Bitcoin Core does not keep the entire UTXO set in memory. Starting from an empty UTXO cache, we only add UTXOs that are needed to validate transactions and new UTXOs created by blocks we process. (The mempool maintain its own implicit UTXO set across the unconfirmed transactions it holds.) Whenever the UTXO cache reaches its limit, all changes are persisted to disk and the UTXO cache is flushed. IIRC, it is also flushed every 24h at the latest. If a newly created UTXO is spent (by a confirmed transaction) before we flush, that UTXO is deleted from the cache immediately and never persisted to disk.

A node would be able to synchronize the blockchain faster and validate blocks with previously unseen transactions more quickly if it could keep the entire UTXO set in memory, but even if you have a set a huge dbcache that can only happen once: when you first start your node and synchronize until your node is restarted (or has run for 24h). Unless your node must absolutely minimize transaction validation and block processing time, it is also completely unnecessary. Once you are caught up with the chaintip, your node learns about unconfirmed transactions throughout. We validate transactions before adding them to the mempool, at which point we have already retrieved all necessary UTXOs from disk, and also cache their script validation. When a block comes in, we only need to retrieve any UTXOs for transactions that the node does not have in its mempool.

It would also take north of 10 years to create that many UTXOs, even if we designated all blockspace to that purpose. So whatever Raspberry Pi successor people run at that point will hopefully be a bit more beefy than today.