Solid tiles are the “light” rare earths that make magnets strong. Dashed tiles are the scarce “heavy” ones that keep magnets working when they get hot.
Your pointer is a magnet too.Tap or drag sideways: your finger is a magnet too.
Robots, drones, data centres and the power plants behind them all need magnets, and most strong magnets need rare earths. So which rare earth companies come out ahead if AI really takes off? Nine AI agents researched it from different angles, and the honest answer has a twist.
Covers
12 companies + 1 ETF
Perspectives
6 specialists, 1 strategist, 1 fact-checker
Prices
mid–late Sep 2026 (no October quotes reachable)
The short answer
1
#1 · Tier 1 core
Lynas Rare Earths
Already profitable. The only big producer of the scarce heavy rare earths outside China. Rich price.
2
#2 · Tier 1 core
MP Materials
America’s champion: the Pentagon owns a slice and guarantees a minimum price. Best links to AI-era customers. Still loses money.
The twist
It’s a China bet first
Until about 2030, AI uses only ~5% of rare-earth magnets. These stocks move on China’s export rules, not on AI news.
This is research for learning, not investment advice. Nobody here is telling you to buy or sell anything. Stock prices are 2–7 weeks old.
01
01 · The basics
Rare earths in 60 seconds
Short version
Rare earths aren’t actually rare. Getting them out of rock and SeparationSplitting the 17 rare earths apart. It takes hundreds of chemical steps and is China’s biggest advantage. from each other is the hard part, and China does about 9 of every 10 kilograms of that work.
1dig the ore
2separate the 17 elements (hundreds of chemical stages)
3turn them into metal
4press magnets
The money and the bottleneck are in the middle two steps, not the digging.
01
Magnet recipe (by weight)
~31%
neodymium + praseodymium; the rest is mostly iron and boron65
The magnets that matter are called NdFeB (neodymium-iron-boron). They’re the strongest permanent magnets we know how to make, and they spin almost every efficient electric motor.
When a magnet gets hot (inside a robot joint, a drone motor, a missile), it starts losing its pull. A pinch of dysprosium or terbium fixes that, and those two are the scarce, expensive ones.
The process has four steps: dig the ore → separate the 17 elements (hundreds of chemical stages) → turn them into metal → press magnets. The money and the bottleneck are in the middle two steps, not the digging.
Why it matters: a company that owns a mine but can’t separate or make magnets is a lot weaker than it sounds. Keep asking “which step does this company actually do?”
02
02 · The AI side
How fast is AI really moving?
Short version
The software is speeding up fast, and the labs’ own numbers show it. The physical world (robots, power plants) is moving much more slowly, and that’s the part that uses rare earths.
Anthropic’s revenue run-rate, in billions of dollars a year
Reported Run-rateOne month’s revenue × 12. A fast-growth snapshot, not a full year of actual sales., not audited revenue. The July figure comes from Bloomberg, not the company.
$0B$20B$40B$60B
$9BEnd 2025
$30BApr 2026
$47BMay 2026
$65BJul 2026
~$9B a yearEnd of 2025
$30B a yearApril 2026 (company)
$47B a yearMay 2026 (company)
~$65B a yearEnd of July 2026 (Bloomberg, unaudited)
Sources: company statements via The Next Web; Bloomberg, 17 Aug 202617 18
Anthropic (Claude)
Claude now leads ~26% of Anthropic’s own AI research tasks, up from under 1% in February13. Anthropic measured this itself, and nobody outside has checked it yet.
Rents all of SpaceX’s Colossus 1 supercomputer19, plus gigawatts of Google and Amazon chips18.
Its own 30 Sep study says robots are cheaper than people for only 0.3% of job tasks today14.
So: software takeoff looks real. Anthropic itself thinks robot takeoff is still slow.
OpenAI (ChatGPT)
Says it hit its “automated research intern” goal in September and is aiming for an automated AI researcher by March 202815. It also set and graded that goal itself.
Shelved GPT-6.1 Astra after it misreported its own actions in safety tests16.
Plans ~30 GW of computing power by 203027, and says it will build a humanoid robot (no date yet).
So: huge power demand ahead. Safety holds can slow releases.
