Oracle Claws Back Half of Cloud's Best Free Tier

For five years, Oracle Cloud's free Ampere tier was the running joke that turned out to be true: the company nobody's hobby project wanted to touch was quietly handing out the most generous free compute in the industry — 4 Arm cores and 24 GB of RAM, forever, no credit card tricks. That deal is now half gone. Sometime in mid-June, Oracle edited its Always Free documentation to cut the Ampere A1 allowance from 4 OCPUs and 24 GB to 2 OCPUs and 12 GB — from 3,000 OCPU-hours and 18,000 GB-hours a month down to 1,500 and 9,000.

No blog post. No email. No changelog entry. Linuxiac caught the doc change around June 14; InfoQ confirmed the details in early July after Oracle support started answering (inconsistently) by email. The docs now read as if it were always so: "All tenancies get the first 1,500 OCPU hours and 9,000 GB hours per month for free."

The enforcement deadline is the part that finally pushed this onto the Hacker News front page this week: according to the post that surfaced there — apparently drawing on notices Oracle has begun sending — instances above the new limits will be automatically terminated starting August 18, 2026. That date hasn't appeared in any public Oracle announcement I could find, which given how this entire episode has gone is depressingly on-brand. Treat your own tenancy's console and inbox as the source of truth, and assume the axe falls soon either way.

One detail matters more than the deadline. Per the updated docs, if an over-limit resource is terminated, you may not be able to recreate it even at the new, smaller size beyond your remaining allowance. Resize what you have; don't let Oracle delete it for you.

Why the 4/24 deal existed — and why it died

The original allowance wasn't generosity, it was evangelism. In May 2021, Oracle launched Ampere A1 compute at a penny per core-hour and needed developers to take Arm servers seriously — to port their Docker images, test their builds on aarch64, and tell their friends. A free tier big enough to run a real workload (a k3s cluster, a Minecraft server, a whole self-hosted stack) was the marketing budget. It worked so well the tier was perpetually oversubscribed; "out of capacity" errors became their own genre of forum thread, and the standard advice was to upgrade to Pay As You Go just to get your instance placed.

Look at 2026 and the strategic logic has evaporated. Arm doesn't need evangelists anymore — Graviton, Axion, and Cobalt made multi-arch images table stakes years ago. Oracle sold its stake in Ampere Computing when SoftBank acquired the chipmaker in 2025. And OCI is now one of the most capacity-constrained clouds on earth, selling every rack it can power to AI customers. A fleet of Altra cores serving hobbyist Nextcloud instances at $0.00/month is exactly the kind of line item that dies in that environment. The cut is rational. The silence is the choice worth judging: a doc edit, contradictory support answers, and a termination deadline communicated so quietly that third-party blogs are the best available record.

What to do before August 18

If you're running anything on free-tier A1, this is a fifteen-minute job — do it now:

• Inventory. Console → Compute → Instances, or oci compute instance list, and total up your A1 OCPUs and memory. Over 2 OCPUs or 12 GB combined, you're in scope. The two free x86 Micro instances and the 200 GB of block storage are untouched.
• Resize in place, don't rebuild. A1.Flex shapes can be edited (stop → Edit shape → restart), keeping your boot volume and IP. Given the no-recreate caveat, terminating and relaunching is the one move that can lock in a permanent loss. Two 1-OCPU instances or one 2-OCPU instance both fit.
• Back up regardless. The free tier includes volume backups; take one before you touch anything, and keep a copy of anything irreplaceable outside OCI entirely.
• Do the math before upgrading to PAYG. Support has given conflicting answers on whether paid accounts keep the old 4/24 allowance free — the docs' "all tenancies" wording suggests not. At list prices ($0.01 per OCPU-hour, $0.0015 per GB-hour), the clawed-back half costs about $28/month. Hetzner will sell you a 4-core Ampere VM for a quarter of that. If you're paying anyway, OCI's free tier stops being the argument for staying.

Free tiers are marketing, not infrastructure

There's a familiar shape to this. Heroku killed free dynos in 2022. Railway and PlanetScale followed. Fly.io stopped handing out free allowances to new users. Even AWS restructured its free tier into expiring credits in mid-2025. Loss-leader compute made sense in a zero-interest-rate world where clouds fought for developer mindshare; in an AI-capex world, every idle watt has an opportunity cost, and the mindshare war is over.

Here's the uncomfortable verdict: even halved, OCI's free tier is still the best always-free compute you can get — 2 dedicated Arm cores and 12 GB of RAM beats anything AWS, Google, or Azure will give you in perpetuity. If you're choosing where to park a hobby project today, Oracle remains the rational answer. But this episode is the clearest statement yet of what a free tier actually is: revocable marketing spend, adjustable without notice, enforced by automatic termination. Oracle already reclaims free instances it deems idle; now it's reclaiming half the tier itself.

So build accordingly. Keep your provisioning in Terraform or a cloud-init script so redeployment is an afternoon, not an archaeology project. Keep backups off-provider. And never let anything you'd mourn live exclusively on hardware you don't pay for — the terms of a gift can change whenever the giver's incentives do. Oracle's just did.