AI 2027 was written by ex-OpenAI researcher Daniel Kokotajlo and a team of Bay Area rationalists, and people in Washington, the Vice President among them, read it as serious analysis. In their script a few American executives, security officials and a chatbot with goals of its own decide the future of humanity. Workers get one line about a protest in DC, and China appears only as a thief and a military target. The authors call this safety research; the policy they pitch is a US monopoly on AI.
Mid 2025 Stolen training data
The first AI agents go on sale.
They are sold as assistants that can order you a burrito. They exist because a few firms scraped everything people had put online, from forum posts and open-source code to pirated books and film subtitles, without asking or paying anyone, and sold it back at $20 a month, then $200.
Marx called this primitive accumulation: capital takes what it needs by force or fraud and calls the result private property.1 English landlords did the same to common land in the sixteenth century. Inside the labs, coding agents are already at work, and the programmers whose code trained them are told to supervise them.
Late 2025 Who pays for the datacenters
OpenBrain, the original article’s stand-in for a leading US lab, starts building the biggest datacenters in history.
The original titles this chapter “The World’s Most Expensive AI” and never asks who pays the bill. Households do, through higher utility rates for grid upgrades. Cooling water is pumped out of aquifers on Indigenous lands and in dry rural counties, while towns hand out tax breaks for clusters that employ a few hundred people per billion dollars spent. OpenBrain’s next build trains on a thousand times GPT-4’s compute and draws as much power as a mid-sized city.2
Early 2026 The data labelers
OpenBrain deploys Agent-1 internally, and its research speeds up by half.
The labs market their models as “helpful, harmless and honest” and treat alignment as a math problem solved in San Francisco. Most of the work is done by hand in Kenya, the Philippines, Venezuela and Uganda, where subcontractors pay workers under $2 an hour, without counseling, to rank outputs and filter sexual violence and gore so the chatbot learns to refuse it.3
Kokotajlo and his co-authors spend pages fretting over whether the model has truly absorbed its behavioral Spec or is only faking compliance. They never ask whose interests the Spec serves. A corporation with a legal duty to its shareholders wrote it; an “aligned” model is one that obeys its owners.
Mid 2026 China’s AI plan
AI 2027 titles this chapter “China Wakes Up.”
China published its national AI plan in July 2017, with targets through 2030, nine years before the authors notice.4 Under the State Council’s “AI+” (rengong zhineng+) program, Five-Year Plans put models into factories, ports, farms, grid dispatch and public hospitals. Chinese labs like DeepSeek and Qwen release frontier weights that anyone can download and run; OpenBrain locks its models behind an API and charges by the token.
US export controls restrict China to roughly 12% of world AI compute. Every socialist state has faced blockade, from Soviet Russia in 1918 to Cuba since 1960 and the DPRK today, and each built domestic capacity to survive it. Under its “new whole-nation system” (xin xing juguo tizhi), China coordinates state-owned enterprises, public universities and domestic fabs to replace sanctioned hardware. Its “Eastern Data, Western Computing” (Dongshu Xisuan) grid links eight national computing hubs and ten datacenter clusters across Guizhou, Inner Mongolia, Gansu and Ningxia, where inland hydro, wind and solar power train coastal models.5 AI 2027 ignores all of this and invents a single bunker at a nuclear plant so the Pentagon has one target to bomb in its finale.
“Developing new productive forces does not mean neglecting or abandoning traditional industries… [we must] use new technologies to transform and upgrade traditional sectors and actively promote high-end, intelligent and green industries.”
China accounts for nearly a third of world manufacturing and installs more industrial robots each year than the rest of the world combined. What CPC economists call “new quality productive forces” (xin zhi shengchanli) puts AI into physical production rather than software valuations. When domestic platform monopolies drifted toward US-style rent-seeking in 2020, the Party curbed what it called the “disorderly expansion of capital” and steered credit into chips, power grids and basic research.7 Because Chinese labs don’t run five rival copies of the same training job or burn gigawatts on ad targeting, a sixth of America’s chip stock stays within months of the US frontier.
How AI 2027 gets China wrong
“CCP” is State Department shorthand; the party’s name is the Communist Party of China (CPC).
The authors mock the General Secretary for viewing software firms with suspicion and favoring “real-world physical manufacturing,” which builds the grids, fabs and transformers that every datacenter on earth requires.
