Hello from Lisboa,
I hope you had an interesting and productive week.
Claude, when prompted, shared the following on the systems thinking lessons one can take from mountains; “The treeline as a tipping point made visible - Climb a mountain and you watch the forest thin, struggle, and then stop - the treeline. Above it, conditions are beyond what trees can tolerate. Below it, they thrive. The treeline is one of the most visible tipping points in nature: a threshold where a small change in conditions produces a qualitative shift in what the system can support. What makes it useful as a systems lens is that the treeline is moving. In most mountain ranges, it’s climbing as temperatures rise — but not smoothly. Trees advance in pulses, cluster around microclimates, and then consolidate. The threshold isn’t a clean line; it’s a zone of uncertainty and volatility where the system is figuring out its new equilibrium. Every organization in transition lives at the treeline — past the old stable state, not yet established in the new one, highly sensitive to local conditions.”
Take Note…keep an eye on the tipping points…they are dynamic…
1 Getting Visual
2 If You Read One Thing Today - Make Sure it is This
3 Consequential Thinking about Consequential Matters
4 Big Ideas
5 Big thinking
6 Consistency
1 Getting Visual
The Long (Long) View - The Power of GPTs - “2% - how an endless series of remarkable technologies averages out to steady economic growth - Figure 1 displays my favorite graph in economics: real GDP per person in the United States for the past 150 years. On a log scale, the time series is roughly a straight line. Indeed, Moore’s Law and US GDP per capita growth are in some sense duals, illustrating the “rule of 70” that the product of the growth rate and doubling time under constant exponential growth equals 70: transistor density doubles every two years and grows at 35 percent per year while per capita US GDP grows at 2 percent per year and doubles every 35 years. Consider the astounding innovations that underlie the graph. In the 1870s, Thomas Edison’s experiments with electric lighting were just getting underway. Fifty years later, electrification had transformed the economy, both in factories and in city life. Throughout the 150 years, innovations such as the internal combustion engine, airplanes, vacuum tubes, antibiotics, transistors, semiconductors, personal computers, and the internet profoundly changed living standards. Many of these innovations are what economic historians call “general purpose technologies,” whose transformative effects extend throughout the economy. Many also automated some of the tasks involved in creating new ideas—say, through improvements in scientific tools and equipment—raising the productivity of research and idea generation. Yet apparently none of these innovations changed the long-run growth rate of the US economy. How can we understand this disconnect? One natural hypothesis is that within any technology field, ideas get harder to find (Bloom, Jones, Van Reenen, and Webb 2020). The steam engine runs out of steam. Without the discovery of the next general purpose technology, one might expect economic growth to slow down. From this perspective, each of these new general purpose technologies did indeed raise the growth rate of the economy in a counterfactual sense: but for their invention, the counterfactual is that growth would have slowed. The continued emergence of amazing new technology classes is what made sustained growth at 2 percent per year possible. Perhaps artificial intelligence is just the latest general purpose technology that lets 2 percent growth continue for several more decades. Notice that even in this scenario, AI potentially has large transformative effects, because the counterfactual is one in which growth would otherwise have slowed.” - Charles I. Jones
The Long (Long) View - Tech-driven disruption - “However, valuations and debt levels are high because the bull case is also strong. Capex is driven by ROI and enormous demand – so much so that tech/AI adoption is putting pressure on labor markets. Demand for AI continues to rise as the models continue to offer more intelligence per dollar. Bain reported that the cost of tokens fell by 50% from Dec 2024 to Dec 2025 while token consumption grew by 4.5x. The demand for tokens is so extreme that electricity itself is often the constraint on installed capacity. This new AI trend extends an old story. Tech has driven decades of rising productivity, resulting in rising profits, supply booms, and deflationary pressure. New and existing markets collide and adjust, and margins find their new equilibria. Tech-driven disruption is probably the single most important economic story of the last 30 years. The average revenue per employee across public companies continues to march higher as rapid innovation suggests this trend is not likely to slow down. Corporate revenues and real output per worker have been growing fast for several decades. Even better, rising revenue and profits capture only part of the picture because much of the gains go directly to consumers through lower prices and new services that would have previously been unavailable. New markets for software-powered services can grow insanely fast. Cursor reached $4B in annualized revenue in June 2026, only three years after launching. Not only are the new technologies being built more quickly, but buyers are increasingly nimble with well-informed and capable CTOs who are under pressure to adapt and adopt. This speed means that disruptions happen fast, bringing growth but also competition. Startups are able to disrupt large legacy businesses faster than ever before, but these startups are also subject to the same risks themselves.”
Broom Ventures
The Long (Short) View - US Federal Debt Dynamics - “From World War Two until the late 2010s, U.S. debt dynamics were helped by the fact that the economy in general grew faster than federal debt, which helped keep borrowing contained relative to the size of the nation’s output of goods and services. That relationship began to shift due to the spending used to counter the 2007-2009 financial crisis and the COVID-19 pandemic roughly a decade later. The sweeping tax cuts pushed by President Donald Trump and passed by the Republican-controlled Congress in his first and second terms in the White House exacerbated the deficits, adding to the total debt pile. One typical use of a government’s financial power is to support the population and economy in the event of a crisis. Deficits expand during recessions as unemployment payments and other “stabilizers” rise. The cash funneled to households helps boost demand and shorten the downturn. During the COVID-19 pandemic, in particular, payments to households hit historic levels, and the economy rebounded quickly and defied fears of a lingering collapse. But deficits, the gap between government spending and tax revenue, have remained near such recession levels even as the economy grew. Part of that is due to Trump’s tax cuts, but it is also tied to the now hard-wired growth in spending on an aging population, and the recent reversal of tariffs that forced the administration to begin sending checks back to importers. Though the U.S. crossed the headline-making $40 trillion debt milestone last week, arguably it’s the higher year-to-year annual deficit, now close to 6% of gross domestic product, that presents the bigger threat to debt sustainability. Economists generally think of the 3% level as manageable.” - Reuters
Spotlight - US Federal Finances - Sub-optimal Sub-prime…
The Long (Short or Long) View - Fed Magic - Fiscal & Monetary Blend - Where to Next?
