The Week Three Physical AI Bets Landed at Once
Three unrelated announcements in five days pointed at the same thing.
Travis Kalanick’s industrial AI company Atoms raised $1.7 billion in a round led by Andreessen Horowitz, with Ben Horowitz joining the board, according to TechCrunch. Two days later, a new AI lab called Prentis — co-founded by Reid Hoffman, Mark Pincus, and serial entrepreneur Ritankar Das — was reported to be in talks to raise $100 million at a $1 billion valuation. In between, a weekend rumor that Anthropic had acquired Physical Intelligence spread fast enough to require a public denial — and it later emerged, per The Information, that Anthropic and Physical Intelligence had in fact held acquisition talks earlier this spring.
These are not three random funding stories. They are the same story: the thesis that AI’s next durable moat sits not in foundation model parameters but in AI systems that control physical equipment, computers, and the messy workflows humans run every day.
Travis Kalanick’s $1.7B Bet: The Bits-to-Atoms Arc
Atoms is Travis Kalanick’s re-entry into infrastructure-scale AI — and a deliberate callback to the ambition he had at Uber before his 2017 exit. “16 years ago, I started a journey to digitize the physical world,” Kalanick wrote on X when the Atoms round was announced. “Understand, predict and control the physical world with software. Building ‘atoms-based’ computers where CPU is manufacturing, storage is real estate, and network is transportation.”
Atoms was built atop Cloud Kitchens, the ghost kitchen business Kalanick ran after leaving Uber. In March 2026, he rebranded and announced the acquisition of Pronto, the heavy industry automation company co-founded by Anthony Levandowski — the same engineer who ran Uber’s self-driving unit. Atoms is also pursuing mining automation, extending Pronto’s existing work in that sector.
The $1.7 billion round included Bain Capital and Fifth Wall alongside Andreessen Horowitz. The most notable participant, however, was Uber itself — the same company that pushed Kalanick out as CEO following complaints of sexual harassment, discrimination, and a toxic workplace, according to TechCrunch’s reporting at the time. The Uber re-investment signals two things: the industrial AI thesis is credible enough to override a complicated history, and Kalanick is back at the center of serious capital.
Kalanick’s stated goal is to build a “wheelbase for robots” — a platform layer for physical automation the way Uber built a platform for transportation. Ben Horowitz called it “unfinished business” in his own announcement post: “It takes a rare kind of entrepreneur to change these old-school, heavy parts of our economy.”
For practitioners, the relevant question is what Atoms’s bet on industrial AI tells you about where AI-driven automation is heading. The Atoms thesis is not that large language models will replace manufacturing workers — it is that AI-directed physical systems (autonomous vehicles in mines, robot arms in logistics, sensor-controlled kitchen equipment) become the profitable applications once the model layer matures. That maturation is what everyone at the table, including a16z, appears to believe has arrived.
Prentis: When AI Controls Your Computer, Not Just Answers Questions
Where Atoms goes after physical machines, Prentis goes after the layer just above: the screen. Prentis, launched in April 2026 and co-founded by Ritankar Das alongside Reid Hoffman and Mark Pincus, is training models to learn how office workers navigate documents and systems — then building AI agents that control computers to automate those same tasks, according to TechCrunch.
The target workflows are specific: insurance claims processing, customs duty refund exceptions, any task where a human currently hunts across multiple applications to find and move information. Prentis says it has signed contracts worth up to $50 million with several customers, including a healthcare management service organization, a manufacturer, and goods and clothing manufacturers, per two people familiar with the discussions. Its own pitch deck estimates a $75 million annualized run rate by Q3 2026 — though Prentis notes those figures are performance-dependent and based on 20% of realized savings, not recognized revenue.
The model Prentis is shipping is called Hive-32B. In its pitch materials, Prentis claims Hive-32B outperforms OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6 on two computer-use benchmarks: WindowsAgentArena, which measures end-to-end task completion on real Windows applications, and ScreenSpot-v2, which tests a model’s ability to locate the correct on-screen control. TechCrunch notes it has not independently verified those benchmark results.
The cost angle is where the Prentis argument gets practically interesting: the company claims roughly 10 times lower cost per task than frontier model APIs. If that holds, it reframes the computer-use category from an expensive curiosity to a deployable operations layer. This is a materially different capability from what you get querying GPT-5.4 or Claude Opus 4.6 via API today — those models can assist with computer-use tasks, but Hive-32B is purpose-built for Windows task completion at scale.
Computer-use agents are the fastest-growing application category that most enterprise teams are still treating as a demo. For context on where the autonomous agent space was even six months ago, our review of Cognition AI’s Devin 2.0 showed what a focused autonomous agent delivers when it is constrained to a single domain (software development). Prentis is applying the same constraint logic — small, cheap, purpose-built — to office work instead of code.
The Anthropic-Physical Intelligence Talks: What the Denial Tells You
On a Sunday in late July, tech blogger Robert Scoble posted on X that Anthropic had acquired Physical Intelligence (PI), the robotics startup co-founded by Lachy Groom. The rumor spread fast — fast enough that it required an official response. Physical Intelligence CEO Karol Hausman told employees via a Slack message containing a gif that the reports were not true, according to The Information. Groom did not respond to TechCrunch’s request for comment.
