What Goldman Sachs' Bear Case Actually Says (and Why You Should Read It Instead of the $38B Headline)

When Goldman Sachs published their revised humanoid robot forecast in late 2024, the number that ricocheted around every pitch deck, every conference keynote, every LinkedIn post was $38 billion by 2035. We covered that base case in detail in our breakdown of the Goldman report. The $38B figure became a kind of industry password. Drop it into a conversation and you signal that humanoid robots are real, that serious people with Bloomberg terminals believe in them, and that there is an addressable market worth chasing.

But the Goldman report does not say the humanoid robot market will be $38 billion. It says the market could be somewhere between $12 billion and $152 billion, and that the analysts themselves are not sure where in that range reality will land. That is a 28x spread. If you told a venture capitalist that your startup’s total addressable market was somewhere between $12 billion and $152 billion, they would politely ask you to come back when you had figured out which one it was.

The $38 billion base case is the most psychologically comfortable number. It is large enough to justify investment but modest enough to seem credible. It sits right in the zone where you can nod along without having to confront the possibility that this entire industry might end up as a footnote, or that it might eat the entire labor market. The bear case and the bull case are the interesting ones. They force you to think about what would actually have to be true for each outcome to materialize. And right now, in early 2026, there is real evidence pointing in both directions.

This article is about the $12 billion bear case. Not because we think it is the most likely outcome, but because almost nobody reads it, and the assumptions behind it deserve serious examination.

The 28x range problem

Before we go any further, sit with this number for a moment. The gap between Goldman’s bear case and bull case is a factor of 12.7x. Expressed differently, Goldman’s own analysts are saying that the 2035 market could be roughly the size of the global dishwasher market ($12B) or roughly the size of the global semiconductor equipment market ($152B). Those are two profoundly different futures. One of them is boring. The other one reshapes civilization.

This range is not a failure of analysis. It is an honest reflection of genuine uncertainty. The humanoid robot industry is at a stage where the outcome depends on several binary questions that nobody can answer with confidence. Will foundation models for manipulation actually generalize? Will actuator costs follow a consumer electronics curve or a aerospace curve? Will regulators let these things work alongside humans in factories? Will the public accept them?

Every forecasting firm faces the same problem. But they handle the uncertainty differently.

Morgan Stanley gives a single point estimate of $29B without publishing a bear/bull range. Citi GPS offers $45B but buries their downside scenario in a footnote. ARK Invest, in their 2026 Big Ideas report, throws out a $24 trillion revenue opportunity figure that makes Goldman’s bull case look conservative by comparison. Only Goldman publishes all three scenarios with enough methodological detail to let you audit their assumptions.

That transparency is the report’s greatest value. Not the $38B number. The framework.

The bear case in full: $12 billion by 2035

Goldman’s bear case is not a doom scenario. It is not “humanoid robots fail.” It is something more subtle and, in some ways, more interesting. It is a scenario where humanoid robots work, sort of, but not well enough or cheaply enough to justify their form factor over simpler alternatives. In the bear case, the humanoid robot becomes a solution looking for problems that existing automation already solves more cheaply.

Here are the five pillars of the bear case.

Pillar 1: AI progress stalls at “good enough for demos, not enough for deployment”

The bear case assumes that the AI capabilities required for humanoid robots to perform useful work in unstructured environments do not arrive on schedule. Specifically, Goldman assumes that foundation models for robotic manipulation plateau at a level where they can handle 60-70% of subtasks in controlled demos but fail at 15-25% of real-world edge cases.

This is a critical distinction. The difference between a 95% success rate and a 75% success rate in a factory setting is not 20 percentage points. It is the difference between a useful tool and an expensive liability. A robot that drops one part in four, or that freezes when it encounters an unexpected obstacle 20% of the time, is not deployable. It requires a human babysitter, which defeats the entire economic rationale.

Goldman’s bear-case AI assumption is not that progress stops entirely. It is that progress continues but hits diminishing returns. The first 80% of capability comes fast (we are arguably there now for simple pick-and-place tasks). The next 15% takes much longer. The final 5%, the part that makes robots reliable enough for unsupervised deployment, may require architectural breakthroughs that have not happened yet.

Pillar 2: Hardware costs refuse to follow the consumer electronics curve

The base and bull cases assume that humanoid robot hardware costs decline by 15-20% annually, following a trajectory similar to flat-panel TVs, lithium batteries, or industrial robotic arms. The bear case assumes costs decline by only 5-8% annually, keeping the per-unit price above $40,000 through 2030 and above $30,000 through 2035.