The Long (Short) View - Signal: A breakdown in global order - “2018 was likely the cheapest year for organized protection in recorded history: Pax Romana and Pax Britannica can’t hold a candle to the supposed “End of History” in the late 2010s. Today, even as regional conflicts have pushed spending higher, allocation to global defense is still well below the peacetime norm, let alone “wartime footing” by historical standards.” - Citrini Research
The Long (Short) View - Growing asymmetry of warfare - “Both Ukraine and Iran drive home two points. First is the growing asymmetry between the expensive traditional military assets and the low-cost but highly capable world of autonomous drones, loitering munitions, and guerrilla tactics – weapons that are far more easily accumulated and deployed by lesser powers. Second, the ability to sustain active warfare requires a reinvigoration of the supply chain from peacetime power projection to actual industrial production. Empty Clip - “Systemic constraints in the munitions industrial base, including limited production capacity, fragile supply chains, long-lead dependencies, and related production bottlenecks, may impair the ability of the United States to produce, sustain, and expand the availability of munitions, missiles, and equipment required for the national defense.” - Presidential Determination and Delegation of Authority Under Section 708 of the Defense Production Act, June 11, 2026. In just five months, the US military has fired more missiles and interceptors than it has in any year since Desert Storm. And unlike bombarding and occupying ill-equipped adversaries in the GWOT, this is not a unilateral (and inherently voluntary) offensive. Rather, the massive expenditure of costly interceptors is a response to advanced Russian and Iranian militaries who have developed both abundant autonomous drones and high-powered ballistic missiles that threaten nearly every military asset or allied infrastructure throughout the region. Military planners have been ringing alarm bells about inventory drawdowns for some time, and (for national security reasons) no one knows exactly what the true numbers are. Prior estimates across a wider range of munitions show that expenditures (through five months) have dramatically exceeded annual production capacity and current delivery cadence. The hardest numbers come from the FY2027 DoW budget proposal. The request shows a nearly 5x increase in dollars allocated to missiles across the spectrum. Interceptor depletion is at the top of mind, not just for the US but for anyone who relies on these systems for their own protection. The most recent estimates from CSIS show that Patriot and THAAD stockpiles have been drawn down by ~66% and ~44%, respectively, adding that there are “no good alternatives to Patriot and THAAD for ballistic missile defense.” This is urgent. The assumption of comprehensive missile interception capability has been central to both the US and its allies. It has allowed US bases and warships to operate comfortably in the Gulf, while also providing the cover for military, civilian, and energy infrastructure to allies in the Middle East (and well beyond). America’s umbrella of protection is leaking. The response demands a multipronged approach – one that addresses both the practical reality of today and technical shifts of tomorrow. Stockpiles of existing weapons platforms will be replenished. But DoW planners face manufacturing lead times measured in years and an asymmetric landscape that demands modern platforms — which means the shift in procurement has to come with long-term demand signals attached.” - Citrini Research
Spotlight - Department of War running out of Defence? - “A Center for Strategic and International Studies estimate said that by July, the United States had used about 65% of its Patriot interceptors and reduced its THAAD (Terminal High-Altitude Area Defense) interceptor stockpile by at least 38%.” - Reuters
2 If You Read One Thing Today - Make Sure it is This
Steve Hsu asked GPT to search his online writing, academic papers, podcast interviews, and media coverage, and to synthesize that material into the essay that follows (and then several other models to review and revise). The model completed the project in about an hour of sustained work across multiple turns. A human historian or biographer would have required much longer to complete the task. As he puts it; “The project is at once a test of deep-research capabilities and an exercise in self-indulgent narcissism.”
Hsu is an interesting person with a broad range of interests and real brainpower to explore them as such it also holds plenty of interesting contradictions, questions and perspectives - go explore it here in full:
Some Takeaways
“A life can be misread by its nouns.
Theoretical physicist. Silicon Valley founder. Computational genomicist. University research executive. Public intellectual. Builder of artificial-intelligence systems. Documentary filmmaker. Set side by side, the titles suggest restless polymathy, a man moving from one absorbing subject to another.
They miss the verb that binds them: to make.
Again and again, Stephen Hsu has been drawn to ideas poised between theory and science fiction—not fantasies, but possibilities waiting upon some missing threshold. Has the cost of measurement fallen far enough? Has the dataset grown large enough? Has the necessary mathematics already been invented in another field? Can an institution be built around the answer? Hsu reconstructs the problem from first principles, searches for the hidden constraint, and crosses whatever disciplinary boundary stands between the idea and its realization.
What unifies the career is a way of moving through the world. In physics, he asks what can be known when quantum mechanics, gravity, and cosmology press against one another. In genomics, he asks how much of a human future lies encrypted in DNA, and how large a dataset is needed to read it. In entrepreneurship, he turns a technical possibility into a working system. In university leadership, he confronts the problem of organizing talent and capital at scale. In artificial intelligence, the object of inquiry begins to answer back, becoming a collaborator in discovery.”
“Hsu possesses an unusual feeling for the ripening hour of a problem. He neither pursues difficulty merely because struggle is noble nor waits until fashion has made an idea safe.
He watches for the moment when an old impossibility becomes newly soluble—when cheaper sequencing, larger biobanks, greater compute, better algorithms, or more powerful models bend the curve. Before committing years to genomics, for example, he estimated whether realistic sample sizes should suffice. When the data crossed the predicted threshold, his group moved quickly and produced accurate predictors. This is ambition governed by calculation: audacity with a theory of when to act.”
“Recalling his father in a From the New World interview, Hsu gives the decisive object an almost talismanic glow:
“He had what then, in the pre-internet era, was—for a precocious kid like me—the magic secret: a library card at the university library.” — Brian Chau interview, From the New World
Before the internet, a library card was not a convenience but a passage into worlds otherwise sealed off by age and geography. Hsu’s intellectual life began not merely with precocity, but with premature access to the archive of adult knowledge.”
