Silicon Valley’s slow robots: why investors are backing robot brains with no business model

Walk into Physical Intelligence’s San Francisco lab today and you won’t see sci‑fi humanoids; you’ll see clumsy arms failing to fold trousers and heroically peeling zucchinis. It looks unimpressive – and that’s exactly why it matters. Behind those cheap arms sits one of Silicon Valley’s most aggressive bets that “foundation models” are coming for the physical world next. The company has raised over a billion dollars to build “ChatGPT for robots” without promising investors any near‑term revenue. In this piece, we’ll unpack why that’s happening, who should care, and what it means for the next decade of robotics.

The news in brief

According to TechCrunch, San‑Francisco–based startup Physical Intelligence has quietly raised more than $1 billion in funding at a valuation of about $5.6 billion. The two‑year‑old company, co‑founded by robotics researchers Sergey Levine, Chelsea Finn, Karol Hausman and former Stripe executive Lachy Groom, is building general‑purpose “robotic foundation models” – essentially large AI models that can control many different types of robots.

Physical Intelligence gathers data from relatively inexpensive robotic arms (around $3,500 off the shelf) deployed in its own lab, test kitchens and partner sites such as warehouses and homes. That data is used to train models that the team explicitly compares to ChatGPT, but for physical manipulation. Most of the company’s spending goes to computing power rather than hardware, TechCrunch reports. Despite backing from firms like Khosla Ventures, Sequoia and Thrive Capital, Groom does not give investors a clear commercialization timeline.

The article contrasts this approach with Pittsburgh‑based Skild AI, which has reportedly raised $1.4 billion at a $14 billion valuation and claims around $30 million in early revenue from its “Skild Brain” deployed in security, warehouses and manufacturing.

Why this matters

Physical Intelligence is important less for what its robots can do today, and more for the bet it represents: that the most valuable part of the robotics stack will be a few general‑purpose “brains” that sit above commodity hardware.

If that bet is right, the winners are obvious. Whoever builds a dominant robot foundation model becomes the default operating system for the physical world: every arm, drone or mobile base becomes just another “client.” That position would resemble what Windows was for the PC era or Android for smartphones – with even deeper lock‑in, because retraining a foundation model is vastly harder than swapping an app.

Investors also stand to benefit. A $5.6 billion valuation for a pre‑revenue, 80‑person company tells you how aggressively capital is chasing the “AGI for the real world” narrative. If Physical Intelligence does manage to ship a general robot brain, it could underpin entire industries: logistics, food, e‑commerce fulfilment, light manufacturing, even elder care.

The immediate losers are robotics startups built around tightly scoped, application‑specific stacks. If one foundation model can handle “any platform, any task” with enough training data – as Physical Intelligence aims to – then purpose‑built perception or control software risks being commoditized.

There’s also a subtler shift: from hardware‑led to intelligence‑led robotics. For decades, industrial players competed on precision mechanics, durability and custom integration. Physical Intelligence flips that logic. Hardware can now be cheap and slightly terrible; good enough intelligence can compensate. That’s a direct challenge to incumbent robot OEMs whose margins rely on proprietary stacks tightly coupled to their arms.

Finally, the company is testing whether Silicon Valley’s appetite for long‑term, capital‑intensive moonshots has really returned after the self‑driving car hangover. Backing a firm that openly refuses to talk about business models is a strong signal that top‑tier VCs believe “robot GPTs” justify another multi‑billion‑dollar R&D cycle.

The bigger picture

The Physical Intelligence story sits at the intersection of three major trends: the rise of foundation models, the commoditization of hardware, and a strategic split in how frontier tech gets developed.

On the AI side, this is the logical next step after large language models. Over the last three years, OpenAI, Anthropic, Google and others have proven that scaling data and compute can unlock surprisingly general capabilities in software. Robotics has always lagged because collecting diverse, high‑quality interaction data is slow and expensive compared to scraping the web. Physical Intelligence and rivals like Skild AI are trying to close that gap by industrializing data collection: cheap arms, 24/7 experiments, and a pipeline that constantly feeds new trajectories back into training.

This also echoes a familiar divide from autonomous driving. One camp (think Waymo) prioritized safety, deep research and constrained pilots, accepting a long march to revenue. Another (closer to Tesla’s approach) threw partially capable systems into the wild to learn from scale. In robotics, Physical Intelligence is playing the “Waymo” role – research‑first, minimal commercial pressure – while Skild is unabashedly in the “ship and iterate” camp, arguing that real‑world deployment creates a superior data flywheel.

