An inexpensive, capable AI model can look like an obvious procurement win. Yet the price of an API call says little about the full risk of connecting that model to customer records, proprietary documents, or operational systems. Data sovereignty can turn a bargain into an unacceptable liability.
The debate around China-based DeepSeek illustrates the collision. Its reported efficiency challenged the assumption that competitive large language models require Silicon Valley-scale budgets. Subsequent concerns about where data is stored, who may access it, and how state influence operates changed the enterprise calculation.
For leaders, model selection is becoming an exercise in jurisdiction and accountability as much as performance.
Low Model Costs Can Hide Expensive Exposure
Generative AI pilots often begin with measurable technical factors: quality, latency, integration effort, and price. Those criteria are useful, but they can exclude the conditions under which a provider processes and retains information.
Bill Conner, Jitterbit’s CEO and a former adviser to Interpol and GCHQ, argues that reported concerns around DeepSeek move the issue beyond ordinary privacy compliance and into national-security territory. The source article cites US government revelations alleging Chinese data storage and sharing with state intelligence services, alongside concerns about military procurement networks and export controls.
Those claims demand careful vendor scrutiny because enterprise AI rarely remains a standalone chat window. A successful proof of concept is likely to connect with data lakes, customer platforms, code repositories, and intellectual property. Any uncertainty about provider access then reaches through the same integrations.
A cheaper model cannot deliver savings if its use triggers sanctions exposure, regulatory action, lost intellectual property, or lasting reputational damage. Cost efficiency must therefore include the downside of the provider’s legal and geopolitical environment.
Sovereignty Requires Answers About Who and Where
Data residency is only one part of sovereignty. Knowing the country where information is stored does not fully explain who administers the infrastructure, which subcontractors can reach it, what laws compel disclosure, how long prompts persist, or whether model improvement uses customer data.
Procurement teams should trace inference locations, encryption control, logging, retention, deletion, incident response, and government-access obligations. They also need evidence that contractual promises match technical architecture. A regional endpoint means little if operational access or backups cross the boundary.
The assessment should extend to the model supply chain. Open weights, hosted services, fine-tuning providers, gateways, and monitoring tools can each introduce a new jurisdiction or data handler. Leaders need a current map rather than a one-time questionnaire completed during a pilot.
Requirements will vary by sector. A public marketing draft is not equivalent to a credit decision, medical record, defence document, or regulated customer interaction. Strong governance classifies the intended data and task before choosing the permissible model.
Governance Must Price Risk Before Deployment
Conner frames the decision as a fiduciary responsibility for CEOs, CIOs, and risk officers. That framing changes who belongs in AI evaluation. Engineering can judge whether a model performs; legal, security, compliance, and business owners must decide whether the operating arrangement is defensible.
Enterprises can create a tiered portfolio instead of making a universal model choice. Low-sensitivity tasks may use a broad set of services, while protected workloads run only through providers with verified controls or within sovereign infrastructure. Routing policies can enforce those boundaries automatically.
Exit planning matters too. Organisations should know whether they can move prompts, evaluations, and workflow logic to another provider without rebuilding the application. Portability reduces the temptation to tolerate new risks simply because switching has become difficult.
The DeepSeek debate is ultimately bigger than one laboratory. AI markets will continue producing models that are cheaper, faster, or surprisingly capable. Each will arrive with an institutional context that benchmarks do not capture.
Mature buyers will not reject efficiency. They will define it more honestly. The winning service is not the one with the lowest visible price, but the one that delivers acceptable performance without placing control of sensitive data, regulatory standing, or strategic assets outside the organisation’s tolerable risk boundary.
