From PaaS Back to Your Own Server: What a Side Project Actually Costs in 2026

The previous guide to free hosting ended on this: no mainstream PaaS will now give you $0 for an always-on web app with a database. There is one exception, and it is not a PaaS - Oracle Cloud's Always Free tier, covered in section 8, along with the reclamation rule that makes it unreliable for exactly the kind of quiet side project you are likely to run on it.

This one answers the next question - where you should go instead, and what that costs.

Every price and plan detail here was checked on 6 August 2026 against the provider's own pricing page, docs and changelog, with no third-party roundups used as a source. That matters more than usual for this article, because its central claim is that most existing writing about these services is out of date - including writing from six months ago.

The lead is Zeabur. In February 2026 this platform did something rare: it publicly announced that it was leaving the compute business, and published its full reasoning. A company that was itself a managed PaaS now sells "we manage the server you rent." The whole industry trend, compressed into one product.

---

The answer first

There is no $0 in that last column. Not because free died, but because free changed shape - the last section covers which kinds of free are still perfectly healthy.

---

1. How free PaaS tiers disappeared

Every entry below is dated and traceable to the company's own page. This is the part of the article you can cite.

Two things worth pulling out of that table.

First, this is not a 2022 history lesson. The last four rows - Netlify's credit rates doubling, Render's free bandwidth cut to a twentieth, Hetzner's June increase, Render's auto-conversion - all happened in the last four months. The final row, Render converting legacy workspaces on 1 August, was five days ago.

Second, the Fly.io entry was done quietly. The date 2024-10-07 appears only inside Fly's own pricing docs; I could not find a corresponding blog post or changelog entry. So it is a primary-sourced policy date rather than a public announcement. Sign up for Fly today and what you get is a trial of "2 hours of machine runtime or 7 days of access, whichever comes first," with machines auto-stopping after 5 minutes of inactivity. That is not a free tier.

Why this happened, from a provider that said so

Most coverage lists what happened and never explains why it clustered in 2022-2023. Koyeb's post of 2023-10-27, "Sustaining free compute in a hostile environment," gives the answer: from October 2022, crypto miners began launching hundreds of deployments to mine during the build stage, because builds are not billed as runtime; by December 2022 an "exponential surge of abuse" forced Koyeb to require payment methods.

Free compute was not killed by cost. It was killed by abuse, after which cost became unbearable. Worth noting that in the same post Koyeb committed to keeping its free tier free - so it also counts as counter-evidence.

The honest part: not everything got worse

If this article only described decline it would be a complaint, not a reference. Three things moved the other way:

• Railway turned a free plan back on on 2025-08-27. It is much smaller - $1/month non-rollover against the $5/month recurring credit removed in 2023 - but the direction is reversed.
• Cloudflare publicly committed to keeping D1's free tier at GA, and the timing is striking: 2024-04-01 is five weeks before Render halved free Postgres lifetime and three weeks before PlanetScale's removal took effect.
• Vercel does not belong in the villain list. Its big 2025-06-25 change (Active CPU pricing for Fluid compute) was a price cut - Vercel's own worked example has a fully utilised Standard machine dropping from about $0.318/hour to about $0.149/hour. I found no primary source showing Vercel reduced Hobby's included free usage. The only verified Hobby reduction is the 30-day deployment retention cap announced 2026-04-27 and effective 2026-04-29, which affects deployment history, not whether your app runs.

So the defensible thesis is narrower than "free hosting died," and it is a better one:

What died is the free, always-on app with a persistent database. Free static hosting and free serverless/edge compute both survived, and in places improved.

Hold on to that distinction and you won't rent a VPS for something that belongs on Cloudflare Workers.

---

2. What Zeabur did: a PaaS company exiting the compute business

Zeabur is one of the better-known deployment platforms in the Chinese-speaking developer world, with full documentation, pricing and support in Traditional Chinese. And essentially every existing guide to it describes a product that has not accepted new projects since 15 March 2026.

That is not an exaggeration. Here are the dates.

What happened

On 23 February 2026, Zeabur co-founder Yuanlin published a changelog post titled "The Shared Cluster Will No Longer Accept New Projects." Three phases:

Zeabur's docs now label it "Shared Cluster (Deprecated)."

Existing services keep running. Zeabur's commitment is unusually firm: services already on the shared cluster will not be affected, will continue to run safely and stably even if you never migrate, and this applies with "no exceptions." But it also said two less comfortable things: existing shared-cluster services still require an active Developer Plan subscription to keep operating, and Zeabur "may adjust the shared cluster's usage-based pricing in the near future" and was "considering removing the $5 shared cluster usage credit currently included in the Developer Plan."

Whether that $5 credit was actually removed, I could not confirm - I found no later changelog or doc. In February 2026 it was stated as under consideration. Do not treat it as done.

