Home Lab vs Cloud: When Does Self-Hosting Make Sense?

The “home lab vs cloud” debate usually gets framed as a binary choice. It shouldn’t be. The real question is where the break-even point falls for each workload you run.

Some things are obviously cheaper to self-host. Some are obviously cheaper in the cloud. The interesting decisions sit in the middle, and that is where people get the math wrong because they forget electricity, replacement drives or their own time.

This guide works through three common workloads: storage, compute and AI inference. Cloud prices come from each provider’s published price list on the date shown in the sources. Hardware prices are planning ranges, and electricity uses $0.18 per kWh, close to the US residential average in EIA’s latest data. Then it covers what cost comparisons miss: privacy, uptime and control.

Storage: NAS vs Cloud

Storage is where self-hosting wins most clearly, but only if you keep the data for years. Drive prices move a lot, so rerun this math with the prices on the day you buy.

The Cloud Side

B2 is straightforward to compare with a NAS: pay-as-you-go object storage with a simple price. AWS S3 and Google Cloud Storage price storage by region and storage class and bill internet egress beyond a small free allowance (AWS includes 100GB a month across its services), so use their calculators if you are already on one of them.

The Home Lab Side

A Synology DS925+ with four Seagate IronWolf 8TB drives in RAID 5 gives you 24TB usable:

Running costs:

• Electricity: Synology rates the DS925+ at 37.91W in access with four drives installed and 12.33W in HDD hibernation. At about 38W around the clock, that is roughly $5 a month.
• Drive replacements: if you budget one drive failure every four to five years at the drive prices above, set aside about $6 to $8 a month.

Monthly operating cost: about $11 to $13 for 24TB.

The Break-Even

Against Backblaze B2 with 16TB stored (about $111 a month), the NAS saves roughly $98 to $100 a month and pays for itself in about 20 to 25 months. Filled to 24TB (about $167 a month on B2), it pays back in about 13 to 16 months. With only 8TB stored (about $56 a month on B2), payback stretches to roughly four to five years.

The gap widens the longer you keep the data. The DS925+ has four bays, so you grow it by replacing drives with larger ones one at a time, or by adding Synology’s DX525 expansion unit. On B2, every extra 8TB costs about $56 a month for as long as you store it.

Verdict: self-host bulk storage you will keep for years. If you store only a few terabytes, a cloud or consumer plan is cheaper and simpler.

Compute: Mini PC vs Cloud VM

Compute is more nuanced. The break-even depends on utilization: how many hours a month the machine actually works.

The Cloud Side

Published on-demand prices (sources below):

Hetzner’s prices are for its EU locations after its June 2026 price adjustment; its US-location plans cost more.

The Home Lab Side

A Beelink Mini S13 (N150) costs about $350 and idles at an estimated 6 to 10W. It matches a t3.xlarge on memory (16GB) and core count, and its four cores have no burst-credit limit like AWS’s T-series instances.

Electricity at 6 to 10W is roughly $0.80 to $1.30 a month. A share of a UPS adds a little more.

The Break-Even

Against a Hetzner CX43 with the same 16GB of memory, the S13 saves about $17 to $18 a month after electricity and breaks even in about 20 months. Against the cheaper CX33 it takes about 38 to 40 months. Against AWS on-demand it pays back in about 12 months versus a t3.medium, or about 3 months versus a t3.xlarge.

This calculation only applies to always-on workloads. If you need a large VM for three hours a week of batch jobs, the cloud wins. A t3.xlarge at about $0.17 an hour costs roughly $2 a month for that, against a few hundred dollars upfront for a machine that sits idle most of the week.

Verdict: self-host always-on services (DNS, reverse proxy, Home Assistant, media server). Use cloud for bursty, occasional workloads. For hardware choices, see our best mini PC for home server guide.

AI Inference: Local GPU vs API Calls

This comparison moves fastest. Hosted model prices change often, and so do the open models you can run at home.

The API Side

Google’s Gemini API price list (updated September 24, 2026) is a useful reference because it spans cheap and premium tiers. Prices are per million tokens:

A typical query of 500 input tokens and 500 output tokens costs a fraction of a cent on Flash-Lite and about $0.007 on the Pro tier. At 1,000 queries a day, that is about $7.50 a month on Flash-Lite, $42 on Flash and $210 on the Pro tier.

The Home Lab Side

A new RTX 5060 Ti 16GB (16GB of GDDR7, 180W board power per NVIDIA) in a basic tower runs 14B models at up to Q6_K. Street prices for the card vary, so price the whole tower on the day you buy.

In our judgment, the models that fit in 16GB compete with the low-cost hosted tiers, not with frontier models. The largest open models are far bigger: Llama 3.1 70B is a 43GB download at 4-bit in Ollama, more memory than any single consumer GPU holds, so running that class at home means a unified-memory system or several cards.

Assume the box draws 250W at the wall for 8 hours a day of inference and idles at 60W the rest of the time (planning assumptions; measure yours). That is about 89 kWh a month, or about $16.

The Break-Even

Take a heavy workload of 3,000 queries a day at 500 tokens in and 500 out:

• Gemini 2.5 Flash-Lite: about $22.50 a month
• Gemini 2.5 Flash: about $126 a month
• Gemini 3.1 Pro Preview: about $630 a month
• Local 16GB box: about $16 a month in electricity, plus the hardware

Against Flash-Lite, the local box saves about $6.50 a month, so the hardware never pays for itself on cost alone. Against Flash, each $1,000 of hardware pays back in about nine months, but only if a 14B-class local model does your job as well. The Pro tier makes the math look fast, but a 16GB local model is not a substitute for it.

