Benchmarks

A Systems Market, Not a Model Market

Physical AI is not simply a model market.

Georgian + NewtonX Physical AI Survey, Wave 1 · Reporting base: Buyers, n=85 unless a benchmark states otherwise · July 24 to Aug. 21, 2026

Benchmark H3.1

The Hardest Part of Making Physical AI Work

Survey question

Thinking about getting physical AI to work reliably in the real world, which of the following is the hardest part today?

Audience
All qualified buyers, n=85
In depth
Hypothesis 3 — A Systems Market, Not a Model Market Hypothesis 4 — Evaluation Discipline as the Signal

Buyers ranked their top three, so the shares add to more than 100%.

The Hardest Part of Making Physical AI Work

69%

of robotics buyers rate integrating the robot into their existing workflow among the three hardest parts of making physical AI work. It is the highest of any option, and the most-named first choice at 32%.

All qualified buyers, n=85

Data for The Hardest Part of Making Physical AI Work
CategoryIn top three (%)Hardest of all (%)
Integrating it into the real workflow / environment69%32%
Operating it live — monitoring, diagnosing failures, updating safely55%20%
Getting enough good data and being able to simulate / test45%9%
Knowing how to measure whether it's actually working (evaluation)34%5%
Safety and security32%13%
The underlying AI model / autonomy32%11%
The robot hardware itself / on-device compute21%6%
Still evaluating / not applicable12%5%
0%20%40%60%80%Integrating it into the real workflow /environmentIntegrating it into the real workflow / environment — In top three (%): 69%69%Integrating it into the real workflow / environment — Hardest of all (%): 32%32%Operating it live — monitoring, diagnosingfailures, updating safelyOperating it live — monitoring, diagnosing failures, updating safely — In top three (%): 55%55%Operating it live — monitoring, diagnosing failures, updating safely — Hardest of all (%): 20%20%Getting enough good data and being able tosimulate / testGetting enough good data and being able to simulate / test — In top three (%): 45%45%Getting enough good data and being able to simulate / test — Hardest of all (%): 9%9%Knowing how to measure whether it'sactually working (evaluation)Knowing how to measure whether it's actually working (evaluation) — In top three (%): 34%34%Knowing how to measure whether it's actually working (evaluation) — Hardest of all (%): 5%5%Safety and securitySafety and security — In top three (%): 32%32%Safety and security — Hardest of all (%): 13%13%The underlying AI model / autonomyThe underlying AI model / autonomy — In top three (%): 32%32%The underlying AI model / autonomy — Hardest of all (%): 11%11%The robot hardware itself / on-devicecomputeThe robot hardware itself / on-device compute — In top three (%): 21%21%The robot hardware itself / on-device compute — Hardest of all (%): 6%6%Still evaluating / not applicableStill evaluating / not applicable — In top three (%): 12%12%Still evaluating / not applicable — Hardest of all (%): 5%5%
In top three (%)Hardest of all (%)

All qualified buyers, n=85

Benchmark H3.2

What Stops Buyers Scaling

Survey question

What are the biggest barriers to deploying or scaling robots / physical AI?

Audience
All qualified buyers, n=85
In depth
Hypothesis 3 — A Systems Market, Not a Model Market

Buyers ranked their top three, so the shares add to more than 100%. A “no significant barriers” option was offered and taken by nobody.

What Stops Buyers Scaling

51%

of robotics buyers name integrating with existing operations and systems among their three biggest barriers to deploying or scaling. It is the most-named barrier.

All qualified buyers, n=85

Data for What Stops Buyers Scaling
CategoryIn top three (%)Biggest barrier (%)
Integration with existing operations and systems51%16%
Upfront cost / capital required44%20%
Reliability / performance in real conditions39%11%
Security and data concerns28%8%
Difficulty changing existing workflows / processes24%8%
Unclear or unproven ROI22%7%
Workforce concerns and change management21%5%
Availability of high-quality training data20%8%
Supply chain or manufacturing capacity constraints19%5%
Safety and risk to people14%7%
Lack of internal expertise13%4%
Vendor dependence / lock-in6%1%
0%20%40%60%Integration with existing operations andsystemsIntegration with existing operations and systems — In top three (%): 51%51%Integration with existing operations and systems — Biggest barrier (%): 16%16%Upfront cost / capital requiredUpfront cost / capital required — In top three (%): 44%44%Upfront cost / capital required — Biggest barrier (%): 20%20%Reliability / performance in realconditionsReliability / performance in real conditions — In top three (%): 39%39%Reliability / performance in real conditions — Biggest barrier (%): 11%11%Security and data concernsSecurity and data concerns — In top three (%): 28%28%Security and data concerns — Biggest barrier (%): 8%8%Difficulty changing existing workflows /processesDifficulty changing existing workflows / processes — In top three (%): 24%24%Difficulty changing existing workflows / processes — Biggest barrier (%): 8%8%Unclear or unproven ROIUnclear or unproven ROI — In top three (%): 22%22%Unclear or unproven ROI — Biggest barrier (%): 7%7%Workforce concerns and change managementWorkforce concerns and change management — In top three (%): 21%21%Workforce concerns and change management — Biggest barrier (%): 5%5%Availability of high-quality training dataAvailability of high-quality training data — In top three (%): 20%20%Availability of high-quality training data — Biggest barrier (%): 8%8%Supply chain or manufacturing capacityconstraintsSupply chain or manufacturing capacity constraints — In top three (%): 19%19%Supply chain or manufacturing capacity constraints — Biggest barrier (%): 5%5%Safety and risk to peopleSafety and risk to people — In top three (%): 14%14%Safety and risk to people — Biggest barrier (%): 7%7%Lack of internal expertiseLack of internal expertise — In top three (%): 13%13%Lack of internal expertise — Biggest barrier (%): 4%4%Vendor dependence / lock-inVendor dependence / lock-in — In top three (%): 6%6%Vendor dependence / lock-in — Biggest barrier (%): 1%1%
In top three (%)Biggest barrier (%)

All qualified buyers, n=85

Benchmark H3.3

Where Deployment Cost Goes

Survey question

Roughly how is the cost of a physical AI deployment split across the following?

Audience
Qualified buyers who allocated, n=77 — the other 8 answered “still evaluating” and are excluded rather than counted as zero
In depth
Hypothesis 3 — A Systems Market, Not a Model Market

Buyers split 100 points across three categories, enforced to sum to 100.

Where Deployment Cost Goes

70%

of deployment cost goes to the robot and the AI model per robotics buyers’ own estimates, leaving 30% for the integration, evaluation, operations, safety and support work that surrounds them

Qualified buyers who allocated, n=77 · Spread across buyers: robot hardware 10–80%, AI model & autonomy 0–70%, everything around them 5–80%

Data for Where Deployment Cost Goes
CategoryMeanMedian
Robot hardware41%40%
The AI model & autonomy28%25%
Everything around them (data, evaluation, operations, integration, safety, support)30%30%
0%10%20%30%40%50%Robot hardwareRobot hardware — Mean: 41%41%Robot hardware — Median: 40%40%The AI model & autonomyThe AI model & autonomy — Mean: 28%28%The AI model & autonomy — Median: 25%25%Everything around them (data, evaluation,operations, integration, safety, support)Everything around them (data, evaluation, operations, integration, safety, support) — Mean: 30%30%Everything around them (data, evaluation, operations, integration, safety, support) — Median: 30%30%
MeanMedian

Qualified buyers who allocated, n=77 · Spread across buyers: robot hardware 10–80%, AI model & autonomy 0–70%, everything around them 5–80%

Benchmark H3.4

Where the Surrounding Spend Goes

Survey question

Within that “everything around them” spend, where does the largest share go?

Audience
Qualified buyers who put at least 30 of their 100 points on the surrounding category, n=43
In depth
Hypothesis 3 — A Systems Market, Not a Model Market

Single-select, so the shares add to 100%.

Where the Surrounding Spend Goes

70%

of the robotics buyers who spend meaningfully on the work around the robot put the largest share of that spend into live operations (monitoring, fleet management, incident response, updates) or workflow integration

Qualified buyers who put at least 30 of their 100 points on the surrounding category, n=43

Data for Where the Surrounding Spend Goes
Category%
Live operations (monitoring, fleet management, incident response, updates)37%
Workflow / environment integration and customization33%
Data collection, simulation and evaluation16%
Safety, security and compliance9%
Field support and human supervision / teleoperation5%
0%10%20%30%40%Live operations (monitoring, fleetmanagement, incident response, updates)Live operations (monitoring, fleet management, incident response, updates) — %: 37%37%Workflow / environment integration andcustomizationWorkflow / environment integration and customization — %: 33%33%Data collection, simulation and evaluationData collection, simulation and evaluation — %: 16%16%Safety, security and complianceSafety, security and compliance — %: 9%9%Field support and human supervision /teleoperationField support and human supervision / teleoperation — %: 5%5%

Qualified buyers who put at least 30 of their 100 points on the surrounding category, n=43

Benchmark H3.5

How Buyers Obtain the Capability

Survey question

How are the robots / physical AI capabilities obtained in the deployments you are closest to?

Audience
All qualified buyers, n=85
In depth
Hypothesis 3 — A Systems Market, Not a Model Market

Select all that apply, so the shares add to more than 100%. Buyers named 1.9 routes each on average.

Directional

How Buyers Obtain the Capability

71%

of robotics buyers take a complete solution from a robotics vendor — but buyers already running at scale add their own assembly on top, with 56% also using a systems integrator and 53% buying hardware separately

Packaged solutions lead across every stage, most often by buyers still piloting: 79%, compared to 69% at scale. What changes is what buyers add around these packaged solutions. 53% of scaled buyers also buy hardware separately and 56% use a systems integrator, as compared to 21% and 25% of those still piloting.

Segment bases are small (sizes shown in each column header); read differences across segments as directional, not precise

Data for How Buyers Obtain the Capability
CategoryScaled (n=32)Limited production (n=29)Piloting (n=24)
Complete solution from a robotics vendor (hardware + software)69%66%79%
Hardware bought and integrated / configured separately53%59%21%
Through a systems integrator56%38%25%
Built in-house25%3%25%
Rented / robots-as-a-service19%17%8%
Other0%0%4%
0%20%40%60%80%Complete solution from a robotics vendor(hardware + software)Complete solution from a robotics vendor (hardware + software) — Scaled (n=32): 69%69%Complete solution from a robotics vendor (hardware + software) — Limited production (n=29): 66%66%Complete solution from a robotics vendor (hardware + software) — Piloting (n=24): 79%79%Hardware bought and integrated / configuredseparatelyHardware bought and integrated / configured separately — Scaled (n=32): 53%53%Hardware bought and integrated / configured separately — Limited production (n=29): 59%59%Hardware bought and integrated / configured separately — Piloting (n=24): 21%21%Through a systems integratorThrough a systems integrator — Scaled (n=32): 56%56%Through a systems integrator — Limited production (n=29): 38%38%Through a systems integrator — Piloting (n=24): 25%25%Built in-houseBuilt in-house — Scaled (n=32): 25%25%Built in-house — Limited production (n=29): 3%3%Built in-house — Piloting (n=24): 25%25%Rented / robots-as-a-serviceRented / robots-as-a-service — Scaled (n=32): 19%19%Rented / robots-as-a-service — Limited production (n=29): 17%17%Rented / robots-as-a-service — Piloting (n=24): 8%8%OtherOther — Scaled (n=32): 0%0%Other — Limited production (n=29): 0%0%Other — Piloting (n=24): 4%4%
Scaled (n=32)Limited production (n=29)Piloting (n=24)

Segment bases are small (sizes shown in each column header); read differences across segments as directional, not precise