Short answer. For most teams, a managed LLM API is cheaper than owning GPUs until volume reaches several billion tokens a month. Plugsky Builder at $60 and Scale at $120 remove per-token billing but keep fair-use concurrency limits. Rent GPUs when you need custom weights, version pinning or physical control, and expect to pay for idle capacity.
There are three ways to host an LLM. A managed API sells tokens and hides the infrastructure. A serverless inference platform runs an open model on shared capacity and bills by token or second. A GPU rental gives you the machine and leaves the serving stack to you. Each has a different cost curve, and the curves cross at a volume most teams can estimate.
This is a 2027 planning guide with evidence checked on 20 November 2026. It does not claim 2027 benchmarks or tests. The arithmetic uses published list rates and states every assumption, including the ones that make GPU ownership look better than it usually is.
How we evaluated this
We weighted five criteria for a team that expects production traffic, not a demo.
- Unit cost at realistic utilization, 25%. Token price at low utilization; GPU price per hour at 100%, then adjusted for the utilization you actually expect.
- Operational burden, 20%. Who owns serving, autoscaling, cold starts, model updates, monitoring and incident response.
- Model control, 20%. Whether the buyer can pin a version, run custom weights or choose an open model without a provider roadmap.
- Scalability, 20%. Burst handling, concurrency, regional capacity and the process for adding throughput.
- Compliance and isolation, 15%. Data-processing terms, private deployment and the ability to keep traffic inside a boundary.
We did not measure tokens per second, time to first token, cold-start time or GPU utilization. We did not rent a GPU or run a load test. We did not verify capacity availability in any region. The cost comparison is arithmetic on published rates, not a performance evaluation.
Eight hosting options compared
Rates are USD. Token prices are per million tokens; GPU prices are per hour unless stated. "Dedicated GPU" means the buyer can reserve a machine, not that every plan includes one.
| Option | Hosting model | Published rate | Billing granularity | Dedicated GPU | Operational burden | Best fit |
|---|---|---|---|---|---|---|
| Plugsky | Managed API, OpenAI-compatible | Builder $60/month; Scale $120/month; no per-token overage | Monthly plan | Private deployment by contract | Low: provider serves the models | Flat-cost production without GPU ops |
| Together AI | Serverless inference, batch, dedicated endpoints | DeepSeek V4.1 Flash $0.30/$1.20; GLM-5.3 $1.40/$4.40 | Per token | GPU clusters and dedicated endpoints | Low to medium | Open-weight breadth and capacity |
| Fireworks AI | Serverless, reserved throughput, private cloud | From $0.05 per 1M input tokens; DeepSeek V4.1 Flash $0.30/$1.20 | Per token; reservation per minute | Reserved throughput and sovereign private preview | Low to medium | Priority and reserved production capacity |
| Groq | Managed inference on LPU/LPX capacity | GPT-OSS 120B $0.15/$0.60 | Per token | No public GPU rental on the pricing page | Low | Latency-sensitive open models |
| Modal | Serverless GPU containers | H100 SXM $0.001097/second; B200 $0.001736/second | Per second | Yes | Medium to high: you build and operate the serving app | Custom code and bursty GPU work |
| RunPod | GPU rental, secure and community cloud | H100 SXM $4.39/hour secure; from $2.69/hour community | Per second | Yes | High: you own the stack | Lowest raw GPU cost with ops capacity |
| Replicate | Model hosting with per-second billing | H200 $5.49/hour; 2x H100 $10.98/hour | Per second | Yes, including multi-GPU | Low to medium | Multi-GPU jobs and a model library |
| AWS Bedrock | Managed serverless and provisioned throughput | Per-token; provisioned throughput published | Per token or committed throughput | No direct GPU rental through Bedrock | Low | AWS governance and procurement |
Flat-rate is not the same as self-hosted. Plugsky Builder and Scale remove the monthly token cap and per-token overage, but they are managed plans. Fair-use concurrency and RPM limits apply: 10 concurrent and 600 RPM on Builder, 16 and 1,200 on Scale. A buyer who needs a dedicated GPU, custom weights or a specific CUDA stack is looking at a different category.
Cost math: when does a GPU beat an API?
Start with a 75/25 input-to-output token mix and a 30-day month. At 100 million tokens per month, the managed API costs are small: Together or Fireworks with DeepSeek V4.1 Flash is $22.50 input plus $30 output, or $52.50. Groq GPT-OSS 120B is $11.25 plus $15, or $26.25. Plugsky Builder is $60. AWS Bedrock with Claude Sonnet 5.5 at $2/$10 is $150 plus $250, or $400.
