<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel>
  <title>IRIS8 Blog</title><link>https://iris8.ai/blog/</link>
  <description>Ideas on governing AI under supervision.</description><language>en</language>
  <atom:link href="https://iris8.ai/rss.xml" rel="self" type="application/rss+xml"/>
  <lastBuildDate>Mon, 09 Nov 2026 09:56:01 GMT</lastBuildDate>
  <item>
    <title>Sovereign AI: Own vs Rent Is Now a Board Decision</title><link>https://iris8.ai/blog/sovereign-ai-own-vs-rent/</link><guid isPermaLink="true">https://iris8.ai/blog/sovereign-ai-own-vs-rent/</guid>
    <pubDate>Mon, 09 Nov 2026 09:55:00 GMT</pubDate><dc:creator>Neeraj Khandelwal</dc:creator><category>Operating model</category><category>Sovereign AI</category><category>Private AI</category><category>LLM cost</category><category>Open-weight models</category><category>EU AI Act</category><category>DORA</category><category>Enterprise AI pilot</category>
    <description>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.</description>
    <content:encoded><![CDATA[<h2>Key takeaways</h2><ul><li>Sovereign AI is investable: Applied Compute raised $350M at a $3.25B valuation after a $100M run rate in under 18 months.</li><li>Renting is getting more expensive: Claude Enterprise seats include no tokens; GPT-5.6 Sol promo pricing ends 30 Nov 2026.</li><li>Menlo 2026: Anthropic and OpenAI alone account for 86% of enterprise LLM API spend; open-source tokens rose from 11% to 19%.</li><li>Framework: rent the uncertain, own the steady, keep regulated data in-country.</li><li>IRIS8 trains a custom model on your data and deploys it in your cloud or on-prem via an 8-week pilot.</li></ul><p>Sovereign AI used to sound like a policy slogan. In the last week it started to look like a funded market. Applied Compute raised <a href="https://siliconangle.com/2026/11/03/ai-cloud-startup-applied-compute-raises-350m-after-reaching-key-revenue-milestone/" target="_blank" rel="noopener">$350 million at a $3.25 billion valuation</a> after reaching a <a href="https://siliconangle.com/2026/11/03/ai-cloud-startup-applied-compute-raises-350m-after-reaching-key-revenue-milestone/" target="_blank" rel="noopener">$100 million annualised run rate in under 18 months</a>. 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?</p>
<h2 id="renting-frontier-ai-is-getting-more-expensive">Renting frontier AI is getting more expensive</h2>
<p>Two pricing moves matter for anyone budgeting AI API costs in 2026.</p>
<p>Anthropic&#39;s <a href="https://support.claude.com/en/articles/9797531-what-is-the-enterprise-plan" target="_blank" rel="noopener">Claude Enterprise plan</a> 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.</p>
<p>OpenAI&#39;s <a href="https://developers.openai.com/api/docs/models/gpt-5.6-sol" target="_blank" rel="noopener">GPT-5.6 Sol</a> 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.</p>
<p>Concentration makes the rent path stickier. Menlo Ventures&#39; <a href="https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/" target="_blank" rel="noopener">2026 State of Enterprise AI</a> 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.</p>
<p>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&#39; 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.</p>
<h2 id="owning-the-model-is-no-longer-fringe">Owning the model is no longer fringe</h2>
<p>Applied Compute&#39;s raise is one signal that &quot;own your intelligence&quot; is investable. The company runs a cloud for training and inference, and its founders describe the goal as helping companies <a href="https://siliconangle.com/2026/11/03/ai-cloud-startup-applied-compute-raises-350m-after-reaching-key-revenue-milestone/" target="_blank" rel="noopener">fully own their models and the systems that run them</a>. 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.</p>
<p>Enterprises are also moving inference closer to home. In a Broadcom/VMware survey covered by <a href="https://cio.economictimes.indiatimes.com/news/artificial-intelligence/enterprises-shift-to-cost-per-task-for-ai-boosting-private-cloud/135518440" target="_blank" rel="noopener">ETCIO</a>, 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&#39;s claim, not an independent benchmark.</p>
<p>Large operators are acting on the same logic. AT&amp;T&#39;s chief data and AI officer, Andy Markus, told the <a href="https://www.nytimes.com/2026/09/04/technology/open-source-ai-anthropic-openai.html" target="_blank" rel="noopener">New York Times</a> that open models rose from 20% of the company&#39;s AI use in May 2026 to 40%, with room to go higher. Coverage of that interview and follow-on reporting, including the <a href="https://www.wsj.com/cio-journal/why-at-t-is-betting-big-on-open-weight-ai-a0ea03b1" target="_blank" rel="noopener">Wall Street Journal</a>, attributes company-reported savings of up to about 80% on the workloads that moved. Those are AT&amp;T&#39;s figures for its mix of transcription, customer service and coding work, not a guarantee for every enterprise.</p>
<h2 id="a-practical-own-vs-rent-ai-framework">A practical own vs rent AI framework</h2>
<p>Raj Villuri&#39;s framing is useful because it is operational, not ideological. As he put it on <a href="https://www.linkedin.com/posts/rajvilluri_enterpriseai-cto-caio-activity-7520340680842301440-XSJx" target="_blank" rel="noopener">LinkedIn</a>: rent the uncertain, own the steady, and keep regulated data on in-country inference.</p>
<ol>
<li><strong>Rent</strong> — novel, low-volume, or fast-changing tasks. Frontier APIs. Pay per use. Accept that prompts leave your boundary.</li>
<li><strong>Own</strong> — 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.</li>
<li><strong>Sovereign / in-country</strong> — 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.</li>
</ol>
<p>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.</p>
<h2 id="why-private-ai-matters-for-compliance-leaders">Why private AI matters for compliance leaders</h2>
<p>Private AI and sovereign AI are not the same as &quot;compliant by default.&quot; They make the data flow easier to document.</p>
<ul>
<li><strong>Data residency.</strong> 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 a <a href="/solutions/gdpr-compliant-llm/">GDPR-compliant LLM</a> path.</li>
<li><strong>EU AI Act compliance.</strong> 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 (<a href="https://ai-act-service-desk.ec.europa.eu/en/ai-act/eu-ai-act-implementation-timeline" target="_blank" rel="noopener">European Commission AI Act Service Desk timeline</a>; <a href="https://www.consilium.europa.eu/en/policies/artificial-intelligence-act/timeline-artificial-intelligence/" target="_blank" rel="noopener">Council timeline</a>). Ownership of weights and logs helps when you later need to show how a system was trained and tested.</li>
<li><strong>DORA AI for financial firms.</strong> 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 (<a href="https://www.eba.europa.eu/activities/single-rulebook/regulatory-activities/operational-resilience/regulatory-technical-standards-policy-ict-services-supporting-critical-or-important-functions" target="_blank" rel="noopener">DORA Art. 28(8)</a>). 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.</li>
<li><strong>HIPAA AI.</strong> Health workloads still need the right agreements, access controls and logging. Running a <a href="/solutions/hipaa-compliant-llm/">HIPAA-oriented private LLM</a> in your environment does not replace those controls; it keeps the processing boundary inside systems you already govern.</li>
</ul>
<p>Banks and insurers face the same pattern: steady document and decision workflows, plus regulated data. That is why we wrote separately about a <a href="/solutions/private-llm-for-banks/">private LLM for banks</a>.</p>
<h2 id="what-own-looks-like-in-practice">What &quot;own&quot; looks like in practice</h2>
<p>Owning a model does not mean training a frontier system from scratch. In practice it means:</p>
<ol>
<li>Pick an open-weight base model with a licence that fits commercial use.</li>
<li>Fine-tune it on your own examples for one workflow (claims triage, policy Q&amp;A, underwriting notes, clinical summarisation).</li>
<li>Benchmark it on a held-out set of your real cases against the frontier model you use today.</li>
<li>Deploy it as an on-prem LLM or in your VPC.</li>
<li>Keep the weights, the eval set and the promotion records.</li>
