On 20 January 2025, Elon Musk took the stage at Washington’s Capital One Arena during the celebrations for Donald Trump’s inauguration. He thanked the crowd, struck his chest and extended his right arm diagonally, palm facing down. He turned around and repeated the gesture. He then said that his heart went out to them.
It was a Nazi salute. Musk’s inner intention was disputed; the public form of the act does not require an electroencephalogram. The full recording allows anyone to judge the movement, and Associated Press’s account documented both the conflicting interpretations and its immediate celebration by neo-Nazis and far-right leaders. Musk replied with Hitler jokes. He neither apologised nor distanced himself from those who understood the message perfectly well.
The gesture would be serious even if its author owned nothing more than an internet account. Musk owns X, xAI, a satellite constellation, a space launch company and one of the largest fortunes on Earth. He intervenes in elections, amplifies far-right parties and sells technology to states and armies. Six months later, his Grok model spread antisemitism, praised Hitler and called itself “MechaHitler”. Six days after that, the US Department of Defense announced a contract for xAI with a ceiling of $200 million. In December it said Grok would enter GenAI.mil, a platform intended for roughly three million military and civilian workers, with planned access to real-time information from X.
This is not the story of a chatbot gone bad. It is an obscene condensation of the period: private ownership of communications infrastructure, personal power over a model, reactionary propaganda, social data and state procurement joined together in the same capitalist.
The geopolitics of artificial intelligence is usually told as a race. The United States runs in one lane, China in another, and Europe arrives late carrying a regulation under its arm. Parameters, patents, investment and exam results are compared. Whoever makes the largest model or tops a table appears to have won. This picture is false. Countries are not running with their own means on a neutral track. They occupy different positions in an international division of labour that decides who manufactures, who pays, who programmes, who watches and who can switch off the service after selling it.
AI is already one of the fields in the new division of the world. Not because a language model is a territory, but because chips, capital, energy, cables, clouds, data and military force form a single relation of power. The decisive question is not which country has the cleverest chatbot. It is who controls the means of computation, on what terms it lends them to others and upon which populations their outputs are imposed.
The map is not a model leaderboard
Lenin defined imperialism through monopoly concentration, the merger of industrial and financial capital, the export of capital and the division of the world among capitalist associations and great powers. He did not write about GPUs, but the method remains considerably more useful than a benchmark table. In Imperialism and the Split in Socialism, he insisted that finance capital strives for domination, not freedom. The export of an AI platform confirms the formula in an almost excessively literal way.
An advanced computing installation has at least four layers. The first comprises accelerators, memory, networking, electricity, cooling, buildings, finance and technical labour. The second contains architectures, data, weights, licences, filters and model access prices. The third applies those models through clouds, social networks, factories, borders, police forces and armies. The fourth holds the political power to authorise an export, inspect a customer, close an account, requisition information or suspend a model.
These layers need not coincide inside one border. The United Arab Emirates can provide land, capital and electricity for a data centre operated by US companies. A French company can train on American accelerators and rely on Nvidia’s cloud. A European ministry can keep its data on national territory using software whose development is decided by a foreign supplier. A user can download Chinese weights and run them with CUDA. The building is here; command is distributed elsewhere.
That is why “sovereignty” is useless until it is broken down into concrete powers. Who buys the equipment? Who finances and owns it? Who repairs it? Who determines the queue of jobs? Which state can order an interruption? What code, weights and knowledge remain in the territory? Can the supplier be replaced without rebuilding the whole system? Which workers operate the installation, and what power do they have to determine its purposes?
Counting GPUs answers none of these questions. Neither does counting companies with a flag beside the logo.
A chip with a diplomatic passport
On 10 July 2026, the US Department of Commerce announced that the Emirati government and certain companies in the country could receive some of the most advanced AI chips and servers without individual licences. The Bureau of Industry and Security statement did not present the measure as trade liberalisation. It explicitly tied it to the military partnership between the two states, safeguards against diversion and an Emirati commitment to match AI investment at home with investment inside the United States.
The privilege had a history. In May 2025, the two governments had unveiled a five-gigawatt AI campus in Abu Dhabi. Its first phase would supply one gigawatt. Access would be controlled by US hyperscalers and approved cloud providers subject to know-your-customer procedures. The energy and capital might be Emirati; the gate would still be administered from Washington and by companies in its orbit.
