The Artificial Engineer Comes for Rare Earths

Highlights

• Prometheus raised $12B at a $41B valuation to build an 'artificial general engineer' trained on physics and industrial data, not internet text.
• Near-term, AGE technology is bullish for rare earth volumes as faster hardware design drives more motors, turbines, and robots requiring NdFeB magnets.
• Long-term, AGE threatens to compress the multi-decade mine-to-magnet timelines that underpin Western rare earth investment valuations.
• China holds a dual advantage: rare earth processing dominance and the richest corpus of separation and metallization operating data needed to train physical AI models.
• Cost-curve dispersion will widen, rewarding data-rich operators and accelerating substitution and recycling from concept to commercial practice.

---

Jeff Bezos has put one of the largest sums ever raised by a young AI company behind a machine that designs physical things. For rare earths, the question is not whether it works, but which of two clocks it runs faster.

When Jeff Bezos surfaced this month as co-CEO of an AI company called Prometheus, (opens in a new tab) the framing in most coverage was that he had backed another large model. He had not. The defining race in artificial intelligence is the pursuit of artificial general intelligence, the AGI that is meant to match or surpass human cognition across open-ended tasks, and nearly every headline raise of the past two years has been some wager on it. Prometheus is chasing a deliberate sibling. It raised 12 billion USD at a 41 billion USD valuation to build what Bezos calls an "artificial general engineer," AGE: software trained not on internet text but on physics, simulation, and industrial data, aimed at the slow, expensive cycle of designing and manufacturing physical objects. (TechCrunch (opens in a new tab)) The name is a conscious echo of AGI, and so is the scale of the claim: where AGI promises to remake the digital economy, AGE promises to be just as consequential in the physical one. For a rare-earth audience the phrase sounds like a tech-page abstraction. It is not. If even part of the promise lands, it lands first on the most stubbornly physical part of the economy: the metals and minerals that everything else is built from.

The reason this matters to anyone trading, mining, or refining rare earths is a timing argument, and it is worth stating plainly before the reporting starts. An artificial engineer affects the rare-earth market through two forces that run on different clocks. In the near term it raises demand, because cheaper, faster engineering means more motors, more turbines, more robots, and more variants of each. Over a longer horizon it attacks the one number on which the entire Western rare-earth investment thesis rests: the two-to-three-decade timeline widely assumed to be required to stand up a non-Chinese mine-to-magnet chain. Cycle-time compression is precisely what an artificial general engineer is built to deliver. The near-term clock is bullish for volumes. The long-term clock is bearish for the terminal value of long-lived assets. Most of what follows is an attempt to say which clock should worry you more, and when.

What an artificial general engineer actually claims to be

The pitch is simple and radical at once. Instead of training a model on the near-free abundance of internet text, you train it on the laws of physics and on data from real experiments and operating plants, then use it to compress engineering work itself. Bezos has framed the ambition as taking a product that might require 100 engineers over a decade and letting 10 engineers finish it in a year. (Inc (opens in a new tab)) That is his claim, not a demonstrated property of any shipped product; Prometheus has disclosed neither results nor a release. The roughly 150-person team sits in San Francisco, London, and Zurich, recruited from OpenAI, DeepMind, Meta, and other frontier labs. (Yahoo Finance (opens in a new tab))

Before going further it is worth naming the strongest reason this might not work, because it sits at the center of the whole approach. The progress that put AGI within conceivable reach ran on one resource available at vast scale and near-zero cost: the public internet. AGE is trying to borrow that playbook without that data. Physical engineering data is the opposite of text. Experimental results are slow to generate, expensive to run, proprietary, and locked inside the firms and plants that produced them. Bezos has been candid that this is a capital-intensive company in large part because of the cost of building specialized training data, not just compute. (GeekWire (opens in a new tab)) Whether physical AI scales the way text models did, or stalls against the cost of generating its own ground truth, is the open question that decides everything downstream, including its effect on your market. Hold that doubt in mind through every bullish claim below.