The denial is technically accurate. The acquisition did not happen. What actually happened, per The Information, is that Anthropic and Physical Intelligence held acquisition talks this spring — talks that ended without a deal.
Physical Intelligence is not an obscure target. The company has raised more than $1 billion to date and was reportedly in discussions this spring for an additional $1 billion round at an $11 billion valuation. Its pi-zero 0.5 (pi0.5) model is among the more widely used robot brains in robotics research. A Lachy Groom-affiliated company at an $11 billion valuation, building models that control physical robots — that is exactly the kind of target that would attract Anthropic at a moment when it is building toward an IPO and competing with OpenAI’s more aggressive acquisition pace.
Anthropic has made four known acquisitions this year. OpenAI has acquired at least 17 companies since 2023, per TechCrunch. Both companies confidentially filed for IPO in June — Anthropic on June 1, OpenAI a week later — in what could be two of the largest U.S. stock debuts in recent history. That context matters: both labs are in an acquisition sprint precisely because they need to show enterprise revenue, not just model capability, before going public. Physical AI and robotics represent the fastest path to defensible, high-margin industrial contracts.
The acquisition talks failing does not mean the thesis failed. It means the price did not clear, or the terms did not work, or Anthropic decided its own robotics research could get there faster. The rumor spreading as fast as it did — and being taken seriously by the market — is itself a signal about what the market believes Anthropic’s next move should be. For more on where Anthropic’s model lineup sits competitively right now, see our July 2026 Claude Opus 5 vs GPT-5.6 vs Kimi K3 breakdown.
Why This Is All Happening Now: The Velocity That Attracted the Money
None of these bets would be happening at this scale without the underlying growth data. Anthropic hit a $47 billion annual revenue run rate by May 2026, up from $9 billion in 2025, according to Matt Murphy, a partner at Menlo Ventures, which led Anthropic’s $500 million Series D. Murphy told TechCrunch’s Equity podcast that this kind of growth trajectory is unlike anything he has seen in 25 years of investing — not in the internet wave, not in mobile, not in the first cloud boom.
Murphy’s view is that what turned Anthropic from a strong model into a platform was not the model itself. Claude Code, Model Context Protocol (MCP), and Claude Skills — the tooling and integration layer — did more to drive enterprise revenue than any individual benchmark improvement. That framing maps exactly onto what Atoms and Prentis are each building in their respective layers: not better base models, but durable infrastructure that makes AI capability actionable in specific, high-value workflows.
Companies like Lovable and Legora are growing faster than any startups Murphy has seen across his 25-year career, he said. That growth is pulling capital up into the application and robotics layers because investors who missed the model wave are not going to miss the infrastructure wave that sits on top of it.
What This Means for You
If you are an AI practitioner building or evaluating automation tools right now, these three stories converge on three operational realities:
Computer-use agents are coming off the demo track. Prentis’s contracts with actual enterprise customers — healthcare, manufacturing, logistics — at a claimed 10x cost advantage over frontier APIs means the ROI math on computer-use automation is starting to work. You do not need to wait for GPT-5.x’s native computer-use to improve. Smaller purpose-built models beating frontier APIs on specific benchmark tasks (WindowsAgentArena, ScreenSpot-v2) is the early signal that the category is real. Begin scoping which high-volume, multi-system workflows in your organization currently require a human to navigate software — those are your first targets. Our guide on building AI agents with n8n and Claude covers the orchestration layer you would use to connect these tools today.
Industrial AI is a separate category from enterprise software AI. Atoms is not building a better Salesforce plugin. Kalanick’s thesis — CPU as manufacturing, storage as real estate, network as transportation — is about AI directing physical systems: mining equipment, logistics vehicles, industrial kitchens. If your organization operates physical infrastructure at scale, the wave is no longer hypothetical. The $1.7 billion a16z commitment is a market signal about timeline, not just technology.
Foundation labs are in a race to extend their reach before IPO. Anthropic’s four acquisitions this year and the failed PI talks both point to a lab that knows model capability alone does not sustain public market valuations. OpenAI’s 17+ acquisitions reflect the same pressure. What this means for buyers: the pricing and terms on enterprise AI contracts are likely to shift as both companies enter public markets and need to show durable recurring revenue. Lock in favorable terms on multi-year Claude API or OpenAI contracts now, before the IPO pressure translates into pricing leverage on the vendor side.
Next Step
The physical AI and computer-use categories will move faster in the next six months than most enterprise AI planning cycles assume. The clearest action is to run a quick audit of your highest-volume, multi-step human workflows — specifically the ones where a person is navigating multiple applications to complete a single task. Those are the workflows where purpose-built computer-use agents will land first, at costs that make traditional RPA solutions look expensive by comparison. Subscribe to the vtechnews digest to track which vendors move from benchmark claims to verified enterprise deployments as this category matures.
