GPU workloads on our EKS clusters started failing after a routine node group upgrade. The error was cryptic: nvidia.com/gpu resource not found on nodes that clearly had GPUs. Every GPU pod stuck in Pending. This is the story of finding and reproducing a regression in NVIDIA Container Toolkit 1.20.0 that shipped in the EKS NVIDIA AMI v20260818.

Credits hide the bill. The day they end, the same usage shows up as cash. The useful move is to split that bill into a baseline you can commit to, and a variable line you refuse to commit to, then buy the commitment only after the baseline is real.

g6 was not there when the session started. Not on Spot, and not as a normal On-Demand launch either. Karpenter would sit on InsufficientInstanceCapacity while a user was already waiting. I used an On-Demand Capacity Reservation to hold the instance before business hours, then release it when the day was over.

The point was the GPU existing at the right moment. The smaller bill was a side effect.

On-prem workstations were reporting health by cron scripts that posted into chat. The replacement ships metrics from macOS and Linux machines into VictoriaMetrics. Planning docs looked at Grafana Alloy over Tailscale. That is not what shipped.

A roughly 31B-parameter multimodal model, quantized to NVFP4, runs only on Blackwell. The 4-bit kernels are not a format you can fall back onto an L40S or an L4. The service is spot-only, with no fallback on purpose. If there is no Blackwell capacity, the pod stays Pending and nothing is billed.

Each GPU workload in my current cluster has its own Karpenter NodePool. The isolation trick is small: keep the taint key as nvidia.com/gpu, and change only the value.

Scale-to-zero is cheap until the first request waits on an image pull. I benchmarked NVMe instance storage, Bottlerocket, SOCI, AL2023 FastImagePull, and EBS throughput on a fresh node every time.

Setup: g4dn.xlarge, a 3.6 GB image, one run per configuration. Treat the numbers as indicative, not a paper.

The first Langfuse design on EKS was full HA and about $3,400 a month at the load I expected. Scenarios ran from about $1.8K to $7.7K. I re-checked it before buying the hardware. The estimate was about 3.5× too high. The tier that shipped is about $1,000 a month, roughly 71% below the original number.

The first Langfuse write-up on this site was an ECS sketch from earlier in the year. What I actually built, in my current role, is Langfuse v3 on EKS. It is validated on a lean tier. Apps are still cutting over from the old Docker Compose box.

These are the seven things that cause real outages, plus how the keys survive the move.

Two embedding models, one physical GPU. That was the May design. By August each pod had its own L40S. Time-slicing had added sharing and no throughput.