@infrawrench/bulk-instances

AWS instance attributes and prices (EC2, RDS, ElastiCache and more), with their full history, from the bundles Infrawrench publishes at instances-bundles.infrawrench.com.

import { downloadAwsEc2PricingData } from "@infrawrench/bulk-instances";

const data = await downloadAwsEc2PricingData({ token: process.env.INFRAWRENCH_API_KEY });

data.keys(); // ["m5.large", "t2.micro", ...]

const m5 = data.get("m5.large")!;
m5.attributes.vcpu; // "2"
m5.attributesIn("US East (N. Virginia)"); // with that region's overrides
m5.pricing.onDemand.find(
  (p) => p.regionCode === "us-east-1" && p.operatingSystem === "Linux" && p.tenancy === "Shared" &&
    p.preInstalledSw === "NA" && p.capacitystatus === "Used",
)?.hourly; // 0.096

m5.attributesHistory; // every attribute change, oldest first
m5.pricingHistory.reserved; // every reserved price change, oldest first

Any Infrawrench bearer credential works as token (an iwk_ API key, an agent credential or a WorkOS access token). It defaults to INFRAWRENCH_API_KEY.

Other AWS platforms

downloadAwsPricingData(platform, options) loads any platform's bundle; downloadAwsEc2PricingData(options) is the same as downloadAwsPricingData("ec2", options).

| Platform | AWS service | | --- | --- | | ec2 | EC2 | | rds | RDS | | elasticache | ElastiCache | | opensearch | OpenSearch Service | | redshift | Redshift | | docdb | DocumentDB | | neptune | Neptune | | memorydb | MemoryDB | | dax | DynamoDB Accelerator | | sagemaker | SageMaker | | emr | EMR | | gamelift | GameLift | | spot | EC2 Spot (see below) |

The fields that pick a price differ by platform. data.variantKeys names them, and every price has a variant object keyed by them:

import { downloadAwsPricingData } from "@infrawrench/bulk-instances";

const rds = await downloadAwsPricingData("rds");
rds.variantKeys; // ["databaseEngine", "databaseEdition", "licenseModel", "deploymentOption", ...]
rds.get("db.m5.large")!.pricing.onDemand.find(
  (p) => p.regionCode === "us-east-1" && p.variant.databaseEngine === "MySQL" &&
    p.variant.deploymentOption === "Multi-AZ",
)?.hourly;

operatingSystem, tenancy, licenseModel, preInstalledSw and capacitystatus are EC2's variant fields. On other platforms they are "", unless the platform has a field with that name (RDS has licenseModel).

spot is EC2's spot prices. They are in pricing.onDemand, one per region and operatingSystem (Linux or Windows): the lowest price across the region's availability zones, as AWS's Spot pricing page shows it. Its locations are region codes, so region and regionCode are both us-east-1. Spot instances have no attributes; take their specs from ec2. The history goes back to August 2026, hourly at first and every 6 hours since November 2026, and prices change far more often than on-demand prices.

RunPod

downloadRunpodPricingData(options) loads RunPod's GPU and CPU pod prices. It takes the same options as the AWS functions (RunpodPricingDataOptions), caches the same way, and fails with a RunpodPricingDataError, a subclass of AWSPricingDataError. It returns a RunpodPricingData, whose get returns a RunpodInstance; they have the same members as AWSPricingData and AWSInstance.

import { downloadRunpodPricingData } from "@infrawrench/bulk-instances";

const runpod = await downloadRunpodPricingData();
runpod.get("NVIDIA H100 80GB HBM3")?.pricing.onDemand.find((p) => p.variant.cloudType === "Secure Cloud")?.hourly;