How Zeabur explained it

Worth quoting directly, because very few companies put it this plainly:

"Providing compute resources means we are an infrastructure provider. We compete with AWS, GCP, Linode, and Hetzner... we have no structural advantage - our servers run on top of these providers, so our pricing can never be lower than theirs."

And, in the same post:

"the current pay-as-you-go pricing on the shared cluster is below our cost - we've been subsidizing this service."

This is why Zeabur is the right lead for this article. When Fly.io, Railway and Render withdrew free tiers, they were still PaaS companies that had become more expensive. Zeabur left the compute business entirely, and said so out loud.

What Zeabur is now

The homepage headline is "Your AI DevOps Engineer / handling all your deployments," and the product nav lists Servers & Clusters, AI Hub, Domains & DNS, Email Service.

In one sentence: Zeabur no longer sells compute. It sells the dashboard, the agent and the DevOps automation that sits on top of compute you buy from AWS, GCP, Hetzner, Linode, DigitalOcean, Aliyun, Tencent Cloud or Volcengine - or hardware you already own.

A new user picks one of three infrastructure routes, none of them Zeabur-owned:

1. Buy a server from Zeabur - a VPS resold from a third-party provider, provisioned automatically after payment and fully managed by Zeabur (no SSH or OS work required, though you do get root).
2. Connect your own server (fixed IP) - you supply a machine with a public IP and root or NOPASSWD-sudo SSH access.
3. Wonder Mesh - you supply a machine with no public IP (a Mac mini, an old laptop, a Raspberry Pi) which joins a Tailscale-based mesh.

There is a heavier fourth option: Clusters (multi-node Kubernetes), previously enterprise-sales-only, opened for self-serve purchase on 2026-04-02.

Does the $0 plan include a server?

No. This is the most important sentence in the section.

Zeabur's plan table leads with "manageable servers": Free 1, Dev 3, Pro 10, Team unlimited. It reads as though the free plan hands you a machine.

Zeabur's own zh-TW pricing tooltip removes all ambiguity. Translated: a server you purchased from another vendor and own yourself (including devices connected via Wonder Mesh), **not sold by Zeabur, but which can be managed centrally from the Zeabur dashboard.

"1 manageable server" is a binding slot, not a machine. You still pay someone else for the compute, or already own the hardware.

Zeabur's English FAQ draws the same line: the server itself provides compute resources, while the subscription plan lets you use and manage those resources more efficiently; a VPS is a server you purchase at a discount from cloud providers through Zeabur, and a subscription plan adds management services on top of it.

The two axes are also fully decoupled: you do not need a paid plan to deploy onto a server. Zeabur's FAQ says you can purchase and use a server for deployment on its own, but without a subscription plan you get only the most basic functionality. The February 2026 announcement said the same - no Developer Plan is needed to deploy services on a Dedicated Server.

Whether servers Zeabur sells you count against the "manageable server" quota, I could not confirm. The tooltip's "not sold by Zeabur" and the FAQ's consistent use of "external server" both strongly imply they do not, but no page states it. Read the quota as "it counts servers Zeabur did not sell you," not as "Zeabur-sold servers are unlimited."

So what do the paid plans buy?

Team includes 3 seats, additional seats $24 each.

Look at the build spec row: Free and Dev are identical. And Zeabur's docs are explicit that the build and the runtime happen on two different machines - build spec only affects how fast the build completes and has no effect on your running service. It is not a Free-versus-Dev differentiator at all.

Also note that the Free plan's domain quota is 0, as are API keys and webhooks. Taken together, Free is a look-around tier rather than a ship-something tier.

That "0 domains" row most likely counts domains registered through Zeabur's domain product rather than the free .zeabur.app subdomain generated for a service. The dashboard is behind login and I could not confirm which, so do not read it as "you cannot get a .zeabur.app URL on Free."

Three practical rules worth knowing:

• Trials: Dev and Pro each carry a 14-day free trial, but one trial per user in total - you pick one, not both. Team has no trial. At trial end the plan auto-renews and bills unless cancelled beforehand.
• Downgrading deletes nothing: if you had 3 servers bound on a paid plan, all 3 remain manageable after downgrading to Free; you just cannot add more until you upgrade again.
• Cancelling keeps your current plan's features until the end of the billing cycle, then switches the account to Free.

The current four-tier structure was only settled on 22 April 2026. Before that the lineup was Free, $5 and a $80 Team plan - Pro is new. The announcement names the new gating axes explicitly: number of server bindings, CI resource allocation, log retention and processing capacity, and observability features. Server-binding count became a headline plan lever four months ago.

Wonder Mesh: the only genuinely $0 path, and its catch

This is the most interesting piece of the new Zeabur, and nothing written about the platform before March covers it.

Wonder Mesh (announced 2026-03-16, beta approvals from 2026-03-20) turns any macOS or Linux device with no public IP into a full Zeabur server via a Tailscale mesh. Minimums: 2 CPU cores, 2 GB RAM, 40 GB disk. The Free plan gives you 1 such binding.

