Goldman Sachs Says $38 Billion by 2035: Breaking Down the Forecast

In early 2024, Goldman Sachs published an 88-page report titled “Rise of the Humanoids” that projected the humanoid robot market would reach $6 billion by 2035. Twelve months later, they revised that number upward to $38 billion, a 6x increase. That revision was not a minor tweak. It reflected a fundamental rethinking of how fast humanoid robots could penetrate labor markets and how quickly costs would come down.

The Goldman report has become the most cited market forecast in the humanoid robotics industry. Every company pitch deck, every investor presentation, every conference keynote references some version of the “$38 billion by 2035” number. But most people who cite it have not read the full report. Here is what it actually says, how they modeled it, and where the assumptions could break.

Why the revision happened

Goldman’s original $6 billion forecast was published before several key developments.

First, the cost curve moved faster than expected. When the original report was written, the cheapest commercially available humanoid robot was around $100,000 (UBTECH’s Walker X enterprise pricing). By mid-2024, Unitree had launched the G1 at $16,000 and AgiBot was shipping units at roughly $28,000. Goldman’s original model assumed per-unit costs of $50,000-$100,000 through 2030. Reality was already undercutting that by 3-5x.

Second, China’s production ramp surprised everyone. Goldman’s original model assumed gradual scaling from a handful of companies. The MIIT policy push and the emergence of AgiBot, combined with Unitree’s aggressive pricing, meant the supply side was scaling much faster than the model predicted.

Third, the AI capability leap was larger than expected. The integration of large language models and multimodal foundation models into robot control systems progressed faster in 2024-2025 than Goldman’s robotics analysts had anticipated. This expanded the range of tasks humanoid robots could perform, which expanded the addressable market.

How Goldman models the market

Goldman’s forecasting methodology follows a top-down labor market approach combined with bottom-up production capacity analysis. Understanding the methodology is crucial for evaluating the forecast’s credibility.

Step 1: Total addressable labor market

Goldman starts with the global labor force in sectors where humanoid robots could substitute for human workers. They focus on four sectors: manufacturing (330 million workers globally), warehousing and logistics (90 million), commercial services (60 million in relevant roles), and eventually consumer/household (not included in the 2035 number but mentioned as upside).

They estimate that by 2035, approximately 4% of addressable manufacturing jobs and 2% of logistics jobs could technically be performed by humanoid robots. This gives a total addressable unit volume of roughly 1.4 million robots.

Step 2: Penetration rate and timing

Goldman does not assume all technically addressable jobs will be penetrated by 2035. Their base case assumes humanoid robots will fill approximately 250,000 positions by 2035, representing about 18% of the technically addressable market. This is the deployment curve, not the theoretical maximum.

Step 3: Unit economics

This is where the revision was most dramatic. The original model assumed per-unit revenue of roughly $25,000 (hardware sale only). The revised model assumes per-unit lifetime revenue of approximately $150,000, including hardware ($30,000-$50,000), AI software subscriptions ($500-$1,000/month), maintenance contracts ($3,000-$5,000/year), and task-specific application licensing.

Step 4: Production capacity constraint

Goldman acknowledges that production capacity, not demand, will be the binding constraint through at least 2030. Their model assumes global production capacity reaches 100,000 units per year by 2030 and 500,000 units per year by 2035. These numbers assume significant capital investment by multiple manufacturers.

The three scenarios

Goldman presents three scenarios: bear, base, and bull. Most people only cite the base case. The range is important.

Bear case: $12 billion

The bear case assumes AI progress stalls, per-unit costs remain above $40,000, and humanoid robots are limited to simple repetitive tasks in structured environments. In this scenario, humanoid robots are essentially expensive fixed-automation alternatives with legs, and most manufacturers conclude that traditional robotic arms are more cost-effective. Global deployment reaches roughly 100,000 units by 2035.