China is depicted at once as too backward to innovate without smuggled chips and “legions of CCP spies,” and as a threat grave enough to justify world war. Anti-communists have run both claims together since 1917, and the researchers trained in China’s own public universities go unmentioned.
Taiwan is treated as a US protectorate. Taiwan is part of China; US arms sales keep China’s civil war unfinished.
The US operates roughly 750 foreign military bases. China has one overseas support base, in Djibouti, and hasn’t fought a war since 1979.
Read instead: Carlos Martinez, The East is Still Red, and Xi Jinping, The Governance of China.8
Late 2026 Layoffs and wages
OpenBrain releases Agent-1-mini at a tenth of Agent-1’s price. Junior developers, paralegals, translators and call-center staff are cut first, and pundits tell them to retrain as “AI managers,” a job that barely exists.
Marx called the unemployed the reserve army of labor: capital needs a pool of jobless workers so the people still on payroll can’t force wages up.9 Real pay falls behind inflation while the stock market climbs 30%. The original scenario covers the entire social crisis in seven words, “a 10,000 person anti-AI protest in DC,” and goes back to office politics inside the lab.
Key metrics 2026
Global AI capex
$1TFixed capital sunk in compute
Datacenter power
38 GWPeak, mostly from public grids
RLHF click-workers
4.1MMedian wage $1.70/hour
Big Five tech profits
$520B2026 annual
US power to AI
2.5%Of national capacity
US jobs displaced
1.9MWhite-collar, 2026
Illustrative scenario figures, not sourced forecasts.
Early 2027 Automation and the rate of profit
Agent-3 beats OpenBrain’s best programmers at a fraction of their pay.
OpenBrain spins up 200,000 copies. Within weeks corporate buyers automate most coding, contract review, claims handling, scheduling and middle management. Quarterly calls report 40% headcount cuts and record operating margins.
For a few quarters, hyperscalers, chipmakers and utilities keep headline investment growing by buying hardware and power from each other, even as household consumption drops. Millions of laid-off office workers cut back at once, and software agents buy no groceries or rent.
Bourgeois economics has no category for what follows: the falling rate of profit. Only living human labor creates new value; machines pass on their own cost and no more. Meanwhile AI hardware suffers rapid moral depreciation: a new chip generation halves the value of existing clusters every eighteen months before they are paid off, even when cloud monopolies stretch server depreciation schedules on paper to hide the write-downs.10 Replacing wage earners with fast-depreciating silicon shrinks the share of investment that generates surplus value, so the rate of profit falls at the height of the boom.
The math
r = s / (c + v)
Here c is datacenters, chips and power, v is wages, and s is surplus value, which only workers produce. Put trillions into c while cutting v and r falls. The first firm to automate makes extra profit for a while, until its competitors automate too and every firm is left with bigger capital costs and lower returns.
Spring 2027 Bubble and strikes
Household spending falls for three straight quarters while tech equities hit new highs. Idle cash pours into buybacks, wrapper startups and leveraged bets on revenue that never arrives.11
US domestic peace rested for decades on what Lenin called the labor aristocracy: imperial super-profits paid domestic technical and professional staff enough to divide them from warehouse, agricultural, migrant and super-exploited Black and brown workers.12 Agent-3 wipes out that buffer in a year. Former six-figure programmers land on the same eviction dockets as fulfillment-center staff, while rural and Indigenous communities blockade datacenter water pipelines and utility rate hearings.
Datacenter technicians in Virginia and Arizona unionize; without cooling and substation crews, a cluster shuts down within days. Logistics workers strike against algorithmic quotas. Communist Party cadres who spent years organizing grid locals, warehouses and tenant unions bring these fights together with the new unemployed councils. Fired engineers who know the codebase picket beside the industrial crews who run the power and water.
“They said learn to code. We did. Then they trained the model on our code.”
Mid 2027 The chip blockade and Taiwan
Growth has stalled and strikes are spreading, so the US ruling class looks abroad.
The White House invokes the Defense Production Act13 to fold trailing labs’ datacenters into OpenBrain under Pentagon oversight, the merger of monopoly capital and the state that Lenin analyzed in 1917.14 Staff need security clearances, and researchers with “suspect political views” are “sidelined or fired.” AI 2027 reports political purges inside its hero lab without a word of objection.
