Vitalik is straight up being dishonest here, because he wants to shill for his PoS. Attacking PoW network costs hell a lot of money, and the argument that it gets cheaper is just rubbish, because if difficulty drops, then honest miners can mine again. It's actually more fitting to say that in PoS system the attacker can attack the network as much as they want, because it costs nothing to sustain the attack.

I seriously doubt there will ever be a power in this world that will be truly capable and willing to 51% attack Bitcoin.

I read a little about this in the past. The first thing that would be needed for a successful 51% attack is millions of ASIC miners. Now we arrive at the problem where the attacker would get these devices from. Companies like Bitmain can't produce enough to satisfy the demand. Everything is pre-ordered and I don't see how someone could get their hands on a few million miners.

Let's say the attacker succeeded in acquiring them. He now needs enough electricity to power them all and cooling equipment to prevent his factory from going up in flames. That's a constant and huge cost. Sooner or later, he will run out of funds to finance the attack, and the network will start recovering. Even with a successful attack, he can't go through the blockchain and randomly steal people's coins. He can double-spend his own transactions depending on what the depth of them are. I am not an expert in 51% attacks, but this is what I remembered from many other threads in the past including one started by myself.    

In a proof of work system, if your chain gets 51% attacked, what do you even do? So far, the only response in practice has been "wait it out until the attacker gets bored". But this misses the possibility of a much more dangerous kind of attack called a spawn camping attack, where the attacker attacks the chain over and over again with the explicit goal of rendering it useless.

In a GPU-based system, there is no defense, and a persistent attacker may quite easily render a chain permanently useless (or more realistically, switches to proof of stake or proof of authority). In fact, after the first few days, the attacker's costs may become very low, as honest miners will drop out since they have no way to get rewards while the attack is going on.

In an ASIC-based system, the community can respond to the first attack, but continuing the attack from there once again becomes trivial. The community would meet the first attack by hard-forking to change the PoW algorithm, thereby "bricking" all ASICs (the attacker's and honest miners'!). But if the attacker is willing to suffer that initial expense, after that point the situation reverts to the GPU case (as there is not enough time to build and distribute ASICs for the new algorithm), and so from there the attacker can cheaply continue the spawn camp inevitably.

In a proof of work system, if your chain gets 51% attacked, what do you even do? So far, the only response in practice has been "wait it out until the attacker gets bored". But this misses the possibility of a much more dangerous kind of attack called a spawn camping attack, where the attacker attacks the chain over and over again with the explicit goal of rendering it useless.

In a GPU-based system, there is no defense, and a persistent attacker may quite easily render a chain permanently useless (or more realistically, switches to proof of stake or proof of authority). In fact, after the first few days, the attacker's costs may become very low, as honest miners will drop out since they have no way to get rewards while the attack is going on.

In an ASIC-based system, the community can respond to the first attack, but continuing the attack from there once again becomes trivial. The community would meet the first attack by hard-forking to change the PoW algorithm, thereby "bricking" all ASICs (the attacker's and honest miners'!). But if the attacker is willing to suffer that initial expense, after that point the situation reverts to the GPU case (as there is not enough time to build and distribute ASICs for the new algorithm), and so from there the attacker can cheaply continue the spawn camp inevitably.