Oct 9, 2025: China's Ministry of Commerce issued Announcements No. 61 & 62, expanding rare earth export controls to 12 of 17 elements and imposing extraterritorial licensing requirements.
This is direct retaliation for U.S. semiconductor export bans announced days earlier.
China controls 70% of global mining, 90% of processing, and 93% of permanent magnet production. Each F-35 requires 417kg of rare earths. China refines 100% of global samarium.
What does this mean for U.S. defense? How will this affect AI data centers? What happens to semiconductor and EV supply chains? Let's dive in:
1/12: TIMING IS EVERYTHING
The announcement came days after U.S. expanded chip export bans (Oct 7, targeting ASML/TSMC) and weeks before two critical deadlines:
• 90-day U.S.-China trade truce expires
• Trump-Xi meeting in South Korea
Strategic retaliation designed to maximize Beijing's leverage in upcoming negotiations.
2/12: RARE EARTHS 101
17 elements (lanthanides + yttrium/scandium) critical for high-tech applications—magnets, lasers, semiconductors.
They're not "rare" geologically, but incredibly hard to process:
• Only 0.1-1% concentration in ore
• Creates radioactive byproducts (thorium), driving up environmental and political costs
China dominates via low-cost mining and vertical integration. The Bayan Obo mine alone produces 70% of global light rare earths.
3/12: WHAT'S ACTUALLY RESTRICTED - ELEMENTS & MATERIALS
China added 5 rare earths to the restricted list: holmium (Ho), erbium (Er), thulium (Tm), europium (Eu), and ytterbium (Yb)—critical for lasers, fiber optics, and defense systems.
That means 12 out of 17 rare earths are now restricted, including neodymium (Nd), praseodymium (Pr), and dysprosium (Dy) from April.
Plus dozens of refining and mining equipment items.
Effective: Nov 8 for elements/equipment, Dec 1 full implementation.
4/12: FOREIGN PRODUCTS
Any product with >0.1% Chinese-sourced rare earths needs Beijing's export license for re-export.
Even if it's made in Taiwan. Or Vietnam. Or Texas.
This is China's version of the U.S. Foreign Direct Product Rule. Extraterritorial control over global supply chains.
5/12: MORE CONTROLS
END-USE BANS:
• No licenses for foreign militaries, or weapons
• Case-by-case review for ≤14nm chips, ≥256-layer memory, AI/military R&D
TECHNOLOGY & LABOR:
• Ban on exporting REE mining/processing/recycling tech
• Chinese citizens need government approval to join overseas REE projects
Full compliance deadline: Dec 1, 2025.
6/12: CHINA'S DOMINANCE - THE NUMBERS
Mining: 70% global share (240k tons vs U.S. 43k tons)
But here's where it gets scary:
• 90% of global separation/refining
• 93% of permanent magnet production
• 44M tons reserves (37% global)
Global demand: 200k tons/year, growing 7-10% annually from EVs/AI/renewables.
Alternatives exist (Australia's Lynas 8%, U.S. MP Materials 15%), but they all still send ore to China for processing.
7/12: DEFENSE IMPACT
This is where it gets serious.
Rare earths are irreplaceable in military hardware due to magnetic/thermal properties:
China refines 100% of global samarium—the element critical for high-temp military magnets.
8/12: THE DEFENSE CAPABILITY GAP
The implications are stark:
• China builds military hardware 5-6x faster than the U.S.
• U.S. has zero domestic samarium refining capacity
• Short-term: Stockpiles last months
• Long-term: 5-10 years to build independent supply chains
With Indo-Pacific tensions rising, Beijing now has leverage over the foundation of U.S. defense production.
9/12: SEMICONDUCTOR CHOKEPOINT
REEs are critical for chip manufacturing:
• Magnets in lithography equipment (ASML's EUV tools)
• Wafer processing equipment
Ban targets ≤14nm chips (Nvidia A100/H100 territory).
TSMC, Samsung, SK Hynix ALL need licenses if using Chinese REEs.
That 0.1% threshold = de-facto veto power over the semiconductor supply chain.
10/12: BROADER ECONOMIC RIPPLES
Markets reacted immediately: Chinese REE stocks surged 9-10%. U.S. miners like MP Materials rose on investment flows.
Short-term: 20-50% price spikes
Key impacts:
• EVs: 30% REE-dependent
• Wind turbines: up to 200kg/MW
• AI data centers: REE magnets for cooling
Historical precedent: China's 2010 embargo on Japan sent REE prices up 10x.
11/12: U.S. COUNTERMOVES
DoD response:
• $400M equity in MP Materials (largest shareholder)
• $150M loan for heavy REE separation
• 10-year offtake for new magnet facility
• Lynas-Noveon partnership for U.S. production
India/Australia ramping exploration.
Recycling emerging: 10-20% recovery potential from e-waste.
Reality check: 5-10 years minimum to scale. U.S. = <5% global processing today.
13/13: THE BOTTOM LINE
China weaponizing 90% processing monopoly to retaliate for U.S. chip bans. Targeting defense (417kg per F-35) and semiconductors ≤14nm.
