— The US Navy awards a historic $76.6 billion package for five Columbia-class ballistic-missile submarines and nine Virginia-class attack boats — a concrete bet on undersea superiority for the next two decades.
— A Reuters review of more than 80 Chinese papers shows PLA-linked researchers systematically distilling outputs from OpenAI and Anthropic models to train domestic systems for surveillance, cyber operations and tactical decision-making.
— Facing acute shortages of simulators, the Space Force creates a $981 million multi-award vehicle to accelerate training tools for space domain awareness, orbital warfare and missile warning.
I.
Washington: locking in the silent service
The US Navy's announcement of a $76.6 billion contract package for Columbia-class and Virginia-class submarines is more than a procurement story. It is the most visible expression of a strategic choice: undersea platforms remain the most survivable leg of the nuclear triad and the most difficult domain for an adversary to contest.
General Dynamics Electric Boat receives the bulk of the funding for five additional Columbia-class boats (Build II), while Electric Boat and Huntington Ingalls Industries–Newport News Shipbuilding share the award for nine Virginia-class Block VI submarines. Roughly $5 billion is earmarked for shipyard productivity and nuclear industrial-base investments. Vice Adm. Robert Gaucher, director of submarine programs, framed the deal as “cementing undersea dominance.”
The Columbia-class is the replacement for the Ohio-class ballistic-missile submarines. The lead boat was originally scheduled for delivery in 2027; schedule pressure on turbines and hull sections has pushed that date to 2028. The Virginia-class, already a mature design, continues to evolve: Block VI will incorporate further improvements in acoustic quieting, payload volume and special-operations support. With seven Columbias already under contract and more than twenty Virginias in the pipeline, the Navy is buying continuous production rather than episodic batches — the classic remedy for the stop-start industrial-base problems that have plagued US shipbuilding.
One vocabulary point for the French reader. English defence writing distinguishes sharply between ballistic-missile submarines (SSBN) and attack submarines (SSN). The former carry strategic nuclear weapons; the latter hunt other submarines, surface ships and land targets. Mixing the two labels is a common error that immediately signals an outsider.
Key terms — et leur équivalent français
undersea dominance / undersea superiority
supériorité sous-marine — la formule doctrinale américaine ; to cement undersea dominance : consolider la supériorité sous-marine
ballistic-missile submarine (SSBN)
sous-marin nucléaire lanceur d'engins (SNLE) — le vecteur de la dissuasion océanique
attack submarine (SSN)
sous-marin nucléaire d'attaque (SNA) — mission classique de lutte anti-sous-marine et de frappe
shipbuilding industrial base
base industrielle de construction navale — productivity investments : investissements de productivité
continuous production
production en continu — opposé à episodic batches (lots sporadiques)
lead boat / lead ship
bâtiment de tête — le premier exemplaire d'une classe
acoustic quieting
discrétisation acoustique, réduction de la signature acoustique
payload volume
volume utile d'emport — capacité d'emport en armements ou en modules
special-operations support
soutien aux forces spéciales — les Virginia-class sont conçues pour embarquer des opérateurs
to lock in / to cement
sécuriser, ancrer — to lock in continuous production : sécuriser une production continue
A Reuters investigation published on 31 July reveals that Chinese military-linked researchers have been systematically using outputs from OpenAI and Anthropic models to train smaller, specialised domestic systems. The technique is called model distillation: a powerful “teacher” model generates answers, reasoning traces or synthetic data that are then used to train a lighter “student” model capable of running on constrained hardware inside Chinese networks.
The review of more than eighty academic papers and patents — some analysed by the Jamestown Foundation — shows the method applied across intelligence, surveillance, cyber operations and tactical targeting. One PLA unit used GPT-3.5 to summarise sensitive source code before training a local model that never leaves classified networks. Researchers at the National University of Defense Technology distilled an image-processing model for real-time video analysis on unmanned aerial vehicles. Academy of Military Sciences teams applied the same approach to target recognition in simulated maritime scenarios involving drones, ships and unmanned submarines.
Washington has long restricted advanced chips; distillation offers a partial workaround. It does not replicate the full capability of a frontier model, but it transfers selected reasoning patterns at far lower computational cost. US officials view the practice as a potential circumvention of export controls and an intellectual-property risk. Chinese developers, for their part, deny systematic dependence on foreign models.
For the open-source analyst the lesson is methodological. When Chinese papers describe “knowledge distillation,” “teacher-student training” or “synthetic data generation from large language models,” the operational implication is often the creation of edge-deployable systems that no longer need continuous access to Western APIs.
Key terms — et leur équivalent français
model distillation
distillation de modèle — technique de transfert de capacités d'un grand modèle vers un modèle plus léger
teacher model / student model
modèle enseignant / modèle élève — la paire classique de la distillation
frontier model
modèle de frontière, modèle de pointe — les systèmes les plus avancés (GPT, Claude, etc.)
synthetic training data
données d'entraînement synthétiques — générées par un modèle plutôt que collectées
edge deployment / edge computing
déploiement en périphérie, informatique de bord — exécution locale, hors cloud
to circumvent export controls
contourner les contrôles à l'exportation
intellectual-property risk
risque de propriété intellectuelle
reasoning traces
traces de raisonnement — les étapes intermédiaires que le modèle expose
military-linked researchers
chercheurs liés à l'armée — formulation prudente ; PLA-affiliated est plus précis
unmanned aerial vehicle (UAV)
drone aérien, véhicule aérien sans pilote — l'acronyme reste standard
The Space Force has created a $981 million indefinite-delivery/indefinite-quantity contract vehicle — NITE-STAR — to accelerate the development of training tools. Fifteen companies, including Lockheed Martin, Northrop Grumman, L3Harris, Boeing and several pure-play space firms, will compete for task orders over five years.
The driver is straightforward. Senior officers, including Lt. Gen. Greg Gagnon of Combat Forces Command, have publicly described critical shortfalls in simulators for space domain awareness and intelligence missions. Missile-warning training is relatively mature; orbital-warfare training is improving; domain-awareness and intelligence deltas remain “red-Xed.” Without high-fidelity simulators, operators must take live sensors offline — an unacceptable operational cost.
The contract supports the broader Operational Test and Training Infrastructure (OTTI), which will eventually combine ground facilities, digital environments and on-orbit assets. The language of the announcement is revealing: the goal is not merely “training devices” but “advanced training capabilities” that prepare Guardians for “engagements against peer adversaries.” In Space Force usage, Guardian is the service member; peer adversary is the doctrinal shorthand for China and Russia.
One stylistic observation. English military writing increasingly treats space as a warfighting domain on the same grammatical footing as air, land, maritime and cyber. Phrases such as “space domain awareness,” “orbital warfare” and “space control” are no longer metaphors; they are technical terms with defined meanings inside doctrine.
Key terms — et leur équivalent français
Space Force
Force spatiale — service indépendant depuis 2019
Guardian
Guardian — le personnel de la Space Force (équivalent de Airman, Sailor, Soldier)
space domain awareness (SDA)
connaissance de la situation dans le domaine spatial — surveillance et caractérisation des objets en orbite
orbital warfare
guerre orbitale — opérations offensives et défensives dans l'espace
missile warning
alerte missile — détection et suivi des tirs balistiques
To stay current on maritime security after the submarine contracts: the Office of Naval Intelligence publishes a monthly Worldwide Threat to Shipping report covering piracy, armed robbery, and other threats to merchant vessels. The latest issue (29 July 2026) remains the most concise official snapshot of the global maritime threat picture. Terminology is precise and shared across navies and the commercial sector.