Lumenrack Functions goes global

Hello from the Lumenrack team. June is the culmination of many months of work on Lumenrack Functions, the serverless container platform for CPU and GPU workloads. Here is what is new.

CPU and GPU checkpointing

Snapshot a fully warmed CPU and GPU container after model weights and kernels are loaded, then restore new replicas from that snapshot instead of starting from scratch. For vLLM, SGLang, and other GPU-heavy workloads this turns multi-minute cold starts into seconds and cuts wasted GPU time during scale-up.

Checkpointing docs →

Global routing

Deploy to a single URL and Lumenrack routes every request to the closest healthy cluster across multiple regions and providers. If a cluster degrades, traffic fails over automatically.

Global routing docs →

Global storage

One persistent storage layer shared across all your apps and regions. Model weights and shared files are stored once and read everywhere, instead of being duplicated per app or per region.

Global storage docs →

Protected compute tiers

Each deployment now chooses its compute tier.

	Interruptible (default) Cheaper preemptible capacity. Nothing changes for existing deployments.
	Protected On-demand capacity Lumenrack does not reclaim while it is serving a request. For real-time and latency-sensitive workloads.

From July 1, protected is billed at 2x the interruptible rate. Interruptible pricing is unchanged, and enterprise customers with negotiated pricing are not affected.

Compute tiers docs →

From the blog

SOC 2 Type II

Lumenrack has completed its SOC 2 Type II audit and supports GDPR, HIPAA, and ISO 27001 programmes.

Beacon, the Lumenrack global router

How Beacon picks a healthy, cheaper, lower-latency cluster for every request in single-digit milliseconds, with automatic failover.

Restoring GPU workloads from memory snapshots

How Lumenrack snapshots warmed CPU and GPU memory inside its sandbox runtime and restores workloads like vLLM in seconds.