Our Investment in Apptronik

The Sixth Wave of Innovation: The Rise of General-Purpose Embodied AI

Across history, each wave of innovation has fundamentally reshaped human productivity and the nature of labor. First observed by economist Nikolai Kondratiev, these long waves of innovation have repeatedly disrupted and restructured economies.

The steam engine enabled mechanized manufacturing, electricity created assembly lines, computing ushered in digital productivity, and the internet catalyzed global coordination. Today, we are entering the sixth wave: one defined by embodied AI. 

Figure 1: Global Innovation Cycles

This new era is driven by necessity. The global economy is facing a historic labor shortfall. By 2030, the world will be short 85 million workers, translating into $8.5 trillion in unrealized GDP. Traditional software automation has reached its limits, with physical tasks remaining stubbornly labor-dependent. The solution is not more software, but software embodied in machines.

Physical labor is one of the largest markets in the world, but has been difficult to productize outside of specialized machines. The general-purpose robot has been the holy grail, allowing labor to be manufactured and sold at enormous scale. That’s why Mechanism Capital invested in Apptronik’s Series A — one of our largest investments to date. Few teams have the deep hardware roots and AI capabilities needed to scale humanoids. Apptronik has both.

Apptronik is not simply building robots, it is engineering a full-stack system for general-purpose physical AI. Apollo, its flagship humanoid platform, is designed from the ground up to be scalable, affordable, and modular. Its technical roadmap departs from industry convention, favoring manufacturability and real-world deployment over speculative demonstrations.

Importantly, Apptronik is not alone in its ambitions of building a humanoid, but it is differentiated in execution. Where others have pursued vertically integrated approaches, Apptronik has partnered with proven operators like Jabil for manufacturing and Google DeepMind for robotic intelligence, allowing it to move faster and de-risk critical functions. It is a full-stack company with a grounded, systems-level strategy.

Our thesis on humanoids is predicated on the belief that mass commercializing an intelligent humanoid robot is one of the most technically difficult challenges in the world. You require deep expertise across computer vision, robot learning, robot behaviours, humanoid control systems, actuator design, hand engineering, fleet orchestration, high rate manufacturing for complex hardware and many other niche domains where the pool of talent is exceptionally limited. We have reviewed over a hundred humanoid companies, but believe few have the capabilities to build a company that can eventually deliver millions of robots to the real world.

In this report, we articulate why we believe Apptronik represents one of the most asymmetric bets in robotics today, and why humanoids, once considered speculative, are emerging as the preferred form factor for operating in human environments.

Why Humanoids, Why Now

The demand for labor is not only growing, but also fragmenting. Aging populations, shrinking workforces, and increasing wage pressures are converging with rising consumer expectations for speed, reliability, and resilience across physical services. These macroeconomic shifts are creating a structural imbalance between available labor and the tasks that must be done. In response, the labor market is being redefined by a foundational change in who or what performs the work.

Enter general-purpose humanoids. Humanoids won’t just fill labor gaps, they’ll do it cheaper, faster, and at scale. At a $50,000 unit price, a humanoid robot is already cheaper per hour than human workers in most parts of the world—including China, Mexico, and even India at $2/hour.

There is a clear path to achieving a manufacturer cost of well less than $50,000 per high-quality robot in the next few years and there is ample room for that cost to compress significantly more over the next decade as the supply chain matures.

Figure 2: Hourly Cost of Humanoid Robot vs. Labor Rates

The math is straightforward. Robots don’t need breaks, health benefits, or labor protections. They can work 22 hours a day with only battery swaps, and the marginal cost of software improvements approaches zero. In a world constrained by labor availability and inflation, humanoids are a cost-down solution. This cost advantage is only part of the story—humanoids also unlock flexibility that other automation systems cannot match.

Unlike task-specific industrial arms or wheeled delivery robots, humanoid robots are designed to operate in environments already built for people. They do not require infrastructure redesigns, nor do they need dedicated integration teams to implement custom workflows. Instead, they promise generality: the ability to perform a wide range of physical tasks in unstructured spaces with minimal reprogramming. 

Sure, but why do they have to look like humans? Because we’ve already built the world for humans. Door handles, shelves, forklifts, stairs—it’s all optimized for two arms, two legs, and a certain height. You can’t beat a form factor that’s natively interoperable with everything. We’ve optimized hundreds of years of infrastructure around human ergonomics. Tools, vehicles, factories, offices—they all assume a certain range of motion, height, and manipulation. That’s why legs, arms, and hands matter. Not for vanity, but for interoperability.