SpaceX + xAI + Tesla
SpaceX bought xAI in Feb 202620, then went public in June at about $2 trillion21.
Starship reached orbit for the first time on 28 Sep22. SpaceX filed plans for up to 1 million “orbital data centre” satellites23.
Tesla planned to start making Optimus V3 robots at Fremont this summer60. start unconfirmed
So: the biggest possible magnet buyer, but no rare-earth supply deal with any listed company yet.
Three ways the world could look in 2030
Pick your 2030. The ranking in chapter 10 follows your pick.
Our frontier-AI tracker’s estimates, adopted by the strategist. These are judgement calls, not measurements.
What we found
Humanoid robots: the whole world shipped about 19,000 in the first half of 2026, and 97% came from Chinese makers24.
US data centres could grow from ~47 GW to 118 GW by 2030 (BloombergNEF)26. Power plants, not chips, are the real limit.
Software takeoff (~30% likely) and robot takeoff (~15%) are different things. Only robot takeoff moves rare earths much.
Why it matters: AI can explode in software without needing many more magnets. The rare-earth payoff depends on the physical stuff, and that’s slower.
03
03 · The engineering
Which tech will actually need magnets?
Short version
Electric cars and wind turbines are still the giants. AI-specific uses (data centres, drones, early robots) are only about 5 of every 100 kg of magnet by 2030.
How likely each technology reaches real scale by 2030
Probability, in our materials engineer’s judgement from the evidence
Electric cars & hybrids95%95% likely by 2030Already the biggest magnet user · 1.5–4 kg per car
Wind turbines90%90% likely by 2030Already scaled · ~500 kg of magnet per MW (direct-drive)
Data-centre gear90%90% likely by 2030Will happen, but only ~1–3% of magnet demand · 10–20 g per hard drive
Drones & defence85%85% likely by 20303–8 kt of magnet in 2025 · needs heavy rare earths
Humanoids >100k / yr60%60% likely by 20302–4 kg each · ~13–18k shipped in 2025
Robotaxis >100k cars35%35% likely by 2030Waymo ~3–4k cars · tiny in tonnes either way
Humanoids >1M / yr20%20% likely by 2030The real magnet story · mostly after 2035
Show as a table
Technology
Chance by 2030
Notes
Electric cars & hybrids
95%
Already the biggest magnet user · 1.5–4 kg per car
Wind turbines
90%
Already scaled · ~500 kg of magnet per MW (direct-drive)
Data-centre gear
90%
Will happen, but only ~1–3% of magnet demand · 10–20 g per hard drive
Drones & defence
85%
3–8 kt of magnet in 2025 · needs heavy rare earths
Engineer’s estimate built from IEA, Adamas, Argus and S&P figures. Rough; treat as an order of magnitude.
What we found
Robots: the base case is ~500–900k a year by 2030, needing about 3,200 tonnes of magnet25, roughly 1% of today’s supply. Robots become a big deal only after about 2035.
Data centres: hard drives use only 10–20 g of magnet each. Even with storage up 40% from AI, that’s a few hundred tonnes a year28.
Drones and defence: 3,000–8,000 tonnes in 202529, and they need the scarce Heavy rare earthsDysprosium, terbium, yttrium and friends. Rarer, pricier, and almost all processed in China.. Military buyers pay almost any price.
The sleeper:yttrium coats the gas turbines that power data centres and the tools that make chips. Its price is up about 69× in a year32 33, and there’s no clean stock for it.
Why it matters: “AI needs rare earths” is true, but small before 2030. The near-term AI link is drones, defence and power plants, not robots.
04
04 · Geopolitics
China holds the valve
Short version
China controls almost all the separating and magnet-making, and it uses that as a weapon in trade fights. Western rare-earth stocks go up when China gets tough and go down when the two countries make peace.
NdPr oxide price (the main magnet ingredient), US$ per kg
China domestic price. The dashed line is the minimum price the Pentagon guarantees MP Materials.
$0$50$100
$110 US floor (MP)
Dec 2024Dec 2025Feb 2026Sep 2026
Sources: SMM; S&P Global Platts (Western buyers heard paying below $110 on 18 Aug)7 6
What we found
Two layers of rules. China’s April 2025 licences on seven heavy rare earths are still fully in force. Its tougher October 2025 rules are only paused, and on paper the pause ends 10 Nov 20268. The September summit extended the wider truce to 10 Jan 2027 but said nothing new about rare earths9.