Base case: $38 billion

The base case assumes steady AI improvement, per-unit hardware costs declining to $20,000-$30,000 by 2030, and humanoid robots performing a meaningful range of manipulation and navigation tasks in semi-structured environments. Software and services revenue accounts for 60% of total market value. Global deployment: 250,000-300,000 units.

Bull case: $152 billion

The bull case assumes rapid AI breakthroughs, per-unit costs below $15,000 by 2030, and the consumer market opening before 2035. In this scenario, humanoid robots become as ubiquitous in factories as forklifts, and home robots begin penetrating the consumer market. Global deployment exceeds 1 million units. Goldman flags this scenario as “requires multiple concurrent breakthroughs” but not impossible.

What the model gets right

Goldman’s framework makes several assumptions that appear well-supported by current evidence.

Labor substitution is the right frame. Humanoid robots will be valued primarily based on the labor they replace, not as standalone products. This aligns with how every major humanoid robot company is pitching to enterprise customers.

Software and services will drive most of the value. Hardware commoditization is already happening (see: Unitree G1 at $16,000). The real margins will come from AI software, task-specific applications, and maintenance contracts. This is consistent with broader tech industry trends.

Production capacity is the near-term bottleneck. Demand for humanoid robots at the right price point likely exceeds supply capacity through at least 2030. Goldman’s supply-side modeling is more conservative than some other forecasts, which is probably appropriate.

What the model might get wrong

Every forecast model has assumptions that could prove incorrect. Here are the most important ones.

The cost decline curve may be too optimistic. Goldman assumes per-unit hardware costs will decline by 15-20% per year, following a curve similar to industrial robotic arms in the 2000s. But humanoid robots are far more complex than robotic arms, with more actuators, more sensors, and more demanding thermal and structural requirements. The cost decline could be slower than historical robotics parallels suggest.

Labor markets may resist. The model assumes that businesses will deploy humanoid robots wherever they are cost-effective. In reality, labor unions, regulations, public backlash, and institutional inertia could slow adoption significantly. Europe in particular is likely to implement regulations that limit the pace of robotic labor substitution.

The software revenue assumption is unproven. Goldman’s base case assumes $500-$1,000 per month in software subscription revenue per robot. No humanoid robot company has yet demonstrated this kind of recurring software revenue model at scale. It is plausible, but it is a projection based on other SaaS markets, not actual humanoid robot revenue data.

The China factor cuts both ways. Goldman’s model assumes Chinese manufacturers will drive production scale and cost reduction, which benefits the total market size. But if Chinese companies dominate production and sell at low margins, the total addressable revenue could be lower than Goldman projects. A market with millions of cheap Chinese robots is a big market by unit volume but potentially not by dollar value.

How other banks compare

Goldman is not the only investment bank with a humanoid robot forecast. The estimates vary widely.

The range across major banks ($18B to $45B) is narrower than Goldman’s own bear-to-bull range, which suggests a rough consensus that the market will be meaningful but that the exact size depends heavily on assumptions about AI progress and cost curves.

What this means for investors and the industry

The Goldman report matters because it gives institutional investors permission to allocate capital to humanoid robotics. Before the report, humanoid robots were considered speculative technology by most institutional investors. After Goldman put a $38 billion number on it with an 88-page methodology behind it, the sector became “investable” in a way it was not before.

This has real consequences. More capital flowing into the sector means more companies can scale production, which drives costs down, which expands the addressable market, which attracts more capital. Goldman’s forecast is partially self-fulfilling in this way.

The most useful way to think about the Goldman forecast is not as a prediction but as a framework. The methodology, breaking the market into labor TAM, penetration rate, unit economics, and production capacity, gives you a structured way to form your own view. Adjust the assumptions to match your own beliefs about AI progress, cost curves, and regulatory environments, and you will get your own number. Goldman’s contribution is the framework, not the specific answer.

What is not in doubt is that the humanoid robot market exists and is growing. The $500 million in 2025 revenue is real. The 12,800 units shipped is real. The question is only about how big and how fast. And on that question, the Goldman report is as good a starting point as any.