The humanoid industry has produced an extraordinary volume of announcements in the past two years: funding rounds, pilot programs, backflips, dance routines, sprint records. What it has produced almost none of is prices. List prices for general-purpose humanoids remain rare, and when they exist they describe aspirations more than transactions. That is what makes the quiet commercial disclosure from Foundation Future Industries worth the close read it has not fully received: the San Francisco company, speaking to Reuters around its appearance at AMD’s Advancing AI 2026 event, said its industrial Phantom MK-1 units lease for approximately US$100,000 per robot per year.
That single figure, bundled alongside the announcement that Foundation’s next-generation Phantom MK-2 will be built on AMD processors, is the first widely reported rental price for a humanoid robot doing claimed paid factory work in the United States. It deserves better than to be filed as a footnote to a chip partnership. A lease price is a different species of number than a list price or a preorder: it implies contracts, service obligations and a customer who has done the internal math on whether the machine earns its keep. This piece takes the number apart, sets it against labor economics, examines the AMD deal that accompanied it and itemizes which of Foundation’s surrounding claims survive scrutiny.
The Number That Slipped Out
Start with what is actually on the record. The lease figure came via Reuters and was disclosed in the same window as Foundation’s July appearance at AMD’s Advancing AI 2026 event at the Moscone Center in San Francisco, where the company participated in a panel on AI-powered robots for real-world environments. An approximate $100,000 annual figure per robot, per year.
Three clarifications matter immediately.
First, it is a lease term, not a purchase price. Foundation has not disclosed a purchase price for Phantom MK-1, and the base robot’s sale cost remains listed as undisclosed in tracking databases. Buyers are not acquiring an asset; they are renting a capability.
Second, the disclosed number is approximate and unaccompanied by its context. Contract duration, service coverage, uptime commitments, integration costs, minimum fleet commitments and redeployment terms are all absent. A $100,000 annual lease on a five-year commitment with full service is a very different proposition from a one-year trial with thin support.
Third, and most important, it is a real-world data point in a market that has been running on hope. When Barclays assessed global humanoid adoption earlier this month, it projected more than 60,000 new units entering service in 2026 alone. Sixty thousand units implies an emerging rental and purchase market, and markets need reference prices. Foundation has now supplied the first one for industrial humanoids in the US.
The Unit Economics: Benchmarking Against Labor
What does $100,000 per robot-year buy, and against what should it be measured?
The honest benchmark is fully loaded labor cost, not wage alone. A US manufacturing worker earning $25 to $30 per hour costs an employer roughly $55,000 to $70,000 per year once benefits, payroll taxes, workers’ compensation, insurance and turnover are included, using standard loading multipliers. In unionized automotive environments or high-cost regions, fully loaded costs above $80,000 to $100,000 per year are common.
Against that benchmark, a $100,000 annual humanoid lease is not obviously cheap. It is roughly at parity with an experienced full-time industrial worker in an expensive market. The pitch, then, cannot be headcount replacement at half the cost, at least not at this price point and not yet. It has to be something else, and Foundation’s deployment profile suggests what: the company says its robots operate 24 hours a day, five days a week in customer deployments. If a leased Phantom genuinely works three shifts rather than one, the effective comparison is not against one worker but against two to three shifts of labor, which reframes the same $100,000 as $33,000 to $50,000 per shift-equivalent. That is a materially different proposition, and uptime is the variable that decides it.
This is why the missing service terms matter so much. Robotics-as-a-service pricing only works if someone carries the maintenance, repair and failure risk. If the lease includes full service and guaranteed uptime, Foundation is absorbing enormous hardware risk across a fleet of machines whose actuator and gearbox longevity under continuous industrial load remains an open question across the entire industry. If it does not, the customer is absorbing it, and the effective cost of ownership is higher than the headline. Either way, the disclosure invites the right question and leaves it unanswered, which is itself informative: uptime terms are the number industrial buyers will demand next, and no humanoid maker currently publishes them at fleet scale.
There is also a strategic reading. Pricing at labor parity rather than at a discount is the behavior of a supplier with more demand than supply, or one that believes its performance justifies parity. Foundation claims demand is real, pointing to what it says is $100 million in contracted annual recurring revenue. That figure, if accurate, would place it among a very small group of humanoid companies with meaningful recurring income. It is also unaudited, uncustomered and undisclosed in duration: the contracts’ length, cancellation terms and deployment schedules are all unknown. Treat it as a claim, not a fact, but a claim that at least has a specific shape.
Why AMD: The FPGA Argument for Hands
The lease disclosure arrived stapled to a compute story, and the compute story is more consequential than it first appears.
Foundation says Phantom MK-2 will be built natively around AMD processors: the new Ryzen AI Embedded X100 series, delivered on Kria AI system-on-modules. The X100, announced at Advancing AI 2026, pairs up to 16 Zen 5 CPU cores with RDNA 3.5 graphics and an XDNA 2 neural processing unit on a single package, drawing on shared memory. It is AMD’s first embedded family aimed squarely at physical AI, and volume production is expected in the fourth quarter of 2026, which means the first MK-2 units shipping on the new silicon are still months out.
The technically interesting part is not the CPU-GPU-NPU bundle, which mirrors what NVIDIA sells, but what sits next to it. Foundation will use AMD field-programmable gate arrays to drive the robot’s hands and sensing. Andrea Esposito, who leads hand development at Foundation, said the FPGAs let its custom hand “actuate and coordinate all 23 degrees of freedom while fusing high-frequency tactile feedback in a single deterministic control loop”.
