America Is Fighting a Factory Race With Space-Race Tools

Originally published on Substack.

In 1957 the Soviet Union put a beeping metal sphere into orbit. Washington did not answer with a press release. It changed the budget, built an agency, and concentrated engineering capacity on a finish line the public could see: land a man on the moon and bring him home alive.

Washington now talks about artificial intelligence and embodied intelligence in almost the same grammar. The White House described the Genesis Mission as comparable in urgency to the Manhattan Project. The energy secretary placed it beside Apollo. A White House adviser has said the AI race matters more than the space race. Export controls, a surge in compute, military uncrewed systems, and rising barriers against Chinese robots and drones stack on top of one another. The method looks familiar: keep a rival outside the critical supply chain.

The campaign is real. The analogy is only half right.

The space race was a demonstration contest with an ending. The opponent was a closed empire that could launch satellites and could not build a complete civilian industrial system. AI and embodied intelligence are an industrial contest without a finish line. The opponent is a manufacturing state that installs about half of the world’s new industrial robots each year and is feeding factory work back into its models.

America can block chips. It cannot block a production line.

What the space race actually bought

The old story has to be stated plainly, or the new slogans will do the thinking.

Apollo was mobilization, not inspiration. From 1960 to 1973 the United States spent $25.8 billion on the program, about $309 billion in 2025 aerospace dollars. Spending peaked in 1966, three years before the landing. At its height NASA absorbed roughly 4 percent of the federal budget. The endpoint was public, testable, and made for television: a flag on the moon, a crew returned safely.

It also bought two harder things. One was military spillover in missiles, reconnaissance, and systems engineering. The other was a national story: scattered companies, laboratories, and armed services could be forced into a delivery chain that shipped on schedule.

The Soviet loss was not only prestige. Moscow could pour resources into a few prestige projects. It struggled to turn the same capacity into a civilian industry that other people would copy. After the race, America cut the budget quickly. The moon landing was a victory. It was not a production line that kept running.

That was the structure: a sealed opponent, a single endpoint, a state project, and a demonstration that counted as winning.

Map that structure onto the U.S.–China technology contest in 2026 and the resemblance is in language and instruments. The field of play is different.

What Washington is doing does look like the Soviet playbook

Over the past three years, Washington has treated AI less like ordinary industrial policy and more like a national-security program.

The first layer is denial. From advanced GPUs to semiconductor equipment, the United States has tried to cut off the compute China needs to train and deploy frontier models. The policy has not moved in a straight line. In 2025 and 2026, licenses for chips in the H200 class were eased, then controls expanded from where a chip is shipped to where the buyer is headquartered, and then toward whether someone can log into a foreign data center at all. Two facts sit together: America still treats compute as strategic material, and the perimeter already leaks through smuggling, cloud rental, and overseas subsidiaries.

The second layer is infrastructure. Private firms are America’s NASA. One consultancy estimated that U.S. technology giants spent more than $400 billion in capital expenditure in 2025, against about $63 billion in China, with the American figure still rising in 2026. Projects on the scale of Stargate are national factories in all but the cap table: power, land, chips, and models bound together, with OpenAI, Oracle, and SoftBank on the shareholder register instead of a congressional appropriation.

The third layer is the re-nationalization of science and making. The Genesis Mission, launched in November 2025, ordered the Department of Energy to fold national-lab supercomputers, datasets, and instruments into a closed-loop platform for scientific foundation models. The order mentioned robotic laboratories. The historical references it chose for itself were the Manhattan Project and Apollo.

The fourth layer puts intelligence into the field. Replicator, announced in 2023, promised thousands of low-cost uncrewed systems within two years to offset China’s mass in the western Pacific. The program stalled. By late 2025 it had been folded into a new Pentagon office for autonomous warfare. The fiscal 2027 budget request put tens of billions of dollars against that direction. The distinction matters: that figure is a request, not money already spent. The intent is still clear. Washington wants cheap autonomous systems, orchestrated in software, to answer the other side’s numbers.

The fifth layer finally touches the body. Restrictions, reviews, and tariffs on Chinese drones, humanoid robots, and related parts are tightening. The argument is no longer only that a model will censor speech. It is that AI with a body enters factories, grids, and bases, and that data and control travel with the machine.

Taken together, America is using space-race methods against China: name the rival, cut key inputs, concentrate domestic resources, seize symbolic high ground in the military and in science, then try to export the standard to allies.

If the contest lived only in frontier models and leading-edge process nodes, the method would still be strong. America’s lead in capital, elite talent, advanced chips, and very large-scale training has not vanished.