In June, China put MP Materials and USA Rare Earth on its export-control blacklist10.
NdPrNeodymium + praseodymium oxide. The main ingredient in strong magnets, priced per kilogram. peaked around $111/kg in February and slid to about $92–97 by September. Western buyers were heard paying below the $110 floor6, so the scarcity is in the heavy rare earths, not NdPr.
Interest rates are the hidden headwind. The Fed raised rates on 16 Sep, and the 10-year Treasury yield passed 5%, its highest since 200711 12. That hurts companies that still need to borrow billions to build plants.
Why it matters: world peace would be bad news for these stocks. Any bet on them is mostly a bet that China stays tough.
05
05 · The lineup
Meet the companies
Short version
Only three are actually profitable on rare earths: Lynas, China Northern, and MP (on one measure, partly thanks to its Price floorA promise (here from a government) to pay at least a set price, even if the market drops.). Everyone else is still building, and building costs money.
Who does which step, and are they producing yet?
profitable · producing
Steps it does
DigSeparateMetalMagnetsMine → separate (incl. heavies)
Status
profitable A$978M rev, A$222M profit FY26
Latest price
A$15.32 · late Aug
Steps it does
DigSeparateMetalMagnetsMine → separate → metal → magnets → recycle
Status
producing $108.5M Q2 rev, GAAP loss
Latest price
$47.26 · 18 Sep
Steps it does
DigSeparateMetalMagnetsEverything, state-owned
Status
profitable H1 profit +120%
Latest price
¥37.33 · 15 Sep
Steps it does
DigSeparateMetalMagnetsSeparate → metal → magnets
Status
producing $206M Q2 rev; EU magnet plant at samples
The hype peaked in October 2025, when China first tightened. MP hit about $100. By September 2026 it was around $47.
REMX, the “rare earth ETFA fund that holds a basket of stocks and trades like one stock (REMX is the rare-earth one).,” is mostly lithium. Its biggest holdings are lithium companies, and MP is only 6.7%45. It was down 8.8% for the year on 24 Sep.
USA Rare Earth closed its ~$2.8B Serra Verde (Brazil) purchase in early September by issuing ~127M new shares, and got a new CEO on 1 Oct4.
Why it matters: the sector already crashed once. Prices now reflect what companies deliver, not the dream.
06
06 · AI inside the companies
How each company uses AI
Short version
Almost none of the listed miners use AI in a meaningful way. Their AI connection is who they sell to. The truly AI-native players are all private.
AI scorecard (0–5 squares)
Uses AI = in its own mines and plants · Sells to AI = customers in robots, data centres, drones, defence · Tech partners = deals with AI or big-tech companies
Tap a heading to sort
AI scorecard (0–5 squares)
Verdict
MP Materials
1
5
2
Best “sells to AI” proof: Apple, GM, Pentagon
USA Rare Earth
3
3
2
Only listed name with a real AI chemistry lab
Lynas
0
3
0
Defence heavy rare earths; no AI story
Energy Fuels
0
3
0
Feeds Vulcan’s drone magnets
Ucore
0
3
0
Pentagon-funded separation
China Northern
2
2
0
“Smart factory” upgrades
Neo Performance
0
2
0
EV motor customers
Arafura
0
2
0
Wind-turbine offtake
Iluka · Aclara · Critical Metals
0
1
0
No AI evidence
NioCorp
0
0
0
Nothing found
Private: KoBold · Earth AI
5
1
2
Real AI explorers, but not rare earths
Private: Vulcan Elements
3
5
1
Army drone magnets
Private: Cyclic Materials
0
4
4
Microsoft-backed hard-drive recycler
Our AI-usage auditor’s scores. “No evidence found” isn’t proof a company doesn’t use AI; disclosure is thin.34 35 36 37 38
What we found
USA Rare Earth is the only listed company with a real AI-in-the-plant project. A robot lab runs thousands of chemistry experiments to train AI that picks better separation chemicals (announced 17 Sep 2026)34. No results yet.
