Bill Gates spent most of the last decade as the most optimistic person in the room on automation. On Aug. 26, 2026, he stopped being that person. In a roughly 6,000-word essay on Gates Notes titled "A Turbulent AI Era and Critical Choices to Make," he argued that AI "will hit these industries rapidly, over the course of a decade rather than a few generations" — and proposed taxing the machines and the model calls in order to slow it down.
The trade press covered the industry's rejection. What almost nobody covered is that this policy is not hypothetical. Exactly one country has pulled a version of this lever. South Korea did it in 2018, never called it a robot tax, and the measured result should unsettle both sides of the argument: it suppressed automation investment substantially, and it did not produce jobs.
What Gates actually proposed
The essay contains three asks, and the tax is the third of them. Gates calls for new national and global institutions to govern the transition; a "Human Reserved" category of jobs that would be preserved for people by policy rather than by economics; and a tax on AI tokens and on robots, with the proceeds routed to retraining and the safety net.
His stated purpose for the tax is modest by design. As TechCrunch reported, Gates frames it as a way to "slow the rush away from human labor a little and raise money for retraining and a stronger safety net." He concedes the point an economist would raise first — that the tax is inefficient — and argues that innovation is running fast enough to absorb a little inefficiency. That is an unusual thing for a technologist to write down, and it is the honest version of the proposal: a deliberate friction, not an optimization.
It is also a reversal. Gates argued as recently as 2023 that AI-driven labor disruption would be manageable. He told Semafor he no longer believes that: "I am in a state of shock that I'm sort of the first one saying, 'This is crazy. This is insane.'"
The mechanical argument operators should actually engage with
Strip away the institutional proposals and Gates's core claim is a tax-code observation, not a philosophical one. Payroll is taxed. Capital equipment is expensed. In his words: "if you buy a robot, you can usually write it off right away as a business expense. The tax system nudges you toward replacing people with machines."
Here is the part the 2017-vintage rebuttals do not account for: that gap is wider in 2026 than when Gates first made the argument. The One Big Beautiful Bill Act, enacted July 4, 2025, permanently restored 100% bonus depreciation under §168(k) for qualifying property acquired and placed in service after Jan. 19, 2025. It added an elective 100% expensing regime under a new §168(n) for qualified production property. And it set the §179 expensing cap at $2.5 million with a $4 million phaseout threshold.
Whatever you think of the remedy, the diagnosis is stronger now than it was in 2017. A capital goods purchase is fully deductible in year one, permanently. A worker generates employer-side payroll tax every year they are on the books. Manufacturers do not need Gates to explain that asymmetry to them; they have been underwriting capex against it since the moment bonus depreciation went permanent.
The industry rebuttal — and a correction the coverage needs
Coverage of the essay has leaned on the International Federation of Robotics' line that a robot tax "aims to solve a problem that does not exist," alongside then-president Joe Gemma's formulation that "Profits, not the means of making them, should be taxed."
Those quotes are real. They are also from the IFR's statement of Feb. 27, 2017, issued in response to Gates's original robot-tax remarks — not a 2026 reply to this essay. Manufacturing Dive's August coverage recycles the language without that framing, and other outlets have followed. Gemma no longer leads the IFR, so any current title attached to the quote is wrong as well.
This matters for a substantive reason, not a pedantic one. The 2017 statement was a rebuttal to taxing installed industrial robots. Gates's 2026 proposal extends to AI tokens — a consumption-metered, software-side tax base that did not exist as a policy object nine years ago. As of this writing, the robotics trade body has not issued a fresh rebuttal addressing the token extension. Treating a nine-year-old statement as a live response papers over the fact that the industry's most-cited counterargument has not yet been updated for the thing being proposed.
The precedent nobody reports correctly: Korea, 2018
South Korea is invoked in nearly every robot-tax story as "the country that taxed robots." It did not. In 2018, Korea reduced its tax credit for investment in productivity-enhancing and automation facilities from 7% to 3% for large firms. No new levy was created. An existing subsidy was cut.
The distinction is not semantic, and operators should hold it precisely, because it changes the incidence. A credit is a direct dollar-for-dollar offset against tax owed. A deduction only reduces taxable income, so its value is the deduction times the marginal rate. Cutting a credit by four points removes four cents of cash per dollar of qualifying investment. Cutting a deduction by the same nominal amount removes roughly the marginal rate times that — far less. Reporting that describes Korea's move as a "deduction" change understates how hard it bit. Korea's quasi-robot tax was, per dollar, a heavier instrument than the shorthand suggests.
The evidence: it worked on capex and failed on jobs
The reason Korea is worth arguing about is that someone measured it. Holtmann, Braun, Cho, Koch and Langenmayr, in "Investment Effects of a Quasi-Robot Tax: Evidence from South Korea" (2025), treat the 2018 change as a natural experiment, benchmarking affected Korean industries against Japanese counterparts. As summarized by ITIF, affected industries cut industrial robot installations by roughly 28% relative to those Japanese comparators.
Two findings inside that result deserve more attention than the headline number:
-
The response came from financially unconstrained firms. The companies that pulled back were the ones with the balance-sheet room to do otherwise. They shifted investment toward employment at the margin.
-
The overall employment effect was statistically insignificant. The policy demonstrably suppressed automation investment. It did not demonstrably create jobs.
That is the spine of this story, and it cuts against both camps. Gates's critics say a robot tax cannot slow automation without wrecking competitiveness; the Korean data says it slows automation, decisively. Gates's supporters say slowing automation protects workers; the Korean data says it did not, at least not measurably. You can buy less capital formation with this lever. There is no evidence you buy headcount with it.
Run it on a plant floor
Take a worked example — illustrative, not a sourced figure, but built on realistic structure. A $1.2 million robotic cell displaces four loaded FTE at roughly $70,000 each, or $280,000 in annual labor cost.
-
Pre-tax: $1.2M ÷ $280k ≈ 4.3-year simple payback.
-
With 100% bonus depreciation at a 21% federal rate: a first-year shield of about $252,000 pulls effective net cost to roughly $948,000 — payback near 3.4 years.
-
Add a 5% equipment surcharge of the kind a robot tax implies: about $60,000, moving payback to roughly 3.6 years.
-
Apply a Korea-style four-point credit haircut instead: about $48,000 of lost offset — a similar order of magnitude.
Note what that arithmetic says. Both policy levers move payback by two to three months. On a capital committee's spreadsheet, that is noise. And yet Korea saw installations fall 28%.
The behavioral response vastly exceeds the arithmetic incidence, and the interesting question for anyone modeling this policy is why. Four candidate explanations, none of them mutually exclusive: a large share of automation decisions cluster right at the approval margin, where two months of payback flips them; the policy change carries a signaling effect that reframes automation as politically disfavored; internal hurdle rates are set well above the true cost of capital, amplifying small deltas; and every incremental point of payback adds friction inside approval committees that are already looking for reasons to defer. If the third and fourth explanations dominate, then a robot tax is not really a price signal at all — it is a procedural tax on getting projects approved.
Re-basing the market: robot demand is broadening, not collapsing
Much of the commentary around Gates's essay assumes an automation buildout running hot and unchecked, or alternatively a robotics market in retreat. Neither is what the order book shows.
Per A3 data, North American robot orders in Q2 2026 totaled 8,940 units worth $622 million — up 4.3% in units and 21.3% in value year over year. First-half 2026 came to 17,995 units and $1.166 billion, up 2.0% in units and 6.6% in value. The weakness is specific and identifiable: automotive OEM orders fell 25% in the first half versus H1 2025. Non-automotive accounted for 56% of Q2 units.
"The first half of 2026 shows how the mix of the robotics market continues to evolve," said Alex Shikany, EVP at A3. "Automotive remains an important driver of demand, while we're also seeing growth across a wider range of industries."
Value rising faster than units means buyers are purchasing more capable, higher-content systems. And the base is widening — food, logistics, life sciences and general industry are absorbing capacity that automotive is not. A tax aimed at slowing automation in 2026 would land hardest on the diversifying middle of the market, not on the automotive programs that are already pulling back for reasons that have nothing to do with tax policy.
Re-basing labor: the decoupling is real but narrower than advertised
The strongest version of the robot-tax case is that manufacturing output and manufacturing employment have come apart. The current data supports a softer claim than the debate assumes.
The July 2026 ISM Manufacturing PMI registered 55.6%, up 2.3 points from June and the highest reading since May 2022 — a seventh consecutive month of expansion. More pointedly for this argument, the ISM employment index came in at 52.8%, up 3.1 points from June's 49.7%. Susan Spence, chair of the ISM Manufacturing Business Survey Committee, noted that this put employment "in expansion territory for the first time in 33 months."
Payrolls tell a more restrained story. BLS data released Aug. 2 shows manufacturing added 5,000 jobs in July, with June revised up to +11,000 from +3,000. Transportation equipment gained 11,900; food manufacturing lost 6,200. The sector outperformed a weakening overall labor market — the total economy shed 23,000 jobs in July.
So the honest framing is this: a four-year high in output sentiment is producing marginal, positive headcount growth — not zero, and not the outright decoupling the robot-tax argument leans on. Manufacturing is hiring slowly while the broader economy sheds. That is a weaker premise for emergency tax intervention than the essay implies, and operators reading policy proposals this fall should notice when a writer skips the July prints.
The bill that is actually moving
Gates's essay is the loud artifact. The legislative one arrived first. H.R. 10044, the AI Tax and Work Protection Act, was introduced Aug. 6, 2026 by Rep. Greg Casar (D-TX-35) with Reps. Valerie Foushee (D-NC-4) and Sara Jacobs (D-CA-51) — three weeks before the Gates Notes post.
The mechanism is more specific than anything in the essay. It would tax large AI companies on the greater of token sales value or AI product revenue, at a rate indexed to the BLS unemployment rate, with proceeds funding a "Work Protection Administration." It has been referred to the Education and Workforce and Ways and Means committees.
Two structural features are worth flagging. First, indexing the rate to unemployment makes the tax explicitly countercyclical to the labor market — it bites hardest exactly when AI adoption is most plausibly displacing workers, and relaxes when it is not. Second, it targets the model layer, not the shop floor. A manufacturer buying a welding cell is not the taxpayer here; the frontier lab selling inference is. That is a materially different incidence than the 2017 robot-tax debate, and it is why the industry's recycled 2017 rebuttal does not actually answer it.
The measurement problem nobody wants to own
The robot tax is a weak answer to a real question. Gates concedes the weakness up front; his argument is that a little inefficiency is affordable and doing nothing is not. That is a defensible position to hold and a poor one to legislate from, because Korea is the only jurisdiction that has generated evidence, and the evidence says the lever hits the wrong target. You can suppress automation investment by roughly 28% and not measurably move employment. Whatever that policy is buying, it is not jobs.
But the industry's position has a gap of its own, and it is time somebody named it. "A problem that does not exist" was written in February 2017, before frontier models, before permanent bonus depreciation, before a token-metered tax base was conceivable. In the nine years since, the sector has not put forward an alternative mechanism for handling displacement at the speed Gates describes. The counterargument has been, functionally, don't.
If the trade bodies are right that a robot tax destroys competitiveness without protecting workers — and the Korean data gives them real support on the second half of that claim — then the burden is theirs to propose what does work. Retraining credits tied to installed automation. Transition obligations at the plant level. Anything with a measurable outcome variable. Otherwise the only proposal on the table with a live bill number and a scoring mechanism will be the one written by people who have never signed a capital appropriation.
Related reading
-
The Strongest Machine-Tool Half in 28 Years — and Aerospace, Not Detroit, Is Signing the POs
-
GM Walked Away From a 680-Acre Battery Plant. Samsung SDI Is Finishing It for the Grid.
Sources
-
A Turbulent AI Era and Critical Choices to Make — Gates Notes, Aug. 26, 2026
-
Bill Gates wants to see a robot tax and 'Human Reserved' jobs to mitigate harms from AI — TechCrunch
-
'This is crazy. This is insane': Bill Gates has changed his mind about AI and jobs — Semafor, Aug. 25, 2026
-
Industry rejects Bill Gates' call for robotics and AI tax — Manufacturing Dive
-
Why Bill Gates' robot tax is wrong — International Federation of Robotics, Feb. 27, 2017
-
Fact of the Week: Industries Impacted by a Quasi-Robot Tax in South Korea Reduced Industrial Robot Installations by 28 Percent — ITIF, Feb. 9, 2026
-
Investment Effects of a Quasi-Robot Tax: Evidence from South Korea — Holtmann, Braun, Cho, Koch & Langenmayr (SSRN, 2025)
-
Robot Orders Increase in Q2 as Automation Demand Broadens Across Industries — A3 via The Robot Report
-
Manufacturing PMI at 55.6%; July 2026 ISM Manufacturing PMI Report — Institute for Supply Management
-
The Employment Situation — July 2026 — U.S. Bureau of Labor Statistics
-
Manufacturing jobs continue upwards, adding 5,000 jobs in July — Manufacturing Dive
-
H.R.10044 — AI Tax and Work Protection Act, 119th Congress — Congress.gov
-
One Big Beautiful Bill Act Expands 100% Depreciation Expensing Opportunities — BDO