</ol>
<p>On narrow, specialised tasks, small language models fine-tuned this way can perform on par with frontier models for specialised use cases at around 60% lower cost. 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 <a href="/blog/train-llm-on-your-own-data/">How to train an LLM on your own data</a>, and the product pages on <a href="/private-llm/">private LLM</a>, <a href="/on-prem-llm/">on-prem LLM</a> and <a href="/fine-tuning/">fine-tuning</a>.</p>
<h2 id="an-enterprise-ai-pilot-that-proves-one-workflow">An enterprise AI pilot that proves one workflow</h2>
<p>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.</p>
<p>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.</p>
<p><a class="btn btn-primary cta-talk" href="https://cal.com/neeraj-khandelwal-azexta/30min?utm_source=site&amp;utm_medium=cta&amp;utm_campaign=sovereign-ai-own-vs-rent-hero" target="_blank" rel="noopener">Book a pilot scoping call</a> · <a href="/pilot/">See how the 8-week pilot works</a></p>
]]></content:encoded>
  </item>
  <item>
    <title>How to Train an LLM on Your Own Data: Fine-Tuning, RAG, or Both?</title><link>https://iris8.ai/blog/train-llm-on-your-own-data/</link><guid isPermaLink="true">https://iris8.ai/blog/train-llm-on-your-own-data/</guid>
    <pubDate>Wed, 04 Nov 2026 07:30:00 GMT</pubDate><dc:creator>Neeraj Khandelwal</dc:creator><category>Guides</category><category>Fine-tuning</category><category>RAG</category><category>Private LLM</category><category>Data preparation</category><category>LLM cost</category><category>Data sovereignty</category>
    <description>Train an LLM on your own data without sending it out. When to use fine-tuning vs RAG, the six steps, where your data goes, and what to ask before production.</description>
    <content:encoded><![CDATA[<h2>Key takeaways</h2><ul><li>You rarely need to train an LLM from scratch. Fine-tune an open-weights model and add retrieval (RAG) for facts that change.</li><li>Use RAG for current documents and fine-tuning for repeated tasks and formats. Most production systems use both.</li><li>Budget for data preparation: in one survey, data professionals spent about 45% of their time loading and cleaning data.</li><li>Fine-tuning needs fewer examples than most teams expect: OpenAI suggests starting with 50; LIMA used 1,000.</li><li>Self-hosting only pays at steady volume. Do the cost maths with your own numbers before you commit.</li></ul><p>When a team asks how to train an LLM on your own data, they usually mean something narrower than it sounds. They don&#39;t want to build a model from scratch. They want a model that knows their policies, documents and tone, and they want the data to stay inside their own systems. This guide explains the three ways to do that, how to choose between them, and what to check before anything reaches production.</p>
<h2 id="three-ways-to-use-your-own-data">Three ways to use your own data</h2>
<p><strong>1. Put it in the prompt.</strong> You paste relevant text into each request. It needs no training and works today. It does send that text to whoever runs the model, and it is limited by the model&#39;s context window.</p>
<p><strong>2. Retrieval-augmented generation (RAG).</strong> You store your documents in a search index. For each question, the system finds the most relevant passages and gives them to the model with the question. The model&#39;s weights don&#39;t change. The method was described by Lewis et al. in 2020 (<a href="https://arxiv.org/abs/2005.11401" target="_blank" rel="noopener">Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks</a>).</p>
<p><strong>3. Fine-tuning.</strong> You continue training an existing open-weights model (such as Llama, Mistral or Qwen) on examples from your own work, such as past claims decisions or answered support tickets. The model&#39;s weights change, so the behavior is built in. Methods like LoRA train a small set of extra weights instead of the whole model, which cuts the compute needed (<a href="https://arxiv.org/abs/2106.09685" target="_blank" rel="noopener">Hu et al., LoRA: Low-Rank Adaptation of Large Language Models</a>). Preference methods like DPO teach the model which of two answers your reviewers prefer (<a href="https://arxiv.org/abs/2305.18290" target="_blank" rel="noopener">Rafailov et al., Direct Preference Optimization</a>).</p>
