I guess it does increase their cost, or rather your share of their hardware depreciation.
AI serving cost is apparently mostly hardware depreciation rather than operating cost (electricity etc), and if your large context request is occupying VRAM for some fraction of a second then you are paying for the depreciation that occurs in that time!
e. H100 costs $20-40K to buy, with a lifetime of maybe 3 years, and will only consume maybe $2K in electricity if run 24x7 for those 3 years.
AI serving cost is apparently mostly hardware depreciation rather than operating cost (electricity etc), and if your large context request is occupying VRAM for some fraction of a second then you are paying for the depreciation that occurs in that time!
e. H100 costs $20-40K to buy, with a lifetime of maybe 3 years, and will only consume maybe $2K in electricity if run 24x7 for those 3 years.