The Mac mini gathering dust, the retired laptop, the Raspberry Pi 4 - all qualify. The fee is $0, electricity aside.

Then the catch, and it is a big one. Wonder Mesh servers sit behind NAT. Public HTTPS access requires the Gateway feature, and Zeabur's docs are explicit: Gateway requires a dedicated server (non-Wonder Mesh) with K3s installed, and if you don't have one you can purchase one through Zeabur.

In other words: without a separate paid server with a public IP acting as reverse proxy, a Wonder Mesh deployment is reachable only over your LAN.

(Wonder Mesh public access shipped 2026-06-15 as Phase 2, limited to one gateway per workspace and IPv4 only.)

The honest read: Wonder Mesh is a real $0 path, but it suits internal tools, home services and things you use yourself - the Home Assistant end of the spectrum. It is not a free way to put a side project on the public internet.

If that is what you want, though, it is the best value in this entire article: hardware you already own plus a control plane that costs nothing.

Three traps that will waste your time

Trap one: zeabur.com/zh-TW/events/aws-taipei is still live. It advertises AWS Taipei (ap-east-2) with 5-8 ms latency from anywhere in Taiwan and data-residency benefits, and it is very appealing. It also states explicitly that this is a shared cluster offering, invite-code gated. Shared clusters stopped accepting new projects on 2026-03-15. The page carries no date and no deprecation notice - it appears to be stale marketing that contradicts the current product. I could not verify from outside whether invite codes are still being honoured, so the claim here is "this page appears stale and contradicts the deprecation notice," not "this is dead."

Trap two: Zeabur's own docs contradict each other. platform-overview (updated 2026-06-22) says new projects can no longer be created on shared clusters. But quick-start still tells you to "skip this if you're using the shared cluster," build-spec (updated 2026-05-08) still lists "Shared Cluster - Zeabur-managed nodes (AWS / GCP / Tencent Cloud), sized per your plan" as a live runtime target, and the English pricing FAQ still says images deploy to "our shared cluster."

Practical advice: on Zeabur, trust the changelog over the docs. The changelog is dated, attributed and versioned. The docs clearly have not caught up.

Related: Zeabur's Free Plan doc still describes "Auto-sleep - services on the Free Plan automatically sleep after a period of inactivity" and "No credit card required." With Free including no Zeabur compute at all, that reads as leftover text from the shared-cluster era. I could not confirm whether auto-sleep is still enforced on Free-plan services running on your own server, which would be odd behaviour. Treat it as stale documentation rather than asserting either way.

Trap three: you cannot see server prices before logging in. Zeabur's server catalog is behind login. The "$2/month" figure that circulates comes from a December 2024 changelog and is a year and a half old. I could not open the current catalog, so this article quotes no Zeabur server prices. Check them in the dashboard.

Does being a Chinese-language-first platform help?

Partly, but less than you would expect.

What genuinely helps:

• Docs, pricing and support all exist in full Traditional Chinese - unique among the platforms in this article.
• Asia-Pacific coverage is real and unusually broad. The server catalog spans AWS, GCP, Hetzner, Linode, DigitalOcean, Aliyun, Tencent Cloud and Volcengine, filterable by continent, vCPU (1-32) and RAM (2-64 GB), and the site's city dictionary includes Taipei, Tokyo, Osaka, Hong Kong, Seoul, Singapore, Johor, Kuala Lumpur, Jakarta, Bangkok and Manila. For a platform this size that is a genuine edge over Fly, Railway and Render for anyone serving Asian users.

But that city list comes from the site's i18n dictionary. It proves these names can appear in the UI; it does not prove a Taipei-region server is purchasable today, nor at what price. The real catalog is behind login. Do not plan around a Taipei price on this evidence.

What does not help - payment is not localised at all:

• Everything is billed in US dollars. Zeabur's payment docs state that all charges are in USD and that your card issuer may add a foreign transaction fee. The zh-TW pricing page shows $0 / $5 / $19 / $79 in USD too. There is no local-currency pricing anywhere.
• Payments run through Stripe. Accepted: Visa, Mastercard, American Express, JCB and UnionPay, all billed in USD. JCB acceptance is a small plus in markets where it is common.
• PayPal is explicitly not supported, and there are no local payment rails - no LINE Pay, no bank transfer, no convenience-store payment.
• The only card-free route is Alipay, which lets you add credits in CNY, and the zh-TW FAQ frames that answer for users in China. There is no card-free path for anyone else.

In practice: you need a credit card enabled for international transactions. That's it.

Zeabur's zh-TW docs nav does list an invoice / tax-ID page, suggesting local tax-invoice support exists, but the page 404'd or redirected when I tried to read it. Do not plan around it without checking.

---

3. What it actually costs to run the thing yourself

This is the practical core. Every price below was checked on 6 August 2026 on the provider's own page.

2023 versus 2026

2023: a small Rails / Node / Go app with a Postgres cost US$0/month. Fly.io's free allowance alone (3 × shared-cpu-1x 256 MB plus 3 GB of volumes) ran both app and database; Railway gave a recurring monthly credit; Render gave a free web service. The only unavoidable cost was a domain, roughly US$12/year (about US$1/month).

2026, the same app, three honest options:

These are arithmetic composites built from the per-provider prices below, not quotes from any provider. Deliberately excluded because they vary too much: VAT, object storage for user uploads (R2 or B2 free tiers usually cover a side project), transactional email, error tracking and off-site backups. Also excluded, and it is the largest real cost: your own time.

The headline: what was $0 in 2023 is roughly $8-21/month in 2026, and the cheapest honest number for a low-latency app with a real database is about US$10-12 on a single Tokyo VPS you administer yourself.

Important correction: Hetzner raised prices twice in 2026, very unevenly

This section was corrected on 2026-08-07. The full version is a separate article: Hetzner's 2026 price rises, in full.

Every "Hetzner is the cheapest" article is now wrong for anyone who needs Asian latency.

There were two increases in 2026, and this section originally covered only the second.

First: announced 2026-02-23, effective 1 April 2026. Cloud rose 30-36.5% across every line, dedicated 3-21%, and Hetzner's own page states it applied "for both new orders and existing products." Bills went up on machines people already had.

Second: announced 2026-05-27, effective 15 June 2026, 08:00 CEST, for new orders and rescales only. The "Was" column in both tables below therefore already contains the post-April prices.

Orders placed before 15 June keep old pricing, with a trap. Hetzner's billing FAQ states that rescaling (up or down), restoring a deleted server, and transferring a server to a project in a different currency all move a legacy-priced server onto current pricing. Rebuilding, and transfers between same-currency projects, do not. Resize your 2025 CX23 downward to save money and you triple your bill.

Germany / Finland:

Singapore:

See the shape of it. In Germany and Finland, CX and CAX went up about a third this round (+30.2% to +37.6%), while CPX and CCX went up 2.1x to 2.75x. Singapore is not that shape: the same June step there was only +51% to +111%, which the Singapore table above shows on its own. And CX and CAX are EU-only (Falkenstein, Nuremberg, Helsinki).

The conclusion is blunt: Hetzner being cheap is now a Europe-only fact.

• Cheapest European box that runs app plus Postgres: CX23, 2 vCPU / 4 GB / 40 GB SSD at €5.49 (Hetzner's own USD list: $6.49). Add the primary IPv4 at €0.50 ($0.60) and it is €5.99 / US$7.09. 4 GB is comfortable for a small Rails / Node / Go app with Postgres on the same machine.
• Update 2026-08-07: you cannot buy a CX23 right now. Every one of the eight CX and CAX rows on Hetzner's own Cost-Optimized plan page carries a "not available" badge, and /cloud/ labels the whole Cost-Optimized plan "Currently not available." That state appeared between 27 July and 2 August 2026, four days before this article went up, and I missed it. Hetzner has published nothing explaining it. The cheapest orderable Hetzner cloud server in the EU is now CPX12, 1 vCPU / 2 GB / 40 GB at €11.49 (USD list $13.49), €11.99 / US$14.09 with IPv4: double the money for half the RAM.
• Cheapest Asian box: CPX12 in Singapore, 1 vCPU / 2 GB / 40 GB at €15.49 (USD list $17.99), about US$18.60 with IPv4. At the same latency that is more expensive than DigitalOcean, Linode or Vultr, for half the RAM.

All Hetzner figures are list prices excluding VAT and excluding the IPv4 charge. **On tax: Hetzner's VAT table has 36 country entries and Taiwan is not one of them, so Taiwanese buyers fall under "All others" at 0%. For them the ex-VAT list price is the price.

On traffic, this article originally said the figures could not be confirmed. That was wrong, and is corrected here. Hetzner publishes them at docs.hetzner.com/robot/general/traffic/: 20 TB/month included on CX, CAX and CPX in the EU, 20-60 TB on CCX; 0.5-5 TB on CPX in Singapore and 1-8 TB on CCX there; 1-5 TB on CPX in the US. Overage is €1 ($1.20) per TB, metered in 100 MB blocks. Only outgoing traffic is billed; incoming and same-network-zone traffic is free. Dedicated root servers get unlimited traffic on the default 1 Gbit uplink. That page is stamped "Last change on 2024-05-21" and neither 2026 adjustment mentioned traffic, so it predates the current prices.

VPS prices that actually matter for Asia

All from the providers' own pages, checked 2026-08-06.

Vultr Cloud Compute (vc2, shared CPU) - read directly from Vultr's own public plans API:

There is also a 512 MB IPv6-only plan at $2.50 and a 512 MB IPv4 plan at $3.50. High Frequency (vhf, NVMe) starts at 1 vCPU / 1 GB for $6, with 2 vCPU / 4 GB at $24. Prices are identical in every region except São Paulo (+50%).

Asian regions: Tokyo, Osaka, Seoul, Singapore (plus India and Tel Aviv). There is no Taiwan region - confirmed from Vultr's regions API. For users in Taipei, Tokyo is normally the lowest-latency choice.

DigitalOcean Basic droplets:

The nearest datacentre to Taiwan is Singapore (SGP1).

Linode / Akamai (Asia-Pacific): Nanode 1 GB $5 (1 vCPU / 25 GB / 1 TB), Linode 2 GB $12 (1 vCPU / 50 GB / 2 TB), Linode 4 GB $24 (2 vCPU / 80 GB / 4 TB). Prices are uniform across Chennai, Mumbai, Osaka, Singapore, Tokyo, Melbourne and Sydney - no Tokyo or Singapore premium.

Akamai's page labels these "legacy hardware" plans, which suggests new users are being steered to newer families. I did not verify what replaces them or whether they are still orderable. Note also that linode.com/pricing now 301s to akamai.com.

Amazon Lightsail (the one AWS product priced like a VPS): $5 (0.5 GB / 2 vCPU / 20 GB / 1 TB), $7 (1 GB / 2 vCPU / 40 GB / 2 TB), $12 (2 GB / 2 vCPU / 60 GB / 3 TB). Managed databases start at $15/month (1 GB / 1 core / 40 GB), then $30 (2 GB) and $60 (4 GB), with high-availability versions at double.

That $15/month for 1 GB of managed Postgres, against $0 for running it yourself on the same $12 box, is the entire trade-off this article is about.

Contabo looks like the most RAM per dollar: Cloud VPS 4 is 4 vCPU / 8 GB / 100 GB SSD at €5.50/month. But that is the effective rate on a 24-month commitment. I could not confirm the month-to-month price, whether a one-time setup fee applies on shorter terms, or the Singapore/Japan regional surcharge - their configurator did not render usable content. Do not treat €5.50 as a no-commitment monthly price.

Taiwan-local infrastructure: what exists

• AWS Asia Pacific (Taipei), region code ap-east-2, launched 5 June 2025 with three availability zones - the first AWS infrastructure region in Taiwan.
• Google Cloud's asia-east1 is physically in Changhua County, Taiwan.

Neither is a cheap side-project option, and each has a specific blocker:

• Lightsail is not available in ap-east-2. The Lightsail pricing page lists Tokyo, Seoul, Singapore, Hong Kong, Jakarta, Malaysia, Mumbai and Sydney for Asia-Pacific - Taipei is absent. So "run a cheap VPS in AWS Taipei" is not currently a route.
• Google's Always Free e2-micro is restricted to us-west1, us-central1 and us-east1, so asia-east1 is never free.

Chunghwa Telecom's hicloud site still rendered no product or price content after its redirect.

Correction, 2026-08-07: this originally said I had found no Taiwan-based VPS provider whose pricing I could confirm from its own page. That was wrong. 遠振資訊 publishes a full NT$ table at host.com.tw, in a Far EasTone datacentre:

One IPv4 included; extra IPs NT$250/month, two maximum; annual billing gives one month free. At NT$32.21 to US$1 the 4 GB plan is about US$39 and the 16 GB plan about US$156. Against Hetzner Germany with IPv4 - CX23 at US$7.09 and CX43 at US$19.09 - that is 5.5x and 8x. This is one provider's price list, not a survey of the Taiwanese market.

Databases: $0 on the same box, or managed

The biggest saving in the self-managed route is not the machine. It is the database.

Postgres on the same VPS as your app costs $0 extra. For contrast:

Fly's $38/month deserves a pause. The same thing was $0 in 2023, and on your own VPS it is $0 plus your maintenance time. (You can still run an unmanaged Postgres machine on Fly for roughly $3.32 plus volume, which is the cheap path most people actually take.)

Neon's Launch plan dropping its monthly minimum entirely is genuine news that most existing coverage will not have - the old figure was $19/month. Both Neon's pricing page and its docs state that Launch and Scale are pay-only-for-what-you-use with no minimum monthly fee.

Free tiers are covered in section 5.

---

4. The self-managed route, honestly

"Just self-host it" saves money and spends your evenings. This section covers what the tools do and, more usefully, what they explicitly do not do.

All five were confirmed alive via the GitHub API on 2026-08-06: every one had a push within the previous two weeks, none archived. This corner of the ecosystem is healthy.

Dokku - the lowest RAM floor, and the only one that fits a $5 box

Best for: you don't mind a CLI and want the machine bill as low as it goes

MIT, currently v0.38.25. It describes itself as "an extensible, open source Platform as a Service that runs on a single server," deploying via git push, building from Dockerfiles or auto-detecting language with buildpacks, routing via nginx, with databases supplied by plugins.

Requirements: Ubuntu 22.04/24.04 or Debian 11+, amd64 or arm64, and 1 GB RAM minimum for the Docker scheduler (2 GB per node for the k3s scheduler).

That 1 GB floor is by far the lowest of the five, and the only one that genuinely fits a $5-6 box. Concretely: Dokku on a $10 Vultr Tokyo 2 GB instance is comfortable; on a $5 1 GB instance it fits but is tight.

Kamal - the right answer for Rails, and it deliberately does nothing else

Best for: Rails, if you already know Docker

MIT, from 37signals, currently 2.12.0. Tagline: "Deploy web apps anywhere." It gives you zero-downtime deploys, rolling restarts, asset bridging, remote builds and accessory service management, deploying Docker containers onto vanilla Ubuntu servers that have nothing but SSH keys - Kamal installs Docker itself. Its companion kamal-proxy (also MIT, also actively maintained) handles routing and the zero-downtime cutover.

Now the honest part. Kamal has no web UI, no dashboard, no monitoring and no managed database. "Accessories" just means it will run a Postgres container for you; backups, upgrades and tuning are entirely yours.

Kamal's own site says it plainly, and it deserves quoting:

"You're probably still better off with a fully managed service if basic Linux or Docker is still difficult."

That sentence belongs at the bottom of every article that recommends self-hosting.

Coolify - the most complete, but the real RAM requirement is 4 GB

Best for: you want something close to the Heroku experience and will buy one size up for it

Apache-2.0, 60k+ stars, still pushing on 2026-08-06. Self-hosted is free forever with no feature restrictions. Coolify Cloud is $5/month covering 2 servers, plus $3/month per additional server - and note that in the cloud version Coolify runs only the control plane, while your apps still deploy onto servers you pay for. That is exactly the shape Zeabur now has.

Self-hosting minimums: 2 CPU cores, 2 GB RAM, 30 GB free disk. The installer pulls in Docker Engine 24+, sets up /data/coolify and configures SSH keys.

But Coolify's own docs warn that if you run both builds and Coolify on the same server you should monitor resource usage, because high usage could make the server unresponsive.

In plain terms: that 2 GB minimum is a lie of omission for any app you build on-box. Coolify itself eats roughly 1 GB. You are actually shopping for a 4 GB machine, and that is a price-tier jump - $10 to $20 on Vultr Tokyo, $12 to $24 on DigitalOcean.

Budget for that before choosing Coolify.

Dokploy - open source, with an asterisk

Best for: you like Coolify's shape and want to compare

~36k stars, very actively developed, positioned as an "open source alternative to Vercel, Netlify and Heroku."

The licensing needs stating precisely: the repo is Apache-2.0 except anything under a /proprietary directory. That code falls under the "Dokploy Source Available License (DSAL) v1.0," which says it "may only be used in production if you... have agreed to, and are in compliance with, a valid commercial agreement from Dokploy."

So "open source" is true for Dokploy with an asterisk, where Coolify is plain Apache-2.0. If licensing affects your project, that difference matters.

Dokploy Cloud: Hobby $4.50/month per server ($3.60/month billed yearly) for 1 server, 1 org, 1 user and 2 environments; Startup from $15/month for 3 servers, 3 orgs and unlimited users and environments; additional servers $4.50/month each.

CapRover - web UI plus one-click apps

Best for: you want to click a button and have a database, WordPress or monitoring appear

Apache-2.0 (copyright BigZee Ventures LLC), ~15k stars, actively developed. Built on Docker + nginx + Let's Encrypt + Docker Swarm, with a web dashboard, one-click app installs, cluster scaling with automatic load balancing and automatic HTTPS. It advertises about 10 minutes to a first deployment.

CapRover does not publish minimum server requirements on its homepage. Its main architectural difference from Coolify and Dokploy (which use plain Docker) is Docker Swarm, which is effectively in maintenance mode upstream. That does not make CapRover unusable, but it is a long-term factor worth knowing when you choose.

What none of the five do for you

This is my observation from reading all five projects' own docs, not a claim any of them makes:

• OS security updates
• Docker engine upgrades
• Postgres major-version upgrades
• The night the disk fills up
• Off-site backups

None of the five advertises managed backups of your data as part of the free self-hosted product.

So the real accounting for self-hosting is: save $10-30 a month, spend roughly one or two evenings a month, and carry an ongoing risk that if you never set up backups properly, one day the data is simply gone. For a lot of side projects that is a good trade, because those evenings were going into the project anyway. But it is not free. Price it honestly and then decide.

---

5. Where free is still genuinely fine

If the previous four sections left you thinking everything now costs money, that is my failure. Free did not die. It changed shape.

Cloudflare Workers: really free, but a different shape

Free plan: 100,000 requests per day and 10 milliseconds of CPU time per invocation. Workers Paid is $5/month minimum with 10 million requests and 30 million CPU-milliseconds included. Cloudflare charges nothing for egress or bandwidth on Workers.

That 10 ms CPU per invocation is the crucial limit. Cloudflare's free tier is genuinely generous, but it is serverless / edge shaped. It does not give you what Heroku's free tier once did: a long-running always-on process you can SSH into and think about.

So the accurate framing is not "free is back." It is "free survived, but only for a different shape of application."

Cloudflare is also one of the few providers with a public, dated commitment not to remove a free tier (Koyeb is another). At D1's GA on 1 April 2024 it stated it had "no intention of removing the free tier for D1 or reducing the 25 billion row reads included in the $5/mo Workers Paid plan." Free-plan D1 limits are 5 GB total storage, 5 million rows read per day and 100,000 rows written per day.

Static sites

Unchanged. Free static hosting is the one category this round of cuts barely touched. Cloudflare Pages still permits commercial use and does not suspend on inactivity. The previous guide covers the per-provider allowances in full.

A note on where the "unlimited bandwidth" claim actually comes from, since this guide previously questioned it. Cloudflare's Pages limits page lists builds (500/month), files (20,000), file size (25 MiB), custom domains (100) and projects (100), and states no bandwidth or request figure at all - so if that is the page you check, you will not find the claim. The support is on the Workers pricing docs, which state plainly that "Requests to static assets are free and unlimited" and that "There are no additional charges for data transfer (egress) or throughput (bandwidth)." That is a Cloudflare page saying it, so the claim stands. What does have a documented ceiling is anything that runs code: Pages Functions requests count against your Workers quota.

Free managed Postgres: three options that really exist

All usable. All sleep.

The difference that matters is how they wake up: Neon comes back in seconds by itself, while a paused Supabase project needs manual restoration. For a side project nobody is looking at, that gap is large.

Aiven's exact inactivity threshold is not stated in its docs, so treat it as the same failure mode as Supabase, less precisely documented.

Two more worth a line each:

• Turso has the largest free tier of the three: 100 databases, 5 GB storage, 500 million rows read per month, 10 million rows written per month, 3 GB of syncs and 1-day point-in-time restore. But it is SQLite / libSQL, not Postgres, so it is not a drop-in for a Rails app that expects Postgres. Its paid prices (Scaler $24.92, Pro $416.58) have odd cents that look like annual figures divided by twelve - verify the monthly-billed number before committing.
• Prisma Postgres free is 100,000 operations/month, 500 MB storage and 50 databases; Starter is $10/month. It bills per operation, where an operation is a Prisma ORM query - if your app doesn't use the Prisma ORM at all, that billing model is very hard to forecast, so for this audience it is a mention rather than a recommendation.

Oracle Always Free: the only genuinely free always-on VM, with strings

Oracle's Always Free tier still includes:

• Ampere A1 (Arm): 1,500 OCPU-hours + 9,000 GB-hours per month, which Oracle's docs describe as equivalent to 2 OCPUs and 12 GB of memory running continuously
• Up to 2 AMD VM.Standard.E2.1.Micro instances (1/8 OCPU, 1 GB RAM each)
• 200 GB total block volume, 20 GB object storage
• 2 Autonomous Database instances (1 OCPU, 20 GB each)

That is the only "$0, always on, real database" combination in this entire article.

Then the catch, and please read it. Oracle reclaims idle Always Free compute instances. The documented rule: over any 7-day window, if 95th-percentile CPU is under 20%, network under 20% and (on A1 shapes) memory under 20%, the instance is reclaimed.

A quiet side project is exactly that profile.

Two more constraints: Always Free resources must be in your home region, and A1 capacity is famously hard to get in popular regions.

The honest verdict: Oracle Always Free is real, but it is a free resource you have to actively keep alive and can lose at any time. Good for experiments and learning environments. Not good for anything you care about, unless your off-site backups are genuinely solid.

Whether Oracle reduced the A1 allowance from the widely-cited 4 OCPU / 24 GB to today's figure, I could not confirm. The current numbers come from Oracle's docs; I found no archived version of the historical figure, so I will not claim they were cut.

---

6. So what should you do

By situation, with actual answers.

"I just want it online so people can see it"

Render's free tier. It spins down after 15 minutes with no traffic and takes about a minute to wake, which is fine for something nobody uses yet. The free Postgres expires 30 days after creation, so on day 25 either upgrade or pg_dump it out. Put the reminder in your calendar now.

Do not run an uptime pinger to keep it awake. Spun-down time does not consume your 750 hours, and keeping one service awake 24/7 uses roughly 730 of them - the pinger is precisely what gets you suspended for the rest of the month.

"My app needs to be always on, and my users are in Asia"

One Vultr Tokyo instance at $10/month (1 vCPU / 2 GB / 55 GB / 2 TB), running Dokku. Postgres goes on the same box for $0.