The demand backdrop, before you add artificial engineers

Even without physical AI, the critical-minerals demand story is intense. The International Energy Agency's Critical Minerals Outlook projects that demand for minerals used in clean-energy technologies rises sharply through 2040, with the pace depending heavily on policy scenario; rare-earth demand for magnets roughly doubles in its stated-policies trajectory and more under accelerated decarbonization, driven by electric vehicles, direct-drive wind, and grid electrification. Treat any single multiplier you see quoted as scenario-dependent rather than a forecast.

AI is now a second load on the same system. Data-center build-out is lifting demand for copper, aluminum, and the rare earths embedded in power electronics and cooling, with some analysts estimating data infrastructure could account for a low-single-digit share of rare-earth demand by 2030. (Carbon Credits (opens in a new tab)) These trends meet a supply base that has barely moved: China accounts for roughly 60 percent of mined rare-earth output and close to 90 percent of separation and refining into oxides and downstream products, a processing chokehold that is the real source of leverage. Europe has trimmed direct exposure but remains structurally dependent, with China still supplying the majority of German rare-earth imports as of the most recent trade data.

Into that picture, an artificial engineer promises to make hardware development faster and more ambitious. The first-order consequence is more, not less: more designs, deployed and retired faster, which means more demand for advanced materials in the medium term.

More hardware, more magnets, at least at first

Consider where rare earths live: high-performance NdFeB motors in EVs and robots, direct-drive wind turbines, precision actuators and sensors, defense and aerospace systems, and increasingly data-center power electronics. These are exactly the systems that benefit most from optimization across coupled physics: electromagnetics, thermal, fluids, and structures at once. A model that can search design spaces no human team could enumerate will find motor topologies and turbine layouts that hit tighter power-density and efficiency targets.

This is where a well-known economic pattern earns a name. The rebound effect, or Jevons paradox, holds that when a resource becomes more productive to use, total consumption often rises rather than falls, because the efficiency gain unlocks new applications faster than it saves material on existing ones. Applied here, better and cheaper engineering plausibly lifts total rare-earth use even as each individual motor uses slightly less per unit of performance. The pattern is not a law, however. It dominates when demand is elastic and applications are still expanding, which describes EVs, robotics, and grid hardware today; it weakens once a market saturates or once a credible substitute caps the willingness to pay for the constrained input. For the next five to ten years the expansionary case looks stronger.

The same tool that drives that demand also hands original-equipment manufacturers a hedge. If you encode material cost, single-country supply exposure, and recyclability directly into the optimization objective, an artificial engineer can search explicitly for designs that minimize the scariest inputs, the heavy rare earths sourced from one country, without giving up too much performance. Over a decade-plus horizon that points toward more aggressive thrifting of dysprosium and terbium, earlier qualification of reduced-rare-earth and rare-earth-free architectures, and products engineered to come apart cleanly at end of life. The hedge and the demand surge are the same technology pointed at different objectives.

The mine as a testbed, with a caveat

The clearest current deployments are not in Silicon Valley but in heavy industry, and disproportionately in China. Huawei's Pangu industrial models are being rolled into mining, power, and other physical sectors. The company describes its Pangu Mine Model, developed with Shandong Energy Group, as the first commercial AI foundation model for the sector, and points to Inner Mongolia's Yimin pit, where it says a fleet of 100 autonomous, cabless, all-electric haul trucks runs on a 5G-Advanced network with vehicle-to-cloud coordination. (Huawei (opens in a new tab)) Those descriptions come from Huawei's own materials and should be read as the vendor's account rather than independent verification; the underlying capability is real, the superlatives are marketing.

One caveat matters for this audience specifically. Yimin is a coal operation, and most of the mature Pangu deployments are in coal and general heavy industry. Extrapolating from autonomous haulage and mine planning to rare-earth separation chemistry is a real leap: hydrometallurgical separation of adjacent lanthanides is a harder, more constrained problem than moving rock efficiently. The direction of travel is credible; the timeline for the hardest rare-earth flowsheets is not the same as for a haul road.

