AWS Lambda MicroVMs pricing vs Northflank: which is cheaper for AI sandboxes?

AWS launched Lambda MicroVMs on June 22, 2026. It is a new serverless compute primitive that runs Firecracker-backed isolated environments per user or session, purpose-built for executing untrusted or AI-generated code. It competes directly with sandbox platforms like Northflank in the AI code execution category.

This article compares Lambda MicroVMs and Northflank sandbox pricing directly: the billing model, the per-unit compute rates, snapshot and storage costs, and the total cost across realistic sandbox workloads. All rates are as published on the AWS Lambda pricing page and Northflank pricing page as of June 2026.

Lambda MicroVMs charges across three dimensions:

You configure a baseline memory allocation. vCPU is allocated in a 2:1 memory-to-vCPU ratio (e.g., 2 GB memory = 1 vCPU baseline). You pay for baseline compute while the MicroVM is running. When workloads exceed the baseline, you pay only for the active duration of the additional resources consumed, not the peak capacity. Rates in US East (N. Virginia), ARM/Graviton:

• vCPU: $0.0000276944 per vCPU-second
• Memory: $0.0000036667 per GB-second

Lambda MicroVMs use snapshots to enable near-instant startup and suspend/resume. You pay for:

• MicroVM image storage: $0.08/GB-month (1-week minimum retention period)
• Snapshot data reads: $0.02/GB (charged on start or resume)
• Snapshot data writes: $0.02/GB (charged on suspend)

Standard AWS data transfer rates apply for outbound internet traffic and for data transferred between Lambda MicroVMs and your Amazon VPC.

When suspended, compute charges stop. This is the key cost optimisation in Lambda MicroVMs. A suspended MicroVM preserves full memory and disk state but incurs only snapshot storage charges, not compute charges. For interactive sessions with long idle periods, this meaningfully reduces effective cost.

Northflank sandboxes charge two components:

• vCPU: $0.01667 per vCPU-hour
• Memory: $0.00833 per GB-hour

Both are billed per second with no minimum. No snapshot fees, no data transfer fees within the platform, no additional charges for start, suspend, or resume operations. BYOC deployments bill the underlying cloud infrastructure at list price with no markup.

Lambda MicroVMs are ~6x more expensive per vCPU at published rates. The suspend feature partially offsets this for workloads with significant idle time.

Using AWS's own published example from the Lambda MicroVMs pricing page: a developer coding environment, 2 GB / 1 vCPU baseline, vertically scales to 8 GB / 4 vCPU during peak usage (compilation, test execution). 8-hour workday, 2.5 hours active (75% at baseline, 25% at peak), 5.5 hours suspended.

Lambda MicroVMs (AWS's published example, US East ARM):

Baseline usage (2.5 hrs × 75% = 1.875 hrs = 6,750 seconds):

• vCPU: 6,750s × 1 vCPU × $0.0000276944 = $0.187
• Memory: 6,750s × 2 GB × $0.0000036667 = $0.050
• Subtotal: $0.237

Peak usage (2.5 hrs × 25% = 0.625 hrs = 2,250 seconds, incremental 3 vCPU / 6 GB above baseline):

• vCPU: 2,250s × 3 vCPU × $0.0000276944 = $0.187
• Memory: 2,250s × 6 GB × $0.0000036667 = $0.050
• Subtotal: $0.237

Snapshot costs (image storage $0.08/GB-month × 2 GB × 1/30 days = $0.005/day + read on start $0.02/GB × 2 GB = $0.04 + write on suspend $0.02/GB × 2 GB = $0.04): approximately $0.085/day.

Total Lambda MicroVMs per developer per day: approximately $0.56 (excluding VPC data transfer charges)

Northflank (same workload, 2.5 hours active, no suspend):

Active hours only (2.5 hrs, no charge while stopped):

• vCPU: 1 vCPU × 2.5 hrs × $0.01667/vCPU-hr = $0.042
• Memory: 2 GB × 2.5 hrs × $0.00833/GB-hr = $0.042
• Peak scaling (0.625 hrs × 3 additional vCPU + 6 GB): ~$0.06

Total Northflank per developer per day: approximately $0.15

At this workload, Lambda MicroVMs are approximately 3.8x more expensive than Northflank per developer per day, even accounting for suspend-based idle cost savings on Lambda.

Using AWS's second published example: security scans running untrusted code, 8 GB / 4 vCPU baseline, scales to 32 GB / 16 vCPU at peak, 10-minute average job duration, 90% at baseline and 10% at peak, no suspend needed (jobs run to completion).

Lambda MicroVMs (AWS's published example):

Baseline (10,000 jobs × 600s × 90% = 5.4M seconds):

• vCPU: 5.4M × 4 × $0.0000276944 = $598.20
• Memory: 5.4M × 8 GB × $0.0000036667 = $158.40
• Subtotal: $756.60

Peak (10,000 jobs × 600s × 10% = 600K seconds, incremental 12 vCPU / 24 GB above baseline):

• vCPU: 600K × 12 × $0.0000276944 = $199.40
• Memory: 600K × 24 GB × $0.0000036667 = $52.80
• Subtotal: $252.20

Total Lambda MicroVMs: $1,008.80/month (excluding snapshot and data transfer costs)

Northflank (same workload, 10,000 jobs × 10 minutes = 1,666.67 vCPU-hours at 4 vCPU baseline, 8 GB):

• vCPU baseline: 4 vCPU × 1,666.67 hrs × $0.01667/vCPU-hr = $111.16
• Memory baseline: 8 GB × 1,666.67 hrs × $0.00833/GB-hr = $111.07
• Peak incremental (10% of time = 166.67 hrs × 12 additional vCPU): 12 × 166.67 × $0.01667 = $33.34
• Peak memory incremental (166.67 hrs × 24 additional GB): 24 × 166.67 × $0.00833 = $33.32

Total Northflank: approximately $289/month

Lambda MicroVMs are approximately 3.5x more expensive than Northflank for this batch workload pattern.