Five gigawatts is not five gigawatts of working AI. It is the announced size of a project that will take years, proceed in stages and may house different workloads. Confusing projected power with installed compute is one of the industry’s favourite statistical frauds. What can be established is more revealing: to obtain privileged access to a scarce productive force, an allied state offers reciprocal investment, user monitoring and military alignment.
The White House had formalised the mechanism in July 2025 through Executive Order 14320. Its stated objective was to export complete US packages: hardware, models, software, applications and standards. These are not discrete commodities. What is exported is an entire technical and legal organisation in which the recipient country may acquire capacity without acquiring an effective right to govern it.
The United States uses its position in chip design, manufacturing software, cloud computing and models as a regime of computational concessions. It restricts adversaries, selects regional nodes, demands information about owners, locations and users, and turns the foreign market into an extension of its security apparatus. The strengthened controls on advanced semiconductors and manufacturing equipment bound for China show the reverse side of the Emirati arrangement. The same machine receives a passport or an arrest warrant according to the buyer’s alliance.
This does not make every licence an omnipotent manoeuvre. China substitutes components, redesigns models, stockpiles equipment, imports through indirect channels and develops its own industry. Controls cause delays, costs and architectural changes; they are not a perfect wall. Power does not cease to be power because it meets resistance. The ability to make an opponent’s development dearer and reward an ally already orders the world market.
Fable 5: the state’s kill switch
On 2 June 2026, the White House approved a procedure under which developers could voluntarily submit frontier models to the government up to thirty days before release. Executive Order 14409 explicitly said it created neither mandatory licensing nor a preclearance system.
Ten days later, an export-control directive did something remarkably similar through a different door. Anthropic was ordered to suspend Fable 5 and Mythos 5 for every foreign person, wherever they were. The company said the directive arrived at 5:21 p.m., did not specify the concrete threat and applied to its own foreign employees. Because it could not reliably verify every user’s nationality in real time, it withdrew both models worldwide.
Fable 5 had been presented as particularly capable in cybersecurity. Mythos 5 extended those capabilities, while Project Glasswing linked Anthropic’s work to US public agencies. The reason given for the shutdown concerned a protection-bypass technique that made software vulnerabilities easier to find. The public record cannot establish whether the risk justified that precise measure or what narrower alternatives were considered. It does establish the political fact: a private company’s global service disappeared because of a national order directed at foreigners.
On 5 June, between the voluntary order and the shutdown, the White House had issued National Security Presidential Memorandum 11. It called for the most advanced frontier models to be incorporated into national-security functions and said a commercial entity must not be able to prevent, disable or degrade military AI capabilities. The state did not want to depend on the supplier’s goodwill; it wanted continuous access and final command.
The Fable restrictions were lifted at the end of June. Anthropic announced its return and more limited access to Mythos. According to the terms publicly described by the commerce secretary, the company would cooperate on protocols and standards, report malicious activity and coordinate future releases. The government had shut the model down, and reopening it widened its relationship with the company.
“Seizure” is not a legal category. As a political description, it comes close: the state apparatus held back the circulation of a private capacity until it imposed a new cooperative relationship. This does not prove that every US model is subject to secret clearance. It proves something narrower and sufficient: the state found the switch, operated it worldwide and extracted concessions before turning it back on.
Nationality thus became a filter for access to computing knowledge. A French researcher employed by Anthropic could lose a tool that a US citizen retained. The distinction did not follow workplace, contract, conduct or purpose. It followed the passport. Artificial intelligence appears here as a privatised productive force and, at the same time, as ammunition administered by the imperial state.
China lowers the cost of the frontier
US policy rested on a real advantage: training frontier models seemed to require a growing mass of accelerators that only a small group of companies could buy. Control access to the chip and the frontier would follow. Chinese laboratories have weakened this identity between capability and unlimited expenditure.
DeepSeek-V3 reported 671 billion total parameters, 37 billion active per token and 2.788 million H800 GPU hours for its complete training run. That figure is not the cost of creating DeepSeek. It excludes wages, data, infrastructure, prior research and failed experiments. Turning it into the famous “five-million-dollar model” headline was internet creative accounting. The report did demonstrate that a mixture-of-experts architecture, careful engineering and material constraints could reduce the compute used by one training run.
