Eka's Robot Picks 1,500 Things an Hour. Its Gripper Is Rated for About 139 Days of That.

Eka Robotics said on October 20 that it has closed a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime, a fund affiliated with Fidelity's parent FMR LLC (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/; https://ekarobotics.com/blog/eka-robotics-raises-150m-to-build-superhuman-intelligence-for-the-physical-world). The headline number is a robot that picks about 1,500 objects of up to 1 kg an hour, against roughly 600 for a human, and that Eka says succeeds more than 95 percent of the time (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/); a16z specifies the 1,500 comes from laboratory induction and bin-picking tasks (https://a16z.com/announcement/investing-in-eka-robotics/).

What has not been disclosed is most of what an operator would ask first. The pilot customers are unnamed; TechCrunch describes "secretive pilots with a number of industrial customers doing pick-and-place and wire manipulation" (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). The 1,500 figure is a lab figure, and a16z's own announcement calls it that, then adds that "sustaining that performance through full production shifts is the next test" (https://a16z.com/announcement/investing-in-eka-robotics/). There is no production-shift data, no uptime number, and no interval between hardware swaps in the launch posts from Eka, a16z, F-Prime or Obvious Ventures.

Our view: the manipulation results look genuinely strong to outside roboticists, and we have no reason to doubt them; Kris Hauser at UIUC talked about a "wow factor" and a robot that adapted to many objects without retraining (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). But Pulkit Agrawal gave TechCrunch a second set of numbers that got far less attention, and they are the ones that matter for a line. At the speed that justifies the price, Eka's gripper uses up its 5-million-cycle rating in about 139 days of round-the-clock work (arithmetic), so the number to watch is cycles per gripper, not picks per hour, and the distance from 5 million to the 20 million Agrawal himself calls the industrial bar is the distance between a demo and a line.

The deployability numbers Agrawal volunteered

TechCrunch reports that Agrawal "is also focused on deployability." Per that report, Eka's gripper hardware costs less than $1,000 and "is rated for 5 million duty cycles, but the ultimate goal is to hit the 20 million cycle bar set by standard industrial arms" (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). It is to his credit that the number is public at all. It also sets up a piece of arithmetic none of the launch posts do.

Divide 5,000,000 cycles by 1,500 picks an hour and you get about 3,333 operating hours (arithmetic). Run those hours continuously and you get about 139 days, a little over four and a half months (arithmetic). A full year of 24/7 operation at the advertised rate is 8,760 hours times 1,500, or 13.14 million cycles (arithmetic): 2.6 times the current rating (arithmetic) and about 66 percent of the 20-million bar (arithmetic).

That year is not an arbitrary interval. Hauser named it: "What's going to be a big challenge is when they work with integrators on real assembly lines, and it needs to run without breaking for an entire year" (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). At the speed Eka is selling, a year of continuous work is more than two and a half grippers. At the 20-million bar, it is two thirds of one (arithmetic).

Eka knows the test. Its own blog says "generality without performance isn't a product. It's a lab demo," and describes the job ahead as "transforming these capabilities into sustained, reliable performance in real operations" (https://ekarobotics.com/blog/eka-robotics-raises-150m-to-build-superhuman-intelligence-for-the-physical-world). Shyam Kamadolli of Fine Structure Ventures put it in operator terms: "a robot's value comes down to two things: its success rate and its speed over thousands of repetitions" (https://ekarobotics.com/blog/eka-robotics-raises-150m-to-build-superhuman-intelligence-for-the-physical-world). Thousands is the wrong order of magnitude. A day of 24/7 at 1,500 an hour is 36,000 repetitions (arithmetic).

Same gripper, three shift patterns

How fast the rating runs out depends entirely on how many hours the robot works, so here is the same 5 million and 20 million against three schedules. All figures are computed from Agrawal's ratings and Eka's 1,500-per-hour claim.

Read the bottom row honestly and the problem looks smaller. On a single shift, today's gripper clears Hauser's one-year line with months to spare (arithmetic). However, a single shift is also where the 2.5x speed advantage is weakest as an economic argument, because the human it replaces works a single shift too. The case for paying a premium for a fast arm usually rests on running it more hours than a person, and every hour added moves the swap date forward. We flagged the same framing issue with Ultra's OP1, where the autonomy figure disclosed was for one shift and remote pilots covered the rest (https://roboterminal.app/posts/ultra-op1-stationary-raas). Shift assumptions are where robot economics hide.

