The Robots Arrived as Demos, the Labor Replacement Didn't - NIA Trend 3, Revisited at the Half-Year

Held against reality: the robots came as demos, and the labor replacement never reached the macro numbers (🟡, the series’ first shortfall) Placing NIA’s two predictions next to six months of what actually happened. Prediction 1, hardware and AI convergence, did proceed but split by domain (🟢). Prediction 2, labor replacement and maximized productivity, stayed in demos and pilots and never surfaced in macro indicators (🟡). The bottleneck moving from hardware to data and unit economics, and value flowing to the software layer, appear nowhere in either prediction (⚪).

As set out in Part 0, I hold each of NIA’s two prediction sentences for Trend 3 against what actually happened. Where the previous two parts landed at 🔵 and 🟢, this one is the series’ first 🟡.

NIA’s Trend 3, “AI Enters the Physical World: The Physical AI Revolution,” carried two predictions.

• Prediction 1: Hardware advances (robots, autonomous vehicles, smart manufacturing) accelerate AI convergence.
• Prediction 2: Labor-intensive sectors reduce their dependence on people, producing structural cost reduction and maximized productivity.

One link runs underneath both: that hardware advancing leads directly to labor being replaced and productivity being maximized. Six months of developments confirmed the front half of that link (hardware advancing) but the back half (labor replacement) has not yet reached the macro indicators. The lag between them opened far wider than in the earlier parts, which dealt with software.

The half-year, in one line: robots did arrive, in the form of demos and pilots, but the bottleneck moved from hardware to data and unit economics, and the labor-replacement result has been pushed out to 2027 and 2028. The world of atoms moves more slowly than the world of bits.

Prediction 1, Hardware and AI convergence - underway, but split by domain (🟢)

Prediction 1 did in fact proceed over the half-year. What it did not do was proceed at one speed. Robotaxis and humanoids went at different rates.

Robotaxis outran the prediction. Waymo took weekly paid rides to 500,000, a tenfold rise in two years from 50,000 a week in May 2024. Its operating cities grew to ten, and on 8 July 2026 it added San Diego, Las Vegas, Tampa, and Denver while announcing entries into the UK and Japan. The company is valued at $126 billion. China’s Baidu Apollo Go passed 20 million cumulative rides globally, handled 3.2 million fully driverless rides in Q1 alone, and entered Seoul in early 2026. Tesla’s Austin robotaxi, by contrast, expanded across the metro in June but runs roughly 20 vehicles in practice, with analyses putting its incident rate at around four times worse than human drivers.

Even in robotaxis, though, the reliability problem from Part 2 shows up in the same shape. Waymo issued a fleet-wide software recall over robotaxis failing to stop for a halted school bus, and highway driving has been suspended since May 2026 pending a software update. Over eight months, as the fleet grew from 2,000 vehicles to 3,791, safety regressions arrived alongside. Capability came faster than predicted, but how far that capability can be trusted and handed work is still being adjusted, recall by recall.

Humanoids sit a step behind robotaxis. The demos and proof-of-concept deployments are impressive. Figure’s robot spent ten months at BMW’s Spartanburg plant helping produce 30,000 X3 units, moving 90,000 parts and clearing 99% placement accuracy per shift. On that record Figure signed a commercial contract for 40 units of its 03 model, priced at roughly $25 per robot operating hour. China’s Unitree shipped 5,500 humanoids in 2025 to take the global top spot (32.4% share) and received final approval for its Shanghai STAR listing on 3 July 2026, at a valuation of about $6.2 billion, with 2025 revenue up 335% and a first profit of $90 million. Across China as a whole, 14,400 humanoids shipped in 2025, or 84.7% of global volume.

But there is a gap between demo and real deployment that still has to be read carefully. Reporting has repeatedly noted that a substantial share of robot demonstration footage is teleoperated by a human, and BMW pushed back on work Figure had presented as “end-to-end” autonomous, saying it took place during downtime rather than on a live line. Boston Dynamics’ Atlas has its entire 2026 volume committed to Hyundai and DeepMind, with no outside customers until 2027, and Apptronik’s Apollo is at a pilot stage moving parts bins at a Mercedes plant. Volumes and contracts accumulate; whether that volume is genuinely working autonomously still needs one layer peeled back.

On the numbers alone this looks like a hit. But the most symbolic case, Tesla’s Optimus, changes the picture. Optimus V3 had still not entered mass production as of July 2026. The Model S and X lines at Fremont ended in May and are being converted for Optimus, with first production set for late July into August. Elon Musk, walking the newly installed line on 1 July, denied observations that output had quietly begun and said it would be “extremely slow at first because everything is new.” A few hundred units exist so far, none have been sold, and analyst consensus puts consumer sales at late 2027 at the earliest and realistically 2028 to 2029. Musk himself conceded on the January earnings call that Optimus is not “being used in our factories in any meaningful way.”

So Prediction 1 lands at 🟢. The direction, hardware advancing and converging with AI, did proceed, but robotaxis ran ahead while humanoids remain in demos, proof-of-concept deployments, and data collection.

Prediction 2, Labor replacement and maximized productivity - absent from the macro (🟡)

Prediction 2 is different. The result it describes, dependence on labor falling and productivity being maximized, is not visible in the macro indicators for the half-year.

Individual success stories are plentiful. Some of the 223 Lighthouse factories identified by the World Economic Forum reported productivity gains of 40% and lead-time cuts of 48%, and Amazon deployed its millionth warehouse robot in July 2025, reaching a point where robot count rivals its roughly 1.5 million human employees. Robots are involved in 75% of orders.

The problem is that all of these cases added together do not move the macro numbers. US total factor productivity (TFP, the efficiency gain not explained by labor and capital inputs) rose only 0.8% in 2025, and in 2024, 52 of 86 detailed manufacturing industries saw productivity fall. Lighthouse factories are a selection of the ones doing well, and surveys by McKinsey and the World Economic Forum put 74% of automation pilots stuck in pilot purgatory, with only around 30% crossing into actual rollout. Research by the economists Acemoglu and Restrepo finds that automation from 1980 to 2016 raised total factor productivity by just 3.4%. For “maximized productivity” to be confirmed at the macro level, there is a long way to go.

---
config:
  look: handDrawn
  theme: neutral
---
flowchart LR
    A["Site-level demo<br/>Lighthouse +40%"] -->|"rollout ~30%"| B["Pilot stall<br/>74% pilot purgatory"]
    B --> C["Macro indicator<br/>US TFP +0.8%"]
    C --> D["Labor-replacement threshold<br/>2027-2028 consensus"]

The labor-side signal is also still partial. Employment in production and lower-wage occupations fell between 2024 and 2025, but that decline came from expanded outsourcing as well as automation. US manufacturing still has over a million unfilled positions, and about 22% of manufacturers said they plan to adopt physical AI such as robot dogs or humanoids by 2027. Internal Amazon documents were reported to contain a goal of avoiding 600,000 hires by 2033, though the company countered that the document was incomplete. The direction matches the prediction, but nothing on the scale you would call “structural cost reduction and maximized productivity” has appeared yet. Hence 🟡: direction right, magnitude and speed short.

What NIA missed entirely - the bottleneck moved and value flowed to software (⚪ unmapped)

Two changes have no name in either prediction. NIA connected hardware advances straight to labor replacement, and six months of actual developments exposed two new layers sitting in between.

The first is that the bottleneck moved from hardware to data and unit economics. Building a robot body is no longer the hardest part. The real bottleneck is the shortage of data for robots to learn from (the so-called absence of a robotics internet) and the gap that opens when what was learned in simulation is transferred to a real environment (sim2real). Unit economics is still a wall too. The target unit cost for a humanoid is $20,000 to $30,000, while Optimus is estimated at around $150,000 today, and roughly 70% of the bill of materials (BOM) is Chinese components, so removing China pushes cost from $46,000 to $131,000. What decides the contest is not whether you can build the body, but the data it can learn from and the cost you can bear.

The second is that value is flowing to the software layer. Capital is already pointing that way. Funding into robot foundation models in H1 2026 came to $3.92 billion across nine deals, a half-year record, and AI that handles the physical world began drawing valuation multiples at software-company levels. Vision-language-action models (VLA, neural networks that take camera images and language instructions together and emit motor commands), which serve as the robot’s brain, turned robots into a software platform. Physical Intelligence raised at an $11 billion valuation, Skild AI was valued at $14 billion to $15 billion at $30 million of revenue, and Nvidia’s GR00T and Cosmos compete with DeepMind’s Gemini Robotics over this layer. There is a telling admission: even Unitree, first in volume, wrote in its listing documents that it had “not applied embodied large models at scale.” Motion ran ahead and the AI lagged. For reference, China’s AgiBot hit its 10,000th unit in March 2026 and leads on volume alongside Unitree, but the same admission applies across the industry.

This software layer then splits into open and closed camps. Public models such as OpenVLA and SmolVLA compete with closed ones such as π0, Gemini Robotics, and GR00T, and both sides carry the same verification problem: much of what is presented as “zero-shot” (performing a new task with no prior training) came out of controlled environments. The gap between running well on a benchmark and running every time on a factory floor has the same shape as the agent reliability problem from Part 2.

