Imagine the world is building 100 hospitals at once. Most of the news, most of the money, and most of the stock chatter is about which hospital chains will be biggest, which doctors will be most famous, and what the fanciest treatments will be.
Meanwhile, in a small town in the Netherlands, there's one factory that makes the MRI scanners. There's two factories on Earth that polish the lenses. There's one Italian foundry that casts the special steel beams. There's three companies that make the giant generators every hospital needs. And the export licenses for the rare metals inside the wires? Those expire on a specific Tuesday in late November 2026.
If you own a piece of those small factories, you don't care which hospital wins. You're getting paid by all of them, and the buildout is happening whether the headlines are bullish on AI or panicking that day. That's the narrow places . That's what every stock on this page is.
Before there's a chip, there's a rock. China owns most of the rocks.
Every AI chip eventually traces back to a handful of raw materials: rare-earth metals (used in magnets, lasers, and high-tech alloys), indium phosphide (the special crystal needed to make the lasers that move data inside data centers), silicon-on-insulator wafers (the foundation every silicon-photonics chip sits on top of), plus antimony and germanium (military-grade strategic minerals). China dominates most of these. The handful of Western companies that can produce them are the absolute first link in the AI supply chain — and the most important date on the calendar that nobody is watching is November 27, 2026 , when China's temporary suspension of export controls on antimony, gallium, and germanium expires.
If the rocks stop flowing, nothing downstream moves. There is no AI without these nine companies — full stop.
Indium phosphide crystal grower — the laser-material monopoly
A small U.S. company that grows the indium phosphide crystals every AI laser is built from . Outside China, there are essentially two suppliers — AXT is one of them. NVIDIA's $4 billion in 2026 InP supply commitments effectively flow through this company. Source thesis calls it 'The Strait of AXTi v2.0.'
RF + photonics + InP photodetectors
MACOM makes the tiny detector chips that catch the light coming out of fiber-optic cables . They also took a strategic stake in IQE, which makes the wafers those detectors are built on. They sit one layer downstream from AXTI in the same supply chain.
Silicon-on-insulator wafers — the photonics foundation
Soitec is the French company that makes the special silicon wafers every silicon-photonics chip is built on top of . As AI shifts from copper to light, every photonics chip needs one of these as the foundation. OTC listing — less liquid for U.S. brokerages.
Mountain Pass — the only operating U.S. rare earths mine
Operates the only meaningful rare-earth mine and refinery in the U.S. — Mountain Pass in California. Building a magnet factory in Texas. Has Department of Defense investment and Apple as a customer. If China cuts off rare earth exports, this becomes a national-security asset overnight.
Ex-China heavy rare-earth metallization
Tiny pure-play on the step that takes raw rare earth oxides and turns them into useful metal alloys . Pentagon contracts in hand. The downstream cousin to MP. OTC — small-cap, illiquid, speculative.
Only U.S. antimony source — Stibnite Mine, Idaho
Reopening the Stibnite Gold Mine in Idaho — the only meaningful U.S. source of antimony (a strategic mineral used in munitions, semiconductors, and batteries; China controls 48% of supply). Department of Defense already gave them $24.8M, and the U.S. Export-Import Bank wrote a letter about a $1.8B loan. After China's antimony ban, the stock jumped 19%. Production isn't until 2028, so the trade is the news cycle, not the cash flow.
Texas rare earths + magnets
Building a U.S. heavy-rare-earth and magnet operation in Texas — a potential second domestic supplier to MP Materials. Pre-meaningful-production. If U.S. policy continues subsidizing domestic rare earth production, smaller names like this can re-rate violently. Higher dilution risk than MP.
Uranium + rare earth processing — two-shot bet
Energy Fuels operates the only U.S. mill that can actually refine rare-earth concentrates into the separated oxides downstream customers need — and they produce uranium on the side. Two simultaneous bets: critical minerals AND nuclear fuel. Rare combo of optionality. Mill ramp execution is the swing factor.
Canadian germanium recovery from zinc
Major Canadian miner that recovers germanium (used in fiber optics and infrared sensors) as a byproduct of zinc smelting . China controls 60% of germanium supply. If Chinese export controls re-impose on Nov 27, 2026, Teck's tiny germanium output becomes strategically priceless. Most of revenue is base metals though — germanium is rounding-error revenue. Move requires the China event.
Designing a chip is the easy part. Printing it is where the bottleneck lives.
Once you have raw silicon and rare materials, you need a $15-20 billion factory ('foundry') and a fleet of giant machines to print the chip onto a wafer. The lithography step alone uses a $400 million machine that only one company on Earth makes . Then the wafer goes through hundreds more steps — etching tiny trenches, depositing thin films, inspecting for defects. The companies that make these machines (ASML, Applied Materials, Lam, KLA) and the foundry that runs them (TSMC) are the most consolidated chokepoint in the entire chain.
The whole AI buildout runs at the speed of the slowest fab tool. There are no second sources for any of these five names.
EUV lithography — the absolute monopoly
ASML is the Dutch company that makes the $400 million machines required to print modern chips . There is exactly one supplier on Earth. Every Rubin GPU, every HBM4 die, every advanced AI chip passes through an ASML machine. The original 'narrow place' of the entire AI thesis.