In fact, if you were to design a general purpose robot from first principles, you would find that it actually looks something like a human. You’d probably have arms and hands with tactile sensing for all different types of manipulation. Bipedal legs with articulated knees, ankles, and hips, each with at least 6–7 degrees of freedom (DoF) per leg for dynamic locomotion and balance would make sense. You’d want at least two cameras on top of the robot (eyes) for 3d vision/depth perception. You’d probably have a central torso housing core components like the main processor, battery, etc instead of organs. Maybe you could have small differences like retractable wheels attached to the feet for more efficient travel, but for the most part, this robot would resemble a human - a humanoid!

The convergence of generalized robotics in the human form factor is driven by the fact that the most complex and adaptable "general-purpose machine" ever created is the human body, operating within environments built by and for humans. Robotic automation technology has existed for decades, but even in highly industrialized settings, we still see humans working across factories, warehouses, construction sites, etc because humans offer adaptability. This same adaptability is what humanoids can offer. The hardware consistency provides a significant advantage for both cost-effectiveness and scalability.

What About Specialized Robotics?

Skeptics argue that generalized robots won’t win; that specialized robotics will go to market faster, do specific jobs better, and take over. And that’s true, specialized devices usually lead the early innings. But history shows that general-purpose platforms tend to capture larger markets over time. They benefit from scale and broader applicability, sometimes matching performance and cost-efficiency of their specialized counterparts. We argue that both markets will be massive, each dominating different segments of the market. 

Remember MP3 players, GPS devices and digital cameras? All of them were replaced by one general-purpose device: the smartphone. It wasn’t the best at any one thing at first, but it was good enough at everything, and ultimately got better. The iPhone moment arrived in 2007 and now, 18 years later, smartphones are ubiquitous and the applications built on them are endless. Humanoids will similarly become platforms, with “developers” building applications that look like human skills.

The same thing happened in computing. ASICs are, as their name describes, application specific, and are built to perform specialized tasks extremely well. They are however, a fraction of the GPU market. In this world, the only constant is change. Changes in algorithms, production lines, work environments, etc. This is the driving force behind the market demand for products built for adaptability.

Figure 3: General Purpose vs. Specialized Market Sizes

Like GPUs and smartphones, humanoids will be able to fit an extremely wide variety of market needs. This is especially useful in places where humans today perform a variety of tasks, for example, in F&B. Workers can be expected to clean, to deliver food, to prepare food, to manage inventory, etc. Some manufacturers need to deal with thousands of SKUs and different product releases many times a year. Humanoids could be the best option if the factory floor and processes are constantly changing. Businesses also have practical financial considerations, it would not make sense or be possible for all businesses (especially SMEs) to make the large capital expenditures typically required for specialized machinery as opposed to paying recurring costs for robot labor as a service. 

According to the World Bank and OECD, global labor accounts for ~60–65% of GDP, or roughly $60–$70 trillion. Conservative estimates from McKinsey and BCG suggest that 60% or more of this labor involves tasks that can be automated with current or near-future technologies. This implies an addressable market of ~$42 trillion for general-purpose robots, orders of magnitude larger than the market for specialized hardware. 

What makes this different from previous automation waves is that it’s not winner-take-all. The humanoid robotics market is shaping up to resemble the auto industry: multiple players, differentiated hardware and software stacks, and diverse use cases. 

Market Size: A Trillion-Dollar Horizon

The real question is how big the market can get and how does that translate into revenue?

The scale of opportunity in humanoid robotics is unlike anything seen before. Conservative projections suggest that we could see 100 million to 1 billion humanoid robots deployed globally over the coming decades. At an average unit price of $50,000, this translates into $5 trillion to $50 trillion in potential market value.

As Elon Musk stated:

“I think we're headed to a radically different world, an interesting world. My prediction for humanoid robots is that ultimately there will be 10s of billions. Everyone will want to have their personal robot. It will unlock an immense amount of economic potential.”

Based on a logistic growth model aligned with Elon Musk’s long-term projections, the expected global deployment of humanoid robots follows the trajectory below:

Figure 4: Projected Global Growth of Humanoid Robots

To contextualize the opportunity, we map these deployment estimates to potential revenue outcomes, using a conservative $50,000 unit price. Even modest adoption at scale implies hundreds of billions in revenue. Applying average S&P 500 revenue multiples (~2–5x), this translates into trillions in potential enterprise value—underscoring why humanoid robotics is being compared to the next trillion-dollar platform shift.

Figure 5: Illustrative Humanoid Revenue Forecast

Deep Roots

Apptronik's foundation is built on longevity and experience. Its co-founders, Jeff Cardenas and Dr. Nick Paine, met as collaborators at NASA’s Johnson Space Center, where they contributed to the development of Valkyrie, a flagship humanoid robotics project developed by NASA. This program would become the intellectual and technical nucleus of Apptronik.

