data centers are up against

Terrestrial AI infrastructure runs into three constraints: power, cooling, and permitting. Modern AI clusters consume a staggering amount of electricity, enough that they require dedicated power generation and water at an industrial scale. A single H100 consumes roughly 700 W at the chip level, and closer to 1,300 W once supporting systems are included. A cluster containing one million H100s would require roughly 700 MW of raw chip power. That’s on a scale comparable to the electricity consumption of more than one million homes.

Then there’s the other problem: people don’t want them. More than 100 U.S. localities have enacted moratoriums on new data center construction, and over 300 bills were filed at the state level within the first six weeks of 2026 alone. In March, Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced a federal AI Data Center Moratorium Act that would halt construction of any facility drawing 20 MW or more until certain safeguards exist. That specific bill probably won’t pass, but it’s a harbinger of what’s to come. This fight isn’t going to stop.

So some people are looking to space. Orbital data center proponents point out that solar panels are about 1.4 times more effective in space than on Earth, where sunlight doesn’t have to fight through the atmosphere. And space is very cold, so you might assume the cooling problem solves itself. Another big piece of evidence that this might work is the H100 operating on orbit today. The team that put it there deserves massive kudos.

The problem is that the physics simply don’t scale. Let’s start with heat, because in my view thermal management is the most underappreciated challenge. In vacuum there is no air to convect into, so heat radiates out. That means you need radiator panels, heat pipes, thermal transfer systems, and other infrastructure that is extremely hard to scale.

An H100 has an operating temperature between 5 °C and 30 °C. To hold that, a single chip requires three square meters of radiator area. This quickly becomes an enormous infrastructure problem.

Now let’s look at power. The sun provides 1,361 W/m² at Earth’s distance. Assuming 33% cell efficiency, that’s 449 W/m². Again, the scale becomes disastrous. Let’s use the ISS solar array structures as a baseline, as they are the largest arrays ever deployed on orbit.

If you wanted to put a Colossus training cluster on orbit, the other thing you need to consider is moving data between nodes. With an optical link, received power falls with the square of distance. That means, with satellites 440m apart, you’d need 1 THz of bandwidth around a 100 mW laser. That’s modeling it with zero margin, so realistically you would need 1–50 W of output power.

What might make sense instead is building small: putting inference on orbit instead of training. Here’s a single inference node.

So who is going to pay for this? In today’s market, the cost to send a kilogram of material to low Earth orbit is $5,000. An H100 weighs about 1.2kg per card, which would bring the cost at launch to $6,000. Ten H100s on the ground cost a few hundred thousand dollars. All together, on orbit, the costs are ten times that.

Companies raise money on predictions

Who in their right mind pays for this? AI companies. Training large frontier models will probably never make sense on orbit, but inference might. Inference now accounts for about half of AI profits and that number continues to grow. If a single terrestrial data center takes three to five years to build, not including delays, permitting fights, or power constraints, then this might very well be worth the cost. This is the same constraint that has pushed any heavily regulated industry with strong local opposition to friendlier places.

This debate reminds me of Falcon 1 and Falcon 9. When SpaceX began development in 2002, the dominant launch market looked very different. At that time, spacecraft were rare, and most were massive, hugely expensive, and optimized to travel to geosynchronous orbit. Launch vehicle development was also long and expensive.

SpaceX made a different prediction. Their bet was that low Earth orbit would become dramatically more important for the future. They bet that spacecraft would become smaller, and that reusability would matter.

Viewed through the lens of the market that existed in 2002, this was a hugely contrarian bet. If you were designing the perfect rocket from first principles, you probably would not design Falcon 9. But if your goal was to radically lower the cost of access to orbit, reuse hardware, and create a business around high-cadence access to space – well, Falcon 9 makes a lot of sense.

Orbital data centers strike me as a similar kind of bet because they’re betting on the world of tomorrow.

Despite the technical challenges, the category has attracted meaningful capital. Starcloud, which flew the first H100 in orbit in November 2025, raised a $170 million Series A at a $1.1 billion valuation. An even stronger signal comes from SpaceX, which has its own plans to launch and operate computing clusters on orbit.

Investors are not funding orbital data centers because the physics are a solved problem, but because the companies are making predictions they believe are directionally correct. Which, at the end of the day, is what building a company is all about.

— Matt