Google Cloud is strengthening its arsenal in the battle for artificial intelligence with the launch of two new in-house chips, designed to meet two now very distinct needs of the market. On one side, the TPU 8t for training the models, on the other, the TPU 8i for inference, that is to say the concrete execution of the models once they have been deployed.
Google wants to offer more computing, for less energy and cost
The Mountain View firm highlights very ambitious gains: Google promises up to three times more speed for training, an 80% improvement in the performance-price ratio and the possibility of running more than a million TPUs within the same cluster. The American giant wants to convince cloud customers that its in-house silicon can offer more power while reducing energy and financial costs, what is commonly called a “strong” argument.
This announcement is part of a broader infrastructure control strategy. By designing its own accelerators, Google hopes to better control costs, differentiate its cloud service and attract companies looking to industrialize their AI uses/services.
Nvidia, however, remains an essential partner
However, it would be premature to see these in-house chips as a frontal break with Nvidia. Google does not suddenly replace the GPUs of the American champion: it complements them. The company even confirmed that it would offer Nvidia’s Vera Rubin systems in its cloud later this year. At the same time, the two groups continue to collaborate on networks and infrastructure optimization.
A competition which remains, for the moment, a cohabitation
This is the paradox of the moment. Hyperscalers like Google, Microsoft or Amazon are developing their own chips to reduce their dependence on Nvidia, but at the same time continue to buy and massively integrate its technologies. In the short term, the growth of Google Cloud in AI can therefore still benefit Nvidia as much as it plays into the competition.
