First— do you mean the 500k soft target or the million byte protocol rule?
The soft target is trivially lifted a node at a time. I expect the default soft limit will change once the network is consistently producing blocks up against that limit.
So I'll assume you mean the protocol rule—
OTOH, if this infrastructure isn't available when the block size limit is bumped up against and transactions start getting delayed and expensive, I doubt developers will be able to resist demands to increase the limit.
I haven't done the benchmarking to fully figure out exactly where a standard PC peters off, but I'm pretty sure they can process somewhat more than the current limit, at least if they're SSD equipped. So even if you're a full card-carrying member of my Church of Forever Decentralization, whos doctrines requires that the maximum block sizes stay quite small, you could still support a bit of a bump.
pushing fees up and txs to occur off the blockchain on, e.g. Open Transactions servers
It's worth mentioning that beyond escaping the limits external things can have other advantages too. For example, even getting a _single_ confirmation in Bitcoin (the minimum required to resist reversal attacks without using a trusted certification serice) can take a long time— 10 minutes is an _average_, but 30 minutes or longer happens about 7 times per day, an hour or longer every 2.8 days, etc. And even though Bitcoin with the block size limits removed could be coerced to insane scaling levels, it would be a fairly storage and computation inefficient way to process all the world's transactions. D'aniel also points out the considerable privacy/anonymity advantages other systems can have over Bitcoin (and add to Bitcoin when used along with it)
Or will it be raised somewhat after some scalability optimizations are implemented?
The limit can't be raised at all without a hardforking change (old nodes will not accept the new chain at all once the first oversized block is mined).
It's not sufficient to change miners, as DeathAndTaxes suggests— the lifting the 1M protocol rule is a change unlike the BIP16/P2SH change which was fully compatible with old nodes. It's technically the same kind of change needed to adjust Bitcoin from 21m total BTC to 42m total BTC (though obviously not politically equal). Every single piece of Bitcoin software produced would have to be updated to allow the oversized blocks
If the Bitcoin system were to take a hardforking change, switching to Ed25519 would remove ECC signature validation as a performance bottleneck, as a fast quadcore desktop from today can do about 50k Ed25519 validates per second, compared to perhaps a thousand for the curve we use... though the random IO is still an issue.
More recently a number of people have independently invented the idea of committing to a merkle tree of open transactions. If we do adopt some form of this it would allow the creation of nodes which are someplace in between SPV and a pruned full node in terms of security and decentralization benefit— so lower operating costs for nodes that validate. (In particular these nodes would have greatly reduced storage requirements)
No. IIRC Mike Hearn supports moving most nodes to SPV. My impression was that Satoshi also expected most nodes to use SPV. Not sure about the opinions of other developers besides gmaxwell.
Indeed, and Mike's position has gotten us (rightfully) flamed as not-decenteralized by e.g. Dan Kaminsky.
Gavin and Jeff have taken less strong positions than I have on the importance (and viability) of maintaining decenteralization in Bitcoin. Although I expect to convince them eventually, I think _everyone_ is in a wait and see mode. Who knows what will happen? At the moment I would aggressively argue against raising the limit— without it I don't see any alternative to Bitcoin becoming a particularly inefficient distributed system of establishment central banks— but I fully admit my position may change as things develop.
I expect most Bitcoin users by count to be not even SPV— I expect most by count to be semi-SPV thin-clients (which may connect to a couple independent services). But expecting most users to be on not nodes does not preclude there being hundreds of thousands of nodes which perform complete validation, but gigabyte blocks surely would.