“Hsu’s mature method is not Feynman’s pure intellectual individualism. He reconstructs from first principles where he can, borrows provisional knowledge where he must, and keeps track of which is which.”
“…identify a foundational question that respectable routines leave aside, bring the right minds into the room, and insist that the strange possibility deserves a hearing.”
“I guess the unifying theme is knowledge versus uncertainty: the attempt to capture the essential aspects of a messy system in a simplified mathematical model.”
This is close to a personal credo. The model must be simple enough to expose the decisive relation, but the scientist must remember that simplification has purchased clarity by discarding detail. Hsu’s confidence comes from finding structure; his best skepticism is directed at the boundary where structure may have been mistaken for the world.”
“…be a scientist: see the world as it really is.”
“…a powerful description of actual scientific reasoning. Pure deduction cannot move through empirical sciences because many premises remain contingent, approximate, or incompletely measured. Pure empiricism cannot distinguish a meaningful anomaly from noise because it lacks a structural model. Hsu’s approach builds a hierarchy of confidence: derive what can be derived, borrow what must temporarily be borrowed, remember which is which, and revise without embarrassment.
The same cognitive style drives his disciplinary mobility. He does not approach a new field by slowly absorbing all its conventions. He looks for its governing variables, scaling relations, information bottlenecks, and unexamined assumptions. This gives him an advantage over insiders whose knowledge is locally deeper but structurally less explicit. It also makes his criticisms sound abrasive. What appears to an insider as accumulated craft knowledge may appear to Hsu as an unjustified prior; what appears to Hsu as a simple information-theoretic question may depend on biological complexities he has compressed away.
His best work occurs when the abstraction preserves what is decisive and discards what is not.”
“Or does the bubble reputation distract you? Keep before your eyes the swift onset of oblivion, and the abysses of eternity before us and behind; mark how hollow are the echoes of applause, how fickle and undiscerning the judgments of professed admirers, and how puny the arena of human fame. For the entire earth is but a point, and the place of our own habitation but a minute corner in it; and how many are therein who will praise you, and what sort of men are they?”
— Marcus Aurelius, quoted in “Happiness”
“Pessimism of the Intellect means, simply, be a scientist: see the world as it really is, not as you might like it to be. Try to identify and overcome hidden biases or prior assumptions. Always ask yourself: What assumption am I making? What if it is incorrect? How do I know what I know? In many cases, the correct answer is: I don’t know. Never be afraid to admit you don’t know.”
— “Pessimism of the Intellect, Optimism of the Will”
“Sisu is a Finnish term loosely translated into English as strength of will, determination, perseverance, and acting rationally in the face of adversity. However, the word is widely considered to lack a proper translation into any other language. Sisu contains a long-term element; it is not momentary courage, but the ability to sustain an action against the odds. Deciding on a course of action and then sticking to that decision against repeated failures is sisu. It is similar to equanimity, except the forbearance of sisu has a grimmer quality of stress management than the latter.”
— “Sisu”
“The distinction between dramatic bravery and sustained action is central to his temperament. A difficult project is rarely conquered in one heroic instant; it is carried through long periods when the reward is distant, the social signal is adverse, and failure repeats itself.”
“Academic culture privileges analytic depth and publication. Startups expose execution, social judgment, risk tolerance, and speed. Later, as an administrator, Hsu would say that startup experience teaches difficult decision-making under pressure. Across these roles he encountered forms of ability that psychometric discussion often leaves out: the capacity to coordinate other minds and reshape an institution.”
“Research advances often pass through the following phases of reaction from the scientific community: It’s wrong. It’s trivial. I did it first.”
“Only one in a thousand people in our society have the privilege to engage full time in discovery—in curiosity-driven research.”
“Is every genius level STEM guy suited for leadership? No, obviously not. But every leader going forward should be genius level STEM.”
“His critique of elite systems is not simply that the wrong individuals possess prestige. It is that institutions increasingly suppress accurate feedback. Credentialism substitutes for ability, narrative for measurement, procedural consensus for responsibility. His startup experience taught him that reality eventually punishes such substitutions: companies fail, systems break, predictions fail to replicate. Politics and universities can defer correction longer.”
“He often portrays China as more technologically capable and strategically serious than American discourse allows, while recognizing the constraints of its political system.
Summarizing a formulation he credits to the pseudonymous analyst Han Feizi, Hsu argued in early 2026: “China leapfrogged Western expectations so fast… that sort of short-circuited the Thucydides trap.”
Rivalry did not vanish. Washington may simply have recognized China’s military-industrial position only after the favorable window for a preventive confrontation had narrowed, producing retrenchment and “Fortress Americas” rather than a classical rising-power war. Whether the forecast proves correct, its form is characteristic of the man: estimate relative capability, identify a phase transition, and revise strategic expectations before public narratives catch up. The underlying issue is not cultural admiration but state capacity—which civilization can identify talent, build infrastructure, pursue long-term goals, and absorb new technology?
This framework produces sharp insights and blind spots alike. It corrects complacency about American primacy and highlights the material bases of scientific power. But a civilization cannot be evaluated only as a research lab or startup. Freedom, loyalty, solidarity, consent, and the distribution of dignity are not noise variables. Hsu’s strongest public analysis treats pluralism as part of the optimization problem rather than as an obstacle external to it.
His Stoicism moderates the elite-centered view in an important way. If fame is a bubble and public applause unreliable, membership in a prestigious hierarchy cannot be the ultimate measure of a person. His emphasis on ability describes differences in capability; it need not imply differences in human worth. Much of the ethical controversy around Hsu arises precisely because that distinction is difficult to maintain socially once predictive technologies and competitive institutions assign consequences to measured traits.”
“Frontier capability is not produced by luminous ideas alone. It rests on evaluation, data hygiene, repeated failure analysis, and people willing to perform unglamorous work with unusual conscientiousness. The AI laboratory joins the startup and the athletic pool as another place where talent becomes real only through sustained practice.”