Why might costs stay high? Several reasons.

Actuators are the biggest cost driver in humanoid robots, accounting for 40-55% of the bill of materials. Unlike semiconductors, which benefit from photolithography scaling, actuators are electromechanical devices that require precision machining, rare earth magnets, and custom gearing. The physics of making a motor smaller, lighter, and more powerful at the same time are genuinely hard. UBS Evidence Lab’s 2025 cost curve analysis found that actuator costs had declined by only 8% year-over-year in 2024-2025, compared to the 18% annual decline Goldman’s base case assumes.

Sensors are the second-largest cost driver. A humanoid robot needs force-torque sensors in every joint (12-20 of them), IMUs, multiple cameras, and often LiDAR. The sensor suite alone costs $3,000-$5,000 per unit at current volumes. Costs will come down with scale, but the bear case assumes that sensor integration complexity keeps total sensor costs above $2,000 per unit through 2035.

Then there is the thermal management problem. Humanoid robots generate significant heat during sustained operation. Current battery and motor technology produces enough waste heat that most humanoids can only operate continuously for 2-4 hours before needing to cool down. Solving this requires either better batteries (expensive), better heat dissipation (adds weight and cost), or accepting limited operating windows (reduces value to customers).

Pillar 3: The form factor question remains unresolved

This is the bear case’s most provocative argument. Goldman’s bear-case analysts explicitly ask: why humanoid? If the goal is to automate factory tasks, why build a robot shaped like a person when you could build a robot shaped like the task?

Industrial robotic arms have been automating factory work for decades. They are reliable, well-understood, and declining in cost. A modern 6-axis robotic arm costs $25,000-$50,000 and can operate continuously for 80,000+ hours. It does not need to balance. It does not need to walk. It does not need to manage the complex dynamics of bipedal locomotion while simultaneously performing manipulation tasks.

The standard argument for humanoid form factors is that they can navigate environments designed for humans and use tools designed for human hands. This is valid for some applications. But the bear case asks: how many of those applications justify a $40,000+ premium over a fixed robotic arm? In a factory where the layout can be redesigned, probably not many. In a warehouse where you could install a rail system, probably not many. The bear case estimates that only 15-20% of the initially addressable market truly requires a humanoid form factor. The rest could be served by cheaper, simpler alternatives.

Pillar 4: Labor markets resist faster than technology advances

The base case assumes that businesses will deploy humanoid robots wherever the economics justify it, with a roughly 3-5 year lag between cost-effectiveness and actual deployment. The bear case assumes the lag is 7-10 years, driven by three forces.

First, regulatory friction. Europe’s AI Act and evolving workplace safety regulations could require extensive certification processes for humanoid robots working alongside humans. Goldman’s bear case estimates that EU certification alone could add 18-24 months to deployment timelines and $5,000-$10,000 per unit in compliance costs. Japan’s industrial safety framework, while more robot-friendly, still requires significant adaptation for humanoid form factors that were not contemplated in existing regulations.

Second, labor union resistance. In markets where organized labor is strong, including Germany, Japan, South Korea, and parts of the United States, unions will negotiate deployment limits, retraining requirements, and phased adoption schedules. This is not speculation. It is already happening. The UAW’s 2025 contract negotiations with Ford and GM included explicit clauses limiting robotic automation in assembly lines.

Third, public sentiment. McKinsey’s automation survey data shows that public acceptance of workplace robots drops sharply when the robots look humanoid. Paradoxically, people are less comfortable working alongside a robot that resembles them than alongside an industrial arm. This psychological barrier could slow adoption in customer-facing and mixed human-robot work environments.

Pillar 5: Software revenue fails to materialize

Goldman’s base case assumes that 60% of the humanoid robot market’s total value will come from software and services, not hardware. This assumption, more than any other, drives the difference between a $12B market and a $38B market. The hardware market is roughly similar across scenarios. The software and services market is where the scenarios diverge dramatically.

The bear case assumes that software revenue per unit reaches only $200-$300 per month, compared to $500-$1,000 in the base case. Why? Because the bear case assumes that most buyers will be large manufacturers who develop their own task-specific software in-house rather than subscribing to third-party platforms. If Toyota is deploying 500 humanoid robots on its assembly lines, it is not going to pay $500/month per robot for off-the-shelf software. It is going to build its own control stack, the same way it built its own production system.

