Follow-up to The future is robotics (data). Somewhat unpolished notes on what will work in robotics in the coming years.
I recently read Eric Jang's Smooth Exponentials for Robotics post. Here, Jang argues that robotics is entering what Dario Amodei called
a "smooth exponential" of AI capabilities, as if machine intelligence "emerge[s] spontaneously from the right combination of data and raw computation."
The idea is that with enough energy going into robotics, progress will happen no matter what, smoothly but rapidly, over time. Jang lists the main unlocks for this exponential growth as
Modern LLMs that can be evaluated for robotic capabilities
A brief expansion, since this one is new: last week, we saw many videos of Astra directly controlling a robot to perform different manipulation tasks. One of the most inspiring examples came from Thijs Simonian, who posted a video of Astra creating a painting. Each painting took 1 to 3 hours, which, aside from being pretty slow, means it required a considerable token budget. However, videos like these show us tasks that are slow and expensive today, but can be cheap and fast tomorrow. That's why they're great! I predict that within 2 years, similar models will be able to make this painting 4-10x faster. One key unlock might be a video-native model that can interact with the world similar to voice interactions in GPT-Live. Another observation supporting this prediction is simply that Astra scored much higher on a robot control task while using 6.2x fewer output tokens at 2.3x lower cost than Fable 5.1.
Jang suggests the main benefit of testing modern LLMs on robotic tasks is that robotics inherits the evaluation practices built for LLMs. I'd also say that, assuming further speedups arrive, LLMs controlling robots lower the barrier to getting started.[1]
Two more trends that unlock the smooth exponential:
Others have covered this well, so I'll be brief [2]. We see more companies offering cheap robotic arms that actually do meaningful things. Cheaper humanoids become available as well. And it's easier to buy them - just add them to your shopping cart and pay online. Why does this all matter? As hardware costs go down, it becomes much easier for people to contribute to robotics progress. Not just in simulation, but by deploying real robots in the world. The result will be more real-world data that will further boost model capabilities.
Last week, Skild shared that they got to 100M ARR within 10 months. Although one can always question reported ARR, I believe Skild when they say they haven't started scaling yet, which leaves a lot of room to grow [3]. Since they are generously funded, I wouldn't be surprised if they 10x their revenue next year. This looks possible: the tasks they mention each address a far larger market than they currently capture, with few competitors. The physical world is vast, and so is the list of tasks waiting to be automated. Hence, it shouldn't surprise us if several companies focused on research and deployment reach 1-10B ARR in the next 3 years. It matters because it proves real-world value can be delivered. Without that proof, capital will stop flowing in.
Progress in AI is almost always smooth. It only looks jumpy when a smooth curve crosses a binary threshold. Schaeffer et al. showed that emergent capabilities are mostly a mirage: pass/fail metrics hide steady improvements underneath. The trend that general models couldn't do manipulation tasks before and now suddenly succeed is likely due to this.
Hardware costs drop, general models perform robot tasks better and better, and big players reach significant revenue. What does that mean for others building in the space? Much of the advice for building around LLMs applies to robotics: build a product that gets better as API and open-source models improve. For inspiration, we can look at a few companies that did well in the LLM space:
| LLM-era company | Description | Outcome | Robotics parallel | Transfer |
|---|---|---|---|---|
| Cursor | Distribution/IDE layer around models | $60B exit | Orchestration/harness between base models and robot deployments | Medium. Physical integration lacks bottom-up developer adoption, so integration will be slower. |
| OpenRouter | Model distribution/routing layer | $7.5B exit | Routing layer for robot foundation models; per task. Not useful yet, but plausibly a big layer later | Weak at the moment. Too few interchangeable models to route between |
| Cognition | Corporate integration of AI; relentless focus on capturing deployment value; pretty full-stack | $48B valuation, $900M ARR | (Full-stack) integrators with an immense drive to capture real-world deployment value | Strong, but slower loops compared to software, one FDE can manage fewer clients. |
| ElevenLabs | A data modality (generative audio) that was undervalued, great GTM | $20B valuation | An undervalued modality with enormous potential: sensory/tactile information? | Medium. Tactile should be important; demand unknown |
| Hugging Face | Community/developer layer | $13B exit | Community and developer layer for robotics | Strong. LeRobot is already doing this to some extent |
| Mercor | Data acquisition | ~$20B valuation | Provide great robot training data at scale | Strong and crowded. Difficult physical labour can still be cheap per hour (Moravec), but requires hardware to record it. |
| Compute providers (many) | Compute, directly or indirectly | Compute for robot model training and inference | Strong. Transfers unchanged, potentially more demand for local inference. |
Now, robotics will also have opportunities with no LLM-era equivalent at this scale. For example, because evaluating robotic policies is inherently harder than evaluating software-only policies, I can imagine that whoever cracks evaluation becomes huge (Lightwheel seems a good candidate).