“The bottleneck is migrating from the ability to write a training loop toward the ability to choose experiments, diagnose failures, evaluate novelty, secure compute, and improve the research process itself. This is the distinction Hsu learned as a founder: execution is never exhausted by possession of the core idea.”
“It’s hard to put a util value on some things that are in the foreseeable future, like machine intelligence and genetic engineering.
These are not ordinary increments whose benefits fit comfortably into a cost-benefit table. They may change the kinds of agents who make the table, the scale of values those agents pursue, and the identity of the civilization doing the choosing.”
“His projects succeeded not because every forecast was correct but because he repeatedly chose domains in which error met data, engineering, or the market soon enough to be corrected. Coherence was built through contingent choices, not granted in advance…”
“The career turns on three virtues. See without illusion. Dare without guarantee. Recognize the hour. Hsu’s deepest talent may be the last: to feel when an idea is no longer merely premature, when the future has drawn close enough to be grasped. His deepest unresolved question is what can be carried through the gate.”
3 Consequential Thinking about Consequential Matters
The CrossCurrents Substack explores how; “Humanoid robotics is often framed as a software race, but the binding constraint may be physical. (…) The relevant bottleneck for humanoids is not “rare earths” in general, but the NdPr-to-magnet supply chain. China’s dominance in midstream processing and magnet manufacturing gives it leverage over the pace, cost curve, and industrial geography of embodied AI.” It’s Consequential Thinking about Consequential Matters - Go explore it full here (Well worth your time - plenty of insights and visuals)…
Some Takeaways
The Binding Constraint is Physical
Humanoid robotics is almost exclusively framed as a software race. The consensus view is that once we solve the brain the robots will appear from low cost producers.
This view is wrong.
While the West obsesses over the cognitive bottleneck, the binding constraint for deploying embodied AI at scale is physical.
Building millions of humanoids requires a vast ecosystem of materials and capabilities, but the single most critical and constrained variable is torque. High-performance actuators are the only way to achieve fluid movement, and you cannot build them at scale without Neodymium-Iron-Boron (NdFeB) permanent magnets.
NdFeB magnets are the efficiency chip of the physical world.
Without them, robots are too heavy to walk and too inefficient to run. And right now, the supply chain that turns raw earth into these high-performance magnets is a veritable monopoly.
China controls >90% of global magnet manufacturing and the complex chemical midstream that makes it possible.
For China, this is a macroeconomic escape hatch. Facing a debt crisis and a shrinking workforce, Beijing is betting that embodied AI can be its “Second Lever”, a way to productize labor itself and export it to the world.”
“Escaping the trap without a hard deleveraging requires two things at once:
A productivity shock large enough to stabilize the debt math, keep nominal GDP growing faster than the effective interest burden. Productivity shocks are positive supply shocks: they lower marginal costs, expand potential output, and increase real income (which is what China needs urgently).
A distribution mechanism that routes a meaningful share of that surplus to households, otherwise you just get margin expansion, retained earnings, and a more state-heavy economy with the same demand constraint.
In my previous piece, I argued China needs an economic “hail mary” in the form of a tech-driven productivity shock, something powerful enough to lift output and keep the system solvent without allowing a balance-sheet clearing. To generate that shock, China has two plausible levers:
Lever #1: frontier compute / sub-7nm chips capital-light, high-margin digital growth, but externally constrained by export controls, tooling chokepoints, and the Taiwan nexus.
Lever #2: embodied AI — robots as compensatory capacity and productivity infrastructure that can be scaled inside China’s existing industrial base.
China needs a replacement engine that can (1) lift trend growth, (2) generate tax base and cash flows, and (3) do it without triggering a balance-sheet clearing that produces mass unemployment. The strategic bet is that humanoid robotics, and the upstream actuator stack, can become the next EV/solar-style national champion complex, but with higher strategic leverage because what’s being productized is labor itself.
If China can dominate the humanoid supply chain, especially the actuator corridor where cost, yield, and reliability are set, it can build a globally traded, high value-add export category that throws off profits, payroll, and fiscal revenue. Those flows help keep nominal growth above the interest burden, which is the only non-painful way to shrink a debt overhang in relative terms.”
“My core claim is that the competition is shifting from the brain (semiconductors, frontier models, training infrastructure) to the body (actuation). If the U.S. does not own that corridor, it will import the bottleneck module that gates volume. And if you import the bottleneck, your learning curve compounds inside the exporter’s cluster, not domestically, which is exactly how China locked in advantage in solar, batteries, and increasingly EV subsystems.
In that framing, NdFeB (core input in actuators) is the strategic material that sets the ceiling on embodied AI throughput. That’s why the “body” matters. The next decade will be decided by who can manufacture the platform the model inhabits.
If the macro strategy provides the motivation, the machine itself provides the means.
But why the humanoid form factor specifically? Why not simply deploy more traditional industrial arms?
Humanoids: Why They Matter
Humanoid robots are a bet that general-purpose labor can be productized.
The reason the humanoid form factor matters is because the world is already built for humans. Doors, stairs, shelves, tools, carts, and factory layouts all assume a human body plan. If a machine can operate inside that environment without requiring the environment to be redesigned, it becomes deployable across many jobs.
That is what makes humanoids economically different from traditional industrial robots. Most industrial robots are task-specific and require a controlled setup. Humanoids aim to reduce that integration burden.”
“…it helps to ground the reader in what a humanoid actually is. This is, of course, a very simplified map of an extremely complex machine (if you want to dive deeper, take a look atHumanity’s Last Machine).
At a high level you can think of five buckets:
Structure is the skeleton (frames, housings, castings) that carries load and survives impacts.
Energy is metabolism (the battery pack and power electronics) that sets runtime and peak power.
Sensing is eyes and ears (cameras and other vision sensors) that let the robot perceive the world.
Signaling is the nervous system (wiring, connectors, and force/torque sensing) that moves information and feedback through the machine.
Motion is the equivalent of joints and muscles, where electrical power becomes controlled movement at every joint.
It’s easy to over-index on the software bottlenecks like quality data, world models, and the long tail of edge cases. But they are ultimately cognitive bottlenecks, they improve with iteration, better data engines (simulation + teleop + self-supervision), and scaling compute.
