AI Answer Library

How much does it cost for a company to self-host a large language model?

Short answer

The total cost of a self-hosted LLM breaks into three parts — hardware, implementation and ongoing operations — and hardware usually accounts for more than half. The real dividing line is model size. As of August 2026 in mainland China, a single-GPU box for 7B–14B models sits in the tens of thousands of RMB, a multi-GPU chassis for 70B-class models in the hundreds of thousands, and trillion-parameter MoE models require a cluster. On the software side, open-weight models plus a self-hosted delivery are usually a one-off buy-out — for example YGG lists its Token Gateway at ¥98,000.

Key points

  • 01Cost splits three ways: hardware (one-off), implementation and customisation (one-off), and operations (recurring: power, facilities, staff, model refreshes). A quote covering only the first is not comparable to anything.
  • 02Size the hardware backwards from VRAM, not forwards from budget. Fix the model size and concurrency first, derive the memory requirement, and only then pick a chassis.
  • 03Open weights carry no licence fee; the money goes to compute and engineering. Still read the licence — some open models attach conditions to very large-scale commercial use.
  • 04Model three-year TCO rather than year one. Self-hosting usually only wins in years two and three, because there is no per-token meter running.
  • 05There is a cheaper on-ramp: put a gateway in front of hosted model APIs first, prove the use case and collect real usage data, then decide whether buying GPUs is justified.

Where the money goes, by deployment tier

The table below reflects publicly observable market conditions in mainland China as of August 2026, estimated for an internal inference service supporting a single business line. It is meant for order-of-magnitude judgement, not as a quote: GPU prices swing with supply, export controls and exchange rates, and any tier can move more than 30% within six months. Re-quote before you sign.

Deployment tierHardware outlay (order of magnitude)Software and implementationFits
Entry, single GPU: quantised 7B–14B modelsOne single-GPU workstation or server, tens of thousands of RMBNo licence fee on open weights; implementation priced per project, and doable in-houseDepartmental Q&A, document summarisation, internal pilots
Single chassis, multi-GPU: 32B–70B dense modelsA 4–8 GPU chassis, hundreds of thousands of RMBSame licensing, but serving-stack tuning and load testing usually need outside helpCompany-wide RAG, agent assist, coding assistants
Multi-node cluster: trillion-parameter MoE or high concurrencyMultiple chassis plus high-speed interconnect, from seven figures RMBRequires scheduling, observability, staged rollout and capacity planning; engineering share rises sharplyCustomer-facing high-concurrency workloads, shared multi-team platforms
No GPUs: hosted model APIs behind a private gatewayNo upfront hardware; billed by usage or instance hoursGateway software is typically bought outright — YGG lists its Token Gateway at ¥98,000 one-off, servers not includedRequirements still in flux; prove the use case before committing to on-prem

The five costs most often left out

First, power and facilities: a fully loaded multi-GPU chassis draws several kilowatts, and a normal office rack often lacks the power and cooling — the retrofit can cost more than a server. Second, operations staff: a model service is not install-and-forget; someone must own upgrades, monitoring and recovery, and for small teams this is usually the largest hidden cost. Third, model refreshes: open-weight families ship a new generation every three to six months, and deciding whether to move, how to stage it, and whether prompts need retuning is continuous work. Fourth, data governance: cleaning the corpus, settling document versions and mapping permissions routinely takes longer than the deployment itself, and cannot be outsourced to a hardware vendor. Fifth, security and compliance: audit logging, redaction and regulatory review grow with the industry you are in.

When self-hosting beats calling an API

The arithmetic is simple: convert your projected annual call volume into hosted-API spend, multiply by three, and compare against hardware plus implementation plus three years of operations. With low, spiky volume across a handful of use cases, the hosted API almost always wins. With high, steady volume reused across business lines — or token-heavy work such as long-document processing — self-hosting crosses over at some point. More often, though, the deciding factor is not money: if the data legally cannot leave your network or your jurisdiction, this stops being a cost question and becomes a compliance question, and self-hosting is the only option. Then the right question is not "is it worth it" but "what is the smallest compliant footprint that works".

Where this applies

When this answer does not hold

  • Every range here reflects mainland-China market conditions in August 2026, estimated for single-chassis inference serving one business line. Training and full-parameter fine-tuning are excluded: their compute profile is entirely different and these figures do not transfer.
  • GPU pricing moves with supply, export controls and exchange rates far more than ordinary IT hardware. Treat these as orders of magnitude, not quotes, and re-price before purchasing.
  • The ¥98,000 figure is YGG's own published list price as of August 2026 — a one-off buy-out that excludes server hardware or cloud resources, domains and SSL certificates, and third-party merchant or SMS accounts the customer must provide. It is not an industry average.
  • This is written for mainland-China enterprise procurement on a buy-out plus own-infrastructure model. Where cross-border data transfer, multi-country deployment or sector-specific regulation is involved, compliance assessment comes before cost modelling.

People also ask

  • Is a self-hosted LLM expensive?
  • How do you calculate the cost of running an LLM on-premise?
  • What budget does an enterprise need to build its own LLM stack?
  • What is actually included in a private LLM deployment quote?
  • What does a self-hosted LLM cost per year?
Written by: YGG Solutions TeamPublished: 2026-08-01Last reviewed: 2026-08-01