In 2017, they spun out aiming to solve the general purpose manufacturing and architecture problem. Between 2017 and 2024, Apptronik has built eight generations of humanoid and mobile robotic platforms, more than any other company in the humanoid sector. Apptronik’s approach has been methodical and technically rigorous, focusing on building architecture that scales, both mechanically and in terms of manufacturability.

Apptronik was founded in 2017 with the goal of solving the full-stack architecture and manufacturing challenges required for general-purpose robots. At the time, the AI ecosystem was not yet mature enough to support fully autonomous embodied systems, and hardware development in robotics received limited investor support. Despite that, the team built eight generations of humanoid and mobile robotic platforms, more than any other company in the sector. As the AI inflection point arrived in 2023–2024, Apptronik was already positioned with the right architecture, the right partnerships, and a matured engineering foundation to capitalize on it.

At the heart of this is a relentless focus on actuation. While most humanoid competitors rely on harmonic-drive rotary actuators, Apptronik took a contrarian path, designing and building its own proprietary linear actuators—over 40 unique variants to date. This has yielded critical advantages in energy efficiency, thermal performance, reliability, and ease of mass production. But the real value lies in system control and integration: owning the actuator stack gives Apptronik precise control over motion dynamics and power efficiency, a foundation for high-frequency task performance in real-world environments.

Actuators Win

First, what are actuators?

In robotics, actuators are the components responsible for producing movement, analogous to muscles in the human body. Every motion, from a simple lift to a complex multi-joint task, is enabled by actuator performance. Actuators fundamentally define everything from motion precision to payload capacity, from durability to safety, nearly every key system metric is downstream of actuator design.

This is why Jeff Cardenas, Cofounder and CEO of Apptronik, put it plainly:

“The actuator is the muscle. If you don’t own that layer, you’re outsourcing your core. If your actuators aren’t efficient, everything else falls apart—battery life, payload, motion control. That’s why we decided early on to own this layer.”

Apptronik recognized early that off-the-shelf actuators—particularly harmonic-drive rotary actuators—were insufficient for general-purpose robots. These legacy systems suffer from several critical limitations:

• ~70% energy efficiency

• High thermal output (heat = wasted energy)

• Complex gearing → higher maintenance burden

• Bulky form factors → lower power-to-weight ratios

• Poor backdrivability → limited compliance and human safety
  
  

Apptronik took a different path by developing its own proprietary linear actuators, with over 40 internal iterations since inception. These actuators were designed in-house, from first principles, to support long-term manufacturability, thermal efficiency, and high-frequency industrial use.

Key performance advantages:

• >90% energy efficiency → longer runtime per battery charge

• Lower thermal output → no active cooling required

• Fewer moving parts → simpler, more reliable maintenance

• Compact design → higher power-to-weight ratio, improved durability, and faster, more agile movement

• Compliant joints → improved safety around humans

• Shock tolerance → handles continuous industrial workloads

• Integrated modularity → supports multiple robot configurations (torso-mounted, bipedal, wheeled base)

These design advantages translate directly into real-world performance:

• Apollo can lift 55 lbs repeatedly without overheating

• Core thigh actuators are load-tested to 1,500 lbs, ensuring long-term durability under industrial stress

• The actuators support modularity across robot formats—humanoid, wheeled, and stationary

• Limited third-party IP risk, and no upstream supplier bottlenecks

The actuator is not just a component—it is the product. A humanoid robot is, fundamentally, a collection of actuators operating in coordination. Apptronik’s decision to build this layer in-house reflects a systems-level understanding of what it takes to commercialize at scale. It is also one of the company’s deepest and most defensible technical moats.

Companies like Unitree have pioneered the first commercially available robots, but these robots are not built for real work. The G1 has load capacity of 2kg, compared to Apollo’s 25kg. Chinese robots on the market have seen broad issues with overheating, actuator/limb durability, battery life, etc. While we believe Chinese companies will deliver high quality humanoids, the engineering challenges of building a robot that can handle the stresses of years of physical labor is not trivial.

There’s no shortcut to building good, durable hardware. Unlike software where you can release updates in days or weeks using simulations and virtual testing, hardware iteration can take months. There are only a handful of humanoid companies in the world that have had more than a few years for iteration.

Building to Scale

In robotics, the transition from prototype to production is often where promising companies fail. Many over-optimize for demos, sacrificing manufacturability, durability, or maintainability in favor of short-term visual appeal. Apptronik has deliberately taken the opposite approach. Its entire architecture is designed not just for performance, but for scalable deployment in the real world.

