Apologies for the slight delay—I included some BitVM alpha which hopefully makes up for it :)

Do you think we will ever see a major non-USD based stablecoin?

As a good enough first approximation I expect all assets that can be digited to eventually make it onchain. In particular, any non-USD fiat currency you may consider major (e.g. CNY or EUR) will eventually be a stablecoin, either via a directly-issued CBDC or a third-party wrapper.

What's the best possible collateral for it in principle? USD certainly isn't

The best possible collateral for a CNY stablecoin would be central bank money directly issued by the central bank of China. In this CBDC model there is arguably no collateral: the stablecoin is central bank money itself, 1-to-1.

For Liquity- and Maker-style overcollateralised decentralised stablecoins my favourite form of collateral is maximally trustless money like ETH or BTC. Unfortunately neither ETH nor BTC have accrued enough monetary premium to provide enough economic bandwidth for tens of trillions of dollars of stablecoins.

For the sake of argument, let's assume that both ETH and BTC will eventually achieve $100T marketcaps—what would be required to build a successful over-collateralised stablecoin? The three key ingredients are:

1. Full programmability: unfortunately BTC doesn't really fit the bill. For a couple weeks I was excited about BitVM to provide two-way-pegged fully-programmable rollups on Bitcoin. Unfortunately there's an awkward list of caveats associated with BitVM:
   a) 1-of-N whitelisted challengers (no usual 1-of-∞ assumption)
   b) optimistic rollups only (no zk-rollups)
   c) log-round fraud proving (no one-round fraud proving)
   d) drawn-out fixed-time challenge rounds (no chess clock time keeping)
   e) only 4MB/10min of L1 DA

It is possible to do zk-rollups on Bitcoin with fancy cryptography like functional signatures and indistinguishability obfusction but it will likely take decades for the cryptography to become practical.
2) Better oracles: my favourite design here is restaking-based oracles that take the median across consensus participants, in some sense reusing the honest majority assumption of the L1. Zahary Karadjov is working on an oracle platform called Blocksense which has a bunch of interesting ideas to use MACI and other fancy tools to mitigate collusion 51% attacks.
3) Cheaper cost of money: ETH today has a high cost of money (~5%/year) which makes it almost untenable for a scalable decentralised stablecoin. One could use an LST like stETH as collateral instead but you'd lose the high-grade pristineness of ETH, defeating the point of a decentralised stablecoin. My favourite solution to the opportunity of cost of ETH is stake capping. The idea is to have an issuance curve which tends to zero (or, even better, negative infinity) as the total stake approaches the cap (say, 25% of all ETH).