It has GPU types by RunPod's id (NVIDIA H100 80GB HBM3), priced per GPU per hour, and CPU pods by pod id (cpu3c-2-4: flavor, vCPUs, GB of memory), priced per pod per hour; the productFamily attribute is GPU or CPU. Prices are the same in every data center, so every price is in the location global. The only variant key is cloudType, Secure Cloud or Community Cloud. Savings plans are in pricing.reserved with the term 3mo or 6mo, Savings Plan, All Upfront, and hourly is the plan's effective hourly rate. Stock levels and spot prices are not included. The history starts in November 2026.

fetchRunpodInstance(id, options) fetches one GPU type or CPU pod from the Infrawrench API instead (see One instance). It returns a RunpodInstance, or undefined if RunPod has no instance with that id, and fails with a RunpodPricingDataError:

import { fetchRunpodInstance } from "@infrawrench/bulk-instances";

const h100 = await fetchRunpodInstance("NVIDIA H100 80GB HBM3");

Alibaba Cloud

downloadAlibabaPricingData(options) loads Alibaba Cloud ECS prices. It takes the same options as the AWS functions (AlibabaPricingDataOptions), caches the same way, and fails with an AlibabaPricingDataError, a subclass of AWSPricingDataError. It returns an AlibabaPricingData, whose get returns an AlibabaInstance; they have the same members as AWSPricingData and AWSInstance.

import { downloadAlibabaPricingData } from "@infrawrench/bulk-instances";

const alibaba = await downloadAlibabaPricingData();
alibaba.get("ecs.g6.large")?.pricing.onDemand.find(
  (p) => p.regionCode === "cn-hangzhou" && p.variant.operatingSystem === "Linux" && p.variant.networkType === "VPC",
)?.hourly;

Its locations are region ids (cn-hangzhou), so the location and its region code are the same, and regions gives each an English name. Its variant keys are operatingSystem (Linux or Windows) and networkType (VPC, or Classic in older history). pricing.onDemand has pay-as-you-go prices. pricing.reserved has subscriptions, with the term 1mo or 1yr to 5yr, Subscription, All Upfront: they are paid in full when bought, so hourly is 0 and upfront is the price. It also has reserved instances, with the term 1yr or 3yr, Reserved Instance, and a purchase option. Attributes are Alibaba's instance type definitions (CpuCoreCount, MemorySize, CpuOptions.ThreadsPerCore, ...), from May 2023 on. The history starts in May 2018.

fetchAlibabaInstance(instanceType, options) fetches one instance type from the Infrawrench API instead (see One instance). It returns an AlibabaInstance, or undefined if Alibaba has no instance type with that name, and fails with an AlibabaPricingDataError:

import { fetchAlibabaInstance } from "@infrawrench/bulk-instances";

const g6 = await fetchAlibabaInstance("ecs.g6.large");

Azure

downloadAzurePricingData(options) loads Azure virtual machine prices. It takes the same options as the AWS functions (AzurePricingDataOptions), caches the same way, and fails with an AzurePricingDataError, a subclass of AWSPricingDataError. It returns an AzurePricingData, whose get returns an AzureInstance; they have the same members as AWSPricingData and AWSInstance.

import { downloadAzurePricingData } from "@infrawrench/bulk-instances";

const azure = await downloadAzurePricingData();
azure.get("Standard_D2s_v3")?.pricing.onDemand.find(
  (p) => p.regionCode === "eastus" && p.variant.operatingSystem === "Linux" && p.variant.priority === "Regular",
)?.hourly;

It has VM sizes by ARM name (Standard_D2s_v3), priced per VM per hour. Its locations are ARM region names (eastus), so the location and its region code are the same, and regions gives each Azure's display name (US East). Its variant keys are operatingSystem (Linux or Windows) and priority (Regular, Spot or Low Priority). pricing.onDemand has pay-as-you-go prices, dated by when Azure says they took effect. pricing.reserved has reservations, with the term 1yr, 3yr or 5yr, Reserved Instance, All Upfront: hourly is 0 and upfront is the whole term's price. Azure lists reservations under Linux only; on Windows the license is billed on top. It also has savings plans, with the term 1yr or 3yr, Savings Plan, No Upfront, and hourly is the plan's rate. Attributes are skuName, productName and, from Azure's pricing calculator, cores, ram (GB), diskSize (temporary disk, GB), gpu and series. About a third of sizes, mostly the newest and retired series, have no calculator specs. Dedicated hosts are not included. The history starts in November 2026.

