Data center compute doubles every seven months as Astra and Muse raise demand
The author says compute in a single AI data center has doubled roughly every seven months, and cites the multi-step agent calls of GPT-6 Astra and Meta Muse as examples of how new models push compute demand higher.
The compute boom is about to accelerate faster than ever before and the newest AI models are already proving why. Since the launch of Colossus 1 in August 2024, the record for computing capacity inside a single AI data center has doubled roughly every seven months. The leading facility went from around 100,000 H100 equivalents in 2024 to more than one million by 2026 and if the trend continues, frontier data centers could contain several million H100 equivalents by 2028. The biggest reason is that every major AI company is racing to release more powerful models and products. OpenAI recently released GPT-6 Astra, while Meta also released Muse, a personal AI agent that can book appointments, manage projects, and continue working after the user leaves the app. Products like Astra and Muse change the compute equation because they do much more that means one request may trigger many separate model actions before the final task is completed. The demand becomes even larger when these products reach mass distribution. OpenAI is rolling Astra out across ChatGPT, its API, AWS, and enterprise plans, while Meta can distribute Muse through mobile apps, the web, WhatsApp, and eventually its AI glasses. Every new model release also pushes competitors to respond with larger training clusters and better products of their own. OpenAI releases Astra, Meta answers with Muse, Anthropic upgrades Claude, and Google improves Gemini and each launch creates another cycle of training, testing, deployment, and customer adoption. More efficient chips and models will reduce the compute required for individual tasks but better products can attract more users and unlock entirely new use cases. If you want to see how I’m positioned around the broader compute buildout, including the neocloud, memory, networking, and power names I think benefit most as these AI clusters keep scaling, check out the link below. https://link.milkroad.com/tx5qip