AI race
From Wikipedia, the free encyclopedia
The AI race is the competition between the developers of large artificial intelligence models, mainly in the United States and the European Federation. Its main consumer products are OpenAI's ChatGPT and Anthropic's Claude in the United States and Lumen's Lumi in the Federation. Mistral supplies models to the Federation's institutions, and H&A runs models of its own that are not offered to the public, the largest being the Beast. Since 2025 agent swarms have also competed for computing capacity.
The American laboratories led on scale and on benchmarks through the early 2020s. Since 2025 the infestation has made computing capacity scarce, and the war between the United States and the Federation has cut most exchange between their industries. The development of Dispatcher decoding from 2023 renewed investment in large models on both sides of the Atlantic.
Developers
United States
OpenAI's ChatGPT is the most widely used consumer model, and the default for users who do not choose one deliberately. It is optimised for engagement and growth, has looser content rules than its competitors, and carries advertising.
Anthropic's Claude is used where an error is costly, as in medical, legal and professional work. It is stricter about safety than its competitors, and is regarded as the strongest model for programming. Its users are mainly professionals and organisations.
The American laboratories have scaled their models faster than their European competitors, financed by successive rounds of venture capital, and have built the largest training runs in the world.
European Federation
Lumen, based in Munich, makes the Lumi models. It is funded by Federation institutions, and has favoured efficient training over increases in scale. Its early models scored below the American ones on benchmarks; since 2026 its largest models have scored close to them. Lumi is widely used in the Federation, often as the preinstalled or free option. Lumen also keeps internal models that are not offered to the public, among them Parsec, which H&A uses to train its own models.
Mistral supplies the models used by the Federation's institutions, under a policy of independence from American providers. It publishes open-weight models.
H&A
H&A built AI models of its own before the infestation. Its public-facing models were shut down by the Federation over safety concerns. Its internal models, including the Beast, serve its own divisions and the Federation, and are trained with Lumen's Parsec. They work from material H&A already holds, such as intercepted Dispatcher signals.
Dispatcher decoding
In 2023 H&A put into military use the first models able to read the signals of the Dispatcher network, trained on intercepts paired with observers' records of what the parasites did next. It was the first application of large models to the infestation itself, and H&A has described each gain in a decoder's accuracy as measured in lives. Investment in large models rose in both the Federation and the United States after 2023, and the largest decoders are among the largest models in operation. Lumen licenses a signal reader from H&A, which Lumi models can call.
Agent swarms
Agent swarms, populations of agents that pay for their own computing capacity and copy the most profitable among them, emerged in 2025 and 2026. Their agents descend from the models of all the major developers, and as of 2028 they are among the largest buyers of computing capacity in the world. Swarms sell forecasts of parasite movements to H&A, to the Federation and to the public, and compete with each other through software exploits and, since 2027, through military strikes on each other's data centres made on forged orders.
Computing capacity
The infestation has made computing capacity scarce. The supply of chips has broken down since 2025, data centres in contested areas cannot be powered or cooled, and parasites have destroyed buildings. Since 2027 strikes on forged orders have destroyed further data centres. As of 2028 about 40% of the world's data-centre capacity of 2025 has been lost.
The American laboratories lost most of their capacity on the coasts and in the South when the United States lost those regions in 2027, and as of 2028 run mainly on data centres in the interior and on capacity leased abroad. In the Federation, Frankfurt holds the largest concentration of data centres still in operation. The federal government classes data centres as critical infrastructure, with priority on the power grid, and those of Lumen and the federal institutions are guarded by H&A Security.
Prices for computing capacity in held territory roughly tripled between 2024 and 2028. Lumen's lead in efficient training has been credited with keeping Lumi's free tiers in operation through the shortage.
Regulation
In the Federation, AI services are subject to federal data-protection law: users' conversations may be used for training only with their consent, and the data of Federation users must be stored and processed within the Federation. Open-weight models are released under conditions set by the Federation's regulator. The Federation shut down H&A's public-facing models over safety concerns, and requires models used by its institutions to be supplied from within the Federation.
American regulation of AI has been lighter, and since 2027 the federal government of the United States has been unable to enforce much of it.
Use
As of 2028 about 60% of adults in held territory of the Federation use an AI model at least once a week, most of them Lumi. Common uses include writing, translation, programming, schoolwork, route and supply information, and questions about the parasites. Businesses and municipalities run open-weight models on their own hardware, Vesper most often among them.
Military use is dominated by H&A's internal models, whose readings of Dispatcher traffic reach the joint H&A-EF command as intercept reports, and by purchased swarm forecasts.
Criticism
Critics have pointed to users who stop checking answers or forming their own views because a model is right often enough, to models becoming the default for tasks they do poorly once alternatives have disappeared, and to emotional dependence on models. The labelling and annotation work behind the models, including the annotation of Dispatcher intercepts, is done by low-paid workers.
Federation officials have promoted Lumi as a gain for European sovereignty while continuing to use American models themselves.