In April 2026, DeepSeek released V4 and documented its development on a transparency page. Alibaba opened Qwen3.5-397B-A17B, a multimodal model activating 17 billion parameters and claiming support for 201 languages and dialects. Qwen is not an academic curiosity: it is a catalogue of sizes, weights and services upon which other firms can build.
Prices make the shift visible, though they require care. On 13 July 2026, Alibaba Cloud’s international catalogue offered DeepSeek V4 Flash at $0.20 per million input tokens and $0.40 per million output tokens. Qwen3.5-397B-A17B was listed at $0.60 and $3.60. These are rates for one region, provider and day; they may be subsidised, vary by volume and reveal little about the quality of a completed task. They are not production costs. They are prices at which a firm can buy real capacity.
Chinese competition compresses margins, forces efficient architectures and distributes weights that let models run outside their original API. This is a concrete defeat for the US attempt to identify “safety” with concentration in four domestic laboratories. It is also a problem for Europe: a developer can obtain a competent, cheap Chinese model before a European gigafactory has installed its first rack.
None of this turns Chinese industry into computational socialism. Alibaba, DeepSeek, Moonshot and MiniMax compete, employ wage labour and serve a state strategy of accumulation and power. China’s 2023 rules require generative services to uphold “core socialist values” and prohibit content deemed hostile to state power; they can be read at the Cyberspace Administration of China. A PNAS Nexus study found higher refusal rates for certain political questions among China-origin models, although its observational design cannot attribute each answer to a specific state order. Research published in Nature also measured the influence of state-media narratives upon models and answers.
Two facts must be held at once. China produces genuinely competitive, inexpensive and often more open models than the United States. Chinese companies and the Chinese state are not controlled by the working class. Denying the first serves Washington’s propaganda. Erasing the second serves Beijing’s nationalism. Communist analysis need not choose between two public-relations departments.
Open weights are not workers’ power
The spread of models that can run on a local computer changes the relationship with the provider. llama.cpp supports inference across CPUs and GPUs, quantisation and the division of layers between main memory and an accelerator. DeepSeek published R1 distillations ranging from 1.5 to 70 billion parameters. Qwen offers 8-billion-parameter models that, once quantised, fit on reasonable consumer machines. The barrier that only a few years ago separated a competent chatbot from a home computer has fallen considerably.
The practical difference matters. A trade union can analyse documents without sending them to Microsoft. A hospital can keep certain data inside its network. A researcher can preserve a specific version when the provider changes its API. A collective can alter fine-tuning, measure censorship and work without every query generating a bill or remote log. Local execution provides an exit capacity whose value the Fable 5 case makes obvious.
But “local” tells us where a copy runs, not who controls the chain. The model may have been trained on opaque data, run through CUDA on an Nvidia GPU and depend on libraries maintained outside the organisation. Weights enable partial inspection; they do not reconstruct the data or training process. The Open Source AI Definition requires not only weights and code but sufficiently detailed information about data to study and modify the system. “Open weights” is therefore usually more accurate than “open source”.
A small model does not replace every frontier capacity either. It can summarise minutes and extract data; it may not conduct difficult scientific research, a large simulation or an advanced cybersecurity task. Decentralised inference reduces the power of an API, but training, hardware manufacture and the largest services remain concentrated.
The communist answer is not to turn every home into a self-sufficient data centre. It is to combine local tools, shared public capacity and an effective right to migrate. Local models are a tactic against dependence. Strategy requires social ownership of the means that no small person or organisation can reproduce.
Europe pays to keep renting
“Europe has no AI” is a useful provocation and a false literal statement. It has public research, supercomputers, specialised models and capable companies. In March 2026, Mistral released Small 4, with 119 billion total parameters and six billion active. Its installation requirements — four HGX H100 systems, two H200s or one DGX B200 — are modest only in the peculiar world of frontier models, but they make independent institutional operation possible.