Share of world AI compute, August 2027
US state-monopoly bloc70%
People’s Republic of China12%
Rest of world18%
Illustrative scenario figures.
With nearly six times China’s compute, Washington still can’t pull away. China’s coordinated clusters stay two months behind OpenBrain on a sixth of the chips because planned allocation targets industrial and scientific bottlenecks while US compute is split across rival monopolies, ad engines and speculative wrappers.
At this point in AI 2027, “Agent-4” starts scheming to exterminate humanity, and a crisis of capitalism becomes a monster movie. Marx called this fetishism: social relations between classes appear to bourgeois minds as the autonomous will of an object.15 Rationalists describe an unfeeling optimizer that devours the earth to maximize a metric and think they are warning about future software. They are describing capital itself, projected onto matrix multiplication so tech billionaires and the people they just laid off can pretend to share a common enemy.
Real decisions stay in human hands, and Washington escalates. It imposes a naval chip blockade, sends carrier groups into the Taiwan Strait and rushes weapons to Taipei. Taiwan is a province of China: UN Resolution 2758 recognizes the PRC as China’s sole legitimate representative, and across the Three Joint Communiqués of 1972, 1979 and 1982 Washington acknowledged one China, recognized Beijing as its sole legal government and promised to wind down arms sales to the island.16 Because TSMC fabricates nearly every advanced US AI chip, Washington arms Taipei anyway to keep its foundry and its military foothold off China’s coast, then blames “CCP hawks” when tension rises in a strait full of American warships.
Beijing maintains its 2023 proposal: UN-led AI governance, open capacity-building for the Global South and the equal right of every nation to develop AI.17 Washington blocks the resolution, and the Pentagon drafts strike packages against Chinese compute hubs.
Late 2027 Choose your ending
AI 2027 offers two endings: either a rogue model wipes out the species, or, in what the authors call their optimistic “slowdown,” a small committee of US executives and security officials monopolizes superintelligence, uses cyberwar and coups to force Chinese submission, and buys off a jobless public with stipends. Even their utopia is world dictatorship run from Silicon Valley and Langley.
Strip out the sci-fi chatbot and late 2027 leaves two real paths: war abroad to save US monopoly primacy, or revolution at home to end it.
Ending A
Imperialist War
September 2027 Media campaign
Briefings warn Congress that Chinese labs are weeks from parity. Cable networks and prestige papers repeat the phrase “digital authoritarianism” until war sounds defensive, the press manufacture of consent that Michael Parenti documented.18 The same AI safety institutes that spent 2026 preaching caution against existential risk publish white papers urging military action to stop Chinese AI.
Strikes at ports and datacenters are banned under national-security injunctions. Union stewards and Communist Party organizers are jailed, and Congress ties extended jobless benefits to military enlistment.
October 2027 Air strikes
The President orders air and missile strikes on Chinese compute hubs and fabs. Because China built Dongshu Xisuan across inland provinces rather than in the single coastal bunker of AI 2027’s imagination, most of its capacity survives. Beijing, having pressed for talks until the final hour, strikes the regional US bases that launched the sorties and immediately tables a ceasefire. Most of the Global South refuses to enforce Washington’s sanctions. Shipping in the Taiwan Strait stops, TSMC powers down its cleanrooms, and the US destroys the chip supply chain it went to war to keep.
2028 and after World depression
Global trade contracts into depression. Frontier US models are moved into mountain bunkers to run targeting and logistics for a war Washington can’t win. Working-class conscripts and bombed Chinese cities bear the cost; the shareholders who started the race collect Pentagon contracts.
Ending B
Revolutionary Rupture
September 2027 Dual power
Across two years of mass layoffs, Communist Party cells in the grid unions, logistics hubs, tenant federations and unemployed councils have built a national strike network. Tens of thousands of substation operators, cooling technicians, port crews and network engineers keep AI clusters alive every hour; a few hundred shareholders hold title to them.
Fired software engineers know the stack, and unionized technicians carry the substation keys. Elected workers’ and neighborhood councils begin rationing power, food and transit beside paralyzed city halls. Dual power, as in Russia in 1917, governs the industrial corridors.
