Showcases sovereign AI customers (L3Harris, McCarthy, NVIDIA, Tampa General, Ondas/Regna) on AIP, Ontology and Foundry

The commercial version of that argument: customers with small open-weight models and their own ontology data beat frontier models in vertical use cases, so the model layer is commoditised and value accrues to the data and ontology layer. The deployments: (1) L3Harris has integrated more than 30 ERP systems and built hundreds of apps on the ontology, with about 3.5 million data connections and more than 5,000 scheduled data streams. Its Chief Transformation Officer claims its own model beat the frontier labs 'within days' at '95% lower cost' in a test environment; the next step is pairing it with field mission systems such as the VAMPIRE anti-drone system. (2) McCarthy Building (one of the largest private contractors in the US) trained a model on more than 100,000 construction RFIs to predict cost and schedule impact, and saves 30,000 engineering hours a year on drafting subcontract clauses. (3) NVIDIA uses AIP to manage its supply chain from wafer to deployment (announced in September, with details now). (4) Tampa General Hospital (a customer since 2021) runs an on-premises fine-tuned model for data-leakage and privacy classification, with 7% better performance and 22x lower latency. (5) Ondas/Regna AI is confirmed as built on AIP (already reported on 13 Oct). Small-scale support for NDR: L3Harris's deployment depth (consolidating more than 30 ERP systems) and Tampa General adding sovereign AI five years in show existing customers adding use cases, not just maintaining them. Commercial penetration is broadening into construction and healthcare. Now the cold water. No contract value, duration or incremental revenue was disclosed. The 'beat frontier model, 95% cost reduction' line is the customer's own test-environment claim, not an independent evaluation. L3Harris and Tampa General are existing customers, not net-new logos. Coming one week before earnings on 2 Nov, this is marketing: the real test on 2 Nov is US commercial revenue growth, NDR and customer count. Separately, the BBC reported on 21 Oct that Google DeepMind researchers resigned over Pentagon contracts, citing the Maven Smart System (built by Palantir, integrating models from multiple AI vendors) used for target identification. Former Maven lead Shanahan criticised 'deployment speed overriding test & evaluation discipline'. The backlash targets Google, not Palantir's contracts, so it does not touch revenue. But Maven becoming shorthand for the military AI ethics controversy joins the UK NHS and HSBC holdings review as a governance and reputation tail risk, at noise level. Counter-read: more instability on the supplier side raises the value of Palantir's model-neutral, swap-anytime architecture (cf. the Anthropic DoD de-platforming precedent).