The middle row is the one most factories would recognize, and it still falls short of Hauser's year: two shifts reach 5 million cycles in about 9.6 months (arithmetic). Two caveats run the other way. A rating is a rating, and parts can outlive them. And nobody outside the company can yet check the number against a real swap log.

Speed and wear are the same bet

The reason to treat cycle life as the central number, rather than a footnote, is that Eka's speed and its wear come from the same design decision.

TechCrunch reports that "the robot arm is strong enough to hurt itself as it moves that quickly," and an anonymous frontier-lab roboticist told the outlet: "they have a lot of trust in their algorithm not breaking the robot. I wonder how many robots they broke along the way?" (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). The 2.4-second pick (arithmetic) comes from a controller trained in simulation to move fast with hardware that, by TechCrunch's account, can damage itself at that speed. Nothing Eka has published says its simulator models wear. Metal keeps that count on its own.

Then there is the retry behavior that reviewers found so compelling. Hauser described a "creepy humanness that it would try over and over again if it failed the first time" (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). In April, WIRED's Will Knight wrote that the robot "swoops down and nips gently at the item a few times before grasping and lifting it up" (https://www.wired.com/story/when-robots-have-their-chatgpt-moment-remember-these-pincers/). Agrawal's description of the method explains why the searching is deliberate: "Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces" (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). The gripper searches by touching. Touching is a cycle.

Or it may be. The industry has a conventional definition, and it is strict. Robotiq's Hand-E manual, a 2021 edition, defines one cycle as "an object picking attempt, successful or not (open or closing onto an object, or closing on itself)" (https://assets.robotiq.com/website-assets/support_documents/document/Hand-E_TM_InstructionManual_PDF_20210708.pdf). Under that definition a 95 percent success rate means about 1.05 attempts per successful pick (arithmetic), so a 5-million-cycle gripper delivers roughly 4.75 million successful picks (arithmetic), not 5 million. Under that definition, each of Knight's nips is also a cycle, and the count per successful pick rises further. We do not know whether Eka counts nips as cycles, whether its 5 million is a mean or a minimum, or how it was established. Those three questions would be our first email.

The same manual covers the Hand-E for one year or 5,000,000 cycles, whichever comes first, lists finger and fingertip wear as uncovered, and shows maintenance columns at 1 million and 2 million cycles (https://assets.robotiq.com/website-assets/support_documents/document/Hand-E_TM_InstructionManual_PDF_20210708.pdf); Robotiq's knowledge base says maintenance "should be done every 2 million cycles" (https://blog.robotiq.com/knowledge/preventive-maintenance-on-the-hand-e-gripper). Translate that cadence to Eka's speed and a round-the-clock gripper is due for service every 55 days or so (arithmetic). Whether Eka's part needs anything similar is undisclosed.

Where the claimed moat sits, and why it runs through the fingers

The investors are explicit that the bet is the controller, not the planner. F-Prime's Sanjay Aggarwal writes that the vision-language-action planning layer "has seen incredible progress over the last 24 months, but it is also getting increasingly commoditized," and names the open releases he has in mind: Physical Intelligence's pi0.5, NVIDIA's GR00T N1.7, and the Allen Institute's MolmoAct2 (https://www.fprimecapital.com/blog/eka-robotics-the-path-to-superhuman-general-purpose-robots/). Earlier this month we ran a simple download test on three other robot-model launches, and only one shipped weights (https://roboterminal.app/posts/open-robot-models-download-test). F-Prime's complaint about them is that their action chunks execute "largely open-loop," which is exactly the gap a force-closed controller is supposed to fill (https://www.fprimecapital.com/blog/eka-robotics-the-path-to-superhuman-general-purpose-robots/).

