Genuinely divided, recorded as a split: whether to buy inference hardware at all is settled only by usage pattern, and price-per-token arguments are explicitly rejected by the local-hardware camp

A thread asking whether to buy 4x Ascend GX10 (39 upvotes, 169 comments) surfaces the whole buy-versus-rent dispute cleanly. Rent side, 46 upvotes: 'This is such a bad financial decision it's almost funny. There's some real Tulip Mania nonsense going on in this community... you are thinking about spending $30k on a computer that runs slower than even the slowest cloud provider, to access to a crippled version of an open weight model... For $30k, you could use essentially unlimited, high-performance SOTA usage on the cloud for years.' Local side, 16 upvotes: 'I'm going to go against the grain here and say do it. Live the dream. If you can afford it, why not. This is localllama, I can't in good conscience say use cloud providers. The nvidia monopoly, industry price fixing/ gouging and high cost of energy has spoiled my passion for doing anything beyond a 2x3090 lab.' The most useful reply in the thread is the one that refuses the price framing, 20 upvotes: 'Remember that you're in localllama. API will ALWAYS be better in speed and prices per token for a single user. So price and performance are not why users here are going local.'

Two non-price reasons keep the local camp local, and both are named in the corpus. (1) Predictability and censorship control: 'You're also at the whims of whatever company/service wants to restrict what you can use the models for... you also have the fact that most (but not all) cloud providers are effectively stealing all of your work.' (2) Variance: a provider may be 'quietly running quantized versions of their models to help save on costs', so the same model gives different speed and quality between sessions. Against these, buying is a fixed cost with a fixed capability ceiling and a model that will be superseded; one commenter does the arithmetic in the buyer's favour - 4x GX10 at $16k breaks even against a $25/hour RunPod 4xB200 config in about four months at 8 hours/day. The disagreement is therefore about which risk dominates: model obsolescence versus provider dependence.

For a single user or an experimental workload, renting wins on every economic measure and the corpus says so. Buy when either the data cannot leave your control or a provider's serving is unreliable enough to be a production risk - not when throughput is the concern, because rented parts are faster.

A useful sub-split on which rented hardware the local camp accepts: the same commenter argues Sparks are dev boxes for people developing for larger clusters and 'IMHO are a poor choice for local inference' (22 upvotes), which sits against his other remark that the cloud serves a 'crippled version' at 10x the token price. Both can be true - the cloud is faster and cheaper, the local box is yours - and that ambiguity is the honest state of this debate.