The denial came out, and nobody entirely bought it

Over the third weekend of July 2026, AI Twitter got loud about a single claim: Anthropic was acquiring the robotics startup Physical Intelligence. Tech blogger Robert Scoble posted it over the weekend, and within hours it had saturated the timeline.

On July 21, Physical Intelligence CEO Karol Hausman addressed it directly. Not through a press release. Not at a briefing. Through a message on the company's internal Slack, accompanied by a GIF from The Office of a character shaking their head. Meaning: not true.

The rumor didn't die. Two reasons. First, as TechCrunch noted, the denial wasn't forceful. It said "we're not being acquired," not "no such conversation ever happened." Second, The Information reported that the two companies did hold acquisition talks in the spring of 2026. It wasn't fabricated from nothing.

A denied rumor usually survives because it touched something people were already expecting. This one did exactly that.

The cast — a company building robot brains, and one that wants a body

Physical Intelligence was founded in San Francisco in 2024. Its founding team includes Karol Hausman, formerly of Google Brain robotics, alongside UC Berkeley's Sergey Levine and Chelsea Finn — genuinely top-tier robot learning researchers. Those names carry real weight academically; a substantial share of the foundational imitation learning and meta-learning literature came out of that group.

The company doesn't build robot hardware. It builds robot foundation models — the π (pi) series, using a Vision-Language-Action architecture. Feed it camera imagery plus a natural language instruction, and it outputs joint movements directly. π0 came first, then π0.5, a roughly 3-billion-parameter model targeting open-world generalization so mobile manipulators can operate in environments they've never seen during training.

The key positioning is hardware neutrality. Not tied to any one robot maker, the same model runs across different arms and mobile platforms. The pitch is to be robotics' Android.

The funding history matters here too. Roughly $70 million came in alongside the first π0 model, then $400 million in a Series A at a $2.4 billion valuation, then $600 million in a November 2025 Series B at $5.6 billion led by Alphabet's CapitalG — with Lux Capital, Thrive Capital, Jeff Bezos, Index Ventures and T. Rowe Price participating. Subsequent rounds have been reported pushing the valuation into the $11 billion range. Two years old and in the ten-figure club.

And there's one decisive complication: OpenAI is an investor in Physical Intelligence. Any Anthropic acquisition would mean buying a company a direct competitor already holds equity in.

On Anthropic's side, this hasn't been a robotics company. Language models, agents, coding tools — that's essentially the whole business. But something has been shifting.

Why this rumor, why now — Project Fetch

In June 2026, Anthropic published results from Project Fetch Phase Two, an internal experiment with a robot dog.

The summary: Claude Opus 4.7 completed a set of tasks programming a Unitree Go2 quadruped in nine minutes. An AI-assisted human team attempting the same tasks took 181 minutes. Roughly a 20× gap. Claude also got the robot walking around and locating a beach ball, which the human team never managed at all.

The number isn't the important part. Per Anthropic's own account, a model that couldn't even connect to the sensors in August 2025 was finishing all four tasks in under ten minutes by June 2026 — and none of that improvement came from any deliberate robotics training. It fell out of general model scaling.

Anthropic's red team put the direction plainly: "the next step for AI models is to start reaching out into the world and affecting the world more broadly," and "this will really require models to interface more with robots."

So here's the picture. Anthropic doesn't build robots, but it has confirmed that its models are getting better at controlling them faster than anyone planned. Connecting that capability to the physical world requires robot data and an embodied control stack. Physical Intelligence has exactly that. Which is why the rumor sounded plausible to everyone who heard it.

Worth being precise about what Project Fetch actually demonstrated, though. What Claude did well was write the code that controls a robot, not pilot a robot in real time. Technically those are different problems. Writing code affords seconds to think; actual manipulation means computing joint torques on millisecond cycles and reacting to physical failure instantly. VLA models like the π series solve the latter; language models are currently good at the former. So "Claude handled robots 20× faster" is a striking sentence that does not imply Anthropic can do what a robot foundation model company does. If anything that gap is why an acquisition rumor is plausible — the two are complementary.

It's worth going one layer deeper on why data is the decisive constraint. Language models grew on the internet, an enormous dataset that already existed. Robotics has no equivalent. Collecting manipulation data means putting real robots in real environments and recording trajectories while humans teleoperate or demonstrate — an hour of data costs an hour. Simulation is the standard workaround, but the physics gap between simulator and reality (sim-to-real) means it doesn't transfer cleanly. A bottleneck that compute cannot buy is the structural reason a cash-rich frontier lab starts thinking about acquisitions.

What each side gains and loses

Scenario Anthropic Physical Intelligence OpenAI
Deal closes Robot data + control stack, entry into physical AI Compute and capital, loses independence Equity value transfers to a rival; strategic loss
No deal (current) Build in-house or find another partner Stays independent, keeps valuation leverage Investment position intact

What Anthropic would gain is data. The scarcest resource in robot learning isn't compute — it's manipulation data collected in the actual physical world. The internet has unlimited text and images; "the joint trajectory while a robot arm opens a drawer" has to be collected by hand. Physical Intelligence has been accumulating that for years, and it isn't an asset money buys quickly.

What Physical Intelligence would gain is compute and stability. Robot foundation models are expensive to train and slow to commercialize. Sitting inside a frontier lab removes fundraising pressure. What it would lose is hardware neutrality — the moment you belong to one lab, other robot makers may hesitate to build on your model, which collides head-on with the Android-of-robotics position. That's plausibly the real reason for the denial.

OpenAI's position is delicate. As an existing investor it would profit financially from a sale, but watching a competitor take the asset is strategically bad. And with OpenAI itself moving toward hardware and physical AI, that stake is unlikely to be purely financial.

Even with the rumor denied, everyone banked something. Physical Intelligence got public confirmation that a frontier lab wants it, which strengthens its next round. Anthropic signaled interest in robotics at zero cost. Rumors like this frequently leave all parties slightly better off after the denial.

How acquisition rumors like this have ended

Some denials precede deals. Early denial is standard in large M&A. Parties hate premature disclosure before terms are locked, and the target especially denies in order to stop employee and customer flight. Markets do not read "they denied it" as "it isn't happening." Which is why TechCrunch flagging that Hausman's denial wasn't forceful was a meaningful observation rather than a stylistic note.

Plenty of denials are simply true, too. Individual influencers floating claims that turn out to be baseless is common in an industry where information circulates far faster than anyone can verify it.

The more instructive pattern is the recent AI acquihire. Over the past two years, frontier labs have repeatedly taken key people and a license instead of buying the company — Microsoft with Inflection AI, Google with Character.AI and Windsurf, Amazon with Adept. The structure sidesteps antitrust review while extracting what's actually wanted. If Anthropic wants something from Physical Intelligence, that shape is far likelier than a full acquisition. And notably, a company can deny "being acquired" in that scenario without technically lying.

Robotics acquisitions have their own history. Google bought eight robotics companies in 2013 including Boston Dynamics, then divested most of them within a few years because monetizing hardware proved far harder than expected. That lesson still holds: building good robot models and building a profitable robot business are entirely different problems.

How other players move

OpenAI is already active on both hardware and physical AI, holds its Physical Intelligence stake, and has reportedly rebuilt an internal robotics team. One effect of this rumor for OpenAI was a public revaluation of its own position.

Google DeepMind develops robot foundation models internally — the RT series and Gemini Robotics line. Given that Alphabet's CapitalG led Physical Intelligence's Series B, Google is running a dual strategy: build inside, invest outside.

Nvidia competes a layer down. Through GR00T, its humanoid foundation model, and the Isaac simulation platform, it supplies the tooling that lets robot companies build their own models. Selling the infrastructure for making models rather than the models themselves.

Chinese robotics matters here too. Unitree — whose robot dog Anthropic used in Project Fetch — and its peers keep driving hardware prices down. Cheaper bodies raise the relative value of the brain software, and that dynamic is a load-bearing part of why companies like Physical Intelligence are valued where they are.

Korean players are in this too. Hyundai Motor Group owns Boston Dynamics and, at this week's San Francisco AI Summit, announced it will co-develop an "AI Robot Reference Platform" with Nvidia. Bodies looking for brains, brains looking for bodies — the pairing phase is visible from every direction.

What actually changes for you

If you work in robotics, the signal isn't whether a deal happens. It's that frontier labs have started wanting robot data. As language model competition matures, they're hunting the next data frontier, and physical-world manipulation is it. The strategic value of any company holding robot data likely climbs for several years.

If you're an AI engineer, Project Fetch is the more substantive news. Robot control improving through general scaling with no robotics-specific training implies that a large portion of robot software could be displaced by general-purpose models. What it means to write robot control code may change materially.

If you're an investor, watch Physical Intelligence's next round terms, not the rumor. How the valuation prices in now that frontier-lab interest is public is the real read on this company's market value — and a decent thermometer for robotics AI funding overall.

If you're a founder, treat this as a case study in information control during negotiations. A single Slack message reached the press within hours. During any deal process, assume internal communication is public communication.

If you're a general reader, the rumor itself doesn't matter much. The direction behind it does. AI has lived entirely inside screens, and frontier labs are trying to step outside them. Nobody knows when that transition actually arrives, but the preparation has clearly begun.

🥄 Three Things You're Probably Wondering

— So did the acquisition happen or not? Not as of now. Physical Intelligence's CEO denied it and neither party has confirmed a deal. But if The Information is right that real talks occurred last spring, "no deal now" doesn't mean "no deal ever." Large M&A routinely collapses and returns when terms improve.

— Does this mean Anthropic is entering the robotics business? Too early to read it that way. Project Fetch was an internal experiment, and Anthropic has never said it intends to ship robot products. What's confirmed is an observation — "our models control robots better than we expected" — plus a stated direction that models should start affecting the world. That's some distance from a product plan.

— Why does a rumor like this generate so much heat? Because robotics is widely seen as the next big field and no winner has emerged. Language models have consolidated around a handful of labs; physical AI has no clear leader. When a field is genuinely open, every move by a large player gets over-read. Worth remembering that it is over-reading.

References

Numbers and criteria are as of announcement and may change.