The second half of the thesis is about data. a16z says Eka trains with reinforcement learning in physics simulation, so that "robot experience can increasingly scale with compute rather than human labor" (https://a16z.com/announcement/investing-in-eka-robotics/). F-Prime's version is sharper: human demonstrations "scale linearly with the number of people collecting the data, and mostly show the task going right" (https://www.fprimecapital.com/blog/eka-robotics-the-path-to-superhuman-general-purpose-robots/). That is a direct shot at the demonstration-data businesses, where Mecka says it passed a $100 million run-rate in June (https://roboterminal.app/posts/robot-data-sellers-vs-unitree). Knight noted in April that "some very smart experts believe that mixing human demonstration with simulation will yield better results than simulation alone," and the founders declined to give training details, citing commercial edge (https://www.wired.com/story/when-robots-have-their-chatgpt-moment-remember-these-pincers/). We cannot adjudicate that from outside.

What we can say is that both halves of the moat terminate in the same part. F-Prime itself writes that "high degree of freedom grippers are prone to failures," credits Eka with a "proprietary low-cost and reliable approach for force sensing," and notes the early demos use "simple two-fingered grippers with articulated fingers" (https://www.fprimecapital.com/blog/eka-robotics-the-path-to-superhuman-general-purpose-robots/). TechCrunch adds that the hand is "more flexible than traditional grippers" with more force sensors (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/). More sensors and more articulation per finger is more that can wear, at a cadence set by the fast controller that is the point of the company. The 5 million rating is where the model meets atoms.

Objection docket

"Grippers are consumables. At under $1,000 for 5 million cycles, this is less than $0.0002 a cycle. Nobody should care." The cost is indeed trivial: under $0.0002 per cycle (arithmetic), under $0.30 per operating hour (arithmetic). We agree the part cost is not the issue. The issue is that every swap is a stop, a technician, possibly a recalibration of a force-sensing hand, and a question about whether the model's behavior on a new gripper matches its behavior on the old one. Integrators price interruptions, not fingers. Hauser's year-without-breaking test is about interruptions.

"5 million is already what a commodity gripper is warranted for. Eka is at industry normal on day one." Partly right, and the Robotiq comparison supports it: one year or 5 million cycles, whichever first, in a 2021 manual (https://assets.robotiq.com/website-assets/support_documents/document/Hand-E_TM_InstructionManual_PDF_20210708.pdf). But a rating in cycles says nothing about calendar time on its own. Eka's pitch is 2.5 times human pace, and at that pace a round-the-clock gripper reaches 5 million cycles in about 139 days (arithmetic), long before a one-year clock would run out. And Agrawal did not say 5 million was fine; he named 20 million as the goal, which is a fourfold gap he chose to disclose (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/).

"These are lab numbers. Everyone says so, including a16z." Yes, and that is our point rather than an answer to it. The 1,500 and the 95 percent are lab numbers that have been quoted everywhere; the 5 million is a lab number that has been quoted almost nowhere. Both deserve the same scrutiny, and the second one has a date attached.

"The valuation is about the model, not a gripper." The round sold roughly 10.3 percent of the company (arithmetic) at a pre-money of about $1.30 billion (arithmetic), and the investors' essays are about the controller and the data loop, so this is fair as a description of what was priced. a16z's own post says "industrial automation has a brutally simple scoreboard: units per hour, error rates, uptime and cost per unit" (https://a16z.com/announcement/investing-in-eka-robotics/). Error rates and uptime run through the gripper as much as through the model.

Three things to check

1. A named pilot customer or integrator, with any duty-cycle or uptime disclosure, by the end of Q1 2027. Raghuram told TechCrunch that "over the next few months there will be superhuman performance in many dexterous manipulation tasks" (https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/); we read "few months" as through March 2027. Obvious Ventures names trash sorting and meat packing among the jobs it expects Eka to automate (https://obvious.com/ideas/why-we-invested-in-eka-robotics/). A customer in either, with hours run and swaps logged, would be the first real data point.

1. March 10, 2027. That is 139 days after October 22, 2026 (arithmetic): the date a gripper started round the clock at 1,500 an hour during announcement week would reach 5 million cycles. The questions to put to Eka around then are whether it discloses swap intervals from pilots, whether the 5 million rating has moved, and what the company counts as a cycle.

1. October 2026 to October 2027. Hauser's one-year test. A year of 24/7 at the advertised speed is 13.14 million cycles (arithmetic), 2.6 times the current rating (arithmetic) and about 66 percent of the 20-million bar (arithmetic). If Eka reports a full year on any customer line, the number we want is grippers consumed, not picks completed.

The robot already feels like the future. The gripper's spec sheet is where you find out when it arrives.