• Suspend/resume with state preservation: This is Lambda MicroVMs' most differentiated feature. A suspended MicroVM preserves full memory and disk state and resumes near-instantly with no compute charges during suspension. For interactive coding environments where developers step away for hours at a time, this is a meaningful cost advantage over platforms that charge for idle compute. Northflank does not currently offer native suspend/resume with state preservation at the same level.
• Deep AWS ecosystem integration: Lambda MicroVMs integrates natively with Lambda's event sources, the Agent Toolkit for AWS, CloudFormation, and the CDK. For teams already building on AWS Lambda's event-driven infrastructure, this is a lower-friction path to adding sandbox execution than adopting a separate platform.
• No infrastructure to manage: Lambda MicroVMs are fully serverless. You call an API to launch a MicroVM, and AWS handles all capacity, networking, and lifecycle management. Northflank also abstracts infrastructure management, but requires account setup and project configuration.

• Significantly lower compute rates: At published rates, Northflank is approximately 6x cheaper per vCPU and 1.6x cheaper per GB. For high-volume batch workloads or always-on sandbox environments, this gap compounds significantly.
• BYOC across multiple clouds: Northflank BYOC deploys into AWS, GCP, Azure, Oracle, CoreWeave, Civo, on-premises, and bare-metal. Teams with multi-cloud requirements or data residency obligations in non-AWS clouds cannot use Lambda MicroVMs as a standalone solution.
• GPU workloads alongside sandboxes: Northflank supports GPU sandbox environments (H100, H200, A100, L4, L40S, B200) in the same control plane as standard CPU sandboxes. Lambda MicroVMs does not currently support GPU workloads.
• Full deployment lifecycle in one control plane: Northflank is a full-stack platform: CI/CD, managed databases, preview environments, secrets management, and observability alongside sandboxes. Lambda MicroVMs is a compute primitive only.
• No region restrictions: Lambda MicroVMs is currently available in 5 regions (US East N. Virginia, US East Ohio, US West Oregon, Europe Ireland, Asia Pacific Tokyo) on ARM/Graviton only. Northflank offers 6+ Managed regions and 600 global regions via BYOC.
• No snapshot fees or data transfer overhead: Northflank has no additional charges for sandbox lifecycle operations. Lambda MicroVMs adds snapshot storage, read, and write charges on top of compute.

At published US East ARM rates: $0.09970/vCPU-hr and $0.01320/GB-hr for active compute. Suspended MicroVMs incur no compute charges. Additional snapshot storage and data transfer charges apply on top.

$0.01667/vCPU-hr and $0.00833/GB-hr, billed per second. No snapshot fees, no data transfer charges within the platform, no additional lifecycle operation charges.

At published compute rates, Northflank is approximately 6x cheaper per vCPU. Lambda MicroVMs' suspend feature reduces effective cost for workloads with significant idle time, but even accounting for suspend savings on the per-developer coding environment workload, Northflank remains approximately 3.4x cheaper per developer per day in the worked example above.

Northflank supports ephemeral and persistent sandbox environments. Native suspend/resume with full memory and disk state preservation, equivalent to Lambda MicroVMs' feature, is not currently available in the same way. For workloads where suspend/resume state preservation is a core requirement, Lambda MicroVMs have a genuine advantage.

Lambda MicroVMs run in your AWS VPC natively but do not support deployment into GCP, Azure, or on-premises environments. Northflank BYOC supports AWS, GCP, Azure, Oracle, CoreWeave, Civo, on-premises, and bare-metal from one control plane.

Lambda MicroVMs uses Firecracker, the same virtualisation technology that powers standard Lambda Functions. Northflank supports multiple isolation models, including microVM-based isolation with Kata Containers and Cloud Hypervisor, Firecracker-based workloads, and gVisor-based sandboxing, depending on workload requirements.

Lambda MicroVMs is a well-designed primitive for AWS-native teams that need VM-level sandbox isolation with suspend/resume state preservation. The suspend feature is genuinely differentiated for interactive sessions with significant idle time. The compute rates are approximately 6x higher per vCPU than Northflank at published prices, and the gap does not fully close even when accounting for suspend savings on interactive workloads.

For high-volume batch workloads, agent execution, and code execution platforms where sandboxes run continuously, Northflank's lower base rates are the dominant factor. Model your active-to-idle ratio before committing to either platform.

• AI sandbox pricing comparison (2026): PaaS rates, BYOC costs, GPU pricing, and a cost breakdown across E2B, Modal, Fly.io, Vercel Sandbox, and Northflank.
• Best platforms for untrusted code execution: Isolation model selection, multi-tenant design, and network controls for platforms running AI-generated code.
• What is a sandbox environment?: How sandbox environments work, when you need them, and how isolation models differ across platforms.