Requiring more confirmations decreases the probability that a transaction will be reverted, under the assumption that an attacker has < 49% of the hashpower. If you attempt an attack with 49%, then you have a fair chance of mining, say, 6 blocks before the rest of the network. If you get unlucky then you sacrifice those rewards. But if you mine with 51% then your attack chain is (probabilistically) guaranteed to eventually become the longest chain, so there won't be any loss of revenue.

That ignores community consensus; such activity is easy to monitor and public sentiment about the illegitimacy of that fork can cause people to devalue it. This is essentially what happened with the DAO hack. This likely does create a situation where both forks have a non-zero value, but also dramatically lowers the rewards, and thus incentive, for such an attack.

Security in crypto is a very slippery concept, and many conclusions are non-obvious, if not outright counter-intuitive.

There's a question that you're missing or ignoring about how long the 51% attack is sustained. Sustaining a 51% attack for one day is very difficult, doing it for one week or one month is proportionally that much more expensive.

Why is mining an attack chain expensive? If you mine on the honest chain, you earn the block reward for each block you mine. If you mine an attack chain with 49% and your attack fails, then you sacrifice those rewards. Suppose you mine say 10+ blocks with 51%+, over the next three hours while the rest of the network only finds 8 blocks. Then your 10 blocks become the longest chain and you earn the block reward for all 10 of them.

you can split the payment to multiple small amounts.

Confused... I thought the claim was about receiving and not sending?

if multiple accounts send multiple small payments (controlled by the same guy), then the exchange wont wait as long for the confirmations.

The exchange can have thresholds for total volume per given time interval then. Once the total amount of pending transactions breaches a given volume, the transaction period for more transactions within the window goes up.