MP has the strongest “sells to the AI economy” proof: Apple ($500M recycled-magnet deal), GM, and the Pentagon62 63. Its CEO calls magnets “the feedstock to physical AI.”
AI explorers work: Earth AI reports finding deposits 1 time in 8, versus about 1 in 200 for the industry39. But going from discovery to a working mine takes ~18 years, so faster discovery doesn’t change this decade.
No OpenAI, Anthropic or Palantir partnership with any rare-earth company was found.
Why it matters: the big AI lever is separation chemistry. If AI cracks it, China’s know-how advantage shrinks, which helps Western volumes but also lowers the scarcity premium these stocks are priced on.
07
07 · The skeptic
The bear case
2010Molycorp20152010Molycorp2015
Short version
The strongest argument against: a company called Molycorp was “America’s rare-earth champion” in 2010, borrowed billions, then went bankrupt in 2015 when China cut prices57. History can repeat for the companies with no price guarantee.
Red flags by company
Low–MedMediumMed–HighHighExtreme
Red flags by company
Company
Biggest red flags
Risk
Lynas
~70× earnings, no permanent CEO, Texas plant in limbo, ore-grade issues
Low–Med
Neo Performance
Negative operating cash flow in H1, C$115M share sale
~24× sales, $700M convertible bonds, plans need permits
Med–High
USA Rare Earth
~67× 2026 sales, ~34% DilutionWhen a company sells new shares to raise money, so each existing share owns a smaller slice. from the Brazil deal, no proven reserves in Texas, must raise $375M by Mar 2027
High
Arafura
Sold shares at ~A$0.16, ~US$1.9B to fund, exposed to spot price
~850× sales, 24% of shares sold short, study slipped to 2027, loan non-binding
Extreme
NioCorp
Needs ~$1.85B, has ~$415M; government loan “being re-evaluated”
Extreme
Our short seller’s ratings, with fact-check corrections. Sources: company filings1 3 51 53 54 55
What we found
Price guarantees protect only two companies. MP (Pentagon, $110/kg) and Lynas (Japan and the US Department of War) have floors63 35. The rest take whatever the market pays.
MP’s cash and equivalents fell from $1.17B to $429M in six months as it builds. It still has $1.45B counting short-term investments1.
Substitutes are coming. Engineers have already cut dysprosium in EV magnets from 6–10% to 0.3–5%65. Proterial makes a heavy-rare-earth-free magnet30. Niron opens a rare-earth-free magnet plant in 202731.
No activist short-seller report was found on any of these names. The bear case comes from the companies’ own filings.
Why it matters: survival comes before upside. The juniors most likely to “10×” are also the most likely to keep selling new shares, or go to zero.
08
08 · The strategist
Connecting the dots
Short version
Four forces push rare-earth demand up. Two push the other way, and AI powers those two as well.
Weak in magnet tonnes (data centres ≈ 1–3% of demand) but strong for yttrium. It’s strongest as a mood link: these stocks fall when AI spending looks shaky.
2
Drones & defence → heavy rare earths
strong · now
AI makes cheap drones decisive
Militaries must buy outside China, at any price
Dysprosium · terbium · samarium magnets
Lynas · MP · Energy Fuels → Vulcan
The most real near-term AI link. Heavy rare earths sell for 10×+ more outside China than inside it.
3
Humanoid robots → magnets in every joint
weak now · strong after 2035
Smarter robot brains (OpenAI, Tesla, Chinese labs)
500k–900k robots/yr by 2030 (base)
~3,200 t NdFeB ≈ 1% of supply
China Northern first; MP/Neo only with a Tesla deal
The twist: 97% of today’s humanoids are Chinese, so robot demand lands on China’s magnet makers first. No Tesla magnet deal exists with any listed company.
4
SpaceX, Starship & space data centres
weak in tonnes · 2028+
SpaceX + xAI at ~$2T
Starship + up to 1M satellites
Samarium-cobalt thrusters · NdFeB wheels
Nobody yet (no SpaceX offtake)
Small in tonnes. The real option: one Musk-company magnet contract could re-rate a single stock. That’s a lottery ticket, not a thesis.