Physical constraints behave differently. Hardware scaling is gated by throughput, yield, heat, wear, and supply chains, constraints that are slow to relax because they require new factories, new processes, and years of ramp. That’s why the most important question for humanoids is if the body can be built reliably, millions of times, at an acceptable cost?”
What Is An Actuator ?
Think of an actuator as the robot’s muscle at each joint. As Humanity’s Last Machine puts it, “a humanoid robot can be thought of as a collection of actuators operating in coordination”.
When a humanoid bends an elbow, lifts a leg, or rotates its torso, something has to (1) create force, (2) control it precisely, and (3) survive repeated stress. That whole “muscle package” is the actuator.”
Electric vs hydraulic vs pneumatic
Actuators can be powered electrically, hydraulically, or pneumatically. Hydraulics offer very high power density and are excellent for brute-force motion; pneumatics can be cheap and clean but struggle with precision and controllability. But for general-purpose humanoids the industry trend is overwhelmingly electric actuation, because it’s compact, programmable, repeatable, and easier to integrate with batteries and onboard electronics.”
“…the motor is the most crucial and typically the most expensive single element to get right, because it sets the ceiling on torque density, efficiency, heat, and size.
And that’s exactly where NdFeB permanent magnets enter.
They’re what enable high torque in a small package at acceptable efficiency. If the magnet is weaker, you can still build a motor, but you pay for it in size, weight, heat, and battery drain, which compounds across dozens of joints.”
Actuation is the cost base
A humanoid is effectively a battery strapped to a distributed network of actuators. The actuator-heavy limb and joint modules add up to ~$52.3k, or ~95% of the total. The exact numbers will vary by design, but the message is clearly that humanoids scale as a repeated stack of motors + reducers/screws + bearings + sensors across the body (with NdFeB magnets alone representing ~15–20% of total build cost).
And because the motor is the most complex and expensive part of each module, and the motor’s performance depends disproportionately on the magnet, the humanoid scaling question quickly collapses into a materials question.”
Where NdFeB fits
At the heart of every actuator lies a NdFeB permanent magnet as discussed previously. These magnets represent a fundamental step-change in energy density, achieving roughly 400 kJ/m³, the highest of any commercially viable magnetic material.
Rare-earth magnets sit in a different performance regime than conventional materials, enabling high torque in compact packages, exactly what humanoid joints and precision actuators require.
If a humanoid robot were built with conventional magnets, its motors would need to be 4x larger to generate the same torque. The robot would become so heavy it could barely lift its own limbs, let alone a payload.
The Value-Volume asymmetry
This also explains why the rare earths debate is usually framed too broadly. We must look at the disparity between volume and economic value. Rare earths are used in a wide range of industrial applications, from catalysts and glass polishing to ceramics, but their strategic importance is not distributed equally.
By volume, permanent magnets represent roughly 35% of total rare earth consumption. However, when viewed through the lens of economic value, they account for a staggering 91% of the market. This massive concentration of value reflects the fact that high-performance magnets (NdFeB) require the highest-purity inputs, specifically Neodymium, Praseodymium, and Dysprosium, which command the highest prices and face the greatest supply constraints.
Scaling Math
All of that explains why magnets matter.
The next step is why they become the binding constraint under scale. The cleanest way to see it is a stress test: if humanoids ever reach mass deployment, which input becomes the outlier? The chart below ranks bottlenecks by showing how many multiples of today’s annual production would be required under an optimistic deployment scenario.
In this scenario, NdFeB magnet output would need to expand by roughly 186×, versus 14× for lithium, 13× for graphite, and single-digit multiples for cobalt, nickel, and copper.
This is the definition of a binding constraint. While the battery supply chain needs to scale by an order of magnitude, the magnet supply chain needs to scale by two orders of magnitude. NdFeB is the outlier input, and therefore, the pace of humanoid scaling will be gated by the pace of magnet production.”
The Demand Landscape: A Zero-Sum Game
As mentioned previously, NdFeB is the only commercially scalable option for humanoids. That immediately makes the demand landscape a zero-sum game: every incremental kilogram of NdFeB pulled into robotics is a kilogram that can’t go into EV drivetrains, wind generators, or industrial motors unless the supply chain expands—which it doesn’t, quickly.”
The Policy Blindspot
The scariest part of this math is that Western policymakers aren’t even tracking it.
Two of the most authoritative frameworks, the Section 232 investigation report (The Effect of Imports of NdFeB Permanent Magnets on National Security) and the U.S. The Department of Energy’s 2022 supply-chain deep dive (Rare Earth Permanent Magnets), don’t even include robotics as a distinct demand category. Their demand tables allocate the future mostly across wind and EVs, while everything else gets buried in broad buckets like “industrial motors” or “other sintered magnets.”
Supply Chain
The NdFeB magnet supply chain is a multi-tiered industrial ecosystem that transforms mineral-rich ores into high-performance permanent magnets essential for the modern energy transition. The system is fundamentally divided into upstream extraction and separation, midstream metallization and alloying, and downstream magnet manufacturing and finishing.16 NdFeB magnets are comprised approximately of 30% rare earth elements (primarily Nd and Pr, with Dy and Tb used for high-temperature grades), 69% iron, and 1% boron.
Mining (Ore Extraction)
Rare-earth ores (like bastnäsite, monazite or ionic clays) containing Nd and other REEs are mined mostly in China. In 2024 China accounted for ~69% of global rare-earth oxide (REO) production, dwarfing all other countries. But mining is the least unequal part of the chain: the U.S. and Australia do produce meaningful ore, and the real asymmetry shows up later, during separation, metal/alloy, and magnet manufacturing.