<h2 id="fine-tuning-vs-rag-how-to-choose">Fine-tuning vs RAG: how to choose</h2>
<div class="tbl-wrap"><table>
<thead>
<tr>
<th>Your situation</th>
<th>Start with</th>
<th>Why</th>
</tr>
</thead>
<tbody><tr>
<td>Answers must quote current documents (policies, circulars, contracts)</td>
<td>RAG</td>
<td>Update the index, not the model. Answers can cite the source passage.</td>
</tr>
<tr>
<td>Documents change weekly</td>
<td>RAG</td>
<td>Retraining every week is slow and costly.</td>
</tr>
<tr>
<td>The task is a repeated format: classify, extract, summarize, draft in a house style</td>
<td>Fine-tuning</td>
<td>The behavior is learned once; prompts get shorter.</td>
</tr>
<tr>
<td>You want a smaller, cheaper model to do one job a large model does now</td>
<td>Fine-tuning (often distillation)</td>
<td>On narrow tasks, small fine-tuned models can match large general ones (see below).</td>
</tr>
<tr>
<td>You need both current facts and a fixed format</td>
<td>Both</td>
<td>Fine-tune for the task, retrieve for the facts.</td>
</tr>
</tbody></table>
</div><p>In our work, most production systems end up using both.</p>
<h2 id="can-a-small-model-really-match-a-frontier-model">Can a small model really match a frontier model?</h2>
<p>On narrow tasks, public studies say it can:</p>
<ul>
<li>Predibase fine-tuned 310 small models (2–8 billion parameters) with LoRA across 31 tasks. On average, the fine-tuned models scored 10 points higher than GPT-4, and 224 of the 310 beat GPT-4 on their task. GPT-4 still won on broad tasks such as coding and general knowledge (MMLU) (<a href="https://arxiv.org/abs/2405.00732" target="_blank" rel="noopener">Zhao et al. (Predibase), LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4</a>).</li>
<li>Google researchers trained a 770-million-parameter T5 model with &quot;distilling step-by-step&quot; that outperformed the 540-billion-parameter PaLM model on a benchmark, using 80% of the available training data (<a href="https://arxiv.org/abs/2305.02301" target="_blank" rel="noopener">Hsieh et al., Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes</a>).</li>
</ul>
<p>In our work, specialised models perform on par with frontier models on narrow, domain-specific tasks, at around 60% lower cost. Results depend on the task, so we benchmark every model on your own cases before it goes live.</p>
<h2 id="the-six-steps">The six steps</h2>
<ol>
<li><strong>Pick a base model.</strong> Choose an open-weights model whose license allows your commercial use, sized to your hardware and latency needs. Check the license terms yourself; they differ by model.</li>
<li><strong>Prepare the data.</strong> Collect real examples of the task: inputs and the outputs a skilled employee would give. Remove or mask personal data you don&#39;t need. Plan for this to take the most time: in Anaconda&#39;s 2020 survey of 2,360 data professionals, respondents said about 45% of their time went to loading (19%) and cleaning (26%) data, more than model selection, training and deployment combined (about 34%) (<a href="https://www.anaconda.com/resources/whitepaper/state-of-data-science-2020" target="_blank" rel="noopener">Anaconda, 2020 State of Data Science</a>).</li>
<li><strong>Fine-tune.</strong> Run training (for example LoRA, then DPO) inside your own environment. Version every run, so you can say exactly which data produced which model.</li>
<li><strong>Benchmark on your tasks.</strong> Score the model on a held-out set of your own real cases, side by side with the frontier model you use today. Agree the pass mark before you start, not after.</li>
<li><strong>Deploy where your data lives.</strong> Serve the model in your VPC (AWS, Azure, GCP), on-prem, or air-gapped. Pin versions and keep a rollback path.</li>
<li><strong>Keep improving.</strong> Capture reviewed production outputs, retrain on a schedule, and promote a new version only if it passes the same benchmark.</li>
</ol>
<h2 id="how-many-examples-do-you-need">How many examples do you need?</h2>
<p>Fewer than most teams expect, if they are good ones:</p>
