He's talking about the overhead due to carrying variable length IP packets
in fixed length ATM cells.
There's beginning to be some expectation that there will be a transmission
capacity crunch in the carrier's Sonet nets, and this ~25% ATM cell tax may
be looked at carefully as packet over Sonet solutions emerge.
Given the bimodailty of IP traffic size distribution (about 40% of packets are
small, like TCP ACKs or telnet/rlogin keystrokes) the ATM "cell tax" is
closer to 32%.
I.e. a dual clearline DS-3 actually carries as much user data as OC-3c ATM.
Which, incidentally, was why SprintLink backbone design is easily expandable
to dual links (that includes carefully considering implications for routing).
Sean presented that design on NANOG a year ago, BTW. Funny thing, the design
is expandable beyond that, too, so OC-3 ATM is already obsolete.