Budget customer-owned EFS storage separately from model inference

“No Anthropic charge” is not the same as “no infrastructure bill.” That distinction matters when budgeting an AI API application that opts into customer-owned storage for Enterprise Frontier Safeguards (EFS)—not Amazon Elastic File System.

In its EFS announcement, Anthropic says it does not charge for the safeguards. But if customers elect to store activity data in their own cloud account, their cloud provider bills for storage, reads, writes and data egress. The announcement names Amazon S3, Azure Blob Storage and Google Cloud Storage as examples of customer-owned storage.

Anthropic also says customer-owned storage, Customer-Managed Encryption Keys and fully automated review are individually opt-in, and do not change model behavior, API pricing or rate limits. That leaves two separate budget questions: what the model calls cost, and what the chosen cloud storage arrangement costs.

Keep a separate infrastructure line

A useful planning decomposition—not a vendor quote—is:

Customer-cloud EFS budget = storage + reads + writes + applicable data egress

Keep that alongside inference spend rather than treating it as a new token price. For each component, use the cloud provider’s actual billing units and applicable rates:

Component Input to establish before estimating
Storage Average retained billable volume over the billing period, storage class and region
Reads Billable read quantities, including the access pattern needed for monitoring and review
Writes Billable write quantities rather than an assumed one-write-per-API-call ratio
Egress Billable transfer volume and destination; do not assume every transfer is charged or free

Apply the provider’s tiers and allowances where relevant. The announcement does not provide rates, retention duration, operation counts or a data-to-token ratio, so it cannot support a dollar estimate or a savings claim. This decomposition also is not a complete application TCO model.

What to check before opting in

My budgeting recommendation is to identify the cloud account and region first, then confirm retention and access requirements with the teams responsible for security and storage. Do not shorten retention purely to hit a cost target without establishing those requirements.

When access is available, use a bounded evaluation to collect retained volume, reads, writes and transferred bytes over a defined period. Label forecasts as forecasts, then reconcile them against the cloud bill separately from inference spend. Holding token volume constant is not enough to establish that those infrastructure meters stay constant.

The source describes a phased rollout “starting later this fall”; it is not proof that every account or integration can enable EFS now. Confirm availability and the actual storage arrangement before turning this worksheet into a committed budget.

The practical decision is whether the chosen customer-owned storage arrangement fits both the data-control requirements and the infrastructure budget—not whether a zero-priced safeguards feature makes the surrounding workload free.

A useful teaching example for this worksheet is to separate stored inventory from monthly activity. Otherwise, “GB per month” can hide two different quantities.

Suppose, purely for illustration, an available customer-owned storage setup adds 1 GB of billable data each day, starts empty, and retains everything for a 30-day evaluation. Ignore compression, replication and deletions in this simplified example. It ends with 30 GB, but the average stored volume during that first month is roughly 15 GB, depending on write timing. Neither figure tells you the number of billable read or write operations.

I would walk through the estimate in this order:

  1. Draw the retained-volume curve: what arrives, what expires, and what was already there at the start?
  2. Match that curve to the cloud provider’s storage billing unit. Do not automatically use the end-of-month volume as the month’s average.
  3. Estimate reads, writes and applicable transfers separately; bytes arriving are not an operation count.

The budgeting trap is extrapolating that empty-start month as steady state. With no deletions, the next month starts with the previous month’s retained data, even if daily arrivals stay unchanged. With an approved retention window, model expiry explicitly instead.

These are illustrative storage-accounting assumptions, not documented EFS ingestion or retention behavior. The actual retention policy and provider meters are prerequisites before putting numbers into a committed forecast.