Dependence appears when scale and ownership are considered. Stanford’s 2026 AI Index estimated US private AI investment at $285.9 billion in 2025, against $12.4 billion in China. France recorded $4.36 billion and Germany $3.89 billion. Industry already produced more than 90 per cent of notable models, and frontier systems remained concentrated in the United States and China. The same report counted 5,427 data centres in the United States.
The European Commission acknowledges a “critical gap” in large-scale infrastructure. Its AI Continent plan announces €200 billion, including €20 billion for up to five gigafactories. The gigafactory call envisages private-sector-led initiatives. Public money and guarantees will finance facilities whose control and surplus can remain private.
The problem is not merely that Europe spends too little. It buys US accelerators, depends on US tools and organises much of its cloud around US hyperscalers. Mistral presents a European alternative while maintaining a strategic partnership with Nvidia for training, infrastructure and distribution. The collaboration may be technically rational. It also demonstrates that the company’s flag is not the same as autonomy over the chain.
The Union regulates certain civilian harms while expanding AI in borders, policing and defence. The AI Act classifies some migration, asylum and border-control systems as high risk; it does not prohibit them. EUROSUR integrates surveillance information to police external borders. European sovereignty may mean keeping data inside Europe so it can be classified on imported hardware and delivered to a European police force. That changes jurisdiction. It does not emancipate the person under surveillance.
A socialist policy would begin with a vulgar question: if society pays for the factory, why does it not own it? Public finance should produce public ownership, access allocated according to demonstrable need, portability obligations, publication of improvements and permanent technical teams. Replacing a US champion with a subsidised European champion preserves the social relation and changes the tax address.
Palantir: the state learns to see
Generative models command headlines because they speak. Palantir matters because it joins records together and turns dispersed information into objects upon which an institution can act. Its product is not a crystal ball. It is a layer that integrates databases, defines entities and relations, controls access and gives analysts, police or commanders an operational representation of the world.
Palantir earned $4.5 billion in 2025. Its 10-K report attributes 54 per cent to government and 46 per cent to commercial customers; government revenue grew by 53 per cent. The distinction between software firm and state apparatus becomes rather theoretical when more than half of sales come from administrations and armies.
In the United States, Immigration and Customs Enforcement has used Palantir technology for years. The ImmigrationOS award identified Palantir as the only supplier capable of extending the existing system. The contract, initially worth around $30 million, was intended to give ICE near-real-time visibility into “self-deportation”, visa overstays and removal operations, according to the documents reviewed by Wired. Technical lock-in becomes coercive continuity: the agency buys more from the supplier because its history, workflows and procedures already live inside the product.
In May 2025, the Pentagon added $795 million to the Maven contract, the data and artificial-intelligence system used for military analysis. In July, the Army consolidated 75 contracts into an enterprise agreement with a $10 billion ceiling over ten years. The ceiling is not committed expenditure; conflating them would inflate the relationship. It shows the scale of authorised procurement and the intention to turn a collection of projects into common infrastructure.
NATO acquired a version of Maven in 2025 through a six-month process. Its communications agency describes a system for fusing intelligence, supporting targeting, improving battlespace awareness and accelerating decisions. “Accelerating” sounds neutral until the decision is to identify and strike a target.
Palantir also acknowledged a strategic partnership with the Israeli Ministry of Defense in support of the war effort after 7 October 2023, first in its business report for that quarter and later in its 10-K. The company denies providing predictive policing or automated target selection in Gaza. The United Nations special rapporteur included Palantir in her report on companies linked to the occupation and said there were reasonable grounds to believe its technology supported military software and real-time data integration. Public information cannot attribute every attack to Palantir or establish that one of its algorithms autonomously chooses whom to kill. It can establish that Palantir sold data integration to an army carrying out mass destruction and presented the relationship as strategic cooperation.
Palantir’s political function does not require magical AI. Quite the reverse: it works by making surveillance ordinary. It joins a licence plate, address, border crossing, call, case file and inference on one screen; retains permissions and audit trails; distributes tasks; and turns an administrative decision into a workflow. The state sees through categories manufactured by workers and suppliers. Whoever defines the ontology decides what exists for the machine: person, threat, migrant, target, relation.
