Can Google Outsmart Microsoft's $100 Billion Bet in AI Race?
Microsoft and OpenAI flex with a mega-supercomputer. Google's plan involves everyone. Which strategy wins the next AI war?
Hello folks, so you probably already know about the ambitious strides technology giants are taking to push the boundaries of artificial intelligence and computing power. The world's media is buzzing with discussions about Microsoft and OpenAI's bold plan to build a $100 billion supercomputer. This project, set to launch in 2028, represents the fifth stage in Microsoft's ongoing effort to build increasingly powerful supercomputers for OpenAI, with the third stage nearing completion and the fourth set for launch in 2026. But how will Google respond to this powerful move in the AI race? Today, we'll explore that very question.
Microsoft and OpenAI $100 Billion Supercomputer
It will consume up to 5 gigawatts of power, so it's likely to be powered by nuclear fusion energy – Microsoft signed a contract a year ago with Helion, a company in which Sam Altman has a significant stake, for the supply of electricity in large volumes, right by 2028.
Most of the money will go towards chips, and since Altman is likely organizing the energy supply, there might be a similar story with the chips.
According to Altman, creating superintelligence will most likely require a significant breakthrough in energy. The project is technically very risky – it's unclear whether it will be feasible to power, connect, and cool such a number of chips, especially considering that, according to rumors, the construction will take place in the desert.
Perhaps they're planning to bury everything underground? The project's implementation is still in question and depends on the results of GPT-5. What are they even planning to train there, after all?
Google's Guide to Making Microsoft-OpenAI's Plans Obsolete
An asymmetric response from Google DeepMind to the ambitious plan of the Microsoft-OpenAI tandem.
The response from Google DeepMind is completely asymmetric: to devalue the $100 billion investment of competitors by creating a globally distributed system for training super-intelligent AI models (kind of a “peer-to-peer torrent" for training models). Google DeepMind plans to do this based on Distributed Path Composition (DiPaCo) – a method for scaling the size of neural networks in geographically distributed computing entities.
Google's Decentralisation Gambit
The long-term goal of the DiPaCo project is to train neural networks around the world, using all available computing resources. This requires a reconsideration of existing architectures to limit overheads on communication, memory constraints, and inference speed. Instead of a centralised behemoth of a supercomputer, Google envisions a vast, planet-spanning network of smaller computers collaborating to train the next big AI leap.
DiPaCo builds upon existing work like DiLoCo, but its purpose is to parallelise the model training process itself.
The Competitive Landscape Shifts
An intriguing situation is unfolding.
Competition between Google DeepMind and the Microsoft-OpenAI tandem forces the former to break the monopoly of the "AI giants" on creating super-intelligent models. Democratizing access to large-scale AI training could empower smaller players and researchers, leading to unexpected breakthroughs;
But at the same time, it will lead to the collapse of all plans by governments (USA, EU, China) to control AI development by controlling the largest model training centers with a huge computational power. AI development could become far more distributed, making it tougher to regulate.
Is this a calculated risk by Google, or a bold move that even they can't fully predict? There are significant implications:
AI Access: If training super-intelligent models becomes less reliant on expensive supercomputers, will we see an explosion of innovation from a wider range of actors?
Control: If AI development moves from a few centralized locations to a worldwide mesh of computers, will governments and corporations struggle to maintain oversight and ethical guidelines
🔎 Explore more: