In a foundry in Imatra, Finland, three graphite electrodes feed a 75-ton electric arc furnace, the kind of installation that turns scrap steel into billets. This month it ingested something no foundry had ever processed: an entire fleet of commercial humanoid robots, walking to the edge and leaping into the melt under their own learned policies. Figure announced this week that it has decommissioned its F.02 fleet, the hardware generation that carried the company’s first BMW deployment, the birth of Helix, its first household chore demonstrations, and its first logistics customer. A few units remain in storage at headquarters. Everything else is now commemorative metal.
The internet saw a Terminator-flavored publicity stunt, and it was that: Arnold Schwarzenegger, who told founder Brett Adcock to “melt them,” was involved, and the resulting steel is being machined into a limited series of collectible artifacts. But underneath the spectacle sits the least glamorous and most consequential story in humanoid robotics right now. A major operator has retired an entire hardware generation roughly two years after it shipped, and the disposal method it chose was annihilation, because the alternatives were worse. Every assumption investors make about humanoid unit economics, from residual value to lease pricing to depreciation schedules, just got its first contact with reality.
What the melt was, and what it was not
Start with what actually happened, because the details matter more than the memes.
Figure says it contacted metal foundries in the United States and Mexico first, and none would accept the job. The problem was not the robots’ aluminum and steel; it was the lithium-ion batteries, which no foundry wants inside expensive melting equipment. The company even recruited former MythBusters staff to hunt for a willing site. The only taker in the world, Figure says, was the foundry in Imatra. That sentence alone should be read twice by anyone planning fleet-scale humanoid deployments: there is, today, effectively no industrial disposal infrastructure for this product category. The first company to face the problem at fleet scale had to export it to Finland.
At the site, the constraint was time. The team had a 24-hour operating window and six melts, and each melt stayed hot enough for only about 20 minutes before a crust formed on the steel. The robots therefore had to be reliable, precise, and fast, which is precisely why the engineering story deserves more attention than it has received.
The decommissioning problem nobody modeled
Why melt a fleet instead of selling it, storing it, or recycling it properly? Figure’s own explanation is unusually candid, and it is the most economically revealing paragraph the company has published this year.
“Decommissioning a fleet of robots full of custom-built actuators and other intellectual property is no easy task,” the company wrote. “We needed to dispose of the robots in a way that protected our proprietary hardware. Disassembling each robot would take our technical staff so much time that it would delay the launch of F.04.”
Unpack that. First, the resale option was foreclosed by IP: a two-year-old humanoid packed with custom actuators cannot go to a recycler or a broker without leaking the family jewels. Second, the teardown option was foreclosed by opportunity cost: Figure’s engineers are its scarcest asset, and hours spent unscrewing old robots are hours not spent shipping F.04. Third, storage was foreclosed by the generation cycle itself; in the company’s words, “as our F.03 fleet grows, maintaining the F.02 fleet no longer makes sense.”
Put together, the F.02 fleet had negative residual value. Not zero: negative. Figure paid shipping, logistics, film production, and machining costs to make the fleet disappear, and monetized the outcome as merchandise. When an asset class reaches end of life with negative value and requires IP-protected destruction, it is not an asset class in any accounting sense. It is a consumable with a disposal fee.
This is the first hard data point on humanoid depreciation, and it lands in the same season the industry’s finance story goes public. Agility Robotics is preparing to list on Nasdaq as AGLT through its SPAC combination. Public filings require useful-life assumptions for robot fleets, and every analyst building a humanoid model must now confront Figure’s implicit answer: about two years, with zero salvage. Traditional industrial arms are commonly depreciated over five to ten years and stay in service for decades with maintenance. A gap that wide does not just change tax treatment; it changes whether the business model closes at all.
Consider the leasing math. Foundation Future Industries disclosed last month that its Phantom MK-1 industrial humanoid leases for roughly $100,000 per robot per year. If the underlying hardware, like Figure’s, has a useful life near two years, the lessor must recover capital, servicing, and downtime inside 24 months of billable hours. That is a dramatically harder equation than the same price spread over a seven-year life, and it explains why the industry’s pricing has been drifting toward subscriptions and outcome-based contracts rather than sales: the risk of premature obsolescence has to live somewhere, and vendors would rather it live in their own depreciation schedule than in a customer’s capital budget.
The accidental robotics experiment
Here is the part the coverage mostly missed. To execute the melt, Figure trained a new AI model entirely in simulation, using stunt-artist motion capture as a reference, so that robots could jump accurately into a bucket of molten steel at a foundry the team had never visited. Training happened at the San Jose campus, with airbags, teaching robots to leap off the second floor of a building. Deployment happened 8,500 kilometers away, in an environment with extreme radiant heat and the severe electromagnetic interference of an operating arc furnace. Figure reports that camera equipment and other electronics failed on site, while the robots’ learned policies “ran seamlessly.”
Strip away the theatrics and that is a serious result on three axes. Sim-to-real transfer: a policy trained purely in simulation for a never-seen geometry executed first-try in the field, which is the core claim of the entire sim-to-real research program and the thing that historically fails most often. Whole-body control: aggressive, precise, ballistic motion, jumping to a small target with a 20-minute window, is exactly the class of dynamic behavior that classical locomotion stacks struggled to produce and that learned policies now handle routinely. And robustness: in an environment where external sensing hardware died, the robots kept executing, which suggests the control stack leaned on proprioception, joint encoders and IMUs, rather than on vision. For a company whose product pitch is “Helix runs the robot,” an involuntary EMI stress test that the policies passed while the cameras failed is not a footnote. It is evidence.
