Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- The new minimums are 1,000 MBps of throughput and 1,500 metadata IOPS. Each is 75% lower, and the two measures are independent.
- Throughput describes supported read and write data speed at the file servers. Metadata IOPS describes file and directory operations per second; it is not data bandwidth or every client I/O.
- The announcement applies to new file systems in Regions where Intelligent-Tiering is available. It does not establish whether an existing system can use the new minimums.
- A 75% reduction in minimum capacity does not mean a 75% reduction in total cost. Consumed storage, cache, requests, provisioned throughput and metadata IOPS, backups and data transfer can all affect the bill.
The new minimums are 1,000 MBps and 1,500 metadata IOPS
On November 12, 2026, Amazon FSx announced that the minimum provisioned throughput for a new FSx for Lustre Intelligent-Tiering file system is dropping from 4,000 MBps to 1,000 MBps. The minimum metadata performance is dropping from 6,000 IOPS to 1,500 IOPS. Each minimum is reduced by 75%.
The update gives HPC and machine learning workloads that begin with smaller datasets and grow over time a lower starting configuration to evaluate. It changes the available minimum settings; it is not a promise about the speed a particular workload will achieve or the amount its bill will change.
Treat throughput and metadata IOPS as separate measures
MBps means megabytes per second. A file system’s throughput capacity determines the read and write data speed supported by its underlying file servers. It does not guarantee that an application, client and network will measure that rate under every access pattern. AWS’s configuration performance table describes designed performance, not a measurement or guarantee for your client workload.
Metadata IOPS is the capacity to process file and directory operations per second, such as creating, reading and deleting files or directories. It describes work on the file-system namespace, not data bandwidth, and it is not the same as all IOPS issued by clients. Workloads with many files or frequent directory operations may need a separate assessment of this setting.
Assess hot and cold data alongside latency sensitivity
Intelligent-Tiering is intended for workloads that mix hot and cold data and do not require consistent SSD-level performance for every access. It can suit elastic HPC or machine learning datasets, but choose throughput and metadata IOPS from expected concurrent reads and writes, file-operation rates and peak demand rather than from the new minimum alone.
Throughput capacity is independent of the amount of data stored. The file system provides an SSD cache for frequently accessed metadata by default, and an optional provisioned SSD read cache can help with frequently accessed data; include cache size and its cost in the design. When a cache miss is served from an Intelligent-Tiering storage tier, the guide describes time to first byte in roughly tens of milliseconds and per-request charges. A read served from the SSD read cache has sub-millisecond latency and no such request charge. More cache can help, but the benefit depends on the hot dataset’s size and how sensitive the workload is to slower reads of colder data.
Do not apply new-file-system minimums to an existing system
The announcement says the lower minimum throughput and metadata IOPS are for new file systems in Regions where the Intelligent-Tiering storage class is available. It does not say existing file systems will be resized, repriced or migrated automatically, and it does not confirm whether an existing system can be modified to use the new minimums.
For an existing deployment, check current documentation and the file-system options available to your account and Region, then compare them with that system’s configuration, workload and change conditions. Do not infer that the new minimums automatically create the same choices or cost changes for an existing system.
Re-estimate the full set of cost components
The 75% figure is the reduction in each minimum, not a discount on the total bill. Even when a new file system can start at a lower minimum, actual cost still depends on the throughput and metadata IOPS the workload needs, the storage it consumes and its read and write request volume.
Include consumed storage in each Intelligent-Tiering access tier, monitoring and automation, optional SSD read cache, read and write requests, provisioned throughput, provisioned metadata IOPS, backup storage and applicable data transfer. Transfers across Availability Zones or VPC peering, between Regions, and through VPN or Direct Connect can also have charges. Recalculate with current prices for the applicable Region and your own usage assumptions, considering cache, cold-data first-byte latency and request charges together.
Sources
Frequently Asked Questions
Why does the performance table show 3,125 MBps of baseline network throughput for the 1,000 MBps row?
The row lists provisioned throughput capacity, baseline network throughput and maximum SSD-cache disk throughput as separate measures: 1,000, 3,125 and 1,000 MBps, respectively. The 3,125 figure belongs to the network-throughput column; it is not extra capacity to add to provisioned throughput, and it is not a client speed that every workload will measure. Read each value by its column name and assess workload results with the application, cache and network conditions in mind.
Does one application read always create one billable Intelligent-Tiering read request?
Not necessarily. The guide says that each block retrieved from an Intelligent-Tiering storage tier generates a read request, and sequential reads can trigger prefetching. A read served from the file server’s in-memory cache or SSD read cache follows a different path. Application-level read-call counts therefore cannot be converted one-for-one into tier requests or charges; include block retrieval, sequential prefetching and cache hits in the estimate.