China Merchants Bank's 60% token-cost cut outpaces the utilization gain it credits

Cutting the cost of processing 1 million tokens by 60% means getting 2.5 times as many tokens per unit of spend ^[12][13]. Moving average utilization from 35% to more than 60% is a factor of at least 1.71 ^[11][14]. Packing more work onto the same cards covers 1.71 of the 2.5, and at most another 1.46x has to come from somewhere else. The CNCF announcement credits the architecture as a whole and does not itemize the saving ^[25][15].

HAMi is in the stack to share accelerator capacity across Kubernetes workloads ^[7]. Sharing is also what makes per-team cost hard: when one physical card backs several pods, the unit the scheduler hands out is a slice and the unit the finance team paid for is a card. Kueue supplies job queueing and quotas ^[4]. A quota caps what a team may consume before the job runs. Pricing what it did consume is a separate question. The remaining three projects do event-based autoscaling (KEDA), monitoring and metrics (Prometheus), and dataset access acceleration (Fluid) ^[5][6][8].

So a platform team copying this architecture inherits the scheduling and gets to write the accounting itself. Several things also have to hold for the money figure to travel. The bank runs nearly 10,000 accelerator cards and they are heterogeneous, meaning not all the same type or configuration ^[9]. On hardware you already own, a utilization gain is cash only if the alternative was buying more cards; on rented capacity it is cash only when you hand instances back. The New Stack put the open question in its own headline: "Kubernetes can run AI inference. But can it count the real cost?" ^[24]. Unified management reached 99% of the bank's AI compute ^[10]. At that fleet size, roughly 100 cards sit outside the common control plane ^[16].

Kubernetes v1.37 shipped 67 enhancements, two of them Alpha storage security features covering new bind mount options and emptyDir permissions ^[17][18]. Red Hat's Nispriha Jagan and Neeraj Krishna wrote on the Kubernetes project blog that multiple security findings have surfaced regarding emptyDir volumes, one of the most common writable volume types ^[19][21]. I would test both on a serving cluster well before any compliance date, because Alpha is where behaviour still changes. "Supporting noexec, nodev, and nosuid gives users a native way to harden volume mounts to match security benchmarks and policy," the authors wrote ^[20].

Gartner published its Magic Quadrant for Server Virtualization Platforms on Monday and placed HPE in the Challengers quadrant alongside Canonical and Oracle ^[22]. The New Stack discloses that HPE is a presenting sponsor of the Road to KubeCon series in which that quadrant was covered ^[23].