If you build Docker images on the machine, or want Coolify's web UI, step up to $20 for 2 vCPU / 4 GB. Same-latency alternatives: DigitalOcean Singapore 2 GB at $12, Linode Tokyo 2 GB at $12.

Do not pick Hetzner Singapore - 1 vCPU / 2 GB is about $18.60, more expensive than all of the above.

"I don't want to touch SSH"

Two routes, and they cost about the same:

• Railway, paid. Genuinely fully managed. But at its posted rates (RAM $10/GB/month, CPU $20/vCPU/month), a 512 MB app plus a 512 MB Postgres is 1 GB of RAM, which is $10/month in RAM alone. The honest number is $10-20, not the $5 on the Hobby plan.
• Zeabur / Coolify Cloud / Dokploy Cloud plus a server you buy yourself. $0-5 subscription plus the machine. You get the web UI and the automation; the machine bill is between you and the cloud provider.

The only real difference in the second route is who is responsible when the machine breaks - you are. In the first route it's Railway. Whether that is worth a few dollars is your call.

"I care about latency and want a Taiwan region"

There is no cheap answer today. AWS Taipei (ap-east-2, live since 2025-06-05) and GCP asia-east1 (Changhua) really are in Taiwan, but neither is priced for side projects, and Lightsail does not support ap-east-2, so even "open a cheap VPS in AWS Taipei" is not available.

The practical answer is Tokyo. Vultr and Linode both have Tokyo regions with no premium.

"I have an idle Mac mini at home"

Zeabur Wonder Mesh, $0. Two CPU cores, 2 GB RAM and 40 GB of disk qualify, and the free plan gives you one binding.

Remember the limit: without a separate paid server with a public IP acting as Gateway, that machine is reachable only over your LAN. This route is for internal tools and things you use yourself, not for a public launch.

"My thing is a static site or an edge API"

Cloudflare. This part did not get worse - in places it got better. Don't let the previous five sections convince you that everything now costs money.

---

What this article could not confirm

Listed here because a confidently wrong price sends you down a wasted evening, which is worse than an omission:

• Zeabur's actual server prices. The catalog is behind login. The "$2/month" figure circulating online comes from a December 2024 changelog and is not current.
• Whether Zeabur supports local tax invoices in Taiwan. The zh-TW docs nav lists a page for it, but it would not load.
• Whether the $5 shared-cluster credit in Zeabur's Dev plan was actually removed. As of February 2026 it was only "under consideration."
• Whether Free-plan auto-sleep is still enforced on Zeabur. The doc text reads as a leftover.
• Whether servers Zeabur sells you count against the "manageable server" quota. The wording strongly implies not, but nothing states it.
• Render's exact paid monthly prices. Their pricing page would not render figures on repeated attempts. The instance-type ladder is confirmed; the dollar amounts are not.
• Hetzner's included traffic per plan and overage rate. Found, corrected 2026-08-07: they are published at docs.hetzner.com/robot/general/traffic/. See section 3.
• Contabo's month-to-month price. The €5.50 shown is labelled "for the first 24 months"; the page does not state what it becomes afterwards.
• Turso's monthly-billed paid prices. The odd cents suggest annual divided by twelve.
• Chunghwa Telecom hicloud pricing. Their pages rendered no usable content. (For other Taiwanese providers see the correction in section 3 - 遠振 publishes a price list.)
• Why Hetzner's CX and CAX lines became unorderable in early August. Hetzner has published no news post, blog post or status entry about it, and its own price feed still marks those products active and returns prices. So the honest statement is "Hetzner's own pages say you cannot order them right now," not that the line was discontinued.
• That developers actually moved back to self-managed servers. I can date the decline of free PaaS precisely. I found no primary source - no provider-published migration data, no survey - showing that hobby developers moved back to cheap self-managed servers as a result. Sections 3 through 6 are my reading of the prices and tooling, not a documented trend.

---

One sentence

Zeabur said something honest: its servers run on top of AWS, GCP, Linode and Hetzner, so its pricing can never be lower than theirs. That is true of every PaaS - the others just didn't say it out loud. Free tiers are marketing budget, and marketing budget does not subsidise your compute forever.

The $0 answer from 2023 costs about US$10/month in 2026. What you get for it is a machine that doesn't sleep and a database that doesn't get deleted. That is a fair price - as long as money isn't the only thing you count.

---

Share Your Deployed Project

Successfully deployed your side project? Share it with the builder community and connect with other developers who understand the journey from idea to launch.

---

Every price, plan detail and date here was verified on 6 August 2026 directly against each provider's own pricing page, docs or changelog, with no third-party roundups used as a source. Passages marked "Update / Correction 2026-08-07" were changed the next day after re-reading Hetzner's primary sources; the full version is here. Prices move - especially Hetzner's two 2026 increases and Zeabur's April plan restructure, all only months old. Before you put in a card, spend two minutes on the provider's own pricing page.