The bear case also assumes higher customer churn. If robots are not reliably performing tasks well enough (see Pillar 1), customers will cancel software subscriptions and redeploy human workers. This is the feedback loop that makes the bear case internally consistent: poor AI performance leads to low reliability, which leads to low software attach rates, which leads to lower revenue per unit, which leads to a smaller total market.

Bear vs. base: the assumptions side by side

Now that we have laid out the five pillars, let us see how the bear case and base case compare on each critical assumption.

And now the base case against the bull case.

What current evidence says about each scenario

Here is where it gets interesting. We are now in early 2026. The Goldman report’s revised forecast was published in late 2024. We have roughly 15 months of new data to test their assumptions against. Let us score the evidence.

Evidence supporting the bear case

AI manipulation is still not reliable enough for unsupervised deployment. As of March 2026, no humanoid robot is operating fully autonomously in a production environment without human oversight. Tesla’s Optimus units at Fremont and Giga Texas are performing supervised material handling tasks. Figure’s deployments at BMW are described as “human-supervised.” Apptronik’s Apollo units at Mercedes are in pilot programs, not full production. The IEEE Spectrum’s 2025 assessment of humanoid dexterity found that even the best systems achieved only 82% task completion rates on novel manipulation tasks, well below the 95% threshold Goldman’s base case requires.

Cost declines have been slower than projected. The Unitree G1 at $16,000 looks like a base-case data point, but there is a catch. The $16,000 G1 is a development platform with limited manipulation capability. A production-ready G1 with upgraded hands and additional sensors costs closer to $25,000-$30,000. The fully featured units that AgiBot and Unitree are shipping to industrial customers are priced at $35,000-$60,000, depending on configuration. This is closer to the bear case’s cost trajectory than the base case’s.

Actuator costs have not broken through. UBS Evidence Lab tracked actuator pricing through 2025 and found year-over-year declines of 8%, not the 15-20% that the base case requires. The rare earth magnet supply chain, concentrated in China, has added geopolitical risk to the cost structure. Neodymium prices rose 12% in 2025 due to export controls and demand competition from EV motors.

Total deployments are modest. As of March 2026, approximately 15,200 humanoid robots have shipped globally across all manufacturers. That is real and meaningful, but it represents a cumulative total, not annual volume. Goldman’s base case requires annual shipments to reach 100,000 units by 2030. Current annual run rates are closer to 8,000-10,000 units. Reaching 100,000 annual units in four years requires roughly 75-85% year-over-year growth, sustained. Possible but aggressive.

Evidence supporting the base case

The cost trajectory is heading the right direction. While costs have not declined as fast as the base case projected, they are declining. Unitree’s G1 pricing represents a genuine step-change from the $100,000+ pricing of 2023. AgiBot’s manufacturing scale in Shanghai is demonstrating that mass production of humanoids is feasible. The question is the rate of decline, not the direction.

China’s industrial policy is accelerating supply. The MIIT’s humanoid robot development plan, combined with provincial subsidies in Guangdong, Zhejiang, and Shanghai, is creating a policy environment that directly targets cost reduction through scale. Goldman’s base case assumed this kind of government support. It is materializing.

AI capabilities are improving, even if not at base-case speed. Google DeepMind’s RT-X, Nvidia’s GR00T, and various open-source robotics foundation models have shown genuine progress in generalization. A Nature paper published in late 2025 demonstrated that transformer-based manipulation models could transfer skills across different robot morphologies with 87% fidelity, up from 60% two years earlier. This is not base-case performance yet, but the trajectory is encouraging.

Enterprise demand is real. BMW, Mercedes, Amazon, BYD, SAIC, and Foxconn have all placed orders or signed pilot agreements for humanoid robots. These are not speculative commitments from startups chasing hype. These are capital allocation decisions by companies with strict ROI requirements. The demand side of Goldman’s model appears to be tracking.

Evidence supporting the bull case

Foundation models are compounding. ARK Invest’s 2026 Big Ideas report argues that robotic foundation models are on the same exponential improvement curve that language models followed in 2020-2023. If true, the gap between current capability and reliable autonomous deployment could close in 2-3 years rather than 5-7 years. This is the bull case’s central bet, and it is not obviously wrong, even if it is optimistic.

Unitree’s consumer play is working. Unitree’s G1 is the first humanoid robot with meaningful consumer sales. While the majority of units are going to researchers and developers, an estimated 10-15% of G1 sales in late 2025 were to individual consumers. If the consumer market opens before 2035, the bull case’s unit volume assumptions become much more plausible.

The cost curve: where the scenarios live or die

If there is a single variable that most determines which scenario materializes, it is the cost curve. Goldman’s own sensitivity analysis shows that a 10% change in the assumed annual cost decline rate shifts the 2035 market size by $8-12 billion. No other variable has this much leverage.

Let us look at why.

The economics of humanoid robot adoption are driven by a simple comparison: does this robot cost less than the human worker it replaces? The answer depends on three things. The robot’s upfront cost, its annual operating cost (maintenance, energy, software), and the fully loaded annual cost of the human worker it displaces.

Goldman’s model assumes that a humanoid robot needs to achieve a 2-year payback period to be attractive to enterprise buyers. This means the robot’s total cost of ownership over two years needs to be less than two years of the human labor it replaces.

In the bear case, with a $42,000 robot and $8,000 in annual operating costs, the 2-year TCO is roughly $58,000. That is cost-effective compared to a German or American manufacturing worker, but not compared to a Chinese, Vietnamese, or Indian one. The bear case’s addressable market is essentially limited to high-wage countries and high-precision tasks.

In the base case, with a $25,000 robot and $5,000 in annual operating costs, the 2-year TCO is $35,000. This opens the Chinese market, which is the largest manufacturing labor force in the world. This is why the base case’s unit volumes are so much higher than the bear case’s, not because more types of tasks are addressed, but because the China market flips from “too expensive” to “cost-effective.”

In the bull case, with a $14,000 robot and $3,000 in annual operating costs, the 2-year TCO is $20,000. At this price point, humanoid robots become competitive even in middle-income manufacturing environments and the consumer market opens for household applications at price points comparable to a used car.

This is why the cost curve is not just one variable among many. It is the variable that determines the size of the addressable market. A $42,000 robot addresses a $12 billion market. A $25,000 robot addresses a $38 billion market. A $14,000 robot addresses a $152 billion market. The technology and the demand are secondary to the question of how fast costs come down.

What the bear case gets right that bulls ignore

It is fashionable in robotics circles to dismiss the bear case as lacking imagination. But several of its assumptions deserve more respect than they get.

The reliability gap is real and underappreciated

The gap between “impressive demo” and “production-ready deployment” is one of the most consistent failure modes in robotics history. Rethink Robotics’ Baxter was impressive in demos. It failed commercially because it was not reliable enough for production environments. SoftBank’s Pepper was impressive in demos. It was discontinued because it could not reliably perform useful tasks. The history of robotics is littered with machines that worked 80% of the time and were therefore useless for 100% of commercial applications.

The bear case assumes that humanoid robots could follow this same pattern, only at a larger scale and higher cost. A robot that can perform a task successfully 80% of the time in a controlled demo might only achieve 60% in a real factory with variable lighting, unexpected obstacles, and parts that do not arrive in the expected orientation. The remaining 40% requires human intervention, which means you need the human worker anyway, which means the robot is a cost addition rather than a cost substitution.

The form factor argument has historical precedent

The bear case’s skepticism about the humanoid form factor echoes a real pattern in industrial automation. When automated guided vehicles (AGVs) were first introduced, early designs mimicked the form factor of forklifts. Over time, the industry discovered that purpose-built AGVs that looked nothing like forklifts were more efficient, cheaper, and more reliable. The forklift form factor was a transitional design that helped people conceptualize the technology, not the optimal engineering solution.

The bear case asks whether humanoid robots are in the same transitional phase. Maybe the answer to “how do we automate warehouse picking?” is not “build a robot that looks like a human picker” but “redesign the warehouse so that non-humanoid robots can do the picking more efficiently.” Amazon has been doing exactly this with its warehouse designs, optimizing layouts for robot mobility rather than human mobility.

The China price problem is double-edged

Bulls love to cite China’s aggressive humanoid robot production as evidence for rapid cost decline. The bear case flips this argument. If Chinese manufacturers drive unit prices to $15,000-$20,000 but Western companies cannot compete at those margins, the market could be large in unit volume but small in revenue. A market where 500,000 units ship at an average selling price of $20,000 is a $10 billion hardware market. Add moderate software revenue and you are at $12-15 billion total. That is the bear case, even with aggressive unit volumes.

