Thirty five thousand people walked off the line in July over a robot that will not show up until 2028.
The stoppage rolled through Hyundai’s Korean plants, touched about half the company’s global output, held up roughly 5,000 cars, and cost something near $135 million in sales. Reporting called it the first auto strike anywhere set off by humanoid robots. The core demand was not wages. It was one sentence: no robot enters a Hyundai workplace without a labor agreement signed first.
Read that demand again, because it is the whole story. Those workers are not trying to stop the machine. They are asking for a rule. They struck because there is not one.
What Atlas actually is now
Boston Dynamics showed the production version of Atlas at CES, the big Las Vegas electronics show, in January 2026 and started building it in March. The machine has 56 joints that rotate through a full circle, a battery that runs about four hours, and the ability to walk itself to a rack and swap that battery in under three minutes with no person involved. It stands just over six feet, weighs about 200 pounds, and lifts 50 kilograms in a single pick.
The entire 2026 build is spoken for. Not sold on the open market: committed to two customers, both of them family. Hyundai’s own robotics application center gets one fleet. Google DeepMind gets the other. Outside buyers wait until early 2027.
Here is the part most of the coverage skipped. Boston Dynamics has never said how many robots the 2026 run contains. “Sold out” means something very different if the number is forty than if it is four hundred, and until the company publishes it, nobody outside the building knows which sentence they are reading. I am going to take the schedule seriously anyway, for a reason I will get to. But the honest version is that we know the date and not the count.
The price nobody has published
Boston Dynamics has never put a price on Atlas either. Analysts guess somewhere between $130,000 and $145,000. A widely repeated $320,000 figure traces back to an anonymous ceiling estimate rather than a quote from anyone. Treat all of it as weather.
The number with a name attached comes from Morgan Stanley, which put a humanoid at roughly $200,000 in 2024 in wealthy countries and expects about $150,000 by 2028. Spread $150,000 across a ten year service life and you get about $41 a day for the hardware.
That figure is honest only about the hardware. It leaves out electricity, maintenance, spare actuators, the engineers who integrate the thing into a line that was designed for people, and every hour the robot is down. Anybody who quotes you a robot day rate without naming those costs is selling something. Load all of it back in and the comparison a plant manager runs in 2028 still is not close, because the machine takes three shifts and does not get hurt.
Morgan Stanley also expects about 13 million humanoids in service by 2035, on the way to a market it sizes at $5 trillion by 2050. The 2050 number is a guess dressed as a forecast. The 2035 number sits inside the working life of almost everyone reading this.
Why Hyundai and not Google
DARPA, the Pentagon’s research arm, commissioned the first Atlas in 2012, and Boston Dynamics unveiled it on July 11, 2013: six foot two, 330 pounds, 28 hydraulic joints, tethered to a power cable, built to walk into a building like Fukushima so a person would not have to. Google bought the company that same year and never disclosed the price. Four years later it sold Boston Dynamics to SoftBank for about $100 million. SoftBank added $37 million in 2019, then sold eighty percent to Hyundai in a deal that valued the company at $1.1 billion and closed in June 2021. This July, Hyundai moved to buy the remaining twenty percent.
Count the owners. A defense research agency, a search company, an investment fund, and finally a carmaker. Only the carmaker got it into production, and the reason has nothing to do with vision. It is inventory.
An electric power steering unit is a motor, a gearbox, a position sensor, and a controller in one housing. A humanoid robot’s joint is a motor, a gearbox, a position sensor, and a controller in one housing. Hyundai Mobis, the group’s parts division, has been mass producing the first thing for years, and at CES in January 2026 it agreed to supply actuators for Atlas. McKinsey puts actuators at 40 to 60 percent of what a humanoid costs to build, the single largest line item on the bill. Hyundai did not invent its way past the hard part. It already had the hard part sitting on a shelf in Korea.
The easy reading is that Google failed because it builds software and not machines. That holds up, and it just got sharper in a way worth sitting with. Google is back on Atlas. The CES deal puts DeepMind’s Gemini Robotics models into the robot and sends DeepMind a fleet of its own. Hyundai supplies the body. Google supplies the mind. Neither one ships without the other, and the company that owns the body is the one that owns the calendar.
Two dates
This next part is a schedule, and the schedule is the argument. You do not need to hold every date. Hold two.
Hyundai’s plan, as announced: Atlas starts parts sequencing at the Metaplant in Georgia in 2028, reaches Kia’s Georgia plant in 2029, and moves into component assembly by 2030. The group has committed to more than 25,000 robots across Hyundai and Kia plants and is building toward the capacity to make 30,000 robots a year by 2028, inside a $26 billion American investment running four years from 2025.
