For most of my working life capacity planning has meant four things: processor, memory, disk and bandwidth. All four got cheaper every year with reassuring regularity, and the physical layer beneath them was somebody else's problem and effectively unlimited. I no longer think the second half of that sentence is safe to plan on.
The evidence is not dramatic and it is not about artificial intelligence, although that is accelerating it. It is that our colocation provider now quotes us per kilowatt rather than per rack unit. That when we asked for two more racks last year the answer was that space was available and power was not. That the dense hardware a vendor wanted to sell us needed cooling our room cannot deliver, which turned a procurement question into a building question. That new regions in the places we would prefer them are appearing more slowly than demand for them, and that grid connection lead times in several countries are now measured in years rather than months.
What I expect this to mean for ordinary corporate IT is fairly mundane, which is why I think it is worth planning for. The steady annual decline in the cost of a unit of compute will become uneven, holding in some regions and reversing in others. Region selection will become a question about available power and carbon reporting as well as latency, cost and law. Efficiency work will get a second justification and become easier to fund, after twenty years in which the wasteful job that nobody wanted to rewrite was genuinely cheaper to leave running. And energy reporting obligations will land on our desks, because we hold the data and nobody else can produce it.
So I have started measuring consumption where the instrumentation exists, keeping an honest record of the physical envelope of what we still run ourselves, and asking suppliers about power in conversations that used to be only about price and performance.
For most of my career the physical layer was invisible and abundant. I do not expect either of those words to survive the next ten years, and I would rather be early than surprised.
– Serguey Shinder
Top comments (0)