Google can physically move a pod (server holding 1000s) of TPUs from one datacenter to another. So if the TPUs become outdated for a hyperscale datacenter, eg one built for 1-3 GW of power, they could move it to a smaller leased datacenter (see Digital Realty or Equinix) to run less intensive inference for consumer queries.

@Mungofitch - I think the issue about new hardware architectures is valid. Certainly Cerebras has generated a lot of buzz for its chips. With regard to new AI architecture, the transformer architecture is currently the cutting edge. There are others but currently the other AI architectures still need lots of GPU/TPU computation.

There's a new FT piece on the SPVs Google is using to fund order of TPUs from Anthropic. It's pretty illuminating on who is holding the residual risk on chips, datacenter leases, and power agreements.

Financial Times: Content 549f2e23 5aa2 49c7 9ea6 a9784ab7087c