Pulse 250606: Rare Earths, AI in Love and War, Space Programs
Every Friday, I share 3 takes on tech, geopolitics, infrastructure, or AI; 2 interesting reads from others; and 1 soundtrack to wind down for the weekend.
3 Quick takes from me
I. Rare Earths
China holds over 80% of the global supply of rare earth elements (REEs)—materials that underpin critical modern technologies such as semiconductors, electric vehicle (EV) batteries, weaponry, and wind turbines.
In April 2025, China deepened its resource leverage by imposing export controls on seven rare earth elements and derivative products (notably magnets). The new policy mandates that exporters obtain licenses, approvals for which Beijing is strategically delaying. These measures disproportionately affect sectors central to the green energy transition. Targeted elements include gallium (used in AI chips and RF semiconductors), graphite (used in lithium-ion batteries), and dysprosium (used in high-strength magnets in wind turbines).
These elements are relatively scarce, and demand is rising due to the global shift towards renewable energy. As such, China has many short and long-term leverage points even amid Western efforts to diversify the REE global supply chain. But scaling alternative supply chains is a multi-year endeavor, meaning that China for now retains asymmetric leverage over industries that define tomorrow’s economy.
Where this could go
Although Chinese authorities have signaled potentially softening these restrictions (and doing so as a gesture of goodwill amid ongoing trade talks), multiple near-term scenarios can unfold, each carrying distinct implications for global supply chains, industrial policy, and geopolitical alignment.
Full Export Ban (Severe Impact, Moderate Likelihood)
China enacts a universal ban or one on specific countries. We’d likely see retaliatory action by the U.S. and its allies (e.g., dual-use export controls, FDI restrictions, sanctions), WTO complaints, and price shocks across mineral markets as other REE-rich countries such as the DRC and Chile emulate China’s tactics and apply additional export royalties and tariffs.
Delay and deny (High Impact, High Likelihood)
This is the most probable near-term path. China restricts REE exports through bureaucratic tools (e.g., delayed licensing, inspections and audits, and annual export quotas). This more balanced strategy maintains leverage while mitigating the risk of full-fledged retaliation, causing price increases and forcing original equipment manufacturers (OEMs, e.g., Tesla, Apple, Boeing) into costly and risky supply chain workarounds (e.g., sourcing from European or Australian REE providers, stockpiling inventory, vertically integrating mine or processor companies).
Feedback loops & critical mass
Given China's dominant role in the processing and supply of these materials, even a calibrated disruption would ripple far beyond the immediate chokepoints, irrespective of the two scenarios outlined above:
Medium-term effects
Inflationary pressure on central banks, which in turn sparks government subsidy races for REE procurement or mining, and shifts industrial policies where states more directly fund and control production capacity.
Clean energy projects are delayed, undermining current net-zero timelines. Poorer countries, lacking or unable to subsidize domestic green companies, could face longer delays and higher costs.
Investor risk premiums increase for firms reliant on Chinese-sourced materials, potentially leading to valuation discounts in EV, defense, and semiconductor firms.
Long-term effects + feedback loops
Though only moderately likely, a hard decoupling (similar to Huawei/SMIC decoupling) in procurement, processing, and refining could permanently realign supply chains away from China. New REE cartels form, just as OPEC was formed in response to Western dominance in oil pricing and refining. The idea has already taken shape around lithium (in 2023, Mexico called for a “Lithium OPEC” and Chile announced nationalizing its lithium sector). An REE cartel, led by China, and including countries like Myanmar, Vietnam, Brazil, could dominate global REE availability through control of production quotas, prices, and export flows.
When opportunity comes knocking
Diversification is already in motion. The U.K. is pushing to build a rare earth processing plant in France, expected to begin production by 2027. U.S.-based MP Materials is partnering with Maaden to develop a rare earths supply chain in Saudi Arabia. Lynas Rare Earths, an Australian company, has successfully produced dysprosium at its Malaysian refinery, marking the first production of heavy rare earths outside of China. As with any crisis, opportunity remains for the willing and discerning. Here are some ideas:
Western governments increase funding for domestic mining, refining, and substitution technologies under national security mandates.
Australia, Canada, and Saudi Arabia sign major trade agreements and benefit from significant capital inflows as allies shift to friend-shoring.
Vertical integration by automakers, chipmakers, and defense primes will see them taking equity stakes or building direct supply relationships (e.g., Tesla in lithium, Apple in cobalt).
II. AI Bubbles
Mary Meeker is a renowned venture capitalist and former Wall Street analyst, widely regarded as one of the most influential voices in technology. She is the founder and general partner at BOND, an SF–based VC.
