Maquina V5: a free cloud gaming PC built on a Google Colab runtime

Running Steam on a borrowed Tesla T4 in five lines

The whole setup is five lines, handed over under a heading the README calls Inicio Rapido. There is no package manager step, no Dockerfile, and nothing to compile locally, because the machine doing the work is a Google Colab runtime that Google provides at no charge:

from google.colab import drive
drive.mount('/content/drive')
!wget -q https://github.com/kmille36/Colab-Cloud-Gaming/raw/refs/heads/main/ColabSteam
!chmod +x ColabSteam
!./ColabSteam

The first two lines are Colab plumbing rather than shell. `from google.colab import drive` imports the drive module that ships inside the runtime, and `drive.mount('/content/drive')` attaches your Google Drive at that path, which is how a saved game install outlives the runtime being discarded. The three lines prefixed with `!` are shell escapes: they fetch a single executable called `ColabSteam`, mark it runnable, and run it.

What that executable does is not in this repository. The download URL points at `kmille36/Colab-Cloud-Gaming`, not at `jephersonRD/Maquina-V5`, and the repository tree holds only `README.md` and an `assets/` folder of screenshots, so there is no source to read for the streaming stack, the game launcher or the Sunshine configuration. A reader is being asked to trust an opaque binary fetched over the network at first run, and that single decision is the real cost of adopting this.

The hardware table is the honest part of the pitch

The README publishes the machine it expects: an NVIDIA Tesla T4, an Intel Xeon with 2 cores at 2.0GHz, 12.67 GB of RAM, and Ubuntu 22.04.4 LTS. A T4 is a data centre card built for inference workloads, with modest single-threaded performance and a video encoder, which is the right shape for remote streaming and the wrong shape for a fast game. Two cores at 2.0GHz is the number that limits how much the CPU can carry, and it is worth reading before assuming a recent title will hold a steady frame rate.

The README then names sixteen games with a screenshot for each: Doom Eternal, Grand Theft Auto V, Resident Evil 2 Remake, Titanfall 2, Metal Gear Solid V, Hades, Forza Horizon 4 and others. What it does not give is a single frame rate, resolution or latency figure. So the game list tells you what the author got running and nothing about what you will get. Screenshots prove a game launched; they say nothing about whether it was playable.

The configuration table is more useful because it states what is missing. Moonlight control support is listed as working, and both virtual gamepad and physical gamepad are listed as not available. For a shooter or a racing game that is the whole experience, so this table is the one to read first.

Getting the picture to your laptop without opening a port

A Colab runtime has no public inbound address, so the client side needs a way in. The README requires two programs on your own machine: Tailscale, which it describes as a secure network connection, and Moonlight, the streaming client. Tailscale solves the addressing problem by giving both ends an address on a private overlay network, which is why no port forwarding appears anywhere in the instructions.

On the Colab side the host software is Sunshine, and the README gives its web interface address as `https://[tu-ip-tailscale]:47990`, with the Tailscale address standing in for the placeholder. Worth noticing: the README writes `https://` in front of port 47990, which is where Sunshine normally serves its unencrypted management page, so a certificate warning there is a plausible first-run experience rather than a broken install.

The session rules are the other thing to internalise before you start a long game. The README tells you to check the remaining time, to disconnect when you are finished, and to say plainly that the four hour allowance resets every 24 hours. It also warns against hiding or switching away from the Colab tab, because that is how the runtime gets reclaimed. A game session here has a known ceiling, and the ceiling belongs to Google rather than to this project.

Backing up a session to Drive to skip the reinstall

Since the runtime disappears, the expensive part of each session is the game download. The README's backup section exists to avoid paying that twice, and it claims a saving of 50% to 75% of the waiting time on your next session. The procedure is four steps: download and install your game from Steam, Epic or another platform, compile the shaders on first launch if the title needs it, take a backup once it is ready to play, and enjoy an immediate start next time.

The storage format is a single `backup.tar.gz`, and the README gives one hard constraint about it: the file must not sit inside any folder. That is the kind of detail that saves an afternoon, because an archive nested one directory deeper is the usual reason a restore appears to find nothing.

The Drive mount is optional, and the README says so directly, telling you to comment the code out with `#` if you do not need it. That also means skipping the mount costs you the backup path, so the decision is really between a fast reinstall and a slower one. Sharing a backup across accounts works by sharing the download link and creating a shortcut to the archive in the destination Drive, which is a manual arrangement rather than an automated sync.

Where the repository name and the README disagree

Two descriptions of this project sit side by side and they do not match. GitHub reports the repository as `jephersonRD/Maquina-V5`, described as a virtual gaming PC v5, with topics covering cloud gaming, GPU gaming, low-end PC and remote desktop. The README is titled Colab Cloud Gaming, and every badge in it points somewhere else: the Colab link, the stars, the forks, the issues and the license all address `kmille36/Colab-Cloud-Gaming`. The project also reports no license, no releases and no detected primary language, and the tree contains a README and images only.

Read together, that pattern fits a re-upload or a fork presented under a new name rather than an original codebase. What you can decide from this: if you only need the software, the upstream repository is where the issue tracker and any history would live, so that is the place to file problems. If you specifically wanted the Maquina V5 name and description, those are cosmetic differences and nothing in the instructions depends on them.

The update note has a similar wrinkle. The README carries a line reading ACTUALIZADO (10/01/2026) announcing that Steam now works, while the last push to this repository is dated 2026-04-20. Spanish date notation reads day first, so 10 January 2026 is the consistent reading; reading it as 1 October 2026 would place the note six months after the final commit. Either way the README's newest claim is older than the last commit, which means it may not describe the current state of the script.

What this trades away against a paid cloud gaming service

The honest comparison is not with another free Colab script but with the two things people actually do instead: subscribe to a commercial cloud gaming service, or buy a machine with a real GPU. A paid service gives you a stable session that does not expire at four hours, a controller that works, and support when something breaks. A local GPU gives you frame rates that do not depend on a network round trip. Maquina V5 wins on exactly one axis, the price, and it pays for that with input support, session length and predictability.

Two further limits are structural rather than fixable. The runtime is a free Colab instance, so its availability, GPU allocation and session length are set by Google and can change without anything in this repository changing. And the README's own advice to verify remaining time and to disconnect when idle is a workaround for a constraint the project does not control.

What remains genuinely useful is the pattern. Tailscale plus Sunshine plus Moonlight is a workable way to reach a machine that has no public address, and that recipe transfers to any GPU you can get hold of. If you already run Colab sessions with a GPU attached, or you want to understand the streaming setup, the instructions are honest enough to follow. If you want to play a long session tonight, they will not carry you there.

Editorial conclusion

Maquina V5 is worth an evening if your goal is to play something demanding this week without buying hardware, and it is the wrong tool the moment a controller, a long session or a stable frame rate matters, because the README's own configuration table rules out gamepads entirely. Nothing here is yours to maintain: the machine is a free Colab runtime and the program is a download from a different account. The first thing to check is that URL, `kmille36/Colab-Cloud-Gaming`, because if that file disappears every other instruction in the README stops working. After that, the client setup of Tailscale plus Moonlight is the only part you control.

Frequently asked questions

Official sources

1. Issues
2. jephersonRD/Maquina-V5 on GitHub
3. Project website
4. README