Short-term: 20-50% price spikes could throttle AI boom (data centers need REE magnets).
Long-term: Forces Western reindustrialization. Diversification takes years.
THE QUESTION: Can alternatives scale faster than China leverages its monopoly?
Negotiations are possible—Trump-Xi could trade chip access ⟷ REE flow.
The decoupling isn't coming. It's here.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
On ProofBench-Advanced—where models prove formal mathematical theorems—GPT-5 scores 20%. Gemini Deep Think IMO Gold hits 65.7%. DeepSeek Math V2 (Heavy) scores 61.9%.
That's second place—but Gemini isn't open source.
This is the best open math model in the world. And DeepSeek released the weights. Apache 2.0.
Here's what they discovered:
1/ Why Normal LLMs Break on Real Math
Most large language models are great at sounding smart, but:
- They’re rewarded for the final answer, not the reasoning.
- If they accidentally land on the right number with bad logic, they still get full credit.
- Over time they become “confident liars”: fluent, persuasive, and sometimes wrong.
That’s fatal for real math, where the proof is the product.
To fix this, DeepSeek Math V2 changes what the model gets rewarded for: not just being right, but being rigorously right.
2/ The Core Idea: Generator + Verifier
Instead of one model doing everything, DeepSeek splits the job: 1. Generator – the “mathematician”
- Produces a full, step-by-step proof.
2. Verifier – the “internal auditor”
- Checks the proof for logical soundness.
- Ignores the final answer. It only cares about the reasoning.
This creates an internal feedback loop:
One model proposes, the other critiques.
Battery storage is already scaling—159 GW deployed globally, 926 GW projected by 2033.
Renewables needed it first. Now AI needs it too.
Tesla is deploying Megapacks at data centers. China is deploying 30 GW this year, integrating storage directly into AI buildout.
Why? Data centers can’t scale without solving three problems:
- 7-year interconnection queues
- power quality GPUs demand
- backup without diesel permits
Batteries solve all three ↓
Why AI Data Centers Need Batteries
Interconnection is broken. Utility connection takes 7+ years. Batteries bypass it. Skip the queue.
GPUs break traditional power. Training loads swing 90% at 30 Hz. Batteries smooth it in 30 milliseconds.
Diesel doesn’t scale. Permitting is hard. For 20-hour backup, batteries are cost-competitive.
The math: ~1% of data center capex.
The Scale
Global capacity: 159 GW by end-2024. Up 85% from 86 GW in 2023. Projected: 926 GW by 2033.
Cost curve: $115/kWh in 2024, down 84% from $723/kWh in 2013. Still falling.
Economics flipped. Solar plus 4-hour storage runs ~$76/MWh. New gas peakers cost $80-120/MWh.
The universe isn’t just expanding — it’s speeding up
13.8 billion years after the Big Bang, astronomers expected gravity to slowly slow cosmic expansion. Instead, when they looked deep into space, they found the opposite: the universe is accelerating.
Whatever drives that acceleration makes up ~70% of the cosmos.
We call it dark energy.
We can measure it. We can see its effects. So what is it, really?
How we figured this out
Cepheid stars: the distance trick
Henrietta Leavitt discovered that certain stars (Cepheid variables) get brighter and dimmer with a regular period — and that period tells you their true brightness → lets us measure distance to faraway galaxies.
Redshift: galaxies on the move
Vesto Slipher used spectra of galaxies to show many had their light stretched to longer, redder wavelengths.
Redder → moving away faster.
Hubble & the expanding universe
Edwin Hubble and Milton Humason combined Cepheid distances with redshift and found a pattern:
>The farther a galaxy is, the faster it’s receding.
That’s the Hubble–Lemaître law: clear evidence that the universe is expanding.
The shock: expansion is accelerating
In the 1990s, two teams studied Type Ia supernovae, stellar explosions so consistent in brightness that they act like “standard candles.”
By comparing how bright they should be to how bright they look, you can get distance.
By measuring redshift, you get how fast they’re moving away.
The surprise:
• The supernovae were dimmer and farther away than expected.
• That only made sense if, over billions of years, the universe’s expansion had sped up instead of slowing down.
This cosmic acceleration is what we now attribute to dark energy.
🚨The White House just launched the Genesis Mission — a Manhattan Project for AI
The Department of Energy will build a national AI platform on top of U.S. supercomputers and federal science data, train scientific foundation models, and run AI agents + robotic labs to automate experiments in biotech, critical materials, nuclear fission/fusion, space, quantum, and semiconductors.
Let’s unpack what this order actually builds, and how it could rewire the AI, energy, and science landscape over the next decade:
1/ At the core is a new American Science and Security Platform.
DOE is ordered to turn the national lab system into an integrated stack that provides:
• HPC for large-scale model training, simulation, inference
• Domain foundation models across physics, materials, bio, energy
• AI agents to explore design spaces, evaluate experiments, automate workflows
• Robotic/automated labs + production tools for AI-directed experiments and manufacturing
National-scale AI scientist + AI lab tech as infrastructure.
2/ The targets are very explicit and very strategic.
Within 60 days, DOE has to propose at least 20 “national challenges” in: