“Sell shovels in a gold rush” is the oldest advice in investing — which is exactly why the obvious shovels (NVIDIA, TSMC, the cooling names) are already crowded. So we asked the question almost nobody asks: who do the shovel-makers themselves depend on? One layer down lives a handful of companies the entire AI, robotics and space buildout quietly cannot function without. This is the map. Hover or tap any underlined term.
Legend: Leader established, real revenue · Emerging small-cap / early · Private not yet public · Hype narrative running ahead of revenue · SPOF single point of failure — one supplier, one process, no substitute · CONCENTRATED a tight oligopoly or duopoly: few suppliers, but more than one. We tagged several of the second kind as the first until 21 September 2026, which made the map look more fragile than it is. A badge sitting next to three named alternatives is not a single point of failure.
We were using one word for two different things, across several pages, and an outside review caught it in September 2026. The distinction we now hold to everywhere:
BINDING — what limits the next megawatt actually energised this year. Today that is power delivery: grid transformers averaging about 128 weeks — roughly 2.5 years, with four years the tail for the largest custom units — plus interconnect queues, substations, and the trades who build them. A binding constraint is slow, not rare, and it sets the pace of the buildout.
FRAGILE — a single point of failure: one supplier, one process, no substitute. EUV lithography (one maker on Earth), EUV photomask inspection, rare-earth separation. A fragile node is rare, not slow, and it sets the tail risk. Everything tagged SPOF on this page is of this kind.
CONCENTRATED — few suppliers, but more than one: advanced substrates, HBM, dielectric fluids, precision reducers. These are usually capacity-constrained, which puts them closer to binding than to fragile.
We got this wrong inside this very box, on the day we wrote it. The first draft defined “binding” using transformers on four-year lead times — the tail figure we had corrected to a 128-week average on the same morning, in the same commit — and listed HBM and advanced substrates as fragile while simultaneously demoting Ibiden to CONCENTRATED for having three named peers. A reviewer caught both on 21 September 2026. We are leaving this paragraph here because a rule that cannot survive its own author is worth less than one that admits where it failed.
Both claims were true; printing them as the same claim was not. They are different questions with different trades: you own the binding constraint for the grind, and you watch the fragile node for the break.
When a trade gets crowded, the edge isn't in the famous name — it's in the unglamorous supplier the famous name can't live without. Everyone knows NVIDIA makes the GPU. Far fewer know that every advanced GPU needs a special substrate from a tiny club of Japanese makers, must be stacked with HBM that only three firms produce, and must be tested on machines from a tiny club of testers. That's where the scarcity — and the pricing power — actually lives.
This is the most-watched layer, so the edge here is narrow — but you have to see it to find what's beneath it. The famous names, then the ones hiding behind them:
| Role | The known names | The chokepoint beneath (less-watched) |
|---|---|---|
| Make the chip | TSMC, Samsung Leader | ASML SPOF — the only maker of EUV machines on Earth. |
| Inspect the masks | — | Lasertec SPOF — near-monopoly on EUV photomask inspection. A tool the whole industry needs and almost nobody names. |
| Cut & stack the wafers | — | Disco (dicing/grinding — gates HBM and advanced packaging), Advantest (HBM/memory-test leader, with Teradyne), Besi (hybrid bonding; AMAT owns a ~9% stake). |
| The board under the chip | — | Ibiden CONCENTRATED — the “hidden CoWoS,” an advanced-substrate maker most investors have never heard of; Shinko, Unimicron, AT&S alongside. Named peers exist, so this is a capacity constraint rather than a single point of failure. |
| The raw wafer & the resist | — | Shin-Etsu / SUMCO (the two largest of ~five wafer makers), JSR/Inpria + TOK + Shin-Etsu CONCENTRATED (EUV photoresist — three named firms, but Japan ~95%, so the geography is SPOF even though the makers are not) — the chemicals without which the machines print nothing. |
A GPU with nothing to feed it is an expensive paperweight. The plumbing that moves data is its own bottleneck stack — and this is where NVIDIA itself has been quietly buying insurance.
| Role | Names | The tell |
|---|---|---|
| HBM memory | SK Hynix (#1), Micron (MU — cleanest US-listed), Samsung Leader | Reportedly sold out into 2027. Final test led by Advantest and Teradyne. |
| Optical interconnect | Broadcom (AVGO), Marvell (MRVL), Coherent (COHR), Lumentum (LITE) Leader | NVIDIA took ~$2B stakes in both COHR and LITE — a flashing arrow at where it sees fragility. |
| The networking shovel | Amphenol (APH) Leader, Arista (ANET), Fabrinet (FN) | APH = arguably the best risk-adjusted “shovel” here — the connector spine, diversified, less hype. |
| The deepest laser part | — | InP wafers SPOF — gates every optical transceiver. The pick under the optical picks. |
| Frontier optics (private) | Ayar Labs Private (backed by NVIDIA + AMD + Intel — the rare trifecta), Eliyan (AMD/Arm/Meta/Coherent), Lightmatter Private | The trifecta cap table is the thesis — three rivals all de-risking the same part. |
The electricity and heat layer is so large it has its own deep dives. The short version, with the verdict that surprised us:
This is the least-watched, most-chokepoint-dense layer of all — and the one with the clearest single message: the bottleneck is almost never the mine or the robot. It's the refining, the gear, and the magnet — all of them fragile rather than binding, in the sense set out above.
The world has plenty of rare-earth ore. What it lacks is the dirty, capital-intensive step of separation and refining, where China runs ~90% of capacity (and ~98% of gallium, a chip material). Which is why Washington is now treating the Western Hemisphere's own resources as a strategic asset — country by country, in the Americas trade.
| Name | Flag | The role & the catch |
|---|---|---|
| MP Materials (MP) | Leader | The US rare-earth flagship; DoD is its largest holder. But priced for perfection — insiders selling. |
| Lynas (LYC) | SPOF | The only at-scale heavy-rare-earth separator outside China. Structurally strategic. |
| Centrus (LEU) | SPOF | HALEU nuclear-fuel enrichment — gates the small reactors meant to power AI. Russia is the only other commercial source. |
| Sunresin (China) | SPOF | The deep one: makes the chelating resin behind ~90% of gallium extraction. A pick under the pick under the chip. |
| Niron Magnetics | Private | Rare-earth-free magnets — backed by GM, Stellantis, Samsung. If it works, it routes around the whole China magnet chokepoint. The kind of solver we watch for. |
Also on the radar: Freeport (FCX) for copper (real cash flow, rare here), and a cluster of emerging US names — USAR, Energy Fuels (UUUU), Perpetua (PPTA, antimony), NioCorp — that are mostly pre-revenue policy plays. Solvers: Redwood Materials (battery urban-mining, founded by ex-Tesla CTO JB Straubel), Lilac (lithium), Phoenix Tailings.
The humanoid-robot OEMs get the headlines and the hype. But every one of them needs precision parts from a tiny set of suppliers — and that's where the durable scarcity is.
| The part | Who makes it | Why it's the real pick |
|---|---|---|
| Precision reducers (the joint gearbox) | Harmonic Drive (~strain-wave leader), Nabtesco CONCENTRATED | Two firms dominate the gears every robot arm needs. Hard to replicate, decades of know-how. |
| Roller screws (linear muscle) | GSA + Rollvis — Swiss-built, now Schaeffler-owned CONCENTRATED | Even tighter than reducers. A constraint most investors have never heard of (and the Swiss duopoly has just been rolled up by Germany's Schaeffler). |
| NdFeB magnets | China-dominated SPOF | Already delayed Tesla's Optimus. The magnet is the fragile link, not the AI — China-dominated separation, which is why it keeps its SPOF tag. |
| The OEMs themselves | Figure Hype ($39B+ pre-revenue), Tesla Optimus, Agility, Unitree | Best-funded OEMs build actuators in-house — a headwind for merchant suppliers, so read each carefully. |
Same pattern a third time. Not the rocket, not the satellite — the obscure component the satellite can't fly without.
| The deep part | Who / status | The tell |
|---|---|---|
| Rad-hard chips (survive space radiation) | BAE, Honeywell, Microchip, Frontgrade CONCENTRATED | Only ~3 US “Trusted Foundry” sites, none new since the 1990s. Four named suppliers across three sites is a tight oligopoly, not a single point of failure — the site count is the constraint, not the vendor. |
| Reaction wheels / star-trackers | Blue Canyon (owned by RTX) CONCENTRATED | Capacity expanding +400% — a constraint that is loosening, which is the opposite of fragile; — the clearest “demand is real” signal in the sector. |
| Launch & imaging | Rocket Lab (RKLB), Planet (PL — rare positive cash flow) Leader | RKLB's Neutron rocket still unflown (debut targeted Q4 2026) — execution risk priced as certainty. |
| The hype tier | AST SpaceMobile (ASTS), and a 2021–24 SPAC graveyard (Virgin Orbit, Momentus, Terran) Hype | The cautionary pattern: story-stocks that ran out of runway before revenue. |
Strip everything back and a handful of single points of failure sit beneath multiple industries at once. These are the true “picks behind the picks” — the places where one disruption cascades across the whole map. This is the list worth taping to the wall:
A map of chokepoints is dangerous in the wrong hands — it's a list of exciting stories, and exciting stories are where retail money goes to die. Three rules keep it honest:
We're building the full version of this into its own piece — how to read a private or emerging company's true integrity, backers and trajectory before Wall Street can (the “x-ray” framework). It's the natural companion to this map: this tells you where to look; that tells you how to judge what you find.
The honest risk with a map like this is falling in love with the supply side. A few deliberate blind-spot checks we keep running:
What does “picks and shovels” mean in AI investing? It's the gold-rush idea: instead of betting on who strikes gold (the winning AI model), you sell the tools everyone needs to dig — chips, power, cooling, memory. This piece goes one step further: to the suppliers the tool-makers themselves depend on.
Which companies supply the components for AI data centers? Layer by layer: chips (TSMC, ASML, Lasertec), memory (SK Hynix, Micron, Advantest), networking & optics (Broadcom, Amphenol, Coherent), power & cooling (Eaton, Vertiv, nVent), and the raw materials beneath all of it. The tables above flag each name leader / emerging / hype.
What's the difference between mining and refining rare earths — and why does it matter? Mining digs up the ore; refining/separation pulls the individual elements apart. The world has plenty of ore — the bottleneck is the refining, and China runs the large majority of it. That's why the real chokepoint is the refinery, not the mine.
Dragonfly Lens follows every trade one layer deeper than the headline — to where the real scarcity, and the real risk, actually sit. Plain English, every claim flagged and sourced.
Join the Lens →A serious honest note — read this: this page is a research map condensed from multiple deep-dive passes, not a set of recommendations and not a buy list. Market shares, cap-table backers, and “sole supplier” claims change — treat every specific figure here as a starting point to verify yourself, not as confirmed fact. Many names are private, pre-revenue, or thinly traded; some will fail. The flags (Hype, SPOF) are our editorial read, not ratings. Strategic-investor backing reduces the odds a company dies; it does not make the stock go up. Nothing here is personalized investment advice — Dragonfly Lens is not a registered investment advisor. Tickers illustrate the supply chain, not buy recommendations. Do your own work, size for the chance you're wrong, and never confuse an interesting story with a good investment. Facts as of June 2026.