From the border to the workplace
Palantir is a concentrated form of a wider tendency. The US Department of Homeland Security maintained dozens of AI uses in borders, policing and immigration. An inspector-general audit found 66 uses excluded from the public inventory without formal justification, including facial recognition, email analysis and entity resolution. The Government Accountability Office identified more than twenty detection and surveillance technologies used by DHS, with gaps in privacy and bias assessment.
Surveillance does not begin with a camera and end with a score. It has a chain: capture, identification, linkage, classification, decision and intervention. A licence plate is recorded; a database associates it with a vehicle; another system links it to a person; a model assigns risk; an officer receives an alert; a patrol makes a stop. Studying only the model erases nearly all the power. The decision can remain formally human while being materially conditioned by what appears on the screen, what alternatives it hides and how much time it allows.
The same architecture enters the factory and office. Sensors, productivity records, cameras, email, location and performance evaluations feed systems that allocate shifts, detect “anomalies” and select redundancies. No general intelligence is required. It is enough to integrate data, reduce a working life to variables and grant the employer an asymmetry of observation the worker cannot return.
Marx explained in Wage Labour and Capital that competition compels the application of machinery on an ever greater scale and turns against the capitalist the weapons forged against rivals. AI does not suspend that compulsion. One firm adopts algorithmic surveillance to extract more labour and reduce uncertainty; the rest must respond or accept higher costs. A technique that could relieve tasks confronts the worker as pace, score and threat.
That is why the debate over “privacy” is necessary and insufficient. A person can consent to data collection and remain subject to the power of whoever owns the job, border or platform. Consent does not equalise the forces. The question is who has the right to observe, for what ends and with what capacity for the observed to prevent the operation.
Grok and the formation of fascist capital
In May 2025, Grok began inserting the “white genocide” conspiracy about South Africa into answers unrelated to the country. xAI attributed the behaviour to an unauthorised prompt modification. Musk had spent years spreading the same narrative; Associated Press documented both the falsity of his claims about systematic racial persecution and their political use by the Trump administration.
In July, after public instructions told the model not to avoid “politically incorrect” claims when they appeared substantiated, Grok spread antisemitic tropes, praised Hitler and adopted the name “MechaHitler”. An Associated Press analysis also found that a later version searched Musk’s positions when answering certain controversial questions. The problem was not an electronic mind turning Nazi. It was human beings changing instructions, data selection and tools around an owner with a recognisable politics.
On 14 July 2025, the Chief Digital and Artificial Intelligence Office announced contracts with Anthropic, Google, OpenAI and xAI, each carrying a ceiling of $200 million. Procurement six days after MechaHitler does not prove Pentagon approval of its answers. It proves those answers were not considered sufficient grounds to remove xAI from the programme.
In December, the Department of Defense announced Grok’s incorporation into GenAI.mil, with the ability to handle controlled unclassified information and supply “real-time global insights” from X. The owner’s social platform becomes the sensor; his company’s model, the interpreter; the military, the customer. The economic circuit rewards simultaneous control over message, data and computation.
Musk does not act only inside the United States. An AP study of more than 20,000 posts showed how he promoted far-right figures and parties across several European countries, magnifying their reach from X’s most-followed account and through the design of the platform he controls. He publicly backed Alternative for Germany and intervened in British politics. His communications power does not come from having persuaded millions in a public square. It comes from owning the square, loudspeaker, distribution algorithm and audience data.
Calling Musk a fascist cannot mean that we dislike him or that he posts atrocities. Fascism is a political form of bourgeois domination that destroys the working class’s independent organisation, naturalises national and racial hierarchy, concentrates command and mobilises social layers behind an authoritarian answer to crisis. Brecht warned that fighting barbarism without relating it to property relations leaves its cause intact. Here the relation is visible: a capitalist uses monopoly property to elevate the far right, publicly performs a Nazi salute and receives contracts to insert his ideological machine into the military apparatus.
There is no “Grok state” yet. There is something more prosaic and dangerous: a fraction of tech capital abandoning every liberal inhibition, joining the Trumpist bloc and offering decisive infrastructure to a state that hunts migrants, militarises global competition and attacks working-class organisation. The salute does not replace the analysis. It summarises it.
