In some part of a national laboratory, there is a time when a prediction comes true. Not on the screen of a meteorologist or in a satellite picture, but deep in the math: billions of calculations moving time forward one box at a time, variable by variable, simulating a world that hasn’t happened yet. It’s weird to think about. There is a supercomputer somewhere right now that is planning the storm for tomorrow.
For decades, machines that most people will never see and barely know exist have been used to quietly predict the weather. The ones we have now run at petaflop scale, which is one quadrillion operations per second. That’s a number that’s so big it doesn’t really mean anything. But even this is just a step in the long history of computers. Zetascale systems are the next big thing. They are a billion times more powerful than current systems and can do calculations that would take a million hours to do now in just one afternoon.
It’s hard to say enough about what that means for predicting the weather. The Earth is now modeled as a three-dimensional grid, with hundreds of variables (like temperature, humidity, sea ice thickness, and soil moisture) stored in each cell. The math that controls how these variables interact moves forward in very small steps of time over and over, simulating how the atmosphere behaves for weeks or months. It seems to work pretty well. Though the detail isn’t very good, the chaos is real, and after a certain point, the predictions become less accurate very quickly. That calculus would be very different with zettascale machines.
In 2018, Chinese scientists said they thought the first zettascale system would be up and running around 2035. If that timeline holds true—and it’s important to note that it took about fourteen years to go from petaflop to exaflop machines—then the 2030s could bring something truly new: accurate weather predictions for two whole weeks in the future. Do not guess. Not windows for chances. Real predictions based on physics that are accurate enough to use.
The effects are felt right away in disaster response, agriculture, emergency management, and city infrastructure. A good forecast for two weeks is more than just helpful. It’s the difference between a planned evacuation and a mad rush. The crop calendar always works. The power grids are getting ready before the heat wave hits, not after it does. A lot of people who work on climate models feel like they’ve been using good, useful tools that aren’t fully finished for a long time.
It’s not just the speed that interests people in zettascale computing. Scientists think that the architecture of these systems will be very different. They will be decentralized and focused on data, and they will be made to handle huge amounts of data so that moving the data costs more than moving the processors to it. Millions of weaker parts that are connected and work together as a whole. It sounds almost biological, which could be right for machines that were made to look like a living planet.
That path isn’t smooth. Moore’s Law says that processing power almost doubles every two years. This law is getting close to its physical limits. At zettascale, huge amounts of energy will be used; one system could use up to 100 megawatts. There are real problems here that don’t have clean solutions yet, and it’s still not clear if the engineering will get there before the climate deadlines get tough.
But the way is clear. The scientific infrastructure for this work is already fully developed. For example, the Community Earth System Model is already being run on petaflop hardware in labs, and dozens of teams can compare and confirm their results using international coordination frameworks. Next, more power will be used to answer the same important question: how can you simulate a planet well enough to keep the people who live on it safe? More and more, the answer is being found on computers that are room-sized and faster than almost anything the human mind can think of.