History suggests neither side has a clean win. Self‑driving remains stubbornly hard despite billions spent, and Tesla’s vision‑only approach is still controversial. But the pattern is clear: as systems become more complex and data‑hungry, there are powerful incentives to commercialize early just to afford the next round of compute.

The other macro trend is that hardware is finally cheap and good enough. A decade ago, the idea of a sub‑$1,000 arm learning delicate manipulation would have sounded laughable to most roboticists. Today, Physical Intelligence can plausibly say: the bottleneck is no longer the metal, it’s the model.

If you squint, you can see where this leads. In the same way web companies stopped caring which exact server their code ran on once the cloud arrived, future application developers may stop caring which arm or gripper they use – as long as it speaks the dominant robot brain’s “language.” That’s a radical simplification of the robotics stack, and it explains why investors are so willing to fund what still looks like a lab experiment.

The European / regional angle

From a European standpoint, Physical Intelligence and its peers are not just another Silicon Valley curiosity; they are potential suppliers of the next generation of automation across some of Europe’s most strategic sectors.

Europe already has some of the highest robot densities in the world in automotive and manufacturing. German, Italian and Central European factories are full of industrial arms from companies like KUKA, ABB and Fanuc. What’s missing is flexible intelligence that can handle short production runs, frequent retooling and mixed‑product assembly – the reality for Europe’s Mittelstand and SMEs.

A truly general robot brain could be transformative here, but Europe’s regulatory environment will shape how and when that happens. Under the EU AI Act, many robotic systems operating in workplaces or public spaces will almost certainly be classified as “high‑risk.” That means strict requirements around data governance, transparency, human oversight and post‑market monitoring.

For a company like Physical Intelligence, this cuts both ways. On the one hand, training on video data from European homes, warehouses or hospitals immediately raises GDPR questions: what is the legal basis for processing, and how are individuals’ rights protected when that data is repurposed for foundation‑model training in a U.S. lab? On the other, if they invest early in compliance, they could become one of the few foreign vendors able to serve EU industry at scale.

European incumbents won’t stand still either. Robotics powerhouses in Germany, Switzerland and Scandinavia have been quietly funding their own manipulation and learning‑from‑demonstration research for years. The open question is whether they will partner with U.S. robot‑brain providers, license their tech, or try to build sovereign alternatives – a topic that’s likely to surface more as the EU AI Act bites and digital sovereignty debates resurface.

Looking ahead

The next two to three years will decide whether Physical Intelligence’s research‑first strategy is genius or hubris.

The most likely path is a gradual, almost reluctant commercialization. Even if the company refuses to be driven by near‑term revenue, compute bills will keep rising, and investors will eventually want proof that the models work outside carefully controlled labs. That pressure will translate into more paid pilots in logistics, food, light manufacturing and possibly consumer robotics.

Watch for three signals. First, deep partnerships with cloud providers or chip makers: whoever controls the GPUs has enormous leverage over robot‑brain companies, and long‑term compute deals will reveal who is betting on whom. Second, “hero” case studies where a general model outperforms narrowly engineered systems on tasks like piece‑picking, kitting or supermarket replenishment. Third, the first serious safety or labour disputes: as robots move from cages to shared spaces, unions and regulators will inevitably get involved.

There is also real downside risk. If expectations outrun capabilities – glossy demo videos that hide brittleness, or pilots that require far more human babysitting than promised – we could see a mini “robotics winter” reminiscent of the post‑2018 self‑driving slump. That wouldn’t kill the field, but it would shift power back toward more incremental, task‑specific automation.

Still, it’s hard to ignore the structural tailwinds. Ageing populations, chronic labour shortages in logistics and care, and relentless pressure on margins all point in the same direction: more automation, closer to humans, doing messier work. Someone will supply the brains for that shift. The open question is whether they will look more like Physical Intelligence’s slow, patient robots – or Skild’s revenue‑chasing omni‑bots.

The bottom line

Physical Intelligence is a high‑stakes wager that the most valuable layer of the robotics stack will be a few general‑purpose “robot GPTs,” and that it’s worth burning billions before worrying about revenue. That’s risky, but not irrational, given how foundation models have reshaped software. For European and global readers alike, the key question is not whether this class of companies will matter, but which approach will win: research purity or commercial pressure. When robots finally leave the cages and enter our kitchens and warehouses, whose brain do you want running them – and on what terms?