That sentence deserves unpacking because it is the strongest engineering argument in the whole announcement. A 23-degree-of-freedom hand running high-frequency tactile feedback does not primarily need raw throughput; it needs timing guarantees. Motor coordination and touch sensing at control rates measured in hundreds of hertz fail catastrophically if the compute pipeline hiccups, and general-purpose operating systems hiccup. FPGAs, reconfigurable logic circuits that AMD acquired the leading franchise in through its 2022 Xilinx purchase, execute such loops with predictable, deterministic timing that a GPU-first stack struggles to match. In other words, AMD’s differentiated asset in humanoid robotics is not faster AI inference. It is determinism, purchased through Xilinx, applied to the part of the robot that is hardest to control: the hand.
Dexterous manipulation is widely regarded within the field as the binding constraint on humanoid usefulness. Legs are largely solved; hands are not. If FPGA-driven deterministic control genuinely improves 23-DoF tactile manipulation, Foundation will have found a real architectural edge in the least commoditized component of the stack. That is a big if, and it is testable: independent manipulation benchmarks, not vendor demos, will settle it.
The Compute War Beneath the Chassis
The strategic significance is that NVIDIA has finally met a credible counter-offer. Since the Jetson module line and the Isaac simulation stack, NVIDIA has been the default brain for most humanoid projects, a gravity well reinforced by its investment positions across the robotics ecosystem. Foundation, a company with robots doing claimed paid work in customer plants rather than running demo reels, is committing its production fleet to the alternative. AMD, for its part, introduced a Kria robotics platform and an open partner network in the same week, bundling CPUs, GPUs, NPUs and FPGAs into a turnkey offering that mirrors NVIDIA’s integrated hardware-plus-software approach.
Foundation claims its AMD architecture delivers approximately 2.5 times faster inference and three times faster training than the NVIDIA platform it compared against. Note what is missing: the specific NVIDIA parts, the models, the workloads, the methodology. AMD’s own lab numbers claim up to three times the peak single-precision throughput of NVIDIA’s Jetson T5000. Both sets of figures belong in the claims column until independent hardware testing exists. The lesson of physical AI so far is that datasheet numbers rarely survive contact with a factory floor.
But the benchmark dispute is almost secondary. Whoever supplies the brain for the first large humanoid fleets shapes the software habits of the industry: the toolchains, the simulation environments, the deployment pipelines, the developer loyalties. That is the prize NVIDIA won in data centers and is defending in robotics, and it is the prize AMD is now contesting with a portfolio argument NVIDIA cannot fully replicate: a CPU, GPU, NPU and FPGA family under one roof. Foundation is a single customer, but it is the right kind of customer to make the argument credible.
The Claims Ledger
A disciplined read of Foundation’s public position requires separating record from assertion.
On the claim side: $100 million in contracted annual recurring revenue; contribution to more than 24,000 vehicles built in 2025; round-the-clock, five-day operation; the 2.5x and 3x performance figures. None of these are independently verified, the automotive customer is unnamed, the robot count is undisclosed, and no external party has measured intervention rates, the fraction of working time a human must step in.
On the record side: the lease price, disclosed to Reuters; the AMD partnership, announced jointly; the MK-2 hardening targets, which are at least specific and falsifiable: tolerance for 100G shocks, an IP67 dust-and-water rating and 360-degree vision; the October target for a first factory with planned capacity of 5,000 robots per year, followed by a second targeting 50,000. Note that factory capacities are plans, not output, and 50,000 units per year would exceed what the entire industry is projected to field in the US this decade without named demand behind it.
Foundation also carries an unusual political profile that belongs in any honest ledger: Eric Trump is an investor and the company’s chief strategy adviser, the company holds roughly $24 million in Pentagon research contracts and has tested units in Ukraine, an arrangement Senator Elizabeth Warren has publicly criticized. None of this bears on the engineering. All of it bears on procurement flows, defense-channel access and the regulatory weather the company will operate in, which for a dual-use industrial-defense platform is material.
What to Watch
The lease price converts Foundation from a story company into a checkable one. Four checkpoints will do the converting.
First, uptime disclosure. The moment any industrial humanoid maker publishes fleet-level uptime and intervention-rate data alongside its lease price, the RaaS model becomes analyzable. Until then, $100,000 per year is a price without a performance denominator.
Second, the fourth quarter. AMD says X100 volume production begins then, and Foundation’s first factory is slated for October. Silicon shipping and a factory opening within the same quarter is an aggressive alignment of timelines; slippage on either will be visible.
Third, named customers. A single automotive customer willing to put its name to the 24,000-vehicle claim would reprice the credibility of everything else in the ledger.
Fourth, the NVIDIA response. The incumbent has not lost a design war in robotics yet, and it now has a named defector making public performance claims against it. Expect counter-benchmarks, and treat them with exactly the same skepticism.
The deeper significance of this week’s news is not that one startup chose one chipmaker. It is that the industry’s economics have become concrete enough to argue about. A price, a revenue figure, a factory target, a benchmark fight: these are the artifacts of a real industrial category forming. The humanoid market spent 2024 and 2025 demonstrating that the machines can move. In 2026 it is starting to discover what they cost. One hundred thousand dollars a year is the opening bid in that conversation, and every competitor, investor and plant manager now has a number to argue with. That, more than any backflip, is how a market begins.