There is also a wink in the direction of data economics. The same week, Figure confirmed its multi-year Nscale partnership covering up to 100,000 Nvidia Vera Rubin GPUs, with an initial $3.5 billion of compute capacity and intent to exceed $6 billion, aimed at training Helix. The company’s stated logic, from Adcock: “Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute.” A generation of hardware that exists to generate training data and is then destroyed once a better sensor and compute platform arrives is a coherent strategy under that logic. The robots are not just products. They are also disposable instruments for collecting the corpus that makes the next model better.
Two strategies for the same constraint
Set the melt beside the other big hardware signal of the week and a pattern emerges. Elon Musk disclosed that Tesla halved the Optimus brain’s memory, to roughly 96GB on AI5, explicitly because memory supply was “the only way” to reach production volume, even as Micron argues humanoids will eventually want 200GB or more each. Tesla is stretching components across as much volume as the supply chain allows. Figure is burning a generation to accelerate the next one.
These look like opposite moves, but they are responses to the same binding constraint: hardware iteration speed, in a supply chain that will not flex. Tesla pays for it in degraded per-unit capability and deferred autonomy; Figure pays for it in disposal costs and sacrificed residual value. Both companies are effectively admitting that the robot’s value lives in the model, and the metal is a rapidly depreciating container for it. That admission has a second-order consequence for the supply chain: component makers being asked to support two-year cycles, from harmonic drives to tactile sensors to DRAM, are looking at a market that behaves like consumer electronics, not industrial machinery. Pricing and capacity planning will follow.
The recycling gap goes vertical
The production side of the industry is industrializing fast. XPeng says it commissioned the first automated production line for general-purpose humanoids in September, robots building robots, and Figure’s own BotQ facility targets 12,000 humanoids per year. China accounted for the overwhelming share of global humanoid installations last year by most counts. The build-up machine is real.
The tear-down machine does not exist. When BotQ-class facilities reach their stated volumes, the industry will be producing tens of thousands of humanoids annually, each carrying lithium batteries, dense electronics, and rare-earth-rich actuators, on a two-to-three-year replacement cadence. Multiply it out and you get fleets of retired robots measured in the tens of thousands per year by the end of the decade, and the current disposal playbook is: find one foundry in Finland. Battery regulations in Europe and North America already treat lithium packs as hazardous waste requiring certified handling; a fleet-scale decommissioning will not be legally permitted to look like a marketing film. The first company to build certified humanoid recycling, battery extraction, actuator re-manufacturing, and certified refurbished sales, is attacking a market the leaders have structurally ignored, and it will be born with guaranteed demand from the leaders’ own churn.
There is also a missed-market angle worth naming. Figure destroyed F.02 partly to protect proprietary actuators, which implies the company assigned roughly zero value to a second life for those machines, say, in research labs, emerging-market deployments, or training rigs behind NDAs. Either that market genuinely does not exist yet, or no vendor has built the legal and logistical wrapper to make it safe. Both answers are informative. The used-humanoid market is currently priced at nothing by the people best positioned to know.
The marketing ledger
None of this is to deny the stunt’s brilliance as marketing. The Schwarzenegger inversion, the actor who spent a career fighting killer machines now presiding over their voluntary self-destruction, is a once-per-generation brand beat, executed with self-awareness and a wink at AI-skeptic audiences (“we really trained robots to jump autonomously,” Figure insists, preempting the fake-footage charge). The artifacts convert a cost center into revenue and turn the fleet’s death into an asset. And for recruiting, the message to elite robotics engineers is unambiguous: this company ships, breaks things, and moves on.
The risk is more subtle. A company celebrated for melting its old robots has publicly priced its own hardware at zero, two years in. Investors who just underwrote a $39 billion valuation on the back of Series C participants like Nvidia, Qualcomm, and Brookfield should notice that the assets being built have the durability of flagships phones, not factories. The defense, and it is a real one, is that Figure is no longer selling metal. It is selling Helix’s labor, with the robot as the razor and autonomy as the blades. The melt is that thesis stated in 1,600 degrees Celsius.
What to watch
Four checkpoints will tell us whether F.02’s furnace was an outlier or a template. First, F.04’s arrival: the decommission post implies the next generation is close, and the cadence from F.03 to F.04 will confirm or break the two-year generational rhythm. Second, Agility’s public filings as AGLT, which should force the first disclosed useful-life assumptions for a deployed humanoid fleet into the open, a number every competitor’s model will be checked against. Third, whether any first-tier vendor establishes a certified refurbishment or resale channel, which would signal that residual values are quietly positive and disposal-by-furnace was a one-off. And fourth, regulators: the first time a multi-thousand-unit fleet retires under EU battery rules, the industry will discover the real unit economics of decommissioning, with lawyers instead of Schwarzenegger in the film.
The humanoid industry has spent three years competing over whose robot can do the most. This week, for the first time, a major player showed us what happens when the robot can no longer do enough. The answer was a furnace in Finland, and it belongs in every investor model from now on.