With that qualifier, the operational logic holds. Outside China, AI is already improving drill targeting in exploration and tuning grinding, flotation, and leach conditions in concentrators and refineries to lift recovery and cut energy per tonne. An artificial-engineer-class system would stitch these point solutions into one model of an operation's physics and economics, from resource block to tailings. The competitive consequence is cost-curve dispersion. Operators that can supply clean data and are willing to let a model explore their whole flowsheet should move down the cost curve through better recoveries and lower energy intensity; technically rigid, data-poor operations risk being stranded faster than in previous cycles.

The race, and the paradox at the center of it

Prometheus is not alone. P-1 AI, led by former Airbus CTO Paul Eremenko, describes its goal as engineering artificial general intelligence and is starting with data-center thermal systems. (arXiv (opens in a new tab)) The naming is not incidental: each of these efforts is an attempt to point the AGI ambition at the physical world rather than the screen. Siemens is building an Industrial Foundation Model meant to speak the language of engineering and manufacturing, trained on CAD, simulation, process, and sensor data rather than text. Add Huawei's Pangu and a clear contest emerges between US-led, European, and Chinese efforts to build the first genuinely general engineering models, systems that could in principle optimize a lithium brine, a separation plant, and a gigafactory with shared foundations.

That contest contains the most consequential idea in this whole story, and it is a paradox. China's dominance of rare-earth processing is not only a supply advantage; it is a data advantage. The world's richest corpus of real operating data on separation, metallization, and magnet-making sits inside Chinese plants, which is exactly the proprietary, expensive-to-generate training data that physical AI most needs. The same chokehold that gives Beijing market leverage also gives it the best testbed for an industrial engineering model in this sector. Yet the very tools that could compound that lead are now being built by Siemens, P-1 AI, Prometheus, and others, which gives non-Chinese refiners a path to close part of the gap, if they move quickly and are willing to instrument and share their operations. Whether China's advantage grows or erodes turns on who builds the usable model first, and on whether the West can manufacture the experimental data it does not yet own.

What it means

Strip away the funding-announcement noise and three durable pressures remain. Demand for rare-earth-intensive hardware should run hotter and more volatile through at least the mid-2030s, because the same tools that design better EVs, turbines, and robots make it easier to design more of them, even as per-unit intensity edges down. Cost-curve dispersion should widen, rewarding operators who treat data infrastructure as a strategic asset and squeezing those who do not. And substitution and recycling should move from slideware to practice, because materials-aware, design-for-disassembly optimization is a natural task for an artificial engineer, which over a decade makes secondary supply and thrifting genuinely competitive in markets with stringent policy.

Return, at the end, to the two clocks. The near-term clock favors the incumbents: more demand, higher prices, a processing chokehold that no model dissolves this decade, and a physical-AI data bottleneck that may slow the whole project. The long-term clock does not. The single variable an artificial general engineer is built to compress, the time from dream to manufacturing at rate, is the same variable that protects every twenty-to-forty-year mine-to-magnet asset on the strength of an assumed multi-decade replacement timeline. You do not need a finished substitute to feel this; a credibly capitalized, credibly staffed attempt is enough to widen the discount a careful underwriter applies to the back half of those cash flows. The artificial engineer is not an epilogue to the rare-earth story. It is the variable that decides how long the current chapter lasts.

There is a symmetry worth holding here. The pursuit of AGI already drives rare-earth demand through the magnets and power electronics of the data centers it runs on; the pursuit of AGE may eventually relieve that same demand by engineering the elements out. The two largest bets of the decade, one on general intelligence and one on general engineering, converge on the same short list of elements, which means rare earths sit at the seam of both. The visible action today is funding rounds and autonomous-truck videos. The real change will show up where it always does in this business: in feasibility studies, flowsheets, procurement spreadsheets, and eventually the shape of the global cost curve for the elements that quietly hold the energy and AI transitions together.

Register today: REEx Marketplace™ (opens in a new tab)