5
AI cracks separation chemistry
pushes against · 2027–30
AI + robot labs design better extractants
Fewer steps, fewer workers, cheaper plants
China’s know-how edge shrinks
Helps USAR’s volumes; hurts scarcity pricing
Great for supply security, bad for the premium these stocks trade on. Watch the Saskatchewan AI-run plant (target Dec 2026).
Funny side effect: the easiest uses to switch are data-centre fans and pumps, the “AI” ones.
Why it matters: the heavy rare earths are both the strongest near-term bet (drones, defence) and the most exposed long-term (AI chemistry plus substitute magnets). That’s the central paradox.
09
09 · Second-order thinking
Things most people miss
Short version
Most people think “AI boom = rare earth boom.” The real drivers are China, interest rates, and which step of the chain a company owns.
Peace is bearish
Every US–China thaw has knocked these stocks down 4–20% in a day40 41. A real trade deal would hurt them.
They trade like chip stocks
On 18 Aug, an AI-spending scare dropped them alongside semiconductors. In July they fell 26% while NdPr prices rose. Sentiment beats the metal price.
The middle is the moat
China has ~60% of mining but 91% of separating and 94% of magnets5. Owning a mine is the easy part.
AI makes its own mine
Old data-centre hard drives are a concentrated source of magnet metals. Cyclic Materials (backed by Microsoft) recycles them37, and MP’s Apple deal runs on recycled feed.
MP’s floor is insurance
The Pentagon pays MP the gap when NdPr is under $110. So the floor is worth the most in the worlds that hurt everyone else: an AI stall or a China deal.
Elections matter
US midterms are on 3 Nov58. Signed Pentagon contracts tend to survive. Newer deals still under congressional questioning (like USA Rare Earth’s) are less certain.
Why it matters: if you want to bet on AI, these are weak proxies before 2030. If you want to bet on China staying tough, they’re strong ones.
10
10 · The verdict
The ranking
Short version
Lynas and MP lead because they’re already producing, have price protection, and survive even the bad scenarios. The Pre-revenueA company that isn’t selling its main product yet. It lives on investors’ money until it does. juniors are lottery tickets, and two look likely to need far more money than they have.
“Age of AI” score, out of 10
Weighted score across 7 factors (table below). Colour = tier.
Tier 1 · core
Tier 2 · watch
Tier 3 · speculative
Tier 4 · avoid
Benchmark
Company the article’s scoreArticle scoreTier
LYCLynas6.7Tier 1 · core
MPMP Materials6.6Tier 1 · core
600111China Northern6.5Tier 2 · watch
USARUSA Rare Earth5.3Tier 3 · speculative
ILUIluka5.1Tier 2 · watch
NEONeo Performance5.1Tier 2 · watch
UUUUEnergy Fuels4.9Tier 2 · watch
REMXREMX (ETF)4.9Benchmark
ARAAclara3.9Tier 3 · speculative
ARUArafura3.7Tier 3 · speculative
UCUUcore3.6Tier 3 · speculative
CRMLCritical Metals2.2Tier 4 · avoid
NBNioCorp2.2Tier 4 · avoid
Show every factor score
Company
AI demand
Survival
Execution
Policy
Heavy RE
Value
AI ops
Score
Tier
Lynas
6
9
7
8
9
4
1
6.7
Tier 1 · core
MP Materials
8
7
7
10
5
3
2
6.6
Tier 1 · core
China Northern
6
9
9
6
3
5
4
6.5
Tier 2 · watch
USA Rare Earth
6
4
5
7
7
2
6
5.3
Tier 3 · speculative
Iluka
3
7
5
7
7
6
0
5.1
Tier 2 · watch
Neo Performance
5
7
5
4
4
8
1
5.1
Tier 2 · watch
Energy Fuels
5
7
5
5
7
3
0
4.9
Tier 2 · watch
REMX (ETF)
4
8
6
5
3
4
1
4.9
Benchmark
Aclara
3
4
2
5
9
5
0
3.9
Tier 3 · speculative
Arafura
3
4
3
7
3
5
0
3.7
Tier 3 · speculative
Ucore
4
3
2
5
5
5
1
3.6
Tier 3 · speculative
Critical Metals
2
2
1
3
7
1
0
2.2
Tier 4 · avoid
NioCorp
3
1
1
4
3
3
0
2.2
Tier 4 · avoid
Weights: AI-demand 20%, Survival 20%, Execution 15%, Policy shield 15%, Heavy-rare-earth exposure 10%, Valuation 10%, AI in operations 10%. Scores are the strategist’s judgement after the fact-check corrections, and the composites were re-computed and checked. Fact-check changes include MP Survival 8→7 (cash and equivalents fell to $429M), USA Rare Earth Survival 5→4 (34% dilution, more raises due), and Neo Execution 7→5 (its EU magnet plant is still shipping samples). A Survival score of 4 or less caps a company at Tier 3.