As seen, other leading producers include the United States and Myanmar (Burma) at far smaller shares. China’s massive Bayan Obo iron-rare-earth mine (Inner Mongolia, by China Northern Rare Earth Group) is the world’s single largest source of REO. Outside China, major operations are fewer. The U.S. has one large rare-earth mine at Mountain Pass, California (operated by MP Materials), a bastnäsite deposit that produced roughly 45,000 t/year of REO in 2024, which is enough to produce ~ 13% of current world Nd/Pr oxide. Australia’s main rare-earth producer is Lynas Rare Earths, which operates the Mt Weld carbonatite deposit and produced ~10,000 tonnes in 2025, about 63% of which was NdPr oxide.”
“…mining is the one stage where the West still shows up, Mountain Pass and Mt Weld prove there’s sufficient ore outside China. But the choke point starts right after mining, in the separation stage, where mixed concentrate gets split into high-purity NdPr (and Dy/Tb) oxides. That’s where China’s control becomes overwhelming.
Separation (Chemical Processing)
After mining, rare-earth ore concentrates are chemically processed to separate individual rare-earth elements (or groups) from each other. In practice this uses solvent-extraction: the ore concentrate is leached (often with acid), then passed through stages of solvent exchange where lighter REEs (La, Ce, Nd, Pr) are separated from heavier ones (Dy, Tb, etc). This is a complex, multistep chemical process (hundreds of mixer-settler operations) that requires large amounts of reagents and water. Uranium and thorium impurities in the ore must also be removed, making the chemistry more challenging.
China controls ~90% of global rare-earth separation capacity, with large state and private separation plants, linked to major mining districts like Bayan Obo handling most of the world’s concentrate. Malaysia is the most important non-China separation node today via Lynas Rare Earths’s LAMP facility in Kuantan, which processes Mt Weld concentrate and has expanded capabilities over time. India has IREL Limited, and Estonia represents the newer “downstream comeback” in Europe. Neo Performance Materials opened a rare-earth permanent magnet manufacturing facility in Narva in 2025 to reduce reliance on Chinese magnet imports, supported by EU funding. Beyond those four anchors, the rest of the non-China ecosystem is still relatively thin and fragmented.
Refining (Oxide-to-Metal Conversion)
Separation yields rare-earth oxides (REO). The refining stage converts these oxides into either pure RE metals or alloy powders for magnets. There are two main routes: reduction (chemical or electrochemical) of oxides to metal, and alloy melting. For NdFeB magnets, Nd and Pr oxides are typically reduced with calcium (or aluminum, etc.) in high-temperature furnaces to make Nd–Pr alloy metals, which are then mixed with iron and boron and melted into NdFeB alloy ingots. (Dy and Tb are often added as ferroalloys if needed for high-temperature performance.)
Nearly all RE metal refining is still done in Asia. China produces again ~ 90% of rare-earth metal and alloy output. A few non-Chinese facilities exist: Vietnam Rare Earth JSC (Vietnam) runs a plant to make NdFeB alloy, and Less Common Metals (UK) and Silmet (Estonia, now part of Canada’s Neo Performance Materials) have limited separation-to-metal lines. Newer efforts to rebuild refining include: U.S. company MP Materials plans to add oxide-to-metal refining at Mountain Pass, and Energy Fuels is studying full refining of its Utah monazite output. South Korea’s LS Cable/LS Eco Energy is investing in a rare-earth metal refinery in Vietnam to integrate imported oxides into NdFeB metal.
Magnet Manufacturing
Finally, refined Nd–Pr–Dy (and other metals) are alloyed with iron and boron to make NdFeB magnets. The alloy is cast and machined into ingots or pellets, then either sintered (pulverized into powder, pressed in a magnetic field, and heat-treated) or bonded/injection-molded into final magnet shapes.
Again, China overwhelmingly dominates, they produce ~92% of the global market. Outside China, Japan is the largest producer (~7% market share), with several long-established firms: Hitachi Metals, Shin-Etsu Chemical, and TDK, which make magnet alloys, powders, and finished magnets. Other notable producers include Vacuumschmelze (Germany/Apollo-owned) and its Finnish arm Neorem, which supply high-performance magnets to Europe, and Neo Performance Materials (Canada), which makes bonded magnet alloys. The UK’s Less Common Metals supplies NdFeB powders (mostly for bonded magnets).
In the United States, domestic NdFeB magnet production has been negligible; until recently the only U.S. plant was Urban Mining Company (Texas), which recycles scrap NdFeB into sintered magnets. New U.S. projects are changing this: in 2021 MP Materials announced a magnet factory in Fort Worth, TX (with General Motors) to make ~1,000 tonnes/year of NdFeB magnets from its Mountain Pass output. In 2023, Vacuumschmelze also agreed to produce magnets in the U.S. for GM.”
The Ends Diversify, the Middle Stays Chinese
At a high level, the NdFeB chain is bifurcated: China can lose share at the ends, but it still dominates the middle.
My base case is that China loses share at the “ends” of the chain but maintains (or even gains) share in the two middle links. Mining is the most globally contestable step because ore bodies exist outside China and incremental supply can be brought online without immediately recreating the downstream chemical stack; that pushes China’s mining share down over time. Magnet production is also relatively onshorable: plants are modular, capex is manageable, and governments can subsidize capacity close to demand in EVs, wind, and defense, so the U.S., Europe, and Asian allies can take some share here.
But the hard constraint is the middle. Separation and refining are the chemistry-heavy, scale-driven steps with the highest compliance burden and the steepest learning curve, and they remain dominated by China. As a result, even if the West wins share in magnet manufacturing, it still needs NdPr (and Dy/Tb) oxides/metals/alloys, which means the dependency often just migrates upstream. In practice, that dynamic can actually increase China’s leverage in separation/refining even as its share slips in magnets, because whoever controls the chemical conversion controls throughput, pricing, and who gets material when markets tighten.
If the chain were elastic, a price spike would trigger a visible ramp in NdPr/Dy/Tb output within a few quarters. Instead, history shows the opposite, when prices surge throughput barely budges.”
The Gravity Well: Clustering Logic
If NdFeB were abundant and globally fungible, robot manufacturing would locate where labor is cheapest, demand is closest, or incentives are highest. Under structural shortage, that logic flips. When a single input becomes quantity-constraining and clears through allocation, the dominant objective is supply assurance. The predictable result is industrial clustering, where the downstream manufacturing migrates toward the geography that controls the constrained input (NdFeB magnets).
