OpenAI models, Azure terms
Microsoft Foundry
SFT, DPO and reinforcement fine-tuning on the gpt-4.1 family and gpt-4o under your Azure agreement — the surviving route now that OpenAI\u2019s own platform is closing.
Guide · Enterprise fine-tuning
Enterprise LLM Fine-Tuning Services (2026)
Enterprise fine-tuning is a different purchase from developer fine-tuning. The training run is the easy part; the real questions are whose boundary the data trains inside, who owns the result, and who grades the model\u2019s homework. This guide maps the routes and the questions, in the order a procurement team will ask them.
The four enterprise routes
Gemini, Google Cloud terms
Vertex AI
Supervised, preference and reinforcement tuning of Gemini, with tuned endpoints billed at base-model inference prices and Google Cloud\u2019s data-processing terms.
Nova & open models, AWS terms
Amazon Bedrock
Fine-tuning, continued pre-training and reinforcement fine-tuning inside the AWS account and IAM model your security team already audits.
Managed end-to-end
Scale AI
When the enterprise wants the post-training program — data, tuning, evaluation, red-teaming — run as one managed contract rather than assembled in-house.
The questions that decide it
Training prices across these routes differ by single-digit dollars per million tokens. The following differ by much more, and none of them appear on a pricing page:
- Weight ownership and exit. On Azure and Vertex, a tuned closed model is an endpoint, not an asset: no weight export, checkpoints confined to the provider. If your strategy needs a portable model, tune an open model on a platform that hands you the artifact — and get the answer in the contract, not the sales deck.
- Data retention and secondary use. The standard to demand is written: training data is not used to train models for anyone else, retention windows are stated, deletion at exit is attested. Fireworks publishes zero retention by default; Google\u2019s data-processing addendum states customer data is not used for training without permission. A provider that hedges here disqualifies itself for regulated data.
- Residency and boundary. Can training and serving be pinned to your region — or run inside your own cloud account entirely (the Fireworks Sovereign pattern)? Global training tiers on some clouds process data outside your region by design; the regional tier costs more for a reason.
- Certifications and instruments. SOC 2 Type II and ISO 27001 are table stakes; health and finance buyers add HIPAA BAAs, data-processing addenda, subprocessor lists and audit rights. Verify against the provider\u2019s trust center, not a logo strip.
- Lifecycle exposure. Cloud-tuned models retire on the provider\u2019s schedule — Azure publishes retirement dates per tunable model, and OpenAI\u2019s wind-down is the standing lesson: a fine-tune is a dependency, and its end-of-life belongs in your roadmap from day one.
The human feedback loop is the program
A domain-specific LLM is not one training run; it is a loop. Domain experts write the demonstrations the model learns from, grade its outputs against rubrics, and review the failures that drive the next iteration. In regulated domains those experts are the scarce input — a clinical model needs clinicians, a legal model needs lawyers — and their time is bought through the same RLHF and expert data market as frontier-lab post-training. Enterprises staff that loop through expert networks when they cannot hire it: Micro1, for example, certifies domain experts through an AI-led interview and matches them to evaluation and demonstration projects, with published per-role pay ranges a procurement team can diligence. Companies can register to post projects: register your company with Micro1 →
Disclosure: this is a referral link. If your company registers with Micro1 through it, we may earn a referral commission at no cost to you. It does not change where Micro1 ranks — see our methodology.
Budget the loop honestly: in our reading of published prices, the training compute for a focused enterprise tune is hundreds to low thousands of dollars, while expert data, evaluation and a quarter of always-on hosting run to tens of thousands. The fine-tuning services guide has the full market table, and the RAG vs fine-tuning comparison covers the prior question — whether the knowledge problem in front of you needs training at all, or retrieval.
Frequently asked questions
What is the best enterprise LLM fine-tuning service?
The one that matches your existing cloud boundary. Enterprises on Azure fine-tune OpenAI models in Microsoft Foundry; Google Cloud enterprises tune Gemini on Vertex AI; AWS enterprises customize Nova and open models on Bedrock. Enterprises that need the work inside their own cloud account look at in-boundary offerings such as Fireworks Sovereign, and enterprises that want the whole post-training program run for them contract a managed provider such as Scale AI. There is no boundary-free answer, and any ranking that offers one is selling something.
Who fine-tunes OpenAI models for enterprises now?
Microsoft. OpenAI is winding down its own fine-tuning platform and closing it to new users; the same OpenAI models (gpt-4.1 family, gpt-4o, o4-mini for reinforcement fine-tuning) remain tunable in Microsoft Foundry on Azure, under Azure’s enterprise terms, with training at published prices (gpt-4.1 at $25 per million training tokens) plus an hourly hosting charge per deployment.
Do we own the fine-tuned model weights?
On closed-model cloud tuning, no: on Azure and Vertex the tuned model exists only as an endpoint on that cloud — there is no weight export, and checkpoints move only within the provider’s environment. On open-model platforms, the artifact is an adapter or model you can download and serve elsewhere. For many enterprises the endpoint-only model is acceptable; for some — exit strategy, on-premise deployment, model-as-asset accounting — it is the deciding question, and it must be asked before the first training run, not at renewal.
How much does enterprise fine-tuning cost?
Training itself follows the same per-token math as everywhere (dataset tokens × epochs; $1.50–$25 per million training tokens at published cloud list prices). The enterprise bill is dominated by everything else: always-on hosting (Azure charges $1.70/hour per fine-tuned deployment, about $1,241 a month), provisioned throughput commitments, the expert time to build demonstrations and preference data, and evaluation. A realistic enterprise pilot — one domain behavior, expert-written data, evals, one tuned model kept live for a quarter — lands in the tens of thousands, most of it people and hosting rather than training compute.
What security questions should we ask a fine-tuning provider?
At minimum: Is our training data retained, for how long, and is it ever used to train models for others? Can training and serving be pinned to a region, or run inside our own cloud account? What certifications stand behind the service (SOC 2, ISO 27001, HIPAA where relevant) and will they sign a BAA or DPA? Who are the subprocessors? How are tuned models and datasets deleted at exit, and is deletion attested? Published answers exist — Fireworks, for example, states zero data retention by default with SOC 2 Type II and ISO 27001 — so a provider that will not answer in writing is answering.
Sources
Compliance and retention claims are quoted from each provider’s published pages; verify current terms in the provider’s trust center before contracting.
- Fireworks AI pricing — zero data retention by default, SOC 2 Type II, ISO 27001, HIPAA BAA, in-cloud Sovereign offering
- OpenAI API pricing — fine-tuning wind-down notice (closed to new users)
- Vertex AI pricing — model tuning prices and tuned-endpoint billing (Google Cloud)
- Azure OpenAI fine-tuning prices and hosting charges from Microsoft’s retail price list — ecorp IT, November 2026
- Vertex AI tuning review — no weight export, data-protection terms (AnchorTerminal, 2026)
- Microsoft Foundry fine-tuning review — methods, limits, deployment retirement dates (AnchorTerminal, 2026)
Facts on this page were drawn from the public sources above, checked November 2026. If you spot an error, tell us at the corrections address in the footer — we fix sourced errors and note the change.
Last updated: November 8, 2026