I suspect that EXP operation (opcode 0x0a) is either under-priced or inefficiently implemented in geth.
The evm program 6100005b60037ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff9800a509061000101806127101190915700
computes EXP(ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff9,ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff9)
for 10000
times.
It consumes around 3M gas (see testnet tx: https://testnet.etherscan.io/tx/0x5322ba8426c957a2abafec0a4473baf703d7bea221c17c0faa213889e22db034), so it could fit in a single block.
However, on my pc it takes 3.5 seconds to excecute (see experiment.go).
If my calculation are current, then a miner could benifit by continiously submitting such tx (costs around 1 USD per tx) and gain 3.5 seconds advantage over the other miners. Since the block interval is around 15 seconds and the block reward is around 50 USD, then a miner with 6% of the total network mining power could benifit from such an attack.
Some remarks: (1) On parity client it only takes 0.3 seconds
(2) I don't know much about go, so maybe I am missing some optimization flags
(3) EXP(fffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff,fffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff)
only takes a third of that time.