As Elon Musk once said “The extreme difficulty of scaling production of new technology is not well understood. It’s 1000% to 10,000% harder than making a few prototypes. The machine that makes the machine is vastly harder than the machine itself.”

The company’s early partnership with Jabil is a critical differentiator. Jabil is one of the largest contract manufacturers in the world, operating in over 25 countries with more than 100 manufacturing facilities and 140,000+ employees. Jabil serves clients ranging from Apple and Tesla to defense and medical firms—managing extremely high-volume and high-quality production pipelines. This partnership ensures Apptronik can scale beyond lab prototypes or limited pilots. It also unlocks rapid iteration cycles on hardware, consistent quality assurance, and access to robust global supply chains.

Just as importantly, Jabil’s domestic production capability supports Apptronik’s strategy to minimize IP leakage risks. In the highly strategic and increasingly geopolitically sensitive sector of humanoid robotics, safeguarding proprietary actuation systems and software is critical. By avoiding outsourced manufacturing to regions with weaker IP enforcement (particularly China) Apptronik protects its core technology stack while aligning with U.S. national security interests.

Most competitors have taken vertically integrated paths, building out in-house fabrication capabilities from scratch. While we believe there are many benefits to the integration, few will have the expertise and resources to pull it off. Apptronik’s approach—outsourcing non-core manufacturing to an industrial giant—frees it to focus on what matters most: actuation, software, integration, and deployment. Similar to how Tesla bootstrapped production, building manufacturing capabilities can come down the line.

This commitment to manufacturability is reflected in Apollo’s modularity. Apollo is not a single configuration robot—it is a platform. The base can be configured as:

• A stationary mounted torso (for fixed manufacturing or inspection tasks)

• A mobile wheeled version (for logistics, warehouse navigation)

• A bipedal walking humanoid (for dynamic or constrained environments)

Apollo is just the embodiment today. The real product is the full-stack control system that can power multiple robotic forms.This flexibility enables deployment in various environments without requiring full reengineering. This modularity extends to the power system. Apollo uses hot-swappable batteries, enabling 22/7 continuous operation with minimal downtime -> maximizing utilization across shifts.

The ability to scale faster than most competitors has many downstream implications. Both in securing a foothold with customers that will have high switching costs but also kickstarting the data flywheel that is integral to building robot intelligence. From what we have seen, robots doing work in the real world is the most effective data source to building the most performant models. Simulation, teleoperation, human video data are all useful for co-training, but for now, there is no substitute for real robot data.

Apptronik has emphasized supply chain resilience. Its proprietary actuators were designed not only for performance, but for ease of assembly, component availability, and low-cost scalability. The company has made deliberate choices to avoid dependence on hard-to-source components or foreign suppliers exposed to geopolitical risk. This positions Apptronik well to scale production in North America and meet onshore demand across defense, manufacturing, and infrastructure sectors.

Building Towards Autonomy: The Google DeepMind Partnership

Apptronik’s partnership with Google DeepMind is one of its most strategically significant advantages—and a major differentiator in the humanoid robotics landscape. In a space where many startups promise future autonomy but lack the infrastructure to deliver it, Apptronik has secured deep alignment with the world’s leading AI research institution to co-develop its software foundation. A robot without intelligence has very limited utility. There are very few teams in the world capable of building a robot foundation model for humanoids and Deepmind is one of them.

In 2024, DeepMind invested in Apptronik, marking one of its largest investments in embodied AI to date. This engagement involves dozens of DeepMind researchers and engineers working jointly with Apptronik’s team to develop a vertically integrated autonomy stack for general-purpose robotics. We believe their relationship spans model development, training infrastructure, hardware-software co-optimization, and fleet-level deployment strategies.

DeepMind’s technical credentials are unmatched. It has pioneered many of the most important breakthroughs in machine learning over the past decade: from AlphaGo and AlphaFold, and its Gemini models are some of the most frontier in the multimodal space.

Google has led with many innovations in the robot foundation model space from RT-1 to RT-2 to PaLM-E. Gemini Robotics and MuJoCo, its general purpose physics engine are staple tools for robotic learning researchers. 

But even the most capable models require embodiment to become useful in the physical world. Apptronik provides that embodiment. The ability for Google to access a large amount of research allow it to build better software and better models and that in turn benefits Apptronik.

The company’s approach to autonomy begins with teleoperation, moves through human demonstrations, imitation learning, reinforcement learning and ultimately enables autonomous execution across high-frequency physical tasks. Every step of this pipeline is designed to generate high-quality data for fine-tuning and reinforcement learning at scale.