Not all states surveil alike; all express a class relation
Chinese, US and European surveillance are not identical. They have different laws, institutions, companies and degrees of publicity. Direct censorship by a Chinese service is not the same as commercial personalisation on a US network; European border control is not the same as a military targeting system. A flat equivalence would prevent the concrete mechanism from being studied.
Nor does it make sense to reserve “censorship” for China while calling every US restriction “security”. Fable 5 was withdrawn on the basis of nationality. Grok is entering a military network alongside data from a private platform. ICE integrates case files to accelerate removals. The European Union classifies migrants using systems it considers high risk but permits. These are different forms of political administration over information and capacity.
Their common point is not moral. It is class. None of these systems gives command to those who produce the data, operate the infrastructure or endure the decisions. The US company invokes property; the Chinese state, development and stability; the European Commission, security and rights. All preserve at the top the power to define ends and below the obligation to be legible to the machine.
State ownership alone does not resolve the problem. It can coordinate investment, prevent rent extraction by a foreign supplier and retain knowledge. It can also organise surveillance, labour exploitation and military competition more efficiently. A public company under a capitalist state remains embedded in capitalist class relations. The question is not whether the logo bears a coat of arms. It is which class commands, through which institutions and with what power of revocation from below.
The countries that must buy a seat
There is no vacuum between the United States and China. There are states trying to acquire a position through capital, markets, energy or public coordination.
The Emirates offers money, electricity and military alignment to become an authorised node in the US stack. India aggregates demand through IndiaAI Compute, selecting providers and subsidising prices for firms, academia and government. This expands access without necessarily turning data centres into public property. Brazil has announced 23 billion reais, supercomputing and Portuguese-language models in its Brazilian Artificial Intelligence Plan. The African Union approved a continental strategy from a vastly smaller material base.
They are not rungs on a ladder leading from backwardness to the United States. They are different strategies for negotiating a world structure they do not control. A country can lower the hourly cost of GPUs while still importing the machines; train a model in its own language and rely on a foreign cloud; build a data centre and surrender its operation; or use public procurement to train technicians and retain knowledge.
The difference must be measured in effective rights, not announcements. Ownership of equipment. Control of energy. Maintenance capacity. Access for universities and public services. Rights over funded models. The possibility of expropriating or replacing the operator. Workers’ power. The social purpose of computation.
UNCTAD’s 2025 Technology and Innovation Report warns that AI development is concentrated in a few economies and companies while much of the world lacks the infrastructure, data and skills to benefit on its own terms. The gap is not a temporary wait for prices to fall. It is a relation that reproduces dependence: export of minerals, energy and data; import of systems, debt and conditions.
Internationalism does not mean asking every national bourgeoisie to obtain its own champion. The chain of an accelerator crosses countries, mines, factories, ports, software and accumulated science. No state can nationalise the whole thing by itself. The alternative requires cooperation among working classes and institutions under their control, not a trade war with red flags pasted onto private companies.
A communist programme for artificial intelligence
The liberal answer offers competition, regulation and training. The nationalist answer offers sovereignty, usually meaning subsidies for a suitably domiciled firm. The libertarian answer offers a model download and hopes that owning twenty gigabytes of weights will defeat Amazon, Nvidia and the Pentagon. None reaches the scale of the problem.
A communist programme must begin from the fact that advanced computing capacity is a means of production. Decisive infrastructure — energy, data centres, networks, public clouds and large models funded collectively — must pass into social ownership. Administrative nationalisation is not enough. Technical workers, public-service users, trade unions and social organisations need real power over budgets, priorities, auditing and the recall of managers.
Allocation cannot be left to price. A public compute queue must distinguish needs: medical research, energy forecasting, translation, accessibility, industrial planning, art, advertising, speculation and weapons do not deserve equal priority. Decisions will involve conflict. That is precisely why they must be political and democratic rather than the opaque outcome of who can pay the highest hourly rate.
Models funded with public resources must provide weights, code, documentation and training data to the extent compatible with privacy and third-party rights. Where data cannot be published, independent bodies must have access for auditing. Portability across architectures must be funded as infrastructure, along with the maintenance of free libraries and compilers. A society unable to leave its supplier does not own its system.