Last week, BOND released its first major trends report since 2019: “Trends – Artificial Intelligence,” a 340-page deck detailing AI’s sweeping impact across sectors. The report is dense, and many of its observations are still being unpacked. As trends evolve and assumptions are tested, six core insights stand out.1
For the sake of brevity and depth, I’ll focus on a single data point that stood out to me:
Inference costs decreased by 99% since 2020,
while training costs increased to USD 1 billion per run
The report highlights a bifurcation of the AI value chain:
Training is capital-intensive and gated by access to powerful compute infrastructure (e.g., OpenAI, Google, etc.);
But inference is heading toward near-zero marginal cost.
This creates a dynamic where (1) fewer firms hold power through proprietary model weights, compute, and CAPEX, and (2) millions of users and developers embed AI into daily tasks via APIs.
This signals that moats are moving downstream—and fast. Per Meeker’s observations, compute + proprietary data + scaling capabilities are increasingly porous moats. In my view, this is a consequence of open-source proliferation (e.g., DeepSeek R1), commoditization of inference, and what Meeker refers to as “global distribution ramps.”
Open-source models like DeepSeek R1, Alibaba’s Qwen 2.5, and Meta Llama 3 show that capability diffuses faster than expected, especially in coding and reasoning tasks critical for building autonomous agents. I think this trajectory will continue as demand grows for cost-efficient infrastructure and low-code/no-code solutions democratize access to building agents for personal, corporate, and commercial use (see below).
Distribution will surpass model quality in moat-building as AI-native mobile interfaces (e.g., ChatGPT, Claude) and satellite networks extend Internet access to the 2.6 billion people still offline. Their first touchpoints may be through agent-driven interfaces that understand local language, context, and intent.
Verticalized deployments/integrations (e.g., Duolingo’s recent AI integration) demonstrate that the tight coupling of domain-specific data + UX + task tuning often outperforms raw generality in value creation. Moats increasingly lie in full-stack architectures, such as Microsoft (Azure + Copilot + M365), Amazon (AWS + vertical AI partners), Apple (on-device AI + platform control), or OpenAI, with its recent acquisition of Jony Ive’s io.
What does this mean? That we could be at an inflection point where (1) cost collapse democratizes access to tools and (2) interface evolution/adoption reorders platform hierarchies.
Agentic architectures become moats as foundational models (not frontier) commoditize, moat defensibility shifts to other AI stack layers, notably orchestration (e.g., autonomous agent frameworks, workflow design, etc.).
Interface providers are the new gatekeepers. New Internet users will skip browsers and search bars altogether for AI agents fluent in local language, context, and culture. This is analogous to China’s internet jump in the 2000s and 2010s, where the country skipped legacy infrastructure and went directly to mobile internet, forging the dominance of super apps (e.g., WeChat, Alipay) and all-in-one services that integrate communication, commerce, and payments into a single mobile UX. In BOND’s context, interfaces “disintermediate” traditional UX as users are increasingly able to express intent in a few words. Interface providers who capture this shift will gain a strong first-mover advantage.
I would argue that these moats are shifting upstream as well:
Compute infrastructure: building data centers at hyperscale requires long-term capex, energy access, and real estate strategy (moat: difficult to replicate).
R&D: Training frontier models like o3 or Opus 4 requires billions in compute and coordination, notably around alignment and security (moat: capital + expertise).
Semiconductors: Foundry access (e.g., TSMC) as a structural chokepoint (moat: capital + supply chain diversification to hedge against geopolitical risk).
III. Fighting at the Edge
Two remarkable events last week signaled a structural shift in modern warfare: Ukraine’s drone strike on Russian airbases and Anduril’s Series G round.
On June 1, Ukraine’s Security Service of Ukraine (SBU) launched Operation Spiderweb, a covert operation that targeted multiple Russian airbases to degrade Russia’s nuclear-capable bomber fleet. The attack targeted 20 bombers—destroying 10—using just over a hundred modified commercial off-the-shelf drones, inflicting an estimated USD 2 billion in damages.
What made the strike exceptional was its geographic dispersion and target acquisition enabled by autonomous, logic-enabled drones. While autonomous weapons themselves are not new, the use of enemy telecoms, open-source software, and low-cost hardware reflects a deeper shift: the battleground is moving from centralized, single-chain command structures to networked (distributed and decentralized) structures that mirror the logic of modern computing. While this is a science of its own, we can simplify this structure as a node/edge model with three core features:
Decision-making moves away from the structural hub and closer to the edge (drone or squad);
Mission objectives are preprogrammed or dynamically updated via shared protocols and policies;
No central point of failure, meaning each node operates autonomously, creating redundancy to ensure mission continuity.
Generated with ChatGPT 4o. Edges can be bidirectional.
This model is particularly resilient in contested combat environments, where modular, interoperable weapons with local autonomy at the edge level become a force multiplier. Ukraine delivered this attack with a swarm of 117 USD 4,000 drones.
ROI = 2,000,000,000 / (117 * 4,000) ≈ 4273.5
For each dollar spent, Ukraine inflicted USD 4,273.50 in damages. This ROI shows the asymmetric advantage afforded by networked command structures.