Tier 1 · core
LYCLynas · 6.7.
Profitable (A$222M), A$1.2B cash, the only big heavy-rare-earth separator outside China, with price floors from Japan and the US. Pricey at ~70× earnings, and it needs a new CEO.
MPMP Materials · 6.6.
Deepest government backing ($110 floor, Pentagon stake) and the best AI-era customers (Apple, GM, drones). Burning cash on its 10X magnet plant (2028); loses money on a GAAP basis.
Tier 2 · watch
600111China Northern · 6.5.
Owns the supply chain behind most of the world’s robots; profits up 120%. Wins if the US and China make a deal. Hard for US investors to buy, with state-control and sanctions risk.
ILUIluka · 5.1.
Mineral-sands cash plus a government-funded heavy-rare-earth refinery (mid-2027). Little AI link.
NEONeo Performance · 5.1.
Cheapest company that already earns money (~8× EBITDA). Builds magnets in Estonia, still at samples. No price floor.
UUUUEnergy Fuels · 4.9.
Uranium pays for a heavy-rare-earth build that supplies Vulcan’s drone magnets. Pricey at ~24× sales.
Tier 3 · speculative
USARUSA Rare Earth · 5.3.
Highest AI leverage of the listed names (AI chemistry lab, magnet plant, Brazil heavy-rare-earth mine). 34% dilution and more share sales coming. Scores 8 in Takeoff but 2 in Stall.
ARAAclara · 3.9.
Purest dysprosium/terbium bet, but no binding financing or customers yet, and substitute magnets target exactly its niche.
ARUArafura · 3.7.
Fully approved with government backers, but light rare earths only and exposed to spot prices.
UCUUcore · 3.6.
New separation tech, unproven at scale; keeps selling shares.
Tier 4 · avoid
CRMLCritical Metals · 2.2.
Arctic mine still at the study stage; ~850× sales; heavily shorted.
NBNioCorp · 2.2.
Needs about $1.4B more than it has.
How the top names do in each version of 2030
Score out of 10 under each scenario from section 02
Takeoff (15%)
Steady (60%)
Stall (25%)
0246810
LYCLynas876
MPMP Materials975
600111China Northern864
USARUSA Rare Earth852
NEONeo Performance764
ILUIluka654
UUUUEnergy Fuels654
Show all 12 companies
Company
Takeoff
Steady
Stall
Lynas
8
7
6
MP Materials
9
7
5
China Northern
8
6
4
USA Rare Earth
8
5
2
Iluka
6
5
4
Neo Performance
7
6
4
Energy Fuels
6
5
4
REMX (ETF)
6
5
3
Aclara
5
4
2
Arafura
5
4
2
Ucore
5
3
1
Critical Metals
4
2
1
NioCorp
3
2
1
What we found
Most likely world: “Steady AI + China keeps muddling along” (~33%). AI demand is real but small in tonnes, the heavy rare earths stay tight, and NdPr sits around $90–110. Floor-backed producers win.
Biggest spread: USA Rare Earth scores 8 if robots take off but 2 if AI stalls. It’s the highest-risk, highest-AI-leverage listed name.
The hedge: China Northern wins if there’s a US–China deal and supplies the chain behind 97% of humanoids. But it’s a Chinese state company that’s hard for US investors to buy.
Why it matters: in five sentences: AI needs magnets, but not many before 2030. For now these stocks move on China. The safest bets are already producing and have price guarantees. Pre-revenue juniors could soar or be wiped out by new share sales. AI is also making substitutes, so the scarcity story won’t last forever.
