this...sounds exactly like proof of work?
so, proof of work?
The global CPU and power use for the VDF calculations doesn't need to exceed what you could fit into a single server rack.
To demystify what a VDF is, consider the delay function (i.e., the majority of the work done to compute a VDF) used by the most prominent proposals:
Let N = p*q be a product of two large primes (so an RSA modulus) and assume that the primes p and q have been immediately thrown away/forgotten after initialization. Then, computing
f(x) = x^(2^T) mod N
is believed (dating back to a paper by Rivest, Shamir and Wagner in 1996) to take T sequential steps provided that T is large enough. For a large T, the only feasible approach seems to be repeated squaring modulo N. That is, compute y = x^2 mod N, y' = y^2 mod N, ... for T times.
Now if a proof of stake includes a VDF that needs to be computed for every block, then a long-range attack needs to recompute the VDF outputs as well. This is infeasible as it will take a long time given the correct choice of VDF parameters.
Notably, the Chia blockchain mentioned in the article would succumb to long-range attacks as well were it not for their usage of VDFs [2, p. 17].
[1] https://eprint.iacr.org/2018/601.pdf [2] https://www.chia.net/assets/ChiaGreenPaper.pdf