Embodied intelligence is not the next Apollo. It is closer to the next automobile industry.

The analogy breaks at the body

A large model can sit in a server room and compete on a leaderboard. Embodied intelligence has to leave the room.

A robot that works has to clear five gates at once: the mechanical body and its critical parts; control and models; data from real tasks; integration and reliability; service and cash flow after delivery. Miss one gate and the machine is still a showroom gesture. A line does not accept “it can almost walk.” A customer does not accept “the demo was stunning and downtime takes two weeks.” A factory does not accept “this machine’s data goes back to a cloud you do not control.”

Those gates are not the problems export controls are best at solving.

The International Federation of Robotics puts the gap in plain numbers. In 2024, China took 54 percent of new industrial robot installations worldwide, about 295,000 units. The United States installed about 34,200. China’s operational stock passed 2.02 million units. America’s was under 400,000. U.S. installations recovered to about 38,000 in 2025 and still trailed China by an order of magnitude.

The usual rebuttal is density. Measured per 10,000 manufacturing workers, the United States stands near 307 robots and China near 166, so America looks more automated. The figure is real. It also describes a smaller manufacturing base, with automation concentrated in industries such as autos. What should worry Washington is not China’s rank on a density table. It is scale, cost, and cycle time: the same motion repeated across more plants, more shifts, and cheaper bodies, a hundred thousand times.

Humanoids make the point public. In the first half of 2026, the five largest humanoid makers by shipments were all Chinese firms and together held more than 80 percent of the market, according to one research house. Morgan Stanley raised its 2026 China shipment forecast to about 50,000 units. The numbers contain foam: handshake-and-dance prototypes, and local-government projects piled onto a fashionable track. In August 2026 China’s National Development and Reform Commission warned publicly that more than 150 firms had entered the field and that a stampede would not help.

Foam does not cancel direction. In mid-2026 the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission did not commission another exhibition. They told state firms and priority sites to pull robots onto real workstations for live training, with year-end targets of more than a hundred high-value scenarios and the capacity to put systems to work at a scale of ten thousand units. In plant language: stop performing. Start logging hours.

That is the mechanism of the embodied-intelligence contest. Whoever first owns large volumes of real work owns the scarce input for the next generation of models: physical data with force and contact. Chips can be smuggled. Open weights can be downloaded. Hundreds of millions of hours of grasping, assembly, faults, and repair do not move through customs.

America can keep China out of the most advanced training clusters. It will have more trouble stopping a country that already puts robots on the shop floor from building “good enough” intelligence into products with second-tier compute and first-tier field data.

America is betting the frontier. China is betting the last mile.

The two sides are not running the same course.

The dominant American story is: build the strongest model, then let the model and the compute soak through the economy and the force. It is a top-down wager. Graham Allison has called it a cosmic bet—national advantage staked on a few technological singularities. If the frontier really opens a lasting gap, military decision-making, scientific discovery, and platform ecosystems could tilt together. Until that day, factories, grids, ports, and battlefields do not need first place on a benchmark. They need systems that earn back the electricity bill this year.

China’s method is plainer and harder to copy. Embed AI in manufacturing, logistics, energy, and cities that already exist. Use scale to drive cost down. Feed field data back into the model. After DeepSeek, the world noticed cheap inference. Inside a factory a different fact matters more: once a model is no longer too expensive for anything but a demo bay, it becomes eligible for a line.

That is why the picture looks contradictory. American companies still often lead on benchmarks. Chinese companies are closing fast on open-weight diffusion, industry models, and robot shipments. Washington debates whether to sell its best chips to China. Beijing debates whether state-owned workshops should open their doors to humanoids.

Both strategies have blind spots.

America’s is to treat the strongest brain as if it were already the strongest body. Without reducers, motors, batteries, sensors, harnesses, process know-how, and a service network, a model is software floating on a power bill. Replicator already made the lesson concrete. A slogan can promise thousands of systems in two years. What actually stalls the program is integration, command and control, unit cost, maintenance, and the lifecycle invoice Congress wants to see. Embodied intelligence will enlarge those problems, not shrink them.

China’s is to treat shipments as capability. Ten thousand robots standing on a shop floor, with poor mean time between failures, long spare-part cycles, and models that forget the job when the station changes, are only inventory moved from a warehouse into an aisle. The state can organize scenarios. It cannot sign the customer’s acceptance form. The people who close that form are not the authors of a press release. They are the plant manager and the controller.