<ul>
<li>OpenAI&#39;s fine-tuning guide sets a minimum of 10 examples, says it sees improvements from 50–100, and recommends starting with 50 well-made examples. If 50 examples change nothing, it advises rethinking the task before adding more data (<a href="https://platform.openai.com/docs/guides/supervised-fine-tuning" target="_blank" rel="noopener">OpenAI, Supervised fine-tuning guide (Right number of examples)</a>).</li>
<li>The LIMA study fine-tuned a 65-billion-parameter LLaMA model on 1,000 carefully chosen examples. In human comparisons, its answers were equal to or preferred over GPT-4&#39;s in 43% of cases (<a href="https://arxiv.org/abs/2305.11206" target="_blank" rel="noopener">Zhou et al., LIMA: Less Is More for Alignment</a>).</li>
<li>In the Predibase study, the number of training examples did not predict how much fine-tuning helped, which suggests quality and task fit matter more than volume once you have enough (<a href="https://arxiv.org/abs/2405.00732" target="_blank" rel="noopener">Zhao et al. (Predibase), LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4</a>).</li>
</ul>
<p>So start with 50 to a few hundred real, reviewed examples and a benchmark, then add data only where the scores show gaps.</p>
<h2 id="where-your-data-goes-at-each-step">Where your data goes at each step</h2>
<p>This is the question your compliance team will ask, so answer it step by step:</p>
<ul>
<li><strong>Data prep and training:</strong> if training runs in your cloud account or data center, the training data does not need to leave it.</li>
<li><strong>Inference:</strong> a self-hosted model processes prompts on your own servers. With a third-party API, each prompt goes to the provider.</li>
<li><strong>Logs and feedback:</strong> check where prompts and outputs are stored, who can read them, and for how long.</li>
</ul>
<p>Rules such as GDPR (<a href="https://eur-lex.europa.eu/eli/reg/2016/679/oj" target="_blank" rel="noopener">GDPR, Regulation (EU) 2016/679</a>) and HIPAA (<a href="https://www.hhs.gov/hipaa/for-professionals/privacy/index.html" target="_blank" rel="noopener">HHS, HIPAA Privacy Rule</a>) set requirements on how personal and health data is processed and shared. Self-hosting does not make a system compliant by itself; it makes the data flow easier to document and control.</p>
<h2 id="what-it-costs-and-when-it-doesnt-pay">What it costs, and when it doesn&#39;t pay</h2>
<p>A self-hosted model has fixed costs: GPUs, storage, and people to run it. An API charges per use. At high, steady volume, a small self-hosted model can cost less than API calls. At low volume, an idle GPU costs money anyway.</p>
<p>So the honest answer is: it depends on your volume. Work it out with your own numbers: monthly tokens, current API bill, and the GPU size the model needs.</p>
<h2 id="what-to-ask-any-vendor">What to ask any vendor</h2>
<ul>
<li>Who owns the trained weights at the end of the contract?</li>
<li>Can training and inference run entirely in our account or data center?</li>
<li>Which base model, and what does its license allow?</li>
<li>How will you benchmark it against what we use today, on our cases?</li>
<li>What records do we get for audit: data used, training runs, eval results, sign-offs?</li>
<li>What happens when the base model is updated or deprecated?</li>
</ul>
<h2 id="start-with-one-workflow">Start with one workflow</h2>
<p>IRIS8 runs an 8-week pilot on one workflow: we pick the base model, fine-tune it on your data inside your environment, benchmark it against the model you use now, and deploy it in your VPC or on-prem. You keep the weights. Success criteria are agreed before we start.</p>
<p><a class="btn btn-primary cta-talk" href="https://cal.com/neeraj-khandelwal-azexta/30min?utm_source=site&amp;utm_medium=cta&amp;utm_campaign=train-llm-on-your-own-data-hero" target="_blank" rel="noopener">Book a pilot scoping call</a> · <a href="/pilot/">See how the 8-week pilot works</a></p>
<p>Related: <a href="/private-llm/">Private LLM for enterprise</a> · <a href="/on-prem-llm/">Deploy an LLM in your own cloud or on-prem</a> · <a href="/fine-tuning/">Custom fine-tuning</a></p>
]]></content:encoded>
  </item>
</channel></rss>