Neither layer featured much in the discourse of late 2025. A method that extracts trends from text frequency will catch visible hardware such as “robots and autonomous vehicles” but is late to the data bottleneck underneath and the shift in where valuation weight sits. After the infrastructure bottleneck in Part 1 and the reliability gap in Part 2, what the prediction caught last in Trend 3 was again the less visible layer below.

One thing worth noting - first in parts, importing the brain

Mapping this comparison onto Korea, the structure seen in Trend 2 repeats once more in the physical world.

Korea’s hand is strong. Robot density is 1,220 units per 10,000 workers, first in the world (Singapore second at 818, Germany at 449), and the strengths in core humanoid components are clear. Joint actuators come from Hyundai Mobis, batteries from LG Energy Solution, SK On, and Samsung SDI, and AI chips from Rebellions and DeepX. In the parts that make up the body, Korea genuinely ranks near the top.

The problem occurs one layer up. Deployment timelines for finished humanoids have sat at 2028 for six months running. Hyundai unveiled an electric-drive Atlas at CES 2026 and said it would phase the robot into US plants from 2028, and Rainbow Robotics, where Samsung became the largest shareholder with a 35% stake, also cites 2028 for commercial use. The government’s K-Humanoid Alliance has committed a combined public and private ₩1 trillion through 2030, targeting global leadership by 2030. That figure falls short of a single funding round at one overseas company. Skild AI took $1.4 billion in one round.

So Korea’s position reduces to one line. The parts and the motion rank among the world’s best, while in the brain layer where the value is actually rising, Korea is the one following. The structure read in Trend 2 as “first in adoption, importing the standards” appears in Trend 3 as “first in hardware, importing the brain.” There are signals of expectation running ahead, too. On anticipation of Samsung raising its stake, Rainbow Robotics’ price-to-earnings ratio (P/E) jumped into the thousands, and Doosan Robotics carries a market capitalization several times revenue while posting operating losses. This is the shape of what gets called an expectation trade: pricing on anticipation rather than results.

Field validation has already begun. Rainbow Robotics’ mobile dual-arm robot RB-Y1 went into a Coupang fulfillment center on a trial basis, moving past showcase demos into industrial validation. One domestic variable attaches to that transition, though. The Korean Metal Workers’ Union came out against Atlas entering factories in January 2026, and the summer 2026 collective bargaining round is seen as the turning point. Separately from labor replacement not showing up in the macro numbers, the pace of adoption on the ground is decided by labor agreements as much as by technology.

On autonomous driving the point is more direct. China’s Apollo Go entered Seoul in early 2026, making Korea its first Asian market, while domestically the detour has been toward autonomous buses rather than robotaxis. Hyundai and Kia are preparing Korea’s first large-scale pilot, but the gap in accumulated data against Waymo and Apollo Go is wide. That mismatch, a country ahead in components trailing in services and in the brain, is the question this six-month comparison puts to Korea.

Closing - the demo arrived, the result comes at the speed of atoms

Held against the prediction, Trend 3 comes out like this.

Prediction 1 (hardware and convergence) proceeded but split robotaxis from humanoids, and Prediction 2 (labor replacement) stayed in demos and pilots without reaching the macro indicators. The two shifts, the bottleneck moving and value moving to software, were outside the prediction.

Just as in Part 1 the contest over infrastructure was decided not by the chip but by the bottlenecks, and in Part 2 the contest over agents was decided not by autonomy but by trust, the contest in Trend 3 had also moved away from the hardware itself toward the data that teaches the robot, the unit cost you can bear, and the software that forms its brain. NIA connected hardware advances straight to labor replacement, but the distance between them is longer in a world of atoms than in a world of bits.

That Trend 3 is the series’ first 🟡 is not an accident. The physical world opens the widest gap inside a six-month window. Software ships weekly, while retooling a factory line, installing a transformer, and teaching a robot from data move on quarterly and annual clocks. That a one-year forecast misses hardest precisely where the world is slowest is the shape of the lag Trend 3 reveals. The demo has already arrived. The result comes at the speed of atoms.

In the next part I hold Trend 4, the convergence of 6G and satellite communications, to the same method.

The compared source is NIA, “The 12 AI and Digital Trends NIA Forecast for 2026” (IT & Future Strategy No. 6, 2025.12.31). Evidence for H1 2026 developments prioritizes primary sources: the International Federation of Robotics (IFR), BofA, Goldman Sachs, McKinsey, the World Economic Forum, and the US Bureau of Labor Statistics, IR from Tesla, Figure, Unitree, Waymo, and Baidu Apollo Go, Nvidia, DeepMind, Physical Intelligence, and Skild AI, Hyundai Motor, Samsung Electronics, Rainbow Robotics, and Doosan Robotics, the Acemoglu and Restrepo papers, and Korea’s Ministry of Trade, Industry and Energy.