The integration with DeepMind ensures that this data becomes fuel for autonomy improvements across the entire Apptronik fleet. As DeepMind trains its models using Apollo-generated experience, the resulting improvements are deployed back into robots operating in the field, closing the loop between experience and policy. 

Real Commercial Demand

The humanoid robotics industry is moving from prototype to production, and here Apptronik has an incredible breadth of partners.

Its lead partners, Mercedes-Benz and GXO Logistics, represent two major verticals: automotive manufacturing and third-party logistics. These pilots are not speculative experiments; they are structured deployments designed to evaluate Apollo’s ability to deliver measurable ROI in existing workflows.

• Mercedes‑Benz entered a joint commercial agreement in March 2024 to pilot Apollo humanoids within manufacturing facilities—specifically automating tasks like parts delivery and quality checks in environments designed for human workers.

• GXO Logistics, the world’s largest pure-play logistics provider, launched a multi-phase R&D pilot with Apollo in mid‑2024. The goal: assess Apollo’s capabilities in warehouse tasks such as tote handling, scanning, and repetitive inventory workflows—feeding real-world data back into Apollo’s learning loop.

This business-first approach reflects a clear go-to-market strategy:

• Focus on structured use cases: where repeatable, labor-intensive tasks allow Apollo to deliver measurable ROI early in deployment.

• Use pilot programs with large enterprises as live engineering labs, refining actuator performance, autonomy capabilities, and workflow integrations.

• Monetize via a Robotics-as-a-Service (RaaS) model, with hourly pricing, which aligns usage-based incentives and removes upfront capital barriers for customers.
  
  

As Jeff Cardenas, Apptronik’s CEO, noted in a February 2024 interview on the What’s Your Problem podcast:

“The demand for these robots is enormous. We have demand for hundreds of thousands of units already today, with the customers we're working with. The demand is enormous.”

Risks and Execution Challenges

Apptronik is operating in one of the most ambitious and technically complex domains in robotics. As with any frontier company, risks exist, but we see them as idiosyncratic, not systemic. 

1. Hand dexterity and manipulation:

Apollo’s current hands are based on prosthetic designs, prioritizing weight, cost, and basic grasping. While they are sufficient for repetitive logistics and light industrial tasks, they lack the fine manipulation and in-hand dexterity required for more complex workflows. Apptronik is actively developing improved end effectors, but the timeline for achieving human-level manipulation remains uncertain across the entire industry.

2. Actuator divergence from industry standards:

Apptronik’s decision to develop proprietary linear actuators offers significant benefits in energy efficiency, serviceability, and shock tolerance. However, this diverges from the industry norm of harmonic-drive rotary systems, which could complicate third-party integration or tooling compatibility. The tradeoff is clear: greater performance and IP defensibility in exchange for higher responsibility across the full stack.

3. Apollo 2 as an execution milestone:

Apollo 2 represents a critical inflection point in Apptronik’s roadmap. It is the platform’s first full-scale deployment candidate and must demonstrate robust performance, repeatability, and maintainability in the field. Delays or underperformance could impact the company's ability to scale pilots into commercial contracts and extend lead time in a market that is quickly evolving.

The Case for Apptronik

The humanoid robotics market is entering its commercialization era—but not all players are built to scale. What will determine long-term winners is not hype or headline-grabbing demos, but execution across three core dimensions: scalability, versatility, and data capture.

Apptronik is one of the few companies showing real traction across all three.

• Scalability: Through its exclusive partnership with Jabil, Apptronik is positioned to scale manufacturing globally without building its own fabrication infrastructure. Jabil's track record with companies like Apple and Tesla derisks both quality and capacity at volume.
  
  

• Versatility: Apollo’s modular design allows reconfiguration across use cases—from logistics and manufacturing to inspection and R&D—without reengineering the entire system.
  
  

• Data Capture: Apptronik is building a teleoperation-to-autonomy pipeline with structured data feedback from real-world pilots. This software loop enables continual improvement in capabilities, efficiency, and reliability.

We invested in Apptronik’s Series A at a $1.8 billion post-money valuation because we believe it is one of the few companies in this space with the technical depth, commercial focus, and execution strategy to scale. Apptronik has real pilots, real partnerships, and real hardware.

In the same way NVIDIA became the infrastructure layer for machine learning, we believe Apptronik has a path to become the infrastructure layer for physical intelligence. The bet here is not on humanoids becoming real—the bet is on who executes at scale, who learns fastest, and who builds the systems that others will one day depend on.

And in that race, Apptronik is leading with depth, speed, and clarity.