Colour theme

Saravanakumar Mohan

Adopting AI, part 2 of 4

The case for sovereign AI: a long-term strategy

AI costs could move sharply, and the providers are expanding into new markets. Keep your data in, reduce your dependence, and treat sovereignty as a transformation, not a project.

Saravanakumar Mohan7 min readAuthor notes

Introduction

AI is here to stay; there’s no doubt about it anymore. Plenty of organisations are now trying to adopt it at enterprise scale, and one theme keeps coming up: dependence on the foundation model providers.

In August 2026 Sequoia Capital’s Sonya Huang published a how-to guide for sovereign AI, arguing that companies should increasingly own their models (“Not your weights, not your product”) to protect their margins and keep proprietary data “within your walls”. Her guide was written for AI application founders, and Sequoia was careful to add that this is “not a recommendation to get off the labs”. But the argument applies just as well to enterprises, and it comes down to two things:

  1. Keep the data in.
  2. Reduce your dependence on the foundation model providers.

Each of these matters on its own. Done together, they set an enterprise up to be AI-ready and protect it from a very volatile AI landscape.

This isn’t to say you need to start buying your own chips and GPUs. It means looking at your entire AI landscape strategically, in a way that is sustainable as a business and still gives you the right results. In Part 1 I looked at the capabilities AI adoption needs. Here I look at where your AI should run, and why that decision is about the long term.

The price you see isn’t the price you’ll pay

Enterprise AI isn’t priced like the consumer plans most people know. Claude’s Enterprise plan, for example, charges a seat fee plus usage that “scales with model and task”, so an organisation’s real bill depends on how much its people and agents use, not on a sticker price.

That usage is expensive to serve. OpenAI made an operating loss of $20.9 billion on $13.1 billion of revenue in 2025, according to leaked statements reported by Fortune. Anthropic tipped into operating profit only in mid-2026, after losing more than $8 billion at the operating level in 2025. Neither has the long, stable profit record that would make today’s prices a safe planning assumption.

Prices could move either way. The cost of a given level of capability has fallen roughly tenfold a year, and competition, including from open-weight models, keeps pushing it down. But usage is growing as agents run longer and do more, and providers under pressure to turn a profit can raise prices or cut what a plan includes. OpenAI did the latter in September 2026, lowering the usage included in new $200-a-month Pro subscriptions.

So treat AI cost as a scenario, not a forecast. Plan for what happens if your spend per person doubles or quadruples, and make sure you could move workloads if it does.

So where else can these companies make their money?

Climbing the value chain

One option is to climb the value ladder: from selling a model to selling expertise.

The providers are expanding in two directions, and they carry different risks.

Moving into their customers’ markets. At the end of January 2026, Anthropic released role-specific plugins for Claude Cowork, including one for legal work that reviews documents, flags risks and tracks compliance. On 3 February, shares in Thomson Reuters, RELX and Wolters Kluwer fell 13–16% in a single session, in New York, Toronto, London and Amsterdam. That matters if you sell expertise or products a model provider could package.

Bringing their suppliers’ work in-house. In September 2026 The Information reported, based on job postings, that Anthropic was weighing which parts of its billing, fraud and payments stack to build in-house rather than buy from Stripe. Anthropic said only that Stripe “has been a strong partner… and we continue to work with them”. That matters if you sell software or services to AI providers: practically anything they buy from a third party today, they’re in a position to build. Use an HCM? They could build their own. Use an ERP? They could build their own.

When your data becomes their product

To do that well, a provider needs to understand how businesses actually work, and every enterprise using its models is showing it.

Today, OpenAI and Anthropic both commit not to train on business, enterprise or API data by default. For enterprise customers those commitments sit in contracts and data processing agreements, not just on policy pages, and they’re worth reading closely. Consumer plans are different: Anthropic changed its consumer terms in 2025 to train on consumer chats when users allow it.

The risk for an enterprise isn’t a silent policy change. It’s leverage. Contracts are renegotiated at renewal, and a provider that learns how your industry works, from the work you run through its platform, is better placed to sell to you, sell as you, or replace you. How much of that knowledge you let concentrate in one provider is a risk to manage, not a certainty to fear. Sovereignty means you set those terms yourself.

Sovereign AI: private cloud for the AI era

So whatever approach we take has to help an enterprise do the two things we started with: keep the data in, so it doesn’t feed an entity that could one day use it to compete with you; and reduce dependence on external providers, because you never know when prices will jump.

The answer shouldn’t come as a surprise. In the cloud era we called it private cloud. In the AI era we call it sovereign AI: AI that you set up, on your premises or in your cloud, under your control.

NVIDIA defines sovereign AI at the national level, as countries building and running AI “using national infrastructure, data, talent and business networks”. The same logic scales down to a company. As with private cloud before it, what matters is control of your data and models, wherever they physically run.

What makes this practical now is open-weight models. Independent trackers such as Epoch AI and Artificial Analysis put the best of them only around four months, or a handful of benchmark points, behind the leading closed models in 2026. Several, such as DeepSeek V4 and GLM-5.2, ship under the permissive MIT licence. One caveat for a sovereignty strategy: most of the leading open-weight models come from Chinese labs, so where a model comes from deserves the same scrutiny as where it runs.

Public or sovereign: decide per workload

Sovereign AI doesn’t mean never using public AI. It means choosing deliberately, workload by workload.

  • Public AI still makes sense for low-sensitivity work, for experiments and pilots where speed matters more than control, and for frontier capabilities that open-weight models haven’t caught up with yet.
  • Sovereign AI makes sense for sensitive data, for your core intellectual property and know-how, and for steady, high-volume workloads where owning capacity beats paying per token.

Most enterprises will run both. The key is an architecture that lets you swap models without rewriting your applications. That’s the cloud lesson again: hybrid by design, not by accident.

Going sovereign isn’t cheap, either. A single eight-GPU Blackwell server costs hundreds of thousands of dollars, and NVIDIA rates it at up to 14.3 kW, more than the 11 kW a typical data-centre rack is built for. Then there are the people to run, update and secure the models, which is an ongoing cost, not a one-off. Renting GPUs (under $3 an hour for an H100 on the open market) can bridge the gap. Whether owning beats renting, or beats an API, depends on sustained utilisation, and published break-even estimates vary too widely to quote a single number. That’s one more reason to decide per workload.

A transformation, not a project

Sovereign AI isn’t a project. It isn’t something you finish in six months or a year; it’s a transformation of how you do everything.

  • You need good, clean data, and for that you need good data practices and governance processes in place.
  • You need an AI adoption framework: one that is goal-driven and shapes your entire enterprise landscape, rather than living in small pockets. The capabilities from Part 1, such as data management, on-premises AI infrastructure, model lifecycle and vendor management, are where that work starts.

What to take away

  1. Treat AI cost as a scenario, not a forecast. Providers’ finances, rising usage and changing plans mean your bill could move sharply. Keep the ability to move workloads.
  2. AI providers are expanding in two directions: into their customers’ markets and into their suppliers’ work. Know which one affects you.
  1. Sovereign AI is private cloud for the AI era: your data and your models, under your control. Open-weight models now make it practical.
  2. Decide per workload, and expect to run both. Build an architecture that lets you swap models freely.
  3. Treat it as a transformation, not a project. Clean data, governance and an enterprise-wide adoption framework come first.

Next, in Part 3: the new roles AI adoption creates, and how to hire for them.

References