This is a scenario where the bears are right about the total market value even though the bulls are right about the number of robots deployed. Both sides could be correct about different metrics and still arrive at very different market size numbers.

What the bull case sees that bears miss

Fair is fair. The bear case has blind spots too.

Exponential AI progress is not a fantasy

Bears tend to assume that AI progress is roughly linear, that each percentage point of task success rate improvement requires the same amount of research effort as the last one. But the history of AI over the past decade suggests otherwise. Language models went from “occasionally coherent” to “pass the bar exam” in about four years. Image generation went from “blurry nonsense” to “photorealistic” in about three years. If robotic manipulation follows a similar curve, the gap between current capability and production-ready reliability could close much faster than the bear case assumes.

The Nature paper on foundation model transfer in robotics, published late 2025, showed that cross-morphology transfer learning improved from 60% to 87% fidelity in just 18 months. If that rate of improvement continues, we reach the 95%+ threshold Goldman’s base case requires by 2028, not 2032.

Cost curves tend to surprise on the downside

The bear case assumes 5-8% annual cost decline for humanoid robots. But historical data from analogous industries suggests that once manufacturing reaches scale, cost declines accelerate rather than decelerate. Solar panels declined in cost by 99% over 40 years, with most of that decline happening in the last 15 years as manufacturing scaled. Lithium-ion batteries declined by 97% over 30 years, with the sharpest drops coming after 2015 when gigafactories came online.

Humanoid robots are currently at the pre-scale phase. Annual production is in the low thousands. The bear case’s 5-8% decline rate might be accurate for this phase. But if production reaches tens of thousands annually (which even the bear case concedes is possible by 2030), historical patterns suggest cost declines could accelerate to 15-25% annually, which pushes the trajectory toward the base or bull case.

The platform effect could be massive

The bear case assumes that software revenue per unit will be modest because large enterprises will build in-house solutions. This is plausible for the first wave of adopters. But the bull case sees something the bear case misses: the platform effect.

If humanoid robots become standardized enough that third-party developers can build task-specific applications for them, the software revenue model could look less like enterprise SaaS and more like the iOS App Store. Imagine a marketplace where a manufacturer can download a “quality inspection” skill for $200/month and a “palletizing” skill for $150/month and a “machine tending” skill for $300/month, each developed by a specialist vendor. In this scenario, software revenue per unit could exceed Goldman’s base case assumptions.

This is not happening today. But neither was the App Store happening in 2005 when the iPhone was still two years away. Platform effects are difficult to see until they ignite, and then they grow faster than anyone expected.

The labor shortage is a forcing function

The bear case’s assumption about labor market resistance overlooks a demographic trend that is already reshaping manufacturing. Germany, Japan, South Korea, and China are all facing manufacturing labor shortages driven by aging populations and declining interest in factory work among younger generations. Japan’s manufacturing workforce has declined by 15% since 2010. Germany’s is projected to shrink by 7% by 2030. China’s working-age population peaked in 2015 and is declining.

In this context, humanoid robots are not replacing unwilling workers. They are filling positions that cannot be filled with human workers at any wage. This changes the adoption dynamics completely. When the alternative to deploying a robot is not “keep the human worker” but “leave the position unfilled and lose production capacity,” the economic calculus shifts dramatically in favor of adoption, even at bear-case cost levels.

Scoring the scenarios: where do we actually stand?

Let us attempt something that most forecast coverage does not. Let us score each of Goldman’s key assumptions against the evidence available in early 2026.

Where 1 = tracking below bear case, 5 = tracking at base case, and 10 = tracking at bull case, here is our reading:

AI progress: 5/10. Right at base-case trajectory. Real progress but not the exponential jump the bull case requires.

Cost decline: 4/10. Between bear and base. Actuator costs are tracking bear case. Total system costs are declining faster thanks to Chinese manufacturing scale, but not as fast as the base case projected.

Unit volumes: 5/10. Roughly at base-case trajectory. The 15,200 cumulative units shipped by early 2026 are consistent with reaching the base case’s 250,000-300,000 by 2035 if growth rates hold.

Software revenue: 3/10. Closer to bear case. No humanoid robot company has yet demonstrated a $500+/month software subscription model at scale. Most revenue is still hardware sales with basic maintenance contracts.

Regulatory pace: 4/10. Between bear and base. The EU is moving cautiously. The US is permissive but inconsistent across states. China is supportive. Japan is supportive. The global average is slower than the base case assumed.