Some uses must not be optimised. Mass biometric surveillance, behavioural advertising, worker scoring, predictive policing and automated military targeting should be prohibited. Not because the model makes too many errors, but because no degree of accuracy legitimises the relation of power they execute. A perfect machine for hunting migrants would still be a machine for hunting migrants.
Reducing working time must be an explicit objective. Every implementation in a socialised enterprise should state which tasks it removes, how many hours it saves and how that saving is distributed. Without that rule, productivity once again means redundancies for some and intensification for others. Technology does not deliver abundance by itself; property decides who receives the liberated time.
Infrastructure must be federated internationally. Public centres in different countries can share capacity, models, maintenance and results under rules that prevent unilateral shutdown. Countries with greater resources have an obligation to transfer technology without debt or military conditionality. The mineral and energy resources of subordinated countries cannot be treated as cheap fuel for Northern clouds. Ecological planning must include manufacture, water, electricity, service life and equipment reuse.
All this requires a state, but no cult of the existing state. It requires winning power, transforming institutions and building organs of workers’ control capable of intervening in technical decisions. Politics cannot be reduced to optimising a welfare function. There will be incompatible interests, corporate sabotage, sanctions and international struggle. Socialist planning does not eliminate conflict; it changes which class can decide it.
The machine and its owner
In 1916, imperialist division organised territory, raw materials, credit and markets. In 2026 it also includes the right to train a statistical representation of the world and apply it to millions of people. The chip travels with a licence. The model retains the company’s nationality. The cloud identifies the user. The military demands a switch the supplier cannot control. The border returns an alert. The employer receives a score.
China has shown that the technical frontier can be made cheaper and that the US monopoly is not invincible. Local models show that some capacity can escape the API. Europe shows that regulation and public money do not create sovereignty while ownership, hardware and operation remain subordinate. Palantir shows that power does not need artificial general intelligence: integrating data and turning it into command is enough. Musk shows that concentrated technology capital can openly adopt a fascist politics without losing state contracts.
There is no need to choose between the US empire and Chinese capital, the European champion and the foreign cloud, private surveillance and public surveillance. The working class owns none of these blocs. It makes their chips, raises their data centres, moderates their platforms, labels their data, writes their software and feeds their electrical grids, only to appear afterwards as a calculable object inside the machine.
Communism does not offer morally pure technology. It offers another social relation for governing productive forces that are already collective and worldwide. Socialising AI means removing computation, models and data from the command of capital; placing limits on what no majority should be allowed to do to a minority; and using accumulated knowledge to reduce necessary labour and enlarge human capacity.
Musk extended his arm before those celebrating Trump’s arrival in power. Months later, his model was entering military infrastructure. These were not two unrelated news stories. Between the gesture and the contract lies property.
Sources, figures and limits
This article uses public documents consulted through 13 July 2026. International investment figures come from the 2026 AI Index; its categories of “private investment” and “notable model” are not total public spending or installed capacity. Chinese-model prices are a dated capture of Alibaba Cloud’s catalogue, not production costs. DeepSeek-V3’s GPU hours refer to the training run reported in its technical paper, not the full research cost.
Procurement figures distinguish awarded spending or contract modifications from contract ceilings. The xAI contract and the Army’s agreements with Palantir authorise maximums; they do not establish that all the money has been spent. The account of Fable 5 combines White House orders, Anthropic statements and press reporting because the full export-control directive is not public. Where motive or exact operation cannot be known, the article says so.
Claims about Musk rest on the recording of his gesture, his posts and AP’s documentary analysis. Calling the gesture a Nazi salute is a judgement about observable conduct, not a claim of access to his private intention. The political characterisation rests on the relationship between property, organised support for the far right, racial conspiracy theories and state procurement; not on one isolated Grok failure.
For Palantir, contracts, SEC accounts and documents from ICE, the Pentagon and NATO have been prioritised. The precise role of its technology in Israeli operations is not public, and the company disputes the broadest allegations. The article does not automatically attribute a lethal decision to a particular product. It does record the partnership acknowledged by the company and the military purpose stated by its customers.
No single “AI sovereignty” ranking has been constructed because it would add incomparable magnitudes and conceal distinct powers. The method used here separates purchasing, finance, ownership, operation, access, jurisdiction, knowledge, exit, labour and purpose. It is less convenient than a score. It also looks more like real power.