On this page
VM.Standard.E2.1.Micro), and unlike the ARM A1 shape they are usually easy to get: ours launched on the first try in Osaka, a region where A1 failed about 1,670 times in three days. The catch is size. Each micro is 1/8 of an OCPU with burst, 954 MiB of usable RAM and about 50 Mbps to the internet. That is enough for a small website, a bot, cron jobs, uptime checks or a personal relay. It is not enough for Docker stacks, databases or anything that needs real memory. If you need those, keep retrying for the 2-core, 12 GB ARM A1; take the micros either way, because they cost nothing. Tested October 11–14, 2026, on one account in ap-osaka-1.Oracle's Always Free tier has two compute offers, and most write-ups only talk about one of them. The ARM A1 allowance (2 OCPUs and 12 GB since June 2026) is the famous one. The two AMD micro instances are the ones you can actually create in a lot of regions. This page is about the micros: what you really get, how they compare with A1 and with other free VMs, and what they are good for.
Which one should you pick?
| AMD micro (E2.1.Micro) | ARM A1 (A1.Flex) | |
|---|---|---|
| Free amount | 2 instances | 1,500 OCPU-hours + 9,000 GB-hours a month (= 2 OCPU / 12 GB running all month) |
| CPU | 1/8 OCPU baseline, can burst; the OS shows 2 vCPUs | Up to 2 dedicated Ampere cores |
| RAM | 1 GB (954 MiB visible to Ubuntu) | Up to 12 GB |
| Internet bandwidth | Up to 50 Mbps (we measured about 53 Mbps) | Scales with OCPUs |
| Getting one | Launched on the first API call in Osaka | "Out of capacity" for 3 days straight in Osaka |
| Architecture | x86-64, so every image and binary works | ARM64; most software works, some images don't |
| Pick it for | One light, always-on service | Docker, databases, self-hosted apps, CPU inference |
They are not either/or. The free limits are separate, so one account can run both micros and an A1 instance at the same time. The only shared budget is disk, covered below.
What we measured
We created the instance with the OCI API on Ubuntu 24.04, in a tenancy whose home region is Osaka (ap-osaka-1, which has a single availability domain). These are the numbers the machine itself reported:
| Check | Result |
|---|---|
CPU model (lscpu) | AMD EPYC 7551 |
| vCPUs shown to the OS | 2, but the shape is 1/8 OCPU with burst, so sustained load gets a small slice of one core |
Usable RAM (free -m) | 954 MiB |
| Boot volume | 47 GB, of which a fresh Ubuntu 24.04 used 3.8 GB |
| Download from the VM (Cloudflare speed test) | About 53 Mbps, in line with Oracle's 50 Mbps limit |
| Time to create | First attempt succeeded |
Two things in that table catch people out.
"2 vCPUs" is not two cores. Oracle counts x86 capacity in OCPUs, and one OCPU is two hardware threads, so the guest sees two vCPUs. But the micro's baseline is 1/8 of an OCPU. Short bursts (a page render, a cron job) feel normal; anything that keeps the CPU busy for minutes, like compiling or apt upgrade on a big package set, slows down noticeably.
The boot volume eats your free disk. Oracle's minimum Linux boot volume is about 47 GB, and it comes out of the same 200 GB of Always Free block storage as every other volume. 200 GB divided by 47 GB means at most four instances in total. Two micros at the minimum size use 94 GB and leave about 106 GB for an A1 boot volume and data. Don't raise the micro boot volumes "just in case" if you still plan to get A1.
Why the micro is easy to get and A1 isn't
On the same account, a script retried the A1 shape (2 OCPU / 12 GB) about 1,670 times over three days. Every attempt came back "Out of capacity". Forum reports suggest Osaka has been short of free A1 capacity since 2021, and other popular regions have the same problem on and off. The AMD micro, created from the same script and the same region, succeeded on the first call.
That matters because your home region is permanent and Always Free compute only runs there. If you picked a region where A1 is scarce, the micros may be the only free compute you get for a while. The A1 guide has the retry advice; this page assumes you want something running today.
What 1 GB and 1/8 OCPU can run
The rule of thumb: one light, always-on process per micro. Things that fit:
- A small website: a static site behind Nginx or Caddy, or a small Node.js app.
- Scheduled jobs: cron scripts, scrapers, backups that push to object storage.
- A chat bot (Telegram, Discord) that mostly waits on the network.
- Uptime monitoring, e.g. Uptime Kuma for a handful of checks.
- A personal VPN or relay, e.g. WireGuard. If you use Xray with REALITY, note that in our test on Xray 26.3.27 every handshake failed with
www.microsoft.comas the target, whilewww.apple.comworked.
