For years, Amazon Web Services (AWS) has built its power on proprietary hardware solutions, carefully protecting its data centre ecosystem from over-reliance on external providers. However, the scale of the race for dominance in generative artificial intelligence is forcing a review of even the most well-established strategies. The just-announced contract to supply one million Nvidia GPUs to AWS is not only an impressive commercial deal, but above all a signal of deep technological integration between the two giants.
Ian Buck, Nvidia’s vice president of hyperscale computing, clarified the timeframe for this agreement: the process of deploying the units will begin later this year and last until 2027. For the market, this sends a clear message – the demand for computing power is not a temporary bubble, but the foundation of long-term investment plans. The agreement fits in with the optimistic vision of Jensen Huang, who estimates the sales potential of new chip families such as Ruby and Blackwell at one trillion dollars.
However, the key to understanding this collaboration lies deeper than the sheer number of GPUs delivered. The most intriguing element of the contract is AWS’ agreement to implement Nvidia’s Spectrum-X and ConnectX networking solutions. The fact that Amazon is opening up its data centres to a third-party partner’s network architecture is testament to how critical data transfer performance has become to AI processes. Buck stresses that modern inference is a ‘monstrously difficult’ process, requiring not a single chip, but the synergy of the entire technology stack.
Groq chips, acquired by Nvidia as part of last year’s $17bn licensing deal, will also play an important role in this context. AWS plans to use them along with six other types of Nvidia processors to maximise the efficiency of response generation by AI systems.