If the exponential is real, capabilities will be cheap and roughly equally available to everyone. Let's see what then remains for different parts of the ecosystem: from integrators, hardware providers to model builders; all have parts of their business that will become a commodity and parts that do not.
| Layer | Commoditises | Where value sticks |
|---|---|---|
| Application | General models do tasks out-of-the-box | Harder tasks, deep integration into process customer, trust |
| Hardware | General-purpose humanoids, cheaper at volume | Throughput, precision, and hostile environments where general hardware won't be able to do well |
| Model | base model itself; LLM labs offer manipulation | Distribution, fast scalable inference, data flywheel via deployment, data partnerships or simulation |
For companies working on integration, the risk is twofold:
For new integrators to offer sustained value, they must keep moving to harder tasks, while still capturing value from tasks others offer out of the box (which by then are high-margin). I think there are many ways to build a moat here, as long as one focuses obsessively on what makes customers' lives easier and provides more than a simple context layer. One should go all the way until the product is fully integrated into a company, providing immense value and stickiness. The difficulty is that a company has no reason to trust a new startup to integrate into systems built on years of specialised experience. Hence, aim to get read permission first, then write permission [4]. Two worth studying: Scott Wu's mindset at Cognition (more full-stack), and the rapid growth of Wonderful.ai (a classic integrator with forward-deployed engineers).
Other ways to build a moat, beyond deep integration into the company's workflow:
Finally, fine-tuning on customer data may beat general models on accuracy for a while, even when those models are already competent at the task. Build something people want, and make the company powerful.
The risk for companies building specific hardware is that custom hardware is much more expensive to build, while general hardware can benefit from economies of scale. An analogy here is renewable energy: solar panels became much cheaper at an incredible pace because they're small and cheap to iterate on, whereas a nuclear plant is neither, and stays expensive as a result [6]. Take cleaning robots. You could imagine a robot with vacuum-cleaner arms and a built-in bin, much more efficient and flexible than current (floor) cleaning robots. However, humanoids using the same cleaning tools humans use today might quickly become cheaper and more useful for a wide range of tasks. So, if you're building sector-specific hardware, you're better off picking a space where efficiency gains over humans are so large that humanoids won't be able to compete anytime soon, even when they're 20x cheaper than today [7]. Some areas where specific hardware should win:
The current thinking is that deployment gives you the most valuable data for building performant models. Hence, companies building foundation models should also deploy robots: it shortens the feedback loop between model and real-world value. Although I think some well-funded companies will do well this way, I don't think that means every foundation model company should focus on deployment. Deployment is incredibly difficult and time-consuming, and there are other ways to train strong base models. Simulation and world modelling both go a long way when executed well. Add strong collaboration with real data providers on top of that, and one might be able to build a great company without focusing on deployment. Now that we're seeing models from traditionally LLM-focused companies do robotic tasks, it might be that these companies will be the main foundation model players here! They have a) the distribution, b) the talent and experience to make inference fast and scalable, and c) the resources to pursue it seriously.
One related question to foundation models is whether models can be served on cloud GPUs or need local inference. If one can do API inference, the LLM business model translates directly. However, robotics has unique challenges here:
As a result, most players now default to on-device inference. The promise of API inference is real as it should bring down hardware costs, but it requires solving a couple of difficult questions. Those with domain expertise have a real opportunity to build something valuable for both inference speed and network reliability. SemiAnalysis wrote a good post on this: A brain too big to carry on device.
Beyond API-level providers, I think open-source models still offer immense value. They will give companies unique ways to deploy robots and will lead to a more exciting ecosystem. Businesses working on open-source models might not capture much revenue directly from providing the models, but should position themselves as the ones supporting the ecosystem and profiting when the entire robotic ecosystem wins. NVIDIA, with its recent acquisition of Hugging Face, is the most natural candidate to lead here. For NVIDIA, building open-source base models would pay off as it makes it easier for others to participate, and participation requires compute. For Hugging Face, it would be a natural continuation of their LeRobot project, and in line with their community support mission.
I hear two things. One: the general robotics companies were funded years ago, so the opportunity is to build the layer around them. Two: for every actual robotics company, ten are already building that layer. Both are true. The reality is that, although there are indeed many general robotics companies, relatively few have deployed robots. So, although they are well-funded, many have not found the key to delivering real value, and the race is still somewhat open. However, Skild's recent progress suggests this window is closing, and may already be half-shut for new players, given how long ramping up takes. When revenue is increasing sharply, players can be the last ones to discover something.
At the same time, since it's much easier to build a pure software company, especially in 2026, we now see many new companies building tooling for robotics companies. The winners here aren't decided yet. My guess is that fewer tools are being built for business owners and operators who want to automate, since this requires more domain knowledge, is technically harder given the need to integrate into current systems, and has to overcome the trust barrier I mentioned above. However, there should be a real opportunity here, not just by optimising business flows, for which there are plenty of companies, but also by putting the robot managing and learning process into the hands of the ones who will see them in their environment every day. This might seem difficult right now, but as we move more towards highly performant few-shot learning, it should be easier for people without AI experience to train and deploy a robot for a specific task.
have talked to a bunch of robotics founders and they all are inundated with outbound from dozens of robotics dev tools / data collection / infrastructure startups constantly at least 10x more people building sw tools for robots instead of robotics companies
When I speak with owners of physical-work businesses who want to automate, it's clear there's a huge amount of value to unlock. Many aren't aware of recent AI advancements or struggle to use them. Although many AI companies are building solutions, those solutions largely haven't been deployed yet. My exploration so far suggests this comes down to a lack of knowledge about the technology and a lack of trust in it, as many pilots fail to translate into long-term real-world value [8]. More specifically:
All of these can slow innovation long-term because they reduce trust in technology. I think whoever wins back that trust, driven by a desire to provide long-term value, takes the market. This means the best technology won't necessarily come out on top, though the two correlate.
Given the above, a few closing thoughts about what I think will work in robotics in the smooth exponential:
Beyond that, I think it will be important for the world to ensure that AI benefits everyone, not just the digital world but also the physical world [9]. If the people doing manual work today become more productive, and no longer have to do the dull, dangerous, and dirty parts, AI will mean much more than it does today.
[1] I do expect that OpenAI trained on robot data.
[2] Many examples exist; Chris Paxton wrote about this here and here.
[3] A revenue breakdown and volume of Skild's work would be useful here; it might be that ARR is not true ARR originating from modern robot deployments and that the world is not moving at exactly the pace one expects by seeing '10X revenue in 10 months, without scaling'.
[4] Thanks to Emerge VC for this point: https://emergecapital.vc/physical-ai-why-the-application-layer-is-the-next-breakout-segment/
[5] An example: identity verification generates >10 billion in revenue, yet no big tech player provides worldwide compliance or competes at that level with the incumbents.
[6] This is such an incredible achievement. From Our World In Data:
The learning rate of solar PV modules is 20.2%. With each doubling of the installed cumulative capacity, the price of solar modules declines by 20.2%. The high learning rate meant that the core technology of solar electricity declined rapidly. The price of solar modules declined from $106 to $0.38 per watt. A decline of 99.6%. To get our expectations for the future right, we ought to pay a lot of attention to those technologies that follow learning curves. Initially, we might only find them on a high-tech satellite out in space, but the future belongs to them.
[7] Similar to A16Z Oliver Hsu's point in Toward a Horizontal Robotics Platform (2024):
Robotics applications with highly specialised task/environment data that may be able to take advantage of future broad improvements in robotic intelligence by building custom solutions with their domain-specific data.
[8] From this perspective, I hope that governments/subsidies will focus less on supporting pilots that do not have a clear stake in delivering long-term value. Pilots that do not provide long-term value make companies less likely to try it again, which is bad for innovation. At the same time, pilots might be necessary to open eyes to what is possible. Aligning participant incentives with long-term value rather than short-term exploration should be a guiding principle.
[9] "When will average people feel AI's impact" - Nathan Lambert. An unfortunate title for a post with an interesting point.