What doesn't fit: Docker Compose stacks with several containers, any database bigger than SQLite, Java services, anything with a build step you run on the box, and local AI models. With 954 MiB, a single misbehaving process can push the VM into swap or the OOM killer. If you install much, add a small swap file first:
sudo fallocate -l 1G /swapfile
sudo chmod 600 /swapfile
sudo mkswap /swapfile
sudo swapon /swapfile
echo '/swapfile none swap sw 0 0' | sudo tee -a /etc/fstab
Network: fine for a service, slow from China in the evening
From the VM outward, the speed test ran at about 53 Mbps, which is the 50 Mbps cap Oracle documents. 10 TB of outbound traffic a month is free, which a 50 Mbps link cannot realistically use up (running flat out for a month is about 16 TB, and nothing light runs flat out).
If you or your users are in mainland China, expect the route to be the bottleneck rather than the VM. From one home connection in China (likely China Mobile) to the Osaka micro, round-trip time was about 205 ms, and a single TCP download measured:
| Time (China, local) | Setting | Single-stream download |
|---|---|---|
| 00:30 | Default congestion control | 0.13 Mbit/s |
| 00:30 | BBR enabled on the VM | 1.1 Mbit/s |
| 07:48 | BBR enabled on the VM | 13.3 Mbit/s |
That is one connection on one ISP, so read it as a data point, not a benchmark. It does show two useful things: turning on BBR is worth doing on any Oracle VM you reach over a lossy route, and with the same settings the VM was more than ten times faster at breakfast time than around midnight, because the congestion is on the cross-border route.
# Enable BBR (Ubuntu 24.04 ships the module)
echo 'net.core.default_qdisc=fq' | sudo tee -a /etc/sysctl.conf
echo 'net.ipv4.tcp_congestion_control=bbr' | sudo tee -a /etc/sysctl.conf
sudo sysctl -p
Free AMD micro vs other free small VMs
| Option | CPU / RAM | Disk | Free egress | Expires? |
|---|---|---|---|---|
| Oracle AMD micro (×2) | 1/8 OCPU burst / 1 GB each | Shares 200 GB block storage | 10 TB/month | Never |
| Oracle ARM A1 | 2 OCPU / 12 GB | Shares 200 GB block storage | 10 TB/month | Never, if you can get capacity |
| Google Cloud e2-micro | Shared 2 vCPU / 1 GB | 30 GB standard disk | 1 GB/month | Never |
| Small paid VPS | 1 core / 1–2 GB | 20–40 GB | Usually 1 TB+ | A few dollars a month |
Against Google's always-free e2-micro, Oracle's micro has a similar amount of CPU and RAM but far more bandwidth, and you get two of them. Against a cheap paid VPS, the micro loses on CPU and has no support, but costs nothing. If a project outgrows 1 GB, moving to A1 or to a paid box is easier than squeezing it.
Things to know before you rely on it
- Home region only. Oracle's docs say Always Free compute must be created in your tenancy's home region, and the home region can't be changed after signup.
- Idle reclamation applies. Oracle may reclaim an Always Free instance that is idle over 7 days: 95th-percentile CPU under 20% and network utilization under 20%. The third test, memory under 20%, only applies to A1 shapes, so on a micro you can't stay safe by holding memory. A service that does almost nothing all week is the case to watch. We have not run ours long enough to see whether a light bot trips this.
- Upgrading to pay-as-you-go doesn't enlarge the free allowance. Oracle's docs give the same A1 allowance (1,500 OCPU-hours and 9,000 GB-hours a month) to all tenancies. Run A1 at the old 4 OCPU / 24 GB size all month on a paid account and you pay for the hours above that, at $0.01 per OCPU-hour and $0.0015 per GB-hour on Oracle's price list: roughly $27–28 a month. The two micros stay free either way.
- Open ports in two places. As with A1, inbound traffic is blocked both in the VCN security list and by iptables on Oracle's Ubuntu images. The commands are in the A1 guide.
Verdict
Create both micros as soon as your account is open. They are free for good, they usually launch on the first try, and they are enough for one light always-on job each. Don't plan anything memory-hungry around them. For Docker, databases or self-hosted apps you need the 2-core, 12 GB ARM A1, which means a retry script and some patience in busy regions. If you can't wait for A1 and the micro is too small, a few dollars a month buys a VPS without Oracle's capacity lottery; I wrote up what living with a budget host is like in this JustHost review.
Sources: Oracle's Always Free Resources documentation (shape limits, 50 Mbps, home region, block storage, idle rules) and price list (A1 per-hour rates), both checked October 15, 2026. Measurements are from one account in ap-osaka-1, October 11–14, 2026.