# Choose AI models the way you choose suppliers By [Dr. Linda Xu](/about) · 24 Oct 2026 · 6 min read If one of your AI models were excluded tomorrow, could you switch without losing a customer or rebuilding a product? Most companies cannot answer that, and a US order published on 20 October 2026 shows why they need to: choosing an AI model is now a sourcing decision, and it deserves the same discipline as choosing a supplier for a critical part. ## What the order means for business The Department of Homeland Security ordered DeepSeek products removed from federal civilian systems and procurement, including "third-party products and services developed using DeepSeek implementations" [1]. Parallel orders cover defence and intelligence systems [2]. Three points matter. - **It reaches federal agencies and their contractors, not the whole market.** Private companies are not directly bound [3]. - **It reaches products built on the model,** wherever it runs. Fine-tuned versions are likely covered too, because their weights inherit DeepSeek's [3]. - **It flows down the supply chain.** Contractors at every tier must check for covered products and report one within three business days [2]. "Which models are inside your product?" is now a contract question. ## Why "the model has no memory" is not the whole answer A model's weights do not send data anywhere. Where a model is hosted decides where your data travels, so a self-hosted open model keeps data inside your own walls. That part of the argument is right. But weights carry behaviour. A 2025 US government evaluation found DeepSeek models easier to jailbreak than US reference models, and more likely to repeat Chinese Communist Party narratives [4]. Supervised fine-tuning and reinforcement learning can change how a model responds, and open weights let you test it as deeply as you like. What they cannot give you is proof that every unwanted behaviour is gone, because the training data is not open. So open weights solve the data-location problem. They leave the lineage problem open. ## The economics are still compelling This matters because tuned open models have become very good at narrow work. In June 2026, Bridgewater's AIA Labs and Thinking Machines trained a model from the open Qwen3-235B with reinforcement learning and distillation. On six financial tasks it reached 84.7% accuracy against 78.2% for the best frontier model they tested, at 13.8 times lower inference cost per task [5]. The tasks were narrow and the teams measured it themselves, but the pattern is familiar to anyone who has tuned a model for one job. In my current role, sending each task to a model suited to it was part of a 38% reduction in AI cost. Then, on 23 October, Moonshot released Kimi K3.1 Pro, an open-weight model it says scores ahead of GPT-6 Astra and within a point of Claude Opus 5.5 on a coding benchmark [6]. Those are Moonshot's own figures, not yet checked independently. In a normal year, a team paying frontier prices would test it straight away. This year, the first question is who your customers are. The savings are real, and so is the other side of the ledger. Every model change means revalidating outputs, retraining people and re-testing controls. A cheaper model that costs you one federal contract, or one customer audit, is not cheaper. The decision is the total cost, not the token price. ## The tariff problem, in a new form I spent years in global supply chain operations, where a change in trade policy could hold shipments at a border overnight. The lesson from tariffs was never "avoid every country". It was: know where each part comes from, qualify a second source before you need it and price the risk into the contract. Models now face the same pressure from both sides. On 16 October, China's Ministry of Commerce and Ministry of Science and Technology published a draft adding advanced AI models and agents to the country's export-control list [7]. Open-source releases and models offered as a service fall outside the draft, which is open for comment until 15 November. Still, governments on both sides now write rules that name model weights, and open models sit closer to trade policy than they ever have. ## Three questions for the leadership team 1. Which models are inside our products and the software we buy, and where did each come from? 2. Which customers or contracts limit the models we can use? 3. How long would it take to move our most important AI feature to another model? If nobody can answer in a day, the work below is overdue. ## What I would do now - **Give model sourcing one owner.** It sits between AI, procurement, legal and security, and falls through the gaps when each holds a piece. Someone has to own it end to end. - **Build a model bill of materials.** For every AI feature, record the base model, who tuned it, where it runs and which products call it. - **Segment by customer.** Design so that different customers can be served by different models without a rebuild. - **Qualify a second source.** Keep every important capability tested on at least two models, ideally from different countries of origin, behind a gateway so switching is a configuration change, not a project. - **Write it into procurement.** Ask vendors to disclose model lineage and to tell you when a regulatory action affects one of their components. - **Price the concentration risk.** If rules ever narrow the choice to a few closed-model vendors, buyers lose leverage just as those vendors face public-market pressure: Anthropic and OpenAI both filed confidentially for IPOs in 2026 [8]. Sign multi-year infrastructure and model commitments only with exit and substitution terms. The leaders who do well will not be the ones who pick the right model this quarter. They will be the ones who can change models next quarter without disrupting a single user. ## References 1. US Department of Homeland Security, [Notice of Order Regarding Exclusion and Removal of DeepSeek Products from Federal Civilian Executive Branch Information Technology Systems and Procurements](https://www.federalregister.gov/d/2026-21447), Federal Register document 2026-21447, 20 October 2026. 2. Crowell & Moring, [FASCSA Orders Require Exclusion and Removal of DeepSeek AI](https://www.crowell.com/en/insights/client-alerts/fascsa-orders-require-exclusion-and-removal-of-deepseek-ai), 21 October 2026. 3. Cloud Security Alliance, [DeepSeek FASCSA Order: What "Developed Using" Means for AI Supply Chains](https://labs.cloudsecurityalliance.org/research/csa-research-note-dhs-deepseek-fascsa-order-20261021-csa-sty/), October 2026. 4. NIST Center for AI Standards and Innovation, [Evaluation of DeepSeek AI Models](https://www.nist.gov/document/caisi-evaluation-deepseek-ai-models-report), 30 September 2025. 5. Bridgewater AIA Labs and Thinking Machines, [Learning to replicate expert judgment in financial tasks](https://thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks/), 30 June 2026. 6. The Next Web, [Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills](https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber), 23 October 2026. 7. Recode China AI, [China's Draft AI Export Controls Don't Block Open-Weight Models](https://www.recodechinaai.com/p/chinas-draft-ai-export-controls-dont), October 2026. 8. TechCrunch, [Following Anthropic, OpenAI files confidentially for IPO](https://techcrunch.com/2026/06/08/following-anthropic-openai-files-confidentially-for-ipo/), 8 June 2026. ## About the author [Dr. Linda Xu](/about) is Director of Data Analytics & AI at MealSuite, where she owns AI end to end. [More about her](/about). --- Source: https://drlindaxu.com. Dr. Linda Xu (publishes academically as Wei Xu).