$1.45B Series A for ChatGPT hands — still secret on the floor

On October 20, 2026, Cambridge-based Eka Robotics closed a $150 million Series A that values it at $1.45 billion post-money, led by a16z and F-Prime. The pitch is a vision–force–action (VFA) model trained in simulation, plus a force-sensing gripper. WIRED called an earlier demo a “ChatGPT moment for the physical world.” The deployments? Still secretive industrial pilots — pick-and-place and wire manipulation — not a commercial fleet.

$1.45B buys the ChatGPT-for-hands story

TechCrunch’s October 20 report and Eka’s own Series A post agree on the headline numbers: $150 million raised, $1.45 billion post-money, co-led by Andreessen Horowitz and F-Prime. Seed leads Obvious Ventures and E14 Fund returned. MFV Partners, Bison Ventures, Glasswing Ventures, Samsung Next and several Fortune-500 strategics joined.

That is a Series A priced like a late-stage physical-AI bet. For context on the same board of “brains for robots”: Skild AI’s January 2026 Series C cleared >$14B (company / TechCrunch). Physical Intelligence’s May 2026 Series C is tracked around ~$11B via private-market filings — labelled claimed, not a company PR. Same day as Eka, Reliable Robotics upsized a Series D extension by $115M past a >$1B valuation for autonomous aviation (company). FAIRINO’s recent China cobot C+/D is media-claimed at roughly $140M combined — round size, not a confirmed post-money.

Our read: markets are paying up for a force-first narrative while floor hours stay private. The valuation is confirmed; the commercial proof is not.

Vision–Force–Action, not another VLA

Eka’s bet is architectural. Instead of training a vision–language–action model on human demonstration video, the company says it builds a vision–force–action (VFA) model inside a physics-rich simulator, then transfers via reinforcement learning at scale. Co-founder Agrawal told TechCrunch that vision gets you “roughly where the object is,” while “the actual grabbing… is happening through forces.” Haarnoja frames the stack as bottom-up physical intelligence rather than top-down planning.

That story has a press pedigree. In April 2026, WIRED’s Will Knight watched the Cambridge lab screw in a light bulb, chase a rolling bulb across a table, and pack chicken nuggets — and wrote that Eka’s claw “feels like we’re approaching a ChatGPT moment for the physical world.” Expert quotes in the Series A coverage (UIUC’s Kris Hauser; an anonymous frontier-lab roboticist) lean emotional: “wow factor,” “creepy humanness,” raspberry-safe contact. Those are anecdotes, not audited KPIs.

Hardware is part of the claim package. Eka designed a flexible, force-sensing gripper it prices under $1,000, currently rated for 5 million duty cycles with a stated goal of 20 million — the bar Agrawal cites for standard industrial arms. Throughput claims: about 1,500 picks per hour on objects up to 1 kg, versus roughly 600 for a human, with pick success “more than 95% of the time.” All company.

Superhuman demos, undisclosed lines

Here is the tension the $1.45B mark does not erase. Eka is “working in secretive pilots with a number of industrial customers” on pick-and-place and wire manipulation (TechCrunch). No customer names. No facility count. No published operating hours. Hauser’s caution in the same piece is the one investors should pin: integrators, year-long uptime, and tasks beyond pick-and-place will decide whether VFA is a product or a lab demo with a Series A price tag.

Stack that against peers who also sell “physical AI” at double-digit billions with their own disclosure gaps — and against Reliable’s same-day aviation raise, where the hard gate is FAA certification rather than secret factory pilots. Different physics. Same capital market habit: price the story before the public proof.

If the robotics decade rhymes with the EV decade, remember the Western frame we use here: German auto OEMs as a class, not a single marque. Manipulation startups will eventually face the same buyer: plants that care about cycles, scrap and mean time between failure — not raspberry demos.

Four signals after the raise

• Named industrial customers or a published pilot-site count — not “a number of” undisclosed lines.
• Independent or customer-reported throughput / uptime over months, not demo-hour claims.
• Gripper cycle life moving from the rated 5M toward the 20M industrial goal under real duty.
• Evidence that VFA generalizes past pick-arrange-place and wire handling into assembly tasks customers will pay for.