Interruptible GPU capacity bills at 18 percent of on-demand on one operator's own account

Before anyone budgets against 18 percent of on-demand, look at what the evidence is: one account, one instance family, and a survey taken across three regions on a single afternoon ^[1][5]. For the number to transfer, your family has to be one of the GPU families where the discount runs near 80 percent; the author reports 55 to 75 percent for CPU workhorse families ^[4][23]. And you have to be in the right zone, because on-demand pricing was identical in every region he checked while interruptible pricing was not: a 27 percent spread between availability zones in one region, 1.2 percent in another ^[6].

The cheapest zone was also the one AWS itself rated lowest for capacity, and it dropped him; four cents an hour of savings bought an outage ^[8]. "The price you do not pay is not free. It is quoted somewhere else, in interruptions," he wrote ^[9]. Four cents is small against the spread it sits inside: 27 percent of the 62.5-cent midpoint of his observed band is about 17 cents an hour, four times the saving he chased ^[10].

The two-minute reclaim notice is the part most teams already design for, and he calls it the least interesting of the three problems, because it is the one you can engineer around ^[7]. The binding one is quota. His account is allowed 64 vCPUs of interruptible GPU capacity, and it was 64 in every region he checked, so moving does not help ^[11]. He asked for 128, got a same-day partial approval at 64, then learned that increases past that threshold have been removed from self-service and now require a review reachable only through an account team ^[12].

His two-card instance is 48 vCPUs ^[13]. Two of them would be 96, which is 32 vCPUs over the approved quota, about 50 percent more than he has ^[14]. One instance leaves 16 vCPUs stranded, well short of a second ^[14]. "I do not scale; I choose a shape and fit inside it," he wrote ^[15].

The fallback is quoted separately. On-demand GPU quota on the same account was 64 vCPUs in one region, 16 in another, and 8 in a third ^[16]. A 48-vCPU shape fits in none but the first, so in two of the three regions the on-demand escape hatch cannot host the machine it is meant to replace ^[17].

The design follows from that. A controller holds a target number of healthy slots and never names a machine or a zone; which zone he is in is an outcome of that, and it changes with every replacement ^[20]. Under it sits a ladder of shapes tried in order, starting with two cards on one machine ^[22]. "Four pieces, none of them clever, all of them necessary," he wrote ^[21]. For batch-shaped work the entry cost is a retry, against a price cut of roughly 4.3x to 5.5x on this family ^[18][2]. For a live session, the 80 percent has to fund the controller and the ladder, and it cannot buy a larger quota at any price ^[19][12].