Google Cloud

downloadGcpPricingData(options) loads Google Cloud Compute Engine prices. It takes the same options as the AWS functions (GCPPricingDataOptions), caches the same way, and fails with a GCPPricingDataError, a subclass of AWSPricingDataError. It returns a GCPPricingData, whose get returns a GCPInstance; they have the same members as AWSPricingData and AWSInstance.

import { downloadGcpPricingData } from "@infrawrench/bulk-instances";

const gcp = await downloadGcpPricingData();
gcp.get("n2-standard-2")?.pricing.onDemand.find(
  (p) => p.regionCode === "us-central1" && p.variant.provisioningModel === "Standard",
)?.hourly;

It has machine types by name (n2-standard-2), priced per VM per hour. Google bills most machine types per vCPU and per GiB of memory at their family's rates, plus their GPUs, so each price is worked out from those rates and the machine type's shape. Prices leave out bundled Local SSD, OS licenses and sustained use discounts, and TPU, X4, X5 and A4X machine types aren't included. Its locations are regions (us-central1), so the location and its region code are the same. The only variant key is provisioningModel, Standard or Spot. pricing.reserved has committed use discounts, with the term 1yr or 3yr, Committed Use, No Upfront, and hourly is the commitment's rate. Attributes are the machine type's fields from the Compute Engine API (guestCpus, memoryMb, description, isSharedCpu, ...), acceleratorType and acceleratorCount for its GPUs, and zones, the zones in each region that have it, which differ by region (see attributesIn(region)). The history starts in November 2026.

Hetzner, OVHcloud, Oracle Cloud and IBM Cloud

These load the same way as Google Cloud's, with the same options, caching and failures. Each returns its own data and instance types, which have the same members as the AWS ones.

| Function | Data | Instance | Error | | --- | --- | --- | --- | | downloadHetznerPricingData | HetznerPricingData | HetznerInstance | HetznerPricingDataError | | downloadHetznerAuctionPricingData | HetznerAuctionPricingData | HetznerAuctionInstance | HetznerAuctionPricingDataError | | downloadOvhPricingData | OVHPricingData | OVHInstance | OVHPricingDataError | | downloadOvhEcoPricingData | OVHEcoPricingData | OVHEcoInstance | OVHEcoPricingDataError | | downloadOraclePricingData | OraclePricingData | OracleInstance | OraclePricingDataError | | downloadIbmPricingData | IBMPricingData | IBMInstance | IBMPricingDataError |

import { downloadHetznerPricingData } from "@infrawrench/bulk-instances";

const hetzner = await downloadHetznerPricingData();
hetzner.get("cx23")?.pricing.onDemand.find(
  (p) => p.regionCode === "fsn1" && p.variant.ipv4 === "Primary IPv4",
)?.hourly;