Defense tech firms have long been building toward this structural shift. For example, Anduril (which just raised a USD 2.5 billion Series G at a whopping USD 30.5 billion valuation) integrates hardware and software into the “warfare stack” to enable the distributed autonomy of a networked structure.
While Anduril produces hardware using a model that has upended traditional procurement models, the real value/force multiplier lies in its core software product, LatticeOS. This OS is the command & control/orchestration layer, allowing for software integration with assets like drones (air and submarine), sentinels, reconnaissance tools (and soon fighter jets). It allows operators to assign objectives using preprogrammed rules and policies (e.g., in extremely simple conditional logic: if an enemy asset is a drone, then intercept it and move on to the next one), while retaining the ability to intervene or authorize critical actions as needed.
Lattice OS allows nodes to coordinate and operate (edge) as autonomous teams, and for operators to outsource decision-making with preprogramming, enabling a seamless networked chain of command (similarly to how sophisticated cyber threat actors use C2 servers to coordinate malware campaigns concurrently).
As I’ve argued before, tech companies that differentiate themselves at the orchestration layer of the hardware stack will capture the greatest value from AI-driven disruptions to the value chain.
2 Compelling reads from others
I. Creation is not birth; it is murder.
The essay below by Maalvika explores ambition as a vector towards self-sabotage. Having read many books on thinking, experimentation, and creativity, I find this essay remarkable for its sharp one-liners and (in hindsight, obvious) insight that over-researching, over-planning, and over-strategizing can paralyze us into inaction. We place ourselves at risk of falling into a cognitive trough, unable to deliver because our brain is already rewarding you as if we were already executing (this misfire is well-documented). Cognitive science concepts like the “taste-skill gap” further underscore why inspiration isn’t enough in the creative process: our ability to judge what is good (i.e., taste) develops faster than our ability to produce it (i.e., skill), resulting in a gap that shapes and derails the process.
II. Can AI find words for our feelings?
This Financial Times article asks whether using AI to articulate our feelings diminishes their authenticity or purity. The author draws a sharp analogy to GPS tech: just as we've outsourced our sense of direction to navigation apps, outsourcing emotional articulation to LLMs may cause a slow erosion of our own fluency and ultimately connection to inner selves.
1 Musical Cultural Note
The Rare Earths story came to me when my phone showed me pictures of an event I was invited to in 2022: Tom Sachs’ SPACE PROGRAM: RARE EARTHS live performance at the Deichtorhallen Hamburg. Part of Sachs’s Space Program series, this performance simulated an interstellar REE mining expedition, where guests engaged in symbolic tasks to earn entry into the mission. The storyline centered on harvesting REEs from Vesta (the brightest and most visible asteroid from Earth) to illustrate humanity's dependence on technology and resource extraction.
Thanks for reading SecondOrder! Subscribe for free to receive new analyses. If you know someone who might enjoy this content, please forward this email.
Have thoughts on my takes? Comment on Substack or send me a message!
© 2025 James Pooler. All rights reserved.
Unprecedented user, usage, and capital expenditure growth. The world is experiencing rapid and transformative technology innovation and adoption: Data center electricity consumption globally more than tripled between 2005 and 2024; Microsoft's Azure AI Foundry processed over 100 trillion tokens in Q1 2025 (up 5x YoY), and ChatGPT reached 800 million weekly users by April 2025.
Open-source and China threaten LLM moats. China’s leading models are narrowing performance gaps with frontier U.S. models, while delivering lower-cost, compute-efficient alternatives. Open-source releases like Llama 3, Qwen 2.5, and DeepSeek’s R1 series have accelerated this trend, with China leading more efficient alternatives, allowing developers to use cheaper models locally or with reduced premiums for reliable outputs. This may, in the long term, undermine U.S. incumbents’ ability to charge premium subscription prices.
AI physical world ramps are fast and data-driven. One of the more interesting stances considering Meeker’s (and this report’s) VC penchant is the observation that “the biggest changes aren’t coming from Silicon Valley. They’re showing up where software meets the physical world.” For example, Kaiser Permanente’s (a California-based healthcare company) AI Scribe usage is surging. FDA approvals for AI-enabled medical devices have surged from just 1 in 1995 to 223 by 2023, with over half approved in the past three years alone. The FDA itself is pursuing full internal AI integration by June 2025. U.S. federal AI R&D totals $14.7B through FY25.
AI is accelerating global internet adoption at record speed through low-cost inference, ubiquitous mobile interfaces, and AI-native UX (e.g., ChatGPT). Combined with satellite networks extending access to the 2.6 billion people still offline, these news users will enter the internet via AI-native interfaces rather than traditional browsers or search bars. This "first experience" could involve interactions with an "agent-driven interface" that manages various tech platforms, understands local language, context, and intent, potentially making interface owners the winners.