Today, that geography is China. China’s dominance in the midstream means it can deliver magnet-grade material at scale, with shorter lead times and higher reliability than anywhere else. And because humanoids are actuator-dense and complex to manufacture, any disruption in magnets is a production stop. When production-stop risk is existential, OEMs rationally bias toward the ecosystem that maximizes continuity of supply, engineering support, and ramp execution. That is a major reason the robot industrial stack increasingly co-locates in China.”
Actuators are the flywheel
Actuators are the humanoid’s scaling unit: they’re the most complex, most expensive subsystem, and the one that ultimately determines whether robots can move reliably at scale. Inside that stack, NdFeB magnets are the most supply-constrained, high-value input. Put differently: if humanoids become a mass market, the industrial gravity well forms around actuators, and around the NdFeB corridor that feeds them.
The key mechanism is manufacturing throughput.
Once the software stack is good enough, the compounding advantage shifts to whoever can manufacture, qualify, and iterate the physical joint modules fastest. That’s why catch up everywhere is not a plan for the U.S. The only viable response is focus. The U.S. should pour disproportionate resources into the one vertical that governs the learning curve, the cost curve, and the reliability curve for humanoids, then force a manufacturing growth loop there through automation, subsidies, and scale.
Robotics has an exponential, self-reinforcing manufacturing curve, just as AI research can create a self-reinforcing intelligence curve. Once a platform is good enough to do real work, deployment feeds back into capability: more robots in the field generate more operational data, more production volume, and more process learning, which lowers cost and improves reliability, which expands the set of economically viable use cases, which pulls forward more deployment. Wright’s law then acts like a tailwind, if costs fall ~20% for each doubling of manufacturing volume, the winner will be the country that can double volume fastest and keep yield high while doing it.”
“China needs to keep moving up the value chain to escape its debt trap. The easiest way to achieve that is to engineer a persistent spread between the global price and the China price” of key inputs.
Inside China: Magnets are abundant and cheap (subsidized energy + lax environmental enforcement + VAT rebates).
Outside China: Magnets are scarce and expensive (due to export quotas, licensing friction, and no midstream).
This strategy turns Western capitalism against itself. Imagine you are a U.S. humanoid startup. Your software is world-class, but your Bill of Materials (BOM) is $5,000 higher than your Chinese competitor simply because you are buying magnets in Detroit instead of Dongguan.
Once you move the factory, the trap snaps shut. China captures the manufacturing value added, the supply chain ecosystem, and eventually the IP. The U.S. retains the “brand,” but the physical economy moves to China.”
4 Big Ideas
Anthropic are up to more innovative things, Big Ideas are being tested - I don’t normally share “marketing materials” but I think the ideas being explore and tested here are well presented and worth thinking about - Intro here via Ruxandra Teslo:
“Last week, when Anthropic announced the binder design news, I said the exciting part isn’t binder design, but the orchestration of models to produce a research assistant - like entity. This is quite cool and it seems Anthropic is continuing down this road of scientific automation. So what’s happening? They propose the Model Hardware Standard (MHS), which is a common standard that lets AI agents safely connect to and control physical scientific and industrial equipment. This would include things such as microscopes, liquid handlers, lasers and manufacturing machinery.
Today, integrating different pieces of hardware can take weeks or months because each device has its own interface and often requires custom software -- which is extremely annoying. MHS aims to reduce that setup to hours by giving devices a standardized software “driver” and a common way to communicate with AI agents. The blog post gives one example, which involves Claude learning to align a laser. Claude adjusted the laser, then observed the result through a camera, then repeated the process to understand the system and finally, it wrote a script that could perform the alignment automatically.
I think this line of scientific automation is actually much more exciting and higher yield potentially than all the stuff that aims more at making specific scientific discoveries (eg using AI to improve binder design). I think it’s underrated how much boring friction slows down biological research, and it’s exciting to potentially reduce that. Of course, it also seems like being along the path to full scientific automation (including of human scientists) -- though still very early (and potentially it will never happen).
I feel much more conflicted about that; I was just debating with some friends about this. In my mind, scientific discovery is an expression of something very core to the human spirit -- a sort of fundamental curiosity. And I’d be sad if that ceased to be an activity humans were involved in.”
Explore it here in full (Note - the source of this is not just drinking the Cool-Aid but selling it):
https://www.anthropic.com/news/model-hardware-standard-research-preview
Some Takeaways
“We’re opening a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical devices, to a first group of scientific research labs and advanced manufacturers. MHS enables AI agents to operate multiple lab and manufacturing instruments, such as microscopes, liquid handlers, and robotic arms, in parallel, and perform intricate tasks ranging from routine drug discovery experiments to laser calibration on a quantum computer. The development of MHS began as a collaboration between Anthropic and HHMI Janelia Research Campus.
It typically takes a lab or manufacturing facility weeks, if not months, to set up and integrate their hardware. Most devices don’t communicate with each other, instead requiring specialists to build bespoke integrations. MHS reduces this integration work to hours or minutes. And by incorporating AI into these tools, MHS also helps researchers and engineers more readily orchestrate autonomous, round-the-clock experiments and workflows, with agents able to reason through each step in an experiment, update parameters in real time, and, in some cases, recover from hardware errors without intervention.”
“Getting multiple devices in a lab or on a factory floor to communicate with one another can be challenging, even setting aside the added difficulty of integrating AI into the setup. Each device tends to have its own programming interface, and so far there has been no standardized way to integrate them. And once the devices are connected, there is no common way for them to share data with an AI agent, nor to let the agent operate them safely.
MHS addresses these challenges by introducing a standardized driver: software that translates between a computer’s operating system and a hardware device. The MHS driver uses a simple set of primitives—commands like “read” (for example, “get temperature”) or “write” (for example, “set temperature”)—that any hardware device can understand and act on. And it makes each device discoverable in a standard format, so that devices and agents can find each other and communicate across networks without needing a bespoke “translator” program in between.