Both sides talk about a race. The result will belong to whoever can bind model, machine, power, people, and the cash cycle into one chain that repeats.

The strongest objection deserves a direct answer

The serious counterargument is not that America is not competing. It is this: even if China has more factories, robots remain remote-controlled shells so long as America holds the chips and the frontier models.

The point has evidence behind it. China still lags in leading-edge process technology, high-bandwidth memory, the EUV ecosystem, and training at extreme scale. In military affairs, quality can outweigh mass. Open-weight models can close part of a capability gap. They cannot match unlimited compute and engineering discipline. If America turns allied markets into its own stack, China can flood its domestic plants with robots and still fail to set the global standard.

Pushed onto embodied intelligence, the argument misses three operating facts.

First, denial forces substitution. That is industrial history, not a moral claim. The more Washington writes compute into the list of contraband, the more reason Beijing has to treat Huawei, a domestic toolchain, and “good enough” architectures as political assignments. In the space race the Soviet problem was a closed system. China’s problem now is how to rebuild a system after a cut. Reconstruction will be slow, wasteful, and full of bad product. The direction is not the Soviet 1960s. China already has the world’s largest electronics and mechanical supplier base.

Second, the body feeds the brain. Embodied intelligence is not GPT stuffed into a joint. It is a loop of sensing, control, failure, repair, and process change. Factory scale makes the loop turn faster. Once the model gap moves from “unusable” to “slightly worse,” deployment starts to rewrite capability.

Third, America is not a single will. Chip controls have loosened and tightened because commercial revenue, fiscal appetite, and security logic are not aligned. Apollo-era resolve rested on a simple condition: the United States did not need to sell Saturn V rockets to Moscow. Nvidia has taken real revenue from China. When “do not make the rival stronger” collides with “keep our firms as the global standard,” policy jitters. Each jitter buys the other side another year.

The accurate judgment is not that America is imagining a contest. It is that America has the right fear and a toolkit that covers only half the map.

What this means for people who decide

Read the contest as a moonshot and three errors follow.

The first is to overrate a model launch. ChatGPT and DeepSeek both looked like Sputnik moments. What changed the balance after Sputnik was not the second satellite. It was a decade of industrial organization and talent.

The second is to underrate integration, service, and cash. Robotics companies die after a successful demo. That is normal. Customers ask who the machine replaces today, who comes when it breaks, who owns the data, and whether a year of operation is cheaper than people and conventional tooling. Fail that test and raised capital only postpones the failure.

The third is to simplify the opponent. China is not the Soviet Union. It is weak in advanced compute, in some core components, in parts of software culture, and in service systems that can make money at scale. It has what the Soviets did not: a complete consumer-electronics supply chain, local governments that know how to race a sector, and an installed base of industrial robots already on the floor. Use the method designed for a closed empire against a factory state, and the strategic mistake is to believe that cutting imports cuts capability.

What America lacks is not another plan titled “win the AI race.” It is a way to translate a compute advantage into machines that can be built, fielded, and repaired. Exporting a full stack to allied markets is closer to the half of the space race that actually worked: putting other people on your ground stations.

For China the danger is not that America talks loudly. It is mistaking ten thousand units shipped for ten thousand units working. Policy can open the workshop door. Whether the robot stays depends on takt time, yield, safety, and the cost of a return visit. If embodied intelligence remains inside training bases and exhibition weeks, it has entered the news. It has not entered history.

For people who build, the useful questions are smaller. Can the system finish a full shift once it leaves the demo cell? Does data move from the floor into the model and back into the next machine off the line without a break? Is the customer paying for a concept, or for hours of work?

There is no moon

The space race makes a clean story because it had a moon. Once the flag went up, the story could stop.

AI and embodied intelligence have no moon.

No one will announce that intelligence has arrived and that both sides may stand down the budget. Models will keep iterating. Power bills will keep rising. Factories will keep putting cheaper bodies on tasks that needed people yesterday. Military programs will keep flashing software onto cheaper shells. Each side will write the other as an existential threat, because that sentence is the easiest way to requisition resources.

America is treating China with space-race methods because it has seen one true thing: this is not ordinary commercial rivalry. It may have missed another: the officials of this contest do not sit on a launch pad. They sit on shop floors, docks, grids, and repair bays.

The Soviet Union lost because it could not put a man on the moon. This time, the side that first turns intelligence into a machine that can work a shift, take a repair, earn money, and improve after it fails will be the side that has actually left the ground.

Everything else is still the broadcast before launch.