Now price one H100 SXM. RunPod's published secure-cloud rate is $4.39 per hour. For 730 hours, that is $3,204.70 before storage, networking, redundancy, monitoring and the engineer who operates it. Modal's H100 SXM rate of $0.001097 per second is $2,883.52 for the same 730 hours. Replicate's H200 at $5.49 per hour is $4,007.70. A 2x H100 Replicate job at $10.98 per hour is $8,015.40.
| Option | 100M tokens | 1B tokens | 6.1B tokens |
|---|---|---|---|
| Plugsky Builder | $60 | $60 | $60 |
| Together DeepSeek V4.1 Flash | $52.50 | $525 | $3,202.50 |
| Groq GPT-OSS 120B | $26.25 | $262.50 | $1,601.25 |
| Bedrock Claude Sonnet 5.5 | $400 | $4,000 | $24,400 |
| One H100, RunPod secure | $3,204.70 | $3,204.70 | $3,204.70 |
The break-even between one H100 at $3,204.70 and DeepSeek V4.1 Flash at $0.30/$1.20 is about 6.1 billion tokens per month at a 75/25 mix, before operations. At 1 billion tokens, the API costs $525 and the GPU costs six times more. At 6.1 billion, they are roughly equal on raw infrastructure; add an engineer, storage and redundancy, and the GPU needs to be busier still to win.
The comparison assumes 100% GPU utilization. At 30% utilization, the effective H100 cost triples to about $9,614 per month for the workload actually served. That is the most common mistake in self-hosting business cases: the hourly rate is not the monthly cost if the GPU idles.
When a competitor is the better choice
Together AI is the better choice when the team wants a broad open-weight catalogue with serverless, batch and dedicated options, and does not want to run Kubernetes. Fireworks AI wins when reserved throughput, priority modes and a 99.9% uptime commitment matter more than the lowest token price. Groq wins when latency is the product feature and an open-weight model fits; its published GPT-OSS rates are among the lowest in this comparison.
Modal is the better choice for custom Python workloads that need a GPU for seconds at a time, with per-second billing and no cluster to manage. RunPod wins when the team has the ops capacity to run its own serving stack and wants the lowest published H100 hourly rate. Replicate wins for multi-GPU jobs and a library of ready-made models. AWS Bedrock wins when the organization already requires AWS procurement, logging and governance. Plugsky wins when the workload fits a managed API and the buyer wants a fixed monthly number without GPU operations.
What we did not verify
We did not verify GPU availability, queue times or capacity in any region. We did not measure tokens per second, model load time, cold starts or concurrency behaviour. We did not test whether a model served by two providers produces the same output or uses the same tokenizer. We did not verify storage, egress, support or committed-use pricing. We did not test 2027 hardware, because it is not available.
For Plugsky, the plans remove the monthly token cap and per-token overage, not the fair-use concurrency and RPM limits. A SOC 2 Type II report is available under NDA; confirm the current scope and audit period. Request the current report under NDA. HIPAA/BAA is available by contract and subject to review. The 99.9% figure is an uptime target, and the contract schedule governs.
Frequently asked questions
Is it cheaper to host an LLM myself?
Only at sustained high utilization or when you need custom weights. One H100 at $4.39 per hour costs $3,204.70 for 730 hours, before storage and operations. At a 75/25 token mix, that is roughly the raw cost of 6.1 billion tokens on DeepSeek V4.1 Flash.
What does Plugsky mean by unlimited?
No monthly token cap and no per-token overage. It does not mean infinite capacity. Fair-use concurrency and RPM limits apply: Builder includes 10 concurrent requests and 600 RPM, and Scale includes 16 and 1,200.
How much does an H100 cost per month?
At RunPod's published secure-cloud rate of $4.39 per hour, 730 hours cost $3,204.70. At Modal's $0.001097 per second, 730 hours cost $2,883.52. Neither includes storage, networking, redundancy or staff.
When does a dedicated GPU make sense?
When utilization is high, the model needs custom weights or a pinned version, the data cannot leave a controlled environment, or the workload needs a GPU shape that managed APIs do not offer. It rarely makes sense for a few hundred million tokens a month.
Do managed APIs support custom fine-tuned models?
Some do, but support varies. Plugsky publishes a catalogue of 39 model IDs and private deployment by contract; custom weights are a contract discussion, not a self-serve feature. Confirm the exact scope before planning around it.
Is serverless inference cheaper than a reserved GPU?
At low or bursty volume, serverless is usually cheaper because you pay only for what you use. At steady high volume, reserved capacity can be cheaper per token. The crossover depends on the provider, model and utilization.
What is excluded from the break-even calculation?
It excludes storage, egress, redundancy, monitoring, engineering time, support contracts, taxes and idle capacity. It also assumes a 75/25 input-to-output mix and list prices with no cache or batch discount.
Sources
- Together AI pricing — checked 20 November 2026.
- Fireworks AI pricing — checked 20 November 2026.
- Modal GPU pricing — checked 20 November 2026.
- RunPod GPU pricing — checked 20 November 2026.
- Replicate pricing — checked 20 November 2026.
Skip the GPU operations
Plugsky Builder is $60 per month with no monthly token cap and no per-token overage. Check concurrency and RPM against your peak before you commit.
Related reading: the model catalogue, LLM load balancing, unlimited API access, AI API pricing questions, compare Plugsky.