The MHS driver also helps an AI agent understand how to use a device it has never seen before, giving it information about machine characteristics that may not be discernable from code alone (for example, the weight of a robot arm, which is important for knowing how to manipulate it safely). To date, much of this information has been stored in paper manuals, on a user’s computer, or as tacit knowledge. But the MHS driver contains tags that let the user write this information directly in natural language (users can either do this themselves, or by chatting to an agent that interviews them about their hardware setup). With the information from these tags, the MHS driver then automatically produces a reference file with information about a device’s general characteristics, such as what it can measure, what can be adjusted, and what safety limits will be enforced. This file gives the agent everything it needs to know to operate the device.
After the devices are connected and the agent knows how to use each one, the agent needs a way to control the hardware. For MHS, there are three such mechanisms: MCP, the command line interface, and code files (APIs). These work together to enable orchestration across multiple devices via a single line of code.
Once the agent can control the devices, it’s able to receive operating data from each one and supervise and direct the work at a high level. The agent can sequence steps across instruments, monitor results, and adjust parameters as conditions change in real time. When the agent needs to execute long-running tasks or operate devices faster than its online reasoning would allow, it can chain together driver commands from one or more devices in code files. This allows the devices to carry out operations themselves, without the agent needing to reason at every step.
As we’ve tested MHS, we’ve found that Claude interacts with experiments and hardware in an exploratory manner, much as a scientist would. For example, we observed Claude make an adjustment to a laser, observe the results through a camera to assess how its adjustment moved the laser beam, and repeat the process, seeking to understand the sequence of events. Claude then packaged what it learned into code files, writing a deterministic script that let it align the laser without having to reason at each step, so the whole process could run as a single command.”
5 Big thinking
The Good People at the FS Blog explores Feedback Loops in their Mental Models series here - it’s worth reading and pondering in full - do it here:
https://fs.blog/mental-model-feedback-loops/
Some Takeaways
“Feedback loops are created when reactions affect themselves and can be positive or negative.”
“Referring to the credit problems in 2008/2009, Vice Chairman of Berkshire Hathaway, Charlie Munger explained:
By the fourth quarter, the credit crisis, coupled with tumbling home and stock prices, had produced a paralyzing fear that engulfed the country. A free-fall in business activity ensued, accelerating at a pace that I have never before witnessed. The U.S. — and much of the world — became trapped in a vicious negative-feedback cycle. Fear led to business contraction, and that in turn led to even greater fear.”
“In The Education of a Speculator, Victor Niederhoffer says:
One of the common features that all life possesses is a mechanism for maintaining orderly conditions. This tendency is called homeostasis. In system theory, it is called negative feedback … A Common homeostatic behavior in humans is temperature regulation. If the temperature rises above the 98.6 F optimum for normal human activity, sensors on the skin detect it and signal the brain that a rise has occurred. The brain relays the information to the effectors that increase blood flow to the skin. This induces perspiration. The loss in heat, caused by evaporation, lowers the body temperature. When the body cools below a certain point, a comparable mechanism is set off, this time reducing blood flow and causing shivering. This activity generates heat through physical activity thus raising the body temperature.”
“In Universal Principles of Design, William Lidwell, Jill Butler, and Kritina Holden write:
Every action creates an equal and opposite reaction. When reactions loop back to affect themselves, a feedback loop is created. All real-world systems are composed of many such interacting feedback loops — animals, machines, businesses, and ecosystems, to name a few. There are two types of feedback loops: positive and negative. Positive feedback amplifies system output, resulting in growth or decline. Negative feedback dampers output, stabilizes the system around an equilibrium point.
Positive feedback loops are effective for creating change, but generally result in negative consequences if not moderated by negative feedback loops. For example, in response to head and neck injuries in football in the late 1950s, designers created plastic football helmets with internal padding to replace leather helmets. The helmets provided more protection, but induced players to take increasingly greater risks when tackling. More head and neck injuries occurred (after the introduction of plastic helmets) than before. By concentrating on the problem in isolation (e.g., not considering changes in player behavior designers inadvertently created a positive feedback loop in which players used their head and neck in increasingly risky ways. This resulted in more injuries which resulted in additional redesigns that made the helmet shells harder and more padded and so on.
Negative feedback loops are effective for resisting change. For example, the Segway Human Transporter uses negative feedback loops to maintain equilibrium. As a rider leans forward or backward, the Segway accelerates or decelerates to keep the system in equilibrium. To achieve this smoothly, the Segway makes hundreds of adjustments every second. Given the high adjustment rate, the oscillations around the point of equilibrium are so small as to not be detectable. However, if fewer adjustments were made per second, the oscillations would increase in size and the ride would become increasingly jerky.
A key lesson of feedback loops is that things are connected—changing one variable in a system will affect other variables in that system and other systems. This is important because it means that designers must not only consider particular elements of a design, but also their relation to the design as a whole and to the greater environment.”
“The power of this collective effect has not been lost on nature. This is where Johnson’s stories about ants come in. How do the ants do it?
Foraging ants depart the nest with one job in mind, to find and retrieve food. They also have the ability to leave and follow chemical trails. At first, they disperse randomly. When the ants that find food come back to the nest, they leave a chemical trail that their sisters can follow. Studies show that this process allows ants to consistently find the shortest path to the food.
Once researchers understood this collective ability, they decided to play a trick on the ants. In a controlled setting, the scientists placed two food sources at identical path lengths from the nest. As it turned out, the ants ended up using just one of the paths, although which one they chose was random. Why? Because they follow chemical trails, a couple more ants going down one path will attract other ants, triggering a positive feedback loop. So instead of finding an optimal solution, the ants have one crowded path and an equidistant, empty path.
Amazingly, though, nature anticipated this problem as well. As it turns out, ants periodically break from the main path and begin a random search process again. The ants are programmed to strike a balance between exploiting a known food source and exploring for the next food source. The ants are hard-wired to seek diversity.”
6 Consistency
Have a Great weekend when You get to that stage,
Sune