Articles by Nick Evanson

Technology

If Microsoft can't source enough electricity to power all the AI GPUs it has, you have to wonder how Amazon is going to cope in its new $38 billion deal with OpenAI

To keep the AI juggernaut rolling ever forward, you might think that the biggest companies in artificial intelligence desperately need ridiculous numbers of AI GPUs. According to Microsoft, though, the issue isn't a compute or hardware limit, it's that there's not enough electrical power to run it all. And if that's the case, it raises the question as to how Amazon is going to cope now that it's signed a $38 billion, multi-year deal with OpenAI to give the company access to its servers. It was Microsoft's CEO, Satya Nadella, who made the point about insufficient power in a 70-minute interview (via Tom's Hardware) with YouTube channel Bg2 Pod, alongside OpenAI's boss, Sam Altman. It came about when the interview's host, Brad Gerstner, mentioned that Nvidia's Jen-Hsun Huang had said that there was almost zero chance of there being a 'compute glut' within the next couple of years (and by that, he means an excess of AI GPUs). "The biggest issue we are now having is not a compute glut, but it’s power," replied Nadella. "It’s sort of the ability to get the builds done fast enough close to power. So, if you can’t do that, you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today. It’s not a supply issue of chips; it’s actually the fact that I don’t have warm shells to plug into." The warm shells that he refers to aren't tasty tacos; they're hubs already connected with the relevant power and other facilities required, ready to be loaded up with masses of AI servers from Nvidia. What Nadella is effectively saying is that getting hold of GPUs isn't the problem, it's accessing cheap enough electricity to run them all. That's not necessarily just a case of there just not being sufficient power stations—it matters a great deal where they are, because while one can just dump a building down almost anywhere, if the local grid can't cope with the enormous power demands of tens of thousands of Nvidia Hopper and Blackwell GPUs, then there's no point in building the shell in the first place. When I heard Microsoft's CEO make that comment, I immediately recalled another news story I'd read today: OpenAI's $38 billion deal with Amazon that will give it access to AWS's AI servers. OpenAI already has such an agreement with Microsoft to use its Azure services, with the latter pumping money into the former as well. "OpenAI will immediately start utilizing AWS compute as part of this partnership, with all capacity targeted to be deployed before the end of 2026, and the ability to expand further into 2027 and beyond," the announcement says. "The infrastructure deployment that AWS is building for OpenAI features a sophisticated architectural design optimized for maximum AI processing efficiency and performance. Clustering the NVIDIA GPUs—both GB200s and GB300s—via Amazon EC2 UltraServers on the same network." Does this mean AWS is effectively handing over all its Nvidia AI GPUs to OpenAI over the next few years, or will it expand its AI network by building new servers? It surely can't be the former, as AWS offers its own AI services, so the most logical explanation is that Amazon is going to buy more GPUs and build fresh data centers. But if Microsoft can't power all the AI GPUs it has now, how on Earth is Amazon going to do it? Research by the Lawrence Berkeley National Laboratory suggests that within three years, the total electrical energy consumption by data centers will reach as high as 580 trillion Wh. According to the Massachusetts Institute of Technology, that would put AI consuming as much electricity as 22% of all US households per year. To meet this enormous demand, some companies are looking at restarting old nuclear power stations, or in the case of Google, building a raft of new ones. While the former is hoped to be ready by 2028, it will only produce 7.3 TWh per annum at best—a mere 1.3% of the total projected power demand. This is clearly not going to be a valid long-term solution to AI's energy problem. With so much capital now invested in AI, it's not going to disappear without a trace (though that would solve the question of powering it all), so the only way forward for Amazon, Google, Meta, and Microsoft is to either build vastly more power stations, design and build specialised low-energy AI ASICs (in the same way that bitcoin mining has gone this way), or more likely, both. This generates a new question: Where is the money required for all of this going to come from? Answers on a postcard, please.

Technology

The 25th anniversary of the ISS operationally in orbit reminded me the ESA once challenged an Assassin's Creed and Far Cry dev to get DOOM running on a satellite

When the original Doom launched over 30 years ago, you needed a pretty beefy PC to run it well. Fast forward three decades, and Id Software's classic can be powered by a TI-84 calculator (with 100 lbs of potatoes), a USB charging station, and even an Apple Lightning dongle. Beating them all for sheer out-of-this-world coolness, though, is Doom on a satellite, orbiting the Earth. Well, it was. The European Space Agency (ESA) deorbited its OPS-SAT device back in May 2024, but before it was turned into dust motes, Georges Labrèche (engineering manager at ESA) contacted Ólafur Waage, a Norwegian software engineer, to see if he fancied a little challenge. Waage is no random coder, though, as he's a highly experienced game developer, working for Massive Entertainment—makers of Far Cry 6, Assassin's Creed: Valhalla, and countless others. The challenge was simple on paper: port Doom to run on OPS-SAT. This satellite was an experiment by the ESA, which could be used by members of the public directly over the Internet. Essentially, if your access was approved, you could upload code to the little satellite (just 10 x 10 x 30 cm in size) and have it do stuff for science. Of course, there's nothing scientific about having Doom run on a satellite in space, but as a final hurrah for the highly successful project, you'd be hard-pressed to come up with something better. Waage freely admits to not being an expert on Doom ports, but it was a presentation he did on Doom on GitHub Actions that caught Labrèche's eye. Between the two of them, they were hard-pressed to get the whole thing to work in the first place, but not because the cubesat wasn't up to the task. The main platform of OPS-SAT was powered by a Critical Link MitySOM-5CSx, where SOM means 'system on module.' Within this module, there's a dual-core Arm Cortex A9 processor, an Altera Cyclone V FPGA, 1 GB of DDR3 RAM, and 8 GB of storage. That Arm CPU only ran at 800 MHz, and with no kind of GPU present, everything had to be software rendered, so at least it was an authentic, OG port. As basic as all that sounds, it's more than enough to run a port of Doom. However, they had an inspired idea: Use the camera on OPS-SAT to take an image of Earth and use that as the backdrop in Doom, instead of a block Martian landscape. The problem is that the image generated is high in resolution and uses far more colours than the 256 of Doom's built-in palette offers when running in software rendering mode. Cue lots of clever coding and computer science (can you tell I don't fully understand it?), and hey presto! Doom running on a satellite, while in space, and applying some neat shots of our planet in the background. And it wasn't one of those Doom-on-a-carrot projects that just use the carrot as a display: OPS-SAT was running a genuinely playable game. As it turns out, the whole process also accelerated the satellite's demise, as to get the required views of Earth, OPS-SAT needed to be oriented in such a way that its solar panels generated a bit too much drag in the thin wisps of Earth's atmosphere, even at its 500 km altitude. What a way to go out, right? OPS-SAT might be long gone, but it will always hold the record of being the first ever satellite to run Doom in space. The only problem now is how to top this. When are we going to see Doom running at the bottom of Challenger Deep, or Doom running in the volcanoes of Io? I guess if there's a processor at those places, anything is possible.