At lighter use the case is weaker still. A box idling around the clock at 60W costs about $8 a month, more than 1,000 daily queries cost on Flash-Lite.

Verdict: use APIs for light and moderate usage. Self-host for private data, predictable fixed costs, offline use or unlimited experimentation. See our AI home lab on a budget guide for the hardware, and best GPU for local LLMs for card choices.

Beyond Cost: Privacy and Control

Cost comparisons miss the factors that push many home labbers to self-host in the first place.

Data sovereignty. Your NAS has no terms of service. Medical records, legal documents, family photos and security camera footage are all data you may prefer to keep on hardware you control.

Vendor lock-in. Leaving a cloud is less expensive than it used to be: AWS waives internet egress charges for customers moving their data off AWS, through a request to AWS Support with 90 days to complete the move. Everyday egress beyond the free allowance is still billed. With a home NAS, moving data costs nothing but time.

Service continuity. Google alone has shut down many consumer services over the years. Cloud providers change pricing, deprecate APIs and retire features on their timeline, not yours. Self-hosted services keep running as long as you maintain them.

Customization. Cloud services give you their interface. Self-hosted services give you config files. If you want Plex to transcode a specific way, DNS to block specific domains or Nextcloud to fit a particular workflow, you can do it without waiting on a feature request.

Beyond Cost: Uptime and Reliability

This is where the cloud genuinely wins.

Cloud uptime is better. Providers replicate data across facilities and publish availability commitments. Your home lab depends on one ISP line and one power feed. A single ISP outage, power flicker or drive failure takes everything offline. A good UPS handles power blips, but not a four-hour ISP outage.

Cloud handles disasters. If your house floods, your NAS is gone. Cloud data survives because it is stored across multiple facilities. Self-hosters need off-site backups to match that, and many never set them up.

Cloud scales without planning. Need ten times the storage tomorrow? Change a setting. On a NAS, you are buying drives and waiting for shipping.

If you need guaranteed uptime for anything public (a website or a business-critical API), use the cloud. Self-hosting suits services where a few hours of downtime is annoying but not catastrophic.

The Hybrid Approach: What to Self-Host, What to Put in the Cloud

Most experienced home labbers run a hybrid. Here is a practical split.

Self-Host These

• File storage and NAS: large long-term savings, full privacy, no egress fees
• Media server (Plex or Jellyfin): a one-time hardware cost in place of several streaming subscriptions, and you control the library
• DNS filtering (Pi-hole or AdGuard Home): trivial to run, no reason to pay for it
• Home automation (Home Assistant): it needs local network access anyway
• Local AI inference: when you process private data or want unlimited experimentation
• Development and test environments: spin them up and tear them down without watching a billing dashboard

Keep in the Cloud

• Off-site backups: Backblaze B2 at $6.95 per TB per month is cheap insurance against a house-level disaster
• Public websites and APIs: uptime matters, and cloud CDNs are hard to beat
• Email: self-hosted email is a maintenance and deliverability burden; use a provider such as Fastmail or Google Workspace
• Burst compute: temporary heavy workloads that would otherwise mean buying hardware you rarely use
• Collaboration tools: Google Workspace and Microsoft 365 are mature; self-hosted alternatives require more administration

The Backup Rule

Even if you self-host everything, keep at least one cloud copy. The 3-2-1 rule applies: three copies of your data, on two different media, one off-site. Backblaze B2 or an encrypted rclone sync to a low-cost provider covers the off-site copy for a few dollars a month per terabyte.

Common Mistakes

Ignoring electricity costs. A used enterprise server pulling 300W costs about $39 a month at $0.18 per kWh, or about $1,420 over three years, often more than the server itself cost. See our home lab power consumption guide for how to estimate and measure your own draw.

Comparing peak specs instead of actual needs. A 64-core cloud instance costs thousands a month, but you probably don’t need 64 cores. Compare cloud pricing against what you would actually provision.

Forgetting your time has value. Debugging a broken Proxmox cluster at midnight has a cost, even if it never shows up on a bill. If you self-host only to save money and don’t enjoy the maintenance, the savings may not be worth it.

Not budgeting for replacements. Hard drives fail and power supplies die. We suggest setting aside 10 to 15 percent of your hardware cost each year for replacements and upgrades. Cloud pricing already includes this; your math should too.

Skipping the UPS. A sine-wave tower UPS like the CyberPower CP1500PFCLCD protects equipment from power events and gives you time to shut down cleanly. Skipping it is a false economy once a NAS or database is involved. See the home lab UPS guide before you finalize the budget.

Wrap-Up

The break-even math favors self-hosting for always-on storage (about two years at 16TB against B2 in our example, faster as you fill the NAS) and always-on compute (roughly 12 to 40 months against the budget VMs in our example, and about 3 months against a same-memory AWS t3.xlarge). For AI inference, a local box rarely pays for itself against low-cost API tiers; choose it for privacy and control.

Cloud wins for bursty workloads, public-facing services, off-site backups, and anything where guaranteed uptime matters more than cost.

The practical move is to start small. A Beelink Mini S13 can replace a small VPS and tells you whether self-hosting fits your workflow. If it does, expand with a NAS and dedicated compute. If it doesn’t, you are out about $350 instead of the cost of a full lab.

For a concrete shopping list, see our home lab under $500 guide.