Date one: 2028. Robots on a line in Georgia.
Now the rule. The European Union’s AI Act is the only major law anywhere that was about to put real obligations on high risk AI systems, and those obligations were scheduled to start on August 2, 2026. On July 27, 2026, six days before that deadline, a package called the Digital Omnibus took effect as Regulation (EU) 2026/1744. It moved the dates. Systems on the Act’s stand-alone list now comply by December 2, 2027. AI built into products that Europe already regulates as products, the category that holds machinery and vehicles, now complies by August 2, 2028. The same package narrows what counts as a safety component and largely carves out AI inside machinery already covered by Europe’s Machinery Regulation.
Date two: August 2, 2028. Rules for AI inside machines, if the machine is not carved out.
The robots arrive the same year the rules do, in a different jurisdiction, under a definition that may not reach them. Nobody planned that collision. Nobody is scheduled to fix it.
What the United States has
Nothing on this. No framework, no deadline, no definition.
Manufacturing employed 12.69 million Americans in December 2025. Transportation and warehousing employed 6.5 million in January 2026, down 1.8 percent from the year before. That is close to one in eight working people in this country standing in the path of a machine with a published deployment date. What we have to meet it is a handful of state AI bills, most of them about deepfakes and chatbot disclosure. Nothing about actuators. Nothing about what a plant owes a person whose job it retires.
What Oklahoma should be doing
Oklahoma had 140,300 manufacturing jobs as of December 2025, and they skew toward exactly the physical work Atlas is built for.
The play is not competing with Georgia for a robot factory. Georgia already won that one. The play is power. A plant that makes 30,000 robots a year, and every factory that buys from it, needs electricity that is cheap, firm, and clean enough to qualify for federal money. Oklahoma has 20,406 documented orphaned oil and gas wells across 76 of its 77 counties, and the Corporation Commission’s own plan says that at the current pace and current funding it would take 235 years to plug them all. Every one of those holes already goes down to heat.
Phoenix Wells, our plan to plug Oklahoma’s dead oil and gas wells and tap the heat still in them for power, is not only an environmental cleanup. It is workforce transition infrastructure that happens to be shaped like a well.
The strongest argument against those workers
Let me state it the way its best advocate would. Thirty five thousand people shut down half a company’s output over machines that are not in the building, will not be in the building for two years, and are aimed first at the jobs that wreck human bodies over a career. Automation has been coming for factory work since the first spot welder. Living standards went up anyway. A union that wins a veto over new equipment is a union that eventually prices its own plant out of existence.
That argument is right about the history and wrong about the clock. Every previous wave gave people something close to a working lifetime to adjust, and the institutions that made the adjustment survivable, public schools and labor law and social insurance, showed up a generation late and were paid for in the meantime by the people the machines displaced. More than two hundred economists said in July, in a letter sixteen Nobel laureates signed, that this transformation could be larger than the Industrial Revolution and run on a fraction of the schedule.
The workers at Hyundai’s Ulsan plant in Korea did not read a Morgan Stanley deck. They read a press release with a date on it, noticed that nothing else in their lives had a date on it, and made one.
What Foundation says about it
Foundation is built for this exact gap. Universal Basic Citizenship is not a monthly check. It is a floor under the parts of your life that a job currently gates: housing you cannot lose, health care that does not vanish with the badge, training that retools you as many times as the economy demands it. Fund the goods, not the gap. A person whose plant installs Atlas in 2028 should not have their housing, their doctor, and their retraining all riding on whether the line still needs their hands.
The energy piece is the same piece. The factories building these machines run on power, and the states that supply it clean and cheap get the next round of advanced manufacturing. That is why Phoenix Wells and skills training are not two programs. They are one program seen from two ends.
What to watch
Three things, all with dates you can check yourself.
Whether Hyundai’s Korean unions win the prior-agreement clause, and whether any American local asks for the same language before 2028. Whether Europe’s machinery carve-out ends up excluding factory humanoids entirely when the guidance lands. And whether a single state legislature, in Oklahoma or anywhere, puts a bill number on automation displacement before the Georgia line starts moving.
We track the Oklahoma side of that in The Inference, every issue, by bill number.
The rule has about two years to exist before the robot does. Ask whoever represents you what happens in 2028, and watch whether the date means anything to them. If it does not, you have just learned something worth knowing before the next election.