Sovereign AI Sells a Label Buyers Cannot Explain

One in three senior AI buyers at firms with more than $1 billion in annual revenue could not describe sovereign AI in their own words. IDC asked them anyway, in a study Cohere paid for and posted on 25 August 2026.

The same sample already runs generative tools at industrial scale. The word they cannot pin down is the one vendors now stamp on regional clouds, private racks, and Formula One sidepods.

One in Three Buyers Cannot Define the Product

IDC surveyed 508 senior AI buyers in Canada, the United States, the United Kingdom, and Germany from April to May 2026. Every firm cleared $1 billion in revenue. The sample sat in healthcare, life sciences, financial services, the public sector, energy, manufacturing, and telecoms, which is the part of the economy that cannot shrug off a leak.

The InfoBrief (Doc #EUR254834226-IB, May 2026) defines sovereign AI as free choice and control over design, development, deployment, access, operation, maintenance, and governance, plus the foundations those systems sit on. Among people who could offer a definition, 52% reached for local or national control and 35% reached for digital independence. Line-of-business bosses treated it as a risk and cost problem. IT bosses treated it as a rules problem, and their awareness ran two times higher.

Only 13% said the idea was very widely understood inside their own shops. Canada came in lowest, with 10% reporting high awareness and 89% reporting low awareness. The US sat at 12%. Germany reached 29%, almost three times Canada’s high-awareness share. Buyers are writing cheques for a product whose name still splits the room.

WHERE COMPETITIVE ADVANTAGE SHOWS UP

Canada is almost twice as likely as the UK to sell sovereignty as an edge, and the least able to say what the edge is. Telecoms lead the industry list at 37%, then manufacturing at 32%, healthcare at 28%, and financial services and energy at 21% each.

Why UK Firms Still Run Regulated AI on Someone Else’s Cloud

Privacy is the fear everyone can name. Data leakage and privacy sit at the top of the worry list in every sector IDC cut, and the largest shops, those with $20 billion in revenue, worry more than smaller ones. The shares below are the people in each industry who put leakage and privacy first.

WHO FEARS A LEAK MOST

That fear has not moved the plumbing. Over 80% of UK decision-makers still run regulated AI processes on public or private third-party clouds. The study’s own readiness question helps explain why: when people named a single blocker, infrastructure led at 17%, then cost at 10%, with alignment and skills each at 9%.

A sovereignty pitch that starts with “keep the data in-country” dies on that 17%. If the GPUs, the power, and the people are not in the building, the workload goes back to a vendor region with a local pin on the map. The UK numbers make the bind plain. The people most boxed in by rules are still renting the room where the model runs.

Sovereignty Washing Has a Technical Test

Joëlle Pineau, Cohere’s chief AI officer and the former head of Meta’s FAIR lab, gave the mess a name. She calls the gap security dependency, and she calls the marketing around it sovereignty washing.

There’s so much sovereignty washing, where solutions that are not at all contributing to diversification or resilience are being marketed as ‘sovereign’ and painted in that light.

Joëlle Pineau, Chief AI Officer, Cohere

Her test is not a postcode. True sovereignty can include cloud, she said, but parking data in a region or a private cloud does not make the system sovereign. Control has to cover the model, the data, access, updates, and deployment, and it has to name every leftover tie to an outside API. Cohere’s own on-premises or isolated VPC setups put the customer on the hook for hardware, drivers, and the containers, usually on Kubernetes, with the data staying inside that boundary. The docs are blunt about the trade: a VPC is similar protection, and a bit less strict than a box you own.

That is a narrower claim than the national story selling next to it. NVIDIA’s validated AI factory design packages chips, networking, and software so a country or a firm can train and serve agents on premises. The brochure language is ownership. The stack still runs on one vendor’s CUDA world. Rhys Oxenham, SUSE’s vice president and general manager for AI, put the honest limit in one line: full sovereignty at every layer is not realistic today, and choice at each layer is the part that is. A region pin without an exit is a label.

THE CONTROLS PINEAU SAYS HAVE TO BE ENGINEERED

• The model: Who is allowed to change weights, prompts, and fine-tunes, and from where.
• The data: Whether prompts, logs, and training sets ever leave the customer’s boundary.
• The access path: Who can reach the system, including the vendor, and under whose keys.
• The updates: Whether a patch, a policy, or a shutdown can arrive from outside.
• The leftover API: Which calls still depend on a third party that can go dark.

Cohere stated the empty version of the product in public two months before it posted the IDC numbers. On 7 July 2026 the company wrote that sovereign AI means access that cannot be revoked overnight, and that anything short of keeping data, weights, and operations in the customer’s control is a buzz word.

https://x.com/cohere/status/2074599983433519384

The June Cutoff That Made the Dependency Real

The survey closed before the event Pineau now uses as the exhibit. Fieldwork ran in April and May. On 12 June 2026 the US Commerce Department’s Bureau of Industry and Security told Anthropic to block foreign nationals from its two most capable models, Fable 5 and Mythos 5, wherever those people sat. Anthropic could not check nationality in real time at API scale, so it switched both models off for every customer.

HOW THE YEAR LINED UP

1. 4 March 2026: Cohere becomes Aston Martin Aramco’s generative AI partner and puts North on every employee desk.
2. April to May 2026: IDC fields the 508-person survey inside billion-dollar firms.
3. 12 June 2026: The BIS order lands; Anthropic disables Fable 5 and Mythos 5 worldwide.
4. 7 July 2026: Cohere posts that access which can be revoked overnight is not sovereignty.
5. 25 August 2026: Cohere publishes the IDC InfoBrief, with one in three buyers still unable to define the term.

Pineau, speaking days after the cutoff, called it a wake-up. People had the technology, she said, and then learned that access can vanish quickly. Her security point is the same idea at a smaller scale: if critical AI runs through an outside provider, the customer inherits that provider’s failures. She pointed to recent incidents at Anthropic and OpenAI as proof that a crack at the model vendor becomes a crack for everyone downstream.

If critical AI runs through an external provider, you’re relying on their security as well as your own.

Joëlle Pineau, Chief AI Officer, Cohere

IDC’s own forward look sits on that nerve. The firm’s FutureScape CIO Agenda says that by 2028, CIOs at multinational firms will lift spending on modular, sovereign-ready cloud and data-localisation environments by 65% to hedge rising sovereignty demands. Spend can rise without the control list above getting any shorter.

Agents Are Already Inside, Ownership Is Not

The awareness hole would matter less if the tools were still in a lab. They are not. In this AI-aware sample, 99% already use generative AI and 86% use agents embedded in applications. Only 12% said the risks that sovereign AI is meant to address are widely understood in their organisations. Thirteen percent currently use prebuilt or third-party agents, and 67% expect to within 12 months. Internally built agents are expected to jump from 5% to 44% over the same stretch.

Someone has been named. Accountability has not. Of the 508 firms, 420 of 508 already have a person responsible for sovereign AI. Only 8% said that job was well defined. Fully formalized and governed sat at 0.4%. Cohere’s reading is that firms are pinning a name on the problem faster than they are writing the job.

WHO HOLDS THE TITLE

• The default owner: A chief AI officer or head of AI leads 51.7% of the shops that assigned the work.
• The IT path: A CIO or CTO leads 36.9% of those shops.
• Germany’s split: CIOs and CTOs lead 70% of German efforts with a named owner, against 24% in the US, 33% in the UK, and 32% in Canada.

Germany’s pattern fits the awareness gap. The country with the highest high-awareness score also puts the work with the people who already run infrastructure, not with a new AI office. The US, where high awareness is 12%, leans the other way. Titles are multiplying while the control list stays fuzzy, which is how a buyer can fund “sovereign” agents that still call an outside model when they act.

Canada Wants an Edge It Cannot Name

Cohere is a Canadian firm, co-headquartered in Toronto and San Francisco, and it raised nearly US$1 billion across four rounds between 2021 and 2024. The home market is the one that talks about sovereign AI as strategy and then blanks on the definition. Thirty-five percent of Canadian respondents said competitive advantage was the main reason to pursue it, the highest country score in the study, against 10% high awareness.

That split is not a branding accident. Canadian policy has spent two years wrapping compute grants, model work, and industrial strategy in the same word the IDC buyers cannot unpack. If the people who sign enterprise contracts treat sovereignty as a way to look different in the market, vendors will sell them difference. Pineau’s warning is that difference can be a paint job: a local region, a private cloud banner, a slide about residency, and the same external API underneath.

IT staff seeing the concept at twice the rate of business staff makes the next fight internal. The people who will be asked to stand up air-gapped clusters and isolated VPCs are not the people who put “competitive advantage” on the business case. When those two groups do not share a definition, procurement buys a label and engineering inherits a dependency.

What Formula One Buys That a Regional Cloud Does Not

Cohere’s commercial answer is North, its agentic platform, which the firm says can run in the customer’s own environment, including air-gapped rooms, with no remote shutdown switch. Nick Frosst, a Cohere co-founder, has said the stack can stand up on as few as two GPUs, “literally on a GPU in a closet.” That is the opposite of a public API. It is also, in the IDC sample, the thing most firms do not have a plan or a rack for.

The showroom for that pitch is Silverstone. On 4 March 2026 Cohere became the official Generative AI Partner in March of the Aston Martin Aramco Formula One team on a multi-year deal. Every employee gets North for generative and search work across performance, engineering, and operations. Branding went on the side of the AMR26 chassis and the front-wing mirror arms from the Australian Grand Prix in Melbourne.

Formula One represents one of the most demanding data environments in the world. Together with the Aston Martin Aramco Formula One Team, we’re excited to push the boundaries of what enterprise AI can deliver under extreme performance conditions.

Francois Chadwick, Chief Financial Officer, Cohere

Pineau uses the garage as a compressed version of a bank or a hospital: huge telemetry, specialist teams, no patience for a slow or leaky tool, and the useful data already sitting behind the fence. Jefferson Slack, Aston Martin’s managing director of commercial and marketing, called the tie-up a bet on precision and speed. The team gets a private assistant. Cohere gets a rolling proof that the product can live next to the most sensitive operational data a company has.

The 508 buyers IDC interviewed still cannot agree on the noun. Over 80% of the UK cohort still runs regulated AI on a cloud they do not own. Silverstone already has the software on every desk. The people who approve the next enterprise deal are still arguing about the word on the invoice.