We usually think only in terms of GPU capacity, at least for the currently mainstream AI architectures. We might need to start adjusting our thinking somewhat.
There needs to be some way of predicting the effect on capabilities, otherwise it’s not a useful thing to track. For total compute and the capabilities of scale-up systems, there’s a story for how that translates to model capabilities. This story is useful for predicting how model capabilities change over time with the somewhat predictable changes in chip technology, AI company funding, and chip production capacity.
A very large deal for CPU-only compute between Anthropic and Akamai:
https://www.akamai.com/newsroom/press-release/akamai-announces-11-6-billion-multi-year-agreement-with-anthropic-to-support-growing-demand
We usually think only in terms of GPU capacity, at least for the currently mainstream AI architectures. We might need to start adjusting our thinking somewhat.
There needs to be some way of predicting the effect on capabilities, otherwise it’s not a useful thing to track. For total compute and the capabilities of scale-up systems, there’s a story for how that translates to model capabilities. This story is useful for predicting how model capabilities change over time with the somewhat predictable changes in chip technology, AI company funding, and chip production capacity.