Key facts
| Managed model | Flat monthly plans with unlimited fair-use usage; no per-token billing on self-serve |
| Own-stack costs | GPU capex or rental, serving software, autoscaling, monitoring and on-call |
| Time to first call | Managed: minutes with an OpenAI-compatible endpoint |
| Model access | 30+ models under one managed API |
| Deployment options | Plugsky cloud, VPC, on-prem and air-gapped for regulated teams |
| Utilisation risk | Self-hosting pays off only when GPUs stay busy |
| Free plan | 2 free models (plugsky-micro and plugsky-lite), no card required |
| Product status | Chat, streaming, function calling, embeddings and RAG are live |
TL;DR
- Managed wins on time to value, model breadth and zero GPU operations.
- Self-hosting wins on absolute data control and high, steady utilisation.
- Most self-hosting proposals undercount on-call and serving-stack maintenance.
- A hybrid split — managed for most workloads, self-hosted for sensitive ones — is common.
- Plugsky adds VPC, on-prem and air-gapped options when residency demands it.
How it works, step by step
- List workloads and classify data sensitivity and latency requirements.
- Estimate managed cost from measured tokens and current published rates.
- Estimate self-hosted cost: GPUs, serving stack, autoscaling, monitoring, on-call.
- Model utilisation — what percentage of the month will your GPUs be busy?
- Compare time to market: weeks of platform work versus an afternoon of integration.
- Run a pilot on the managed API to validate quality and latency first.
- Adopt a hybrid split if a subset of workloads genuinely requires self-hosting.
Try it yourself
Open the self-hosting break-even calculator →
What building your own stack really involves
Self-hosting is not one line item. A production inference stack includes GPUs or reserved instances, a serving runtime, model weights and updates, autoscaling and queueing, health checks, observability, security patching, capacity planning and on-call rotation. Each layer has failure modes that land on your team at 3am.
The hardware is often the cheapest part. Engineering time, idle GPU hours and the opportunity cost of platform work are what decide the economics.
Where a managed API wins
- Time to value. An OpenAI-compatible endpoint is live in minutes, not quarters.
- Model breadth. 30+ models behind one API, so routing decisions stay in code.
- No idle hardware. Flat-rate plans with unlimited fair-use usage mean cost tracks users, not utilisation.
- Operational surface. No serving stack, no GPU fleet, no on-call for inference.
- Residency without hardware. VPC, on-prem and air-gapped deployment options exist when regulation demands them.
Where self-hosting wins
- Absolute data control. When weights and prompts may never leave your infrastructure, self-hosting is the only answer — though sovereign managed deployments cover many of these cases.
- High, steady utilisation. GPUs that run near capacity for most of the month can beat metered pricing.
- Customisation. Fine-tuned or modified weights that no provider hosts.
- Latency engineering. When you need to co-locate inference with application logic at the edge.
If two or more of those do not apply, the managed route is usually cheaper once operations are counted honestly.
The hybrid pattern
Most organisations converge on a split: managed APIs for general product workloads, and self-hosted or sovereign deployments for a small set of sensitive or specialised models. The managed side provides breadth and elasticity; the self-hosted side satisfies data rules.
Plugsky supports that pattern directly with deployment options ranging from cloud to VPC, on-prem and air-gapped, all behind the same OpenAI-compatible API. Start managed, measure real usage, and move only the workloads that justify the operational cost.
Honest comparison
| Factor | Managed Plugsky API | Self-built inference stack | Hybrid |
|---|---|---|---|
| Time to first call | Minutes | Weeks to months | Fast start, staged build |
| Cost shape | Flat monthly plan, no per-token billing on self-serve | GPU capex plus operations | Mixed fixed and capex |
| Model breadth | 30+ models, one API | Only what you host | Host the few you need |
| Scaling | Fair-use limits, enterprise committed capacity | You size and autoscale GPUs | Burst to managed |
| Operations | Provider-run | Serving stack, monitoring, on-call | Smaller on-call surface |
| Data control | Region choice, VPC, on-prem, air-gapped | Absolute | Absolute for sensitive workloads |
Frequently asked questions
Is a managed AI API cheaper than self-hosting?
Usually yes once operations are counted. Self-hosting only wins with high, steady GPU utilisation or absolute data-control requirements. Compare TCO including engineering time, not just hardware.
What does building an inference stack cost?
GPUs or reserved instances, serving software, autoscaling, monitoring, security patching and on-call. The hidden cost is engineering time and idle capacity, which often exceeds the hardware line.
Can I get data residency without self-hosting?
Yes. Plugsky offers region selection plus VPC, on-prem and air-gapped deployment options, so regulated teams can keep data in approved environments without operating their own serving stack.
How long does it take to integrate a managed API?
Usually an afternoon: change the base URL and model name with your existing OpenAI SDK, then run your tests. The free plan includes two models with no card.
When does self-hosting make sense?
When weights or prompts may never leave your infrastructure, when utilisation is high and steady, when you run custom fine-tuned models, or when latency requires co-location.
Does Plugsky charge per token?
No. Self-serve plans are flat monthly with unlimited fair-use usage, and there are no per-token charges or overage fees. See the live pricing page for current plans.
Can I start managed and self-host later?
Yes, and that is the common path. The API stays OpenAI-compatible, so workloads can move between cloud, VPC, on-prem and air-gapped deployments without rewriting applications.
What utilisation makes self-hosting break even?
It depends on hardware and operations cost, so run the numbers with your own quotes. The break-even calculator models GPU cost, utilisation and operations against measured API usage.