Team briefing · 13 August 2026

Physical AI
Landscape Report
launch

Two parts · 22 figures · Notes and slide captures: Angru

Two halves of the same competitive map, built from two completely different datasets.

Part 1 counts patents

A record of what has been claimed. It lags the actual work by 12 to 18 months, and says nothing about whether the thing was built or sold.

Part 2 counts resumes

A record of who can actually build. Current and self-reported. It runs ahead of patents as a signal, and carries whatever people write about themselves.

The audience Q&A on how the research was produced sits inside Part 1, since it describes the same study.

Part 1

Physical AI from an intellectual property perspective

Speaker  Dr Tai Cheng, Co-Founder and CEO, GoVeda · goveda.com

Host  Yao Chen, Google

Includes the audience Q&A on how the study was produced and what it cost. All 22 photographed slides come from this part.

What this part counts
237,680 Physical AI and robotics patent families
~3 million Relevant patents screened to get there
12–18 mth Publication lag behind the actual filing
A family, not a patent

One invention counted once, however many countries it was filed in. That is why the totals here are smaller than any headline "patents filed" number you will see elsewhere.

At a glance · Part 1 · 1 of 2

Counted from Physical AI and robotics patent filings

  1. Japan holds the most, China files the fastest. Japan leads accumulated volume, 87,956 families to China's 74,160. But China drove the entire rise since 2015 and now files more each year than anyone.
  2. China's patents stop at its own border. Only 7% of its families are filed outside China, against 35% for Japan and 60–69% for Europe and Korea.
  3. Japan owns the body, China and the US own the brain. 76% of Japan's families are body technologies. 72% of China's and 81% of the US's are brain.
  4. Every layer of the brain has a different leader. China leads perception, the US leads compute and models, China leads learning, Japan leads planning and runtime control by a distance.
  5. Recent body filings have already flipped to China, across all four components. That includes motors, where it has just edged past Japan, 5,991 to 5,495.
At a glance · Part 1 · 2 of 2

Counted from Physical AI and robotics patent filings

  1. Two integrators with similar counts can be opposite companies. UBTech holds 2,086 families, Apptronik 18. Unitree's portfolio is motion hardware, Figure AI's and Sanctuary's are manipulation and AI.
  2. Patents predict revenue in Japan, Europe and the US, and not in China (ρ +0.56, +0.54, +0.44, against +0.09 and not significant). Valuation tracks nothing anywhere.
  3. Six ways to monetise IP, and only one of them is selling parts. Integrated modules, platforms and orchestration, factory deployment, services and licensing, data and models, partnerships and ownership.
  4. The cost of this research is compute, not people. GoVeda ran its IPOS patent examination engine instead of hiring analysts, roughly USD 300,000 in tokens for five reports, and is offering the same API to anyone.
1.1

Three million patents, and no analysts

Patents reveal structure other market reports miss. Instead of keyword and classification searches, the team used large language models with frontier model cross-validation, and pulled in adjacent enabling technologies alongside the direct ones.

Brain

Perception, compute, models, learning, planning and control.

Body

The physical parts. Hands, motors, transmissions, power and thermal.

Integrators

The companies putting both together into products.

Physical AI was framed broadly here, not just humanoids. Much of the technology transfers across applications, so a narrow search would have missed it.

“For humans, we cannot cover that many topics. So we basically use our engines, the same engine we provide for the intellectual property office in Singapore for patent examination work.”
Audience Q&A

Time was not the bottleneck. Compute cost and methodology were. A conventional consulting approach was never on the table. The comparison offered was a Morgan Stanley humanoid report that reportedly ran on about fifty analysts.

Stated limitation

The study measures patent volume and where it sits. It does not measure patent quality, or what any single patent is commercially worth. Read every count that follows as a signal of activity and positioning.

1.2

Japan holds the most, China files the fastest

Accumulated Physical AI patent families, by holder region · all years · 237,680 total

Japan
87,956 37.0%
China
74,160 31.2%
Europe
32,112 13.5%
United States
26,483 11.1%
Korea
13,776 5.8%
Other
3,193 1.3%

Accumulated volume lags. On annual filings China drove the entire rise since 2015 and files more than anyone now. The dip at the right edge of the original chart is publication lag, not a slowdown.

1.3 · The finding with the most weight behind it

China's patents stop at its own border

Share of each region's portfolio filed internationally · full track = 100% of that region's portfolio

Other
80%
Korea
69%
Europe
60%
United States
52%
Japan
35%
China
7%

Patents are jurisdictional. Protection requires filing in each country separately, so a domestic-only portfolio protects nothing in the markets those companies want to sell into.

1.3 continued
2 in 3 Foreign applicants choose the United States. It remains the most open and attractive market.
~1 in 3 Filings into China come from foreign applicants.
  • The big cross-border flows all start in Japan, going out to the US, Europe and China.
  • Universities and research institutions account for about half of China's recent filings, with the private sector growing.
  • Those university portfolios are almost entirely domestic, at 1% to 5% international.
1.4

Japan owns the body. China and the US own the brain.

Body Brain Bars on a common axis, so composition and absolute size are both comparable
Japan
88.0k 76% body
China
74.2k 72% brain
Europe
32.2k 53% brain
United States
26.5k 81% brain
Korea
13.7k 61% brain

The US holds 4,900 body families against Japan's 67,200. This is what the brain and body split was built to expose, and none of it shows up in a single headline count.

1.4 continued

Each region concentrates somewhere specific

PerceptionComputeModels,
learning,
planning
& control		Hands &
contact
sensing		Motors &
actuators		Trans-
mission		Power &
thermal
China1.5×0.6×1.0×1.2×0.4×0.7×1.1×
United States1.3×4.5×1.4×0.5×0.3×0.1×1.6×
Europe1.1×0.5×0.4×0.6×1.1×0.9×0.5×
Japan0.4×0.1×1.1×1.2×1.7×1.7×0.6×
Korea1.0×3.2×1.1×0.5×0.8×0.1×3.0×

1.0× = in line with that region's overall share of Physical AI patents. Above 1.0 means over-represented. The US at 4.5× on compute and Korea at 3.2× compute / 3.0× power are the standout concentrations.

1.5

There is no single leader in the brain

Perception · China

48,649 families, roughly three times Europe's 15,961.

By organisation: Bosch 8,078 and Samsung 5,306 lead, then Chinese universities (Beihang, BIT, UESTC, Tsinghua, Southeast, Harbin), all at 1–5% international.

Compute · United States

4,267 families, against China 1,712 and Korea 1,600.

Intel/Altera 1,427 and Samsung 1,410 lead, roughly three times NVIDIA's 450. Moore Threads has filed 100% of its 162 families since 2021, Biren 87% of 308.

Models & control · Japan

4,533 families combined, ahead of China 3,403 and the US 1,797.

Split further it changes again: the US leads models (214), China leads learning (508), Japan leads planning (2,277) and runtime control (1,788) by large margins.

Caveat worth holding onto

Plenty of firms, particularly in the US, keep model work as trade secrets rather than patents. The visible counts understate them.

1.6

Japan leads the body, but not for much longer

Accumulated
  • Japan leads across hands, motors and transmissions.
  • Motors and actuators is the single largest category anywhere in the study: Japan 56,789 against China's 12,884.
  • The one exception is power and thermal, where China leads at 2,937, driven by CATL, LG Energy Solution and BYD.
Recent filings: the opposite story
  • China now leads all four body components.
  • Motors included, where it has narrowly overtaken Japan, 5,991 to 5,495.
  • It also holds a large share of hand manipulation and compact sensing patents, most of them out of universities.

FIG 16: Harbin Institute of Technology leads hands and manipulation on volume at 1,006 families with 0% international. FANUC combines 649 families with 76% international reach.

1.7
“There are really these quiet giants that are actually quietly dominating this space.”
Dr Tai Cheng

Leading organisations overall · families, then share filed internationally

Bosch
11,075 57%
Hitachi
10,673
Samsung
8,622
Harbin IT
top 10 · 1% intl
Beihang
top 10 · 2% intl

Three different kinds of organisation at the top: European and Japanese industrial groups, Asian electronics companies, and Chinese universities.

1.8

Same count, opposite company

2,086UBTech families
700AgiBot
36Agility Robotics
18Apptronik
  • Composition varies as much as size. Unitree leans on motion hardware, with little AI and control.
  • Sanctuary AI, Figure AI and Agility Robotics skew towards manipulation and AI.
  • Above 70% international: Samsung, ABB, Boston Dynamics, Hyundai Motor.
  • Below 20%: UBTech, Tencent Robotics X, Amazon Robotics, Xiaomi Robotics, XPeng IRON.

Headline counts mislead here. The only way through it is to read a layer down.

1.9

Patents track revenue. They do not track valuation.

Patent to revenue correlation · Spearman ρ

Japan
+0.56 n=54
Europe
+0.54 n=45
United States
+0.44 n=42
China
+0.09 n=61 · n.s.

China in neutral grey: at ρ +0.09 across 61 companies, reported as not statistically significant.

Valuation: no relationship at all
  • Skild AI: USD 14bn, almost no visible patents.
  • Figure AI: USD 39bn, roughly 50 families.
  • UBTech: USD 5.04bn, about 2,086 families.

Two groups: highly valued firms holding zero patents, and more established firms where count and valuation do move together.

1.10

Six ways strong IP owners monetise, and only one is selling parts

Integrated modules

Increase content and value per machine.

Schaeffler sells actuator plus sensor modules. Nidec commercialised Smart FLEXWAVE gearsets.

Platforms and orchestration

Own the control interface and the switching costs.

FANUC with NVIDIA connects robots to Isaac and GR00T. KUKA's iiQKA runs one OS across robot types.

Factory deployment

Use your own plants as lead customer and testbed.

BMW with Figure logged 1,250 robot-hours and 90,000+ parts.

Services and licensing

Convert installed assets into recurring revenue.

Formic's robot-as-a-service recorded 645,520 hours at 99.3% uptime. FANUC sells monitoring subscriptions.

Data and models

Turn deployment into a learning loop.

Toyota Research and Boston Dynamics are running large behaviour models on Atlas.

Partnerships and ownership

Acquire the missing layer, or monetise the portfolio.

SoftBank's pending USD 9.375bn purchase of ABB Robotics. Yaskawa acquired Tokyo Robotics.

IP creates option value. Products, factories, contracts, data and capital decide which option turns into revenue. One risk alongside it: litigation exposure to strong holders like FANUC is real, and buying premium licensed components is the least you can do about it.

1.11

They want other people to run this method

USD 300k Roughly, in tokens, for sixty billion tokens across five reports. Proprietary models bring that down a long way.

GoVeda is offering APIs and tools so others can run their own landscape reports, bigger, smaller or bespoke, and says it will help set them up and bring the cost down.

“Who will win? What happens next? How will the game evolve? The honest answer is I don't know. The research we do is try to open the discussion.”
Dr Tai Cheng, closing

The constraint on commissioning this kind of study stops being how many analysts you can afford, and becomes compute and methodology.

Part 2

Physical AI talent landscape and global talent pool signals

Speaker  Olofsson & Company, a Singapore tech and AI executive search firm · olofsson.ai

In partnership with Tai Cheng and Oliver De Gouveia / GVA

No slides were photographed for this part. The individual speaker is not named in the capture.

What this part counts
84,447Physical AI and robotics profiles analysed
~500,000Candidates in the underlying database
56,000 / 28,000Asia Pacific, and a Western benchmark set
What a profile is

Current and self-reported. It runs ahead of patents as a signal, and it carries whatever people choose to write about themselves.

“Patents show what is already owned. Talent shows where the capability to build sits right now.”
Olofsson & Company
At a glance · Part 2

Counted from Physical AI and robotics resumes

  1. 400,000 globally is not the pool you can hire from. In APAC, 56,000 profiles narrow to 5,223 people holding both AI/ML and embodied skills, and about 1,000 working at the frontier.
  2. 72% of companies say they cannot find the talent they need. The speaker had not seen a figure that high in 25 years.
  3. The one real East–West gap is embodied AI and world models, where US and EU density is close to double. Everywhere else the two pyramids are the same shape.
  4. Singapore is small and senior. 4,350 professionals, but a leadership bench that leads Southeast Asia and much of APAC. ST Engineering scores highly regionally and globally.
  5. The window is 12 to 18 months. 51% of APAC talent is open to moving, against roughly 30% in the US and Europe, and the money coming into the field is closing that gap.
  6. For volume, look past the product and AV companies to India's services firms. Keep specialist search for the frontier tier. Reaching Chinese talent needs a route.
2.2

Large in aggregate, thin where it counts

56,000 APAC profiles analysed
5,223 holding both AI/ML and embodied skills · 9.3% of the pool
~1,000 in-house embodied AI and simulation specialists · 1.8% of the pool

Bars are linearly scaled. The narrowness is the finding. The global pool is estimated at around 400,000, but it fragments the moment you ask for the combinations Physical AI work actually needs.

2.3

India is large and junior. Singapore is small and senior.

  • India has the largest visible APAC population, concentrated in software, services and offshore work at firms such as TCS and Infosys.
  • Singapore, at 4,350 professionals, leads Southeast Asia and much of APAC on seniority and leadership capability rather than raw numbers.
  • Korea shows the highest frontier density at 3.5%, but low absolute numbers.
  • China ranks second on both density and total, and is probably higher once domestic non-English speakers are counted.
  • Korea and Japan skew towards body and robotics; India and Southeast Asia skew towards software and digital.
  • Key Western hubs: the Bay Area, Boston, Austin, and parts of France, Germany, Poland and Denmark.
Stated gap in the data

Non-English-speaking markets including China, Japan and Korea carry significant dark numbers that the data underrepresents. A good proportion of the profiles that do appear show English proficiency, framed as useful for firms building global credibility.

2.4

Nobody can hire, and the window is closing

72%of companies report they cannot find talent. Highest in 25 years, per the speaker.
51%of APAC talent is open to moving, against roughly 30% in the US and Europe.
  • Major employers: the large Indian services firms, a Western AV cluster including Cruise, Aurora and Wave, and Tesla, nearly as large as that whole cluster combined.
  • ST Engineering, headquartered in Singapore, scores highly both regionally and globally.
  • Money coming in is making it worse, including a USD 300m OpenAI jobs initiative expected to absorb around 200 people.
“These next twelve months could be a window. Move them while the window is still open.”
Olofsson & Company
Before quoting any of this externally

Three things that do not reconcile

Corpus size and token count

The talk says ~3 million patents and ~60 billion tokens. The Q&A says ~3 billion patents and ~6 billion tokens.

At least one set is a speech-to-text error. GoVeda's own site claims coverage of 220 million patent publications, which rules out 3 billion and makes 3 million the plausible reading. The USD 300,000 figure is the steadiest of the four numbers.

Compute leaders

The transcript names NVIDIA, Intel, IBM, Qualcomm, AMD, Samsung and Google as an undifferentiated group.

FIG 10 shows Intel/Altera 1,427 and Samsung 1,410, against IBM 551, Qualcomm 512, AMD/Xilinx 459, NVIDIA 450, Google 273. The slide figures are used throughout this deck.

The Part 2 speaker's firm

Identified from Angru's own record of the event, not from the transcript, which says only "a specialist tech and AI talent search firm".

Olofsson & Company, Singapore, founded 2017, consistent with the speaker's account of a firm founded roughly eight and a half years ago. The individual is not named.

Regional and domain rankings are unaffected. Only the study's own methodology numbers are in dispute.

Appendix

The original
slides

All 22 figures as photographed on the day. Every chart redrawn in this deck traces back to one of these.

Appendix · 1 of 2

Source figures 1–11

Photographs of the projected slides, taken during the talk. Titles are transcribed verbatim, so American spelling and the original wording are preserved. Click any figure to enlarge.

Appendix · 2 of 2

Source figures 12–22

Photographs of the projected slides, taken during the talk. Titles are transcribed verbatim, so American spelling and the original wording are preserved. Click any figure to enlarge.

Sources
  • Part 1 talk  Physical AI Landscape Report Launch 2026-08-13 - IP Perspective
  • Part 1 Q&A  Physical AI Landscape Report Launch 2026-08-13 - Methodology QA
  • Part 2  Physical AI Landscape Report Launch 2026-08-13 - Talent Landscape
  • Figures  22 slide photographs, _attachments/physical-ai-landscape-2026-08-13/

A fourth session from the same event, a patent SaaS product launch, is not covered here. Charts in this deck are redrawn from the reported figures, not reproduced from the original slides.