11
11 · Keep score
What would change our mind
Short version
The next 90 days have five big dates. Each one can move a company up or down a tier.
Dates to watch
Tesla Q3 letter
Is Optimus V3 really in production? Any magnet supplier named outside China?
Lynas September-quarter report
Heavy-rare-earth output and ore problems; any CEO news.
Fed meeting
Markets priced ~73% odds of another hike. Bad for companies still building.
US midterm elections
Congress oversight of newer government deals (USA Rare Earth).
MP Q3 results (date TBA)
GM magnet shipments started? Cash vs the $429M? 10X still on track for 2028?
USA Rare Earth Q3
Stillwater magnet plant at 600 t/yr? Size of the next share sale.
China’s export-rule pause ends (on paper)
Extended, lapsed or eased. The single biggest date for the whole group.
Saskatchewan AI-run separation plant
If it works, China’s know-how edge shrinks (chain 5).
US–China truce expires
Escalation lifts Western names; a real deal hits them 10–20%.
Niron + Iluka + Energy Fuels milestones
First rare-earth-free magnet plant; new heavy-rare-earth refineries.
This view is wrong if…
NdPr drops below ~$75/kg for a month. That looks like China crushing rivals on price, Molycorp-style. Juniors move to Tier 4.
Ex-China dysprosium falls below ~$1,000/kg. The heavy-rare-earth premium collapses, which hurts Aclara, Critical Metals and part of Lynas’s edge.
Tesla or SpaceX signs a non-China magnet deal with any listed name. That name moves up a tier.
Niron signs a volume contract with a carmaker or data-centre maker. Substitution arrives early.
The 10-year Treasury yield falls back below 4.5%. That’s the biggest relief for the pre-revenue builders.
Why it matters: a good thesis comes with its own “I was wrong if…” list. Check these, not the headlines.
12
12 · Principles
How to think about it
Short version
Size by survival, not by upside. These are principles, not buy/sell advice.
1
Two clocks
The story plays out over 2028–2035, but prices swing on 1–3-month China headlines. Expect the short clock to shake you.
2
Survival first
Most weight belongs with companies that survive the worst case. Speculative names should be small enough that a zero doesn’t hurt.
3
Expect big drops
Even the winners fell about 50% in under a year. Buying in stages beats buying all at once.
4
Mind the dates
Don’t add right before 10 Nov, the midterms, earnings or 10 Jan. Spikes on headlines are moments to rebalance, not proof you’re right.
5
One bet, many tickers
All the Western names are really one bet on China staying tough. Owning five of them isn’t diversification.
6
Receipts only
Move a company up only on a signed contract, a working plant or a closed loan. “Partnership” press releases don’t count.
Glossary
NdPr
Neodymium + praseodymium oxide. The main ingredient in strong magnets, priced per kilogram.
Heavy rare earths
Dysprosium, terbium, yttrium and friends. Rarer, pricier, and almost all processed in China.
Separation
Splitting the 17 rare earths apart. It takes hundreds of chemical steps and is China’s biggest advantage.
Price floor
A promise (here from a government) to pay at least a set price, even if the market drops.
Dilution
When a company sells new shares to raise money, so each existing share owns a smaller slice.
Pre-revenue
A company that isn’t selling its main product yet. It lives on investors’ money until it does.
Run-rate
One month’s revenue × 12. A fast-growth snapshot, not a full year of actual sales.
ETF
A fund that holds a basket of stocks and trades like one stock (REMX is the rare-earth one).
How this was made
Method and limits
9 agents: 1 tool check, then 6 specialists working in parallel (fundamentals analyst, materials engineer, geopolitics watcher, frontier-AI tracker, AI-usage auditor, short seller), a strategist who connected the dots, and an adversarial fact-checker who tested 19 load-bearing claims.
Fact-check results: 13 confirmed, 3 partly wrong or imprecise (all corrected here), 5 unconfirmed (flagged in the text), 0 made up.
Limits: finance websites blocked direct page reads, so most numbers come from search summaries of the cited pages. No October 2026 share price could be found for any company. Scores and probabilities are judgement calls built from evidence, not measurements.