IRIS8 Talk to us
Operating model

Sovereign AI: Own vs Rent Is Now a Board Decision

Sovereign AI is now a funded category. Why enterprises are shifting from rented frontier APIs to owned models, private cloud inference and an 8-week pilot.

· · 6 min read

Sovereign AI used to sound like a policy slogan. In the last week it started to look like a funded market. Applied Compute raised $350 million at a $3.25 billion valuation after reaching a $100 million annualised run rate in under 18 months. At the same time, frontier API pricing for enterprises is shifting toward seat fees plus metered usage, with fewer included tokens. For CTOs and compliance leaders, the practical question is now simple: which workloads should you rent, and which should you own?

Renting frontier AI is getting more expensive

Two pricing moves matter for anyone budgeting AI API costs in 2026.

Anthropic's Claude Enterprise plan bills the seat fee for access only. All usage across Claude, Claude Code and Cowork is charged separately at standard API rates. There is no included token allowance in the seat price. Admins can set spend limits, but the meter still runs on every request.

OpenAI's GPT-5.6 Sol currently lists promotional pricing of $4 per million input tokens and $20 per million output tokens. That promotional pricing ends on 30 November 2026. Teams that built forecasts around the promo need a post-promo plan.

Concentration makes the rent path stickier. Menlo Ventures' 2026 State of Enterprise AI estimates Anthropic and OpenAI alone account for 86% of enterprise LLM API spend, up from 67% the year before. Anthropic, OpenAI and Google together take about 69% of enterprise AI deployment dollars. That is AI vendor lock-in measured in share of spend, not in a single contract clause.

Open-weight models are taking more of the work even while most of the money still goes to closed labs. In the same Menlo report, open-source models' share of enterprise tokens rose from 11% to 19% between September 2025 and September 2026, while their share of model spending stayed around 4%. Usage is moving. Budgets have not fully followed yet.

Owning the model is no longer fringe

Applied Compute's raise is one signal that "own your intelligence" is investable. The company runs a cloud for training and inference, and its founders describe the goal as helping companies fully own their models and the systems that run them. Customers named in the coverage include Microsoft and DoorDash. The contrast with IRIS8 is deployment: Applied Compute runs models on its own cloud; IRIS8 trains a custom AI model on your company data and deploys it in your cloud or on-prem.

Enterprises are also moving inference closer to home. In a Broadcom/VMware survey covered by ETCIO, 56% of respondents already run or plan to run production AI inference on a private cloud. The same coverage describes a shift from cost per token to cost per task. Broadcom claims organisations can save 40% to 90% by moving suitable workloads to private cloud; treat that range as the vendor's claim, not an independent benchmark.

Large operators are acting on the same logic. AT&T's chief data and AI officer, Andy Markus, told the New York Times that open models rose from 20% of the company's AI use in May 2026 to 40%, with room to go higher. Coverage of that interview and follow-on reporting, including the Wall Street Journal, attributes company-reported savings of up to about 80% on the workloads that moved. Those are AT&T's figures for its mix of transcription, customer service and coding work, not a guarantee for every enterprise.

A practical own vs rent AI framework

Raj Villuri's framing is useful because it is operational, not ideological. As he put it on LinkedIn: rent the uncertain, own the steady, and keep regulated data on in-country inference.

  1. Rent — novel, low-volume, or fast-changing tasks. Frontier APIs. Pay per use. Accept that prompts leave your boundary.
  2. Own — high-volume, stable tasks with clear quality tests. Open-weight or small language models fine-tuned on your data. Run as a self-hosted LLM or private cloud AI inference. Measure cost per task.
  3. Sovereign / in-country — regulated data and workflows that need data residency, audit logs, and an exit path. On-prem LLM or client-VPC deployment. Document where training data, prompts and logs live.

Most enterprises will run all three. The mistake is putting steady, high-volume work on an uncapped frontier meter, or putting regulated personal data on a path you cannot explain to an auditor.

Why private AI matters for compliance leaders

Private AI and sovereign AI do not meet regulatory obligations by default. They make the data flow easier to document.

  • Data residency. If training and inference run in your account or data centre, you can say where the data was processed. That matters for GDPR-style obligations and for sector rules. See our notes on private LLMs and GDPR.
  • EU AI Act compliance. Under the Digital Omnibus updates, rules for high-risk AI systems in Annex III apply from 2 December 2027, and rules for high-risk AI embedded in regulated products under Annex I apply from 2 August 2028 (European Commission AI Act Service Desk timeline; Council timeline). Ownership of weights and logs helps when you later need to show how a system was trained and tested.
  • DORA AI for financial firms. The Digital Operational Resilience Act requires financial entities to have documented, tested exit strategies for ICT services that support critical or important functions, so they can end those arrangements without disrupting operations or client services (DORA Art. 28(8)). It does not name any particular cloud brand. It does require a real exit path. A model you can re-host is easier to exit than a closed API with no weights to take with you.
  • HIPAA AI. Health workloads still need the right agreements, access controls and logging. Running a HIPAA-oriented private LLM in your environment does not replace those controls; it keeps the processing boundary inside systems you already govern.

Banks and insurers face the same pattern: steady document and decision workflows, plus regulated data. That is why we wrote separately about a private LLM for banks.

What "own" looks like in practice

Owning a model does not mean training a frontier system from scratch. In practice it means:

  1. Pick an open-weight base model with a licence that fits commercial use.
  2. Fine-tune it on your own examples for one workflow (claims triage, policy Q&A, underwriting notes, clinical summarisation).
  3. Benchmark it on a held-out set of your real cases against the frontier model you use today.
  4. Deploy it as an on-prem LLM or in your VPC.
  5. Keep the weights, the eval set and the promotion records.

On narrow, specialised tasks, small language models fine-tuned this way can perform on par with frontier models for specialised use cases. Results depend on the task, so every model should be scored on your own cases before it goes live. For the method choices behind that path, see How to train an LLM on your own data, and the product pages on private LLM, on-prem LLM and fine-tuning.

An enterprise AI pilot that proves one workflow

IRIS8 is a San Francisco sovereign AI company. We train a custom AI model on your company data and deploy it in your cloud or on-prem. The standard entry is an 8-week enterprise AI pilot on one workflow: agree success criteria up front, fine-tune inside your environment, benchmark against what you run today, and leave you with the weights. The pilot is offered at a fixed price, and that fee is credited to year one if you continue.

That is how you test own vs rent without rewriting your whole stack. Start with one steady, high-volume task. Keep frontier APIs for the uncertain work. Move regulated data onto infrastructure you control.

Book a pilot scoping call · See how the 8-week pilot works

Frequently asked questions

What is sovereign AI in an enterprise context?

Sovereign AI means you control the model weights, where inference runs, and where training data and logs are stored. It is the opposite of depending on a single external API for a core workflow.

Should we stop using frontier APIs entirely?

No. Rent them for novel, low-volume or fast-changing work. Own the steady, high-volume tasks where quality can be tested and cost per task matters.

Is a self-hosted LLM cheaper than an API?

At high, steady volume, a fine-tuned small model often is. At low volume, idle GPUs waste money. Do the maths with your own monthly tokens and GPU size.

Does private cloud AI inference meet our regulatory obligations?

No. It makes data residency and audit easier to document. You still need the contracts, access controls and process evidence your regulators expect.

How do we start without a multi-year build?

Run an 8-week pilot on one workflow with clear pass marks. If the model wins on your cases, expand. If it does not, you have spent weeks, not years, learning that.

See governed AI on your own traffic.

A 30-minute sovereign AI readiness call with the IRIS8 team.

Talk to us