Enterprise demand: 7/10. Tracking above base case. The quality and number of enterprise pilot programs has exceeded what Goldman projected for this stage. BMW, Mercedes, Amazon, BYD, and Foxconn are not waiting for perfect technology. They are deploying now and iterating.

Consumer market: 3/10. Closer to bear case. Unitree’s G1 has consumer sales, but these are overwhelmingly tech enthusiasts and developers, not mainstream consumers buying a home robot. The consumer market is not opening on the bull case timeline.

The five questions that will determine which scenario wins

Rather than predicting an outcome, here are the five binary questions whose answers will determine whether we end up closer to $12B, $38B, or $152B. Track these over the next three years and you will have a better forecast than any investment bank can give you today.

Question 1: Do manipulation foundation models generalize by 2028? If yes, the base case becomes the floor, not the ceiling. If no, the bear case becomes increasingly likely. The specific threshold to watch is whether any humanoid robot can perform a novel manipulation task (one it was not explicitly trained on) with 90%+ success rate in a real factory environment. As of today, we are at roughly 70-75%.

Question 2: Does any manufacturer ship 10,000 units in a single calendar year by 2028? This is the production scale threshold that triggers meaningful cost curve effects. AgiBot and Unitree are both plausibly on track for this, but neither has done it yet. If this happens by 2028, the base case cost assumptions become more plausible. If it does not happen until 2030 or later, the bear case cost trajectory is more likely.

Question 3: Does software subscription revenue exceed $300/month per unit at any company by 2028? This is the test of Goldman’s base case revenue model. If some company demonstrates $300+ monthly recurring revenue per deployed robot at scale (meaning 1,000+ units), the base case software assumptions are validated. If nobody achieves this, the bear case revenue model is more accurate.

Question 4: Does China export humanoid robots to Western markets in volume by 2029? Chinese cost advantages are real. If Chinese humanoids are available in the US and Europe at $15,000-$25,000, it accelerates the cost curve and expands the addressable market. But trade policy, tariffs, and national security concerns could block or restrict Chinese humanoid imports. If China’s cost advantages remain bottled up in the domestic market, the global cost curve is slower.

Question 5: Does any humanoid robot achieve a consumer application that generates word-of-mouth demand by 2030? The bull case depends on the consumer market opening. This does not require a $15,000 home robot doing laundry. It requires any humanoid robot doing any task in a home or consumer environment that makes regular people say “I want one of those.” If that moment happens before 2030, the bull case timeline is in play. If it does not happen until after 2035, it is not.

Why you should read the bear case

We titled this article “Why You Should Read the Bear Case” and we want to close by explaining exactly why.

Reading the bear case does not make you a pessimist. It makes you a better-calibrated thinker. The bear case forces you to articulate what would have to go wrong for the humanoid robot market to disappoint. And when you articulate those failure modes, you discover something surprising: most of them are plausible. Not certain. Not likely. But plausible enough that any honest analysis must account for them.

The $38 billion base case has become so ubiquitous that it has started to function as a consensus estimate rather than a probability-weighted midpoint. People treat it as “the Goldman number” rather than “one of three Goldman numbers.” This creates a risk of groupthink, of an entire industry planning around a specific outcome that the forecasters themselves assign less than 50% probability to.

Goldman’s base case is the single most likely scenario, but “most likely” in a distribution this wide means something like 40% probability. The bear case might be 25% probable. The bull case might be 20% probable. And outcomes that do not fit neatly into any of the three scenarios account for the remaining 15%.

If you are an investor deciding whether to put capital into humanoid robotics, you need to know what the downside looks like. A $12 billion market is still investable, but it demands different companies, different strategies, and different return expectations than a $38 billion market. If you are a manufacturer deciding whether to build humanoid robots or specialized automation, the bear case tells you which market segments are robust to scenario risk and which ones only work if the base case materializes.

And if you are simply someone who wants to understand where this technology is going, the range from $12B to $152B tells you something that the base case alone cannot: we genuinely do not know. The people whose job it is to predict this, who have access to proprietary data and sophisticated models, are telling you that the outcome could vary by more than an order of magnitude. That uncertainty is not a bug in the analysis. It is the analysis.

The Goldman report’s real contribution is not a number. It is a framework for thinking about which assumptions matter and which evidence to track. The $38B headline is useful shorthand. The $12B to $152B range is useful knowledge. Read the bear case. It will teach you more about the future of humanoid robots than the headline ever could.