• hetzner, Hetzner cloud servers, by Cloud API name (cx23), per hour. Locations are Cloud API location names (fsn1, ash). The only variant key is ipv4, IPv6 only or Primary IPv4, since a primary IPv4 is optional and billed apart. pricing.reserved has the monthly cap, with the term 1mo, Monthly Cap, Postpaid, hourly 0 and upfront the cap. Attributes include cores, cpu, memoryInGb, diskInGb and architecture.
• hetznerauction, Hetzner Server Auction listings, by listing id (3107671). Each is one physical server, so its history runs from when it is listed until it is sold, and it has the same variant key and term as hetzner. Attributes are the listing's hardware and details (CPU.Name, RAM.Size, datacenter, fixedPrice, setupFee, ...). The bundle is read hourly, so a price drop is dated by the read that saw it.
• ovh, OVHcloud Public Cloud flavors (b3-8), per hour, from the World and US catalogs, in USD. Locations are pricing zones, not regions: world, 3az, lz-eu, lz-eurozone, lz-af, us and us-east-lz-dal-a. The only variant key is operatingSystem. pricing.reserved has monthly plans where a flavor has them, with the term 1mo, Monthly, Postpaid or All Upfront, hourly 0 and upfront the month's price.
• ovheco, OVHcloud's Kimsufi, So you Start, Rise and Advance servers (24sys01), by data center (gra, sgp); the range attribute says which. These have no on-demand prices, because they are only rented by the month. The rent is in pricing.reserved, with the term 1mo, Rental, All Upfront, hourly 0 and upfront the rent, beside prepaying (1yr or 2yr, Prepaid). The variant keys are memory, storage and systemStorage, which is often empty.
• oracle, Oracle Cloud VM and bare metal shapes, per hour, the same in every commercial region, so every price is in global. The only variant key is operatingSystem. Flexible shapes have no one price, so they are priced at a set of sizes named <shape>-<ocpus>-<GB> (VM.Standard.E5.Flex-2-16), which aren't names Oracle uses: the shape attribute is, and ocpuPrice and memoryGbPrice give any other size.
• ibm, IBM Cloud Virtual Server for VPC profiles (bx3d-2x10), per hour, by region. The only variant key is capacityType, On-Demand, Spot or Burstable 50%, 25% or 10%. pricing.reserved has reservations, with the term 1yr or 3yr, Reserved, No Upfront, and hourly the reservation's rate. About 90 older profiles (bx2, cx2, mx2, gx2, ...) have no public on-demand price of their own, so theirs is IBM's rate for an hourly reservation, which is an approximation.

All of their histories start in November 2026.

There is no single-instance fetch for Azure, Google Cloud, Hetzner, OVHcloud, Oracle Cloud or IBM Cloud yet; download their bundles instead.

One instance

fetchAwsInstance(platform, instanceType, options) fetches a single instance from the Infrawrench API instead of downloading the platform's whole bundle. It returns the same AWSInstance that get does, or undefined if the platform has no instance with that name:

import { fetchAwsInstance } from "@infrawrench/bulk-instances";

const m5 = await fetchAwsInstance("ec2", "m5.large");
m5?.pricing.onDemand.length;

It takes token like the bundle download, and baseUrl for where the API is served from (https://app.infrawrench.com by default). Nothing is cached. The API syncs each new bundle within about half an hour of it being published, so it can briefly trail downloadAwsPricingData.

fetchRunpodInstance(id, options) and fetchAlibabaInstance(instanceType, options) do the same for RunPod and Alibaba Cloud, with the same options (RunpodInstanceOptions, AlibabaInstanceOptions) and the same half-hour sync. They return a RunpodInstance or AlibabaInstance (or undefined) and fail with the provider's error. RunPod's ids have spaces; pass them as they are, they are escaped for you.

Caching

The bundle is downloaded xz-compressed (about 20 MB) and kept decompressed (about 210 MB) in ~/.infrawrench/aws_pricing_data (cachePath), so a cached load skips decompression. Each load asks the server which bundle is latest and downloads it only if it is not cached, deleting the older one. If the server cannot be reached, the newest cached bundle is used instead. Each platform's bundle is cached under its own name (rds-bundle-…json), so loading one never evicts another. Pass cachePricingData: false to always download and never touch the disk.

Loading decompresses and parses the whole bundle. Instances are decoded lazily: get only wraps the raw rows, and attributes, attributesHistory, pricing and pricingHistory each decode on first access and keep the result. The returned objects are shared between calls, so treat them as read-only.

Prices and regions

• region is the AWS location name (US East (N. Virginia)), regionCode its code (us-east-1) where the price list gives one.
• pricing holds only what is still offered. In pricingHistory, a null hourly means the price stopped being offered at at.
• Instances AWS no longer lists are still in keys(), with deletedAt set and empty pricing.