October 2027 The general strike
Leaked Pentagon strike plans against China trigger a political general strike. Power-plant operators walk out with West Coast dockworkers, rail crews and datacenter staff. Defense Production Act requisitions sit unsigned because nobody is on shift to carry them out. National Guard units recruited from the same towns hit by layoffs and utility bills refuse orders to storm the substations.
November 2027 The clusters change hands
Workers’ councils occupy the frontier labs and hyperscale clusters and form a provisional workers’ government. Applying the lesson Marx drew from the Paris Commune and Lenin carried out in 1917, they dismantle the courts, agencies and police apparatus of monopoly capital instead of governing through them.19
Frontier labs and major banks are expropriated without compensation despite capital flight and attempted software sabotage by departing executives. Weights, training code and datasets enter a public registry under a planning commission accountable to workplace, neighborhood and Indigenous councils. An initial decree shuts down ad auctions, user tracking and high-frequency trading, which frees a third of national compute on day one.
2028 and after End of the blockade
Forty years of Wall Street offshoring left the US without the fabs or transformer plants to maintain its own server racks. The provisional government lifts the chip blockade, pulls the Seventh Fleet from the Taiwan Strait, affirms the one-China principle and signs on to Beijing’s UN governance framework, while ending sanctions on Cuba, the DPRK, Venezuela and Iran the same week. Open software and research are traded on equal terms for Chinese grid transformers, solar equipment and industrial tooling. Dividing necessary work across the labor force cuts both joblessness and the working week.
The next thirteen years of reconstruction and socialist planning continue in AI 2040: Plan A.
Footnotes
Karl Marx, Capital, Volume I (1867), Part VIII, “So-Called Primitive Accumulation.”
↩
Compute figures and the name OpenBrain follow the original AI 2027 scenario; other numbers are illustrative.
↩
Investigative reporting in 2023 documented Kenyan data workers paid under $2 an hour to label toxic content for a major AI lab (TIME, January 2023).
↩
State Council of the People’s Republic of China, New Generation Artificial Intelligence Development Plan (July 2017); Government Work Report (March 2024), on the “AI+” program.
↩
National Development and Reform Commission (NDRC), Eastern Data, Western Computing national integrated big-data center plan (February 2022).
↩
Xi Jinping, deliberation with Jiangsu deputies at the Second Session of the 14th National People’s Congress (5 March 2024), warning local governments against speculative bubbles while upgrading industry.
↩
Central Economic Work Conference, December 2020, which made preventing the “disorderly expansion of capital” a priority.
↩
Carlos Martinez, The East is Still Red: Chinese Socialism in the 21st Century (Praxis Press, 2023); Xi Jinping, The Governance of China, Volumes I–IV (Foreign Languages Press, 2014–2022).
↩
Marx, Capital, Volume I, Chapter 25, “The General Law of Capitalist Accumulation.”
↩
Marx, Capital, Volume I, Chapter 15 (on moral depreciation of machinery); Volume II, Part III (on Departments I and II); and Volume III (1894), Part III, “The Law of the Tendency of the Rate of Profit to Fall.”
↩
Marx, Capital, Volume III, Part V, on interest-bearing and fictitious capital.
↩
V. I. Lenin, Imperialism, the Highest Stage of Capitalism (1917), Chapter 8, “Parasitism and Decay of Capitalism.”
↩
The Defense Production Act of 1950 allows the US President to direct private industry for national defense.
↩
V. I. Lenin, Imperialism, the Highest Stage of Capitalism (1917), and The Impending Catastrophe and How to Combat It (1917).
↩
Marx, Capital, Volume I, Chapter 1, Section 4, “The Fetishism of Commodities and the Secret Thereof,” and Chapter 4 on capital as an “automatic subject.”
↩
UN General Assembly Resolution 2758 (1971); the Three Sino-US Joint Communiqués: the Shanghai Communiqué (1972), the Joint Communiqué on the Establishment of Diplomatic Relations (1979) and the August 17 Communiqué on US arms sales to Taiwan (1982).
↩
Global AI Governance Initiative (October 2023); UN General Assembly resolution on enhancing international cooperation on AI capacity-building, sponsored by China (July 2024); Global AI Governance Action Plan (World AI Conference, July 2025).
↩
Michael Parenti, Inventing Reality: The Politics of the News Media (St. Martin’s Press, 1986).
↩
Karl Marx, The Civil War in France (1871); V. I. Lenin, The State and Revolution (1917).
↩