The world's foundry — where the chips are actually printed
TSMC actually manufactures the chips that NVIDIA, AMD, Apple, and Broadcom design. They control roughly 70% of the world's leading-edge chip output and run the entire CoWoS advanced-packaging pipeline that NVIDIA's GPUs require. There is no second source.
Etch + deposition for HBM and 3D NAND
Lam makes the machines that carve out the tiny vertical tunnels (called 'TSVs') that connect stacked memory chips . Every layer added to a memory tower means more passes through Lam's machines. Reportedly also a suitor for BESI.
Process control — the chip industry's quality cop
KLA makes the inspection machines that check chips for defects at every step of manufacturing. As packages get more complex (3D, hybrid bonded), defect rates explode and KLA tools become more critical. Near-monopoly in process control.
Largest U.S. chip-equipment maker
The biggest U.S. chip-equipment company. They own 9% of BESI (the hybrid-bonder monopoly) and partner with them on the integrated 'Kinex' system. You're not just buying equipment — you're buying optionality on the BESI relationship and leadership in the next inflection (Gate-All-Around transistors, HBM, 3D-DRAM).
Glass-substrate laser drilling — niche monopoly
German company specializing in the laser systems used to drill holes through glass substrates — the next-generation replacement for the organic substrates in today's AI chip packages. Glass substrates promise better thermal performance and tighter chip stacking. Intel, TSMC, and other foundries are all moving toward glass for next-gen packaging. Called out by @aleabitoreddit before its run. Small-cap European, illiquid for U.S. retail.
These are the companies that decide what the AI brain looks like.
The chip designers don't make the chips themselves — they design them, then send the design to TSMC. NVIDIA dominates the training market (the part where AI models learn from data). Broadcom and Marvell own the custom-silicon market (where Google, Meta, and Amazon get their own bespoke chips). AMD is the credible challenger. Intel is the strategic U.S. backstop. Arm collects a royalty on essentially every advanced chip shipped. Six companies design the brain that powers the entire AI economy.
Every dollar spent on AI hardware lands here first — but consensus owns NVIDIA already. The asymmetric edge is in the custom-silicon names around it.
The default AI brain — GPU king
NVIDIA designs the GPUs that train and run nearly every major AI model . CUDA software lock-in keeps developers captive. The source's note is honest: this is the most-owned, most-talked-about AI stock on Earth — own it as an anchor, but the asymmetric upside is in the smaller chokepoint names around it.
Custom-ASIC duopoly — Tomahawk + Jericho networking
Broadcom designs custom chips for hyperscalers (Google's TPUs, Meta's MTIA, etc.) and dominates AI networking switches. When Google says 'we don't need NVIDIA,' the chip they're using was probably designed with Broadcom. Half the AI router market lives here.
The other custom-ASIC house
Marvell is Broadcom's main competitor in custom AI silicon . Has 18 active hyperscaler programs in the pipeline — these are multi-year, multi-hundred-million-dollar engagements. Every program that ramps to production is a step-function for revenue.
Only credible merchant GPU alternative to NVIDIA
AMD's MI series GPUs are the only realistic 'plan B' if you don't want to buy NVIDIA . Hyperscalers are increasingly using them as a pricing lever and a hedge. Software stack still trails CUDA, but it's improving fast and the price-performance is real.
U.S. government's de facto semi arm
Intel is now strategically backed by the U.S. government as the only domestic leading-edge fab. The bullish twist: AI inference (running models, not training them) increasingly happens on CPUs, and Intel just flipped the AI server CPU-to-GPU ratio from 1:8 to 1:1 — meaning more CPUs per server, more revenue.
Royalty model — riding the custom-silicon wave
Arm licenses chip designs and collects a royalty on every chip shipped . Every custom AI chip from Broadcom and Marvell, every smartphone, and increasingly every AI server CPU pays Arm. As edge inference grows, so does the toll.
Modern AI is memory-starved. The bottleneck isn't compute anymore — it's feeding the compute.
Every NVIDIA GPU has a tower of memory chips glued right next to it . This is called HBM (high-bandwidth memory). Without enough HBM, the world's GPUs literally sit idle waiting for data to arrive. Three companies make HBM — Micron (US), Samsung, SK Hynix (both Korea). Around them sit a layer of specialty toolmakers (Camtek, Onto for inspection; Entegris for materials; TOWA for sealing) and storage companies (SanDisk for SSDs that buffer the data). Six names own this whole layer.
Wall Street obsesses over GPU shipments. The actual rate-limiter on AI training is HBM volume, and HBM volume runs through these six companies.
3D metrology — inspecting every HBM stack
Camtek makes the inspection machines that measure every layer of a stacked memory tower (HBM) down to the nanometer. The source thesis is direct: 'tool of reference at every HBM/CoWoS player.' Every memory stack at every leading manufacturer goes through a Camtek tool.
Only Western HBM maker with U.S. capacity
Micron is the only American HBM maker (Samsung and SK Hynix are Korean). HBM is the towers of memory glued onto every GPU — and there's a structural shortage. Micron's HBM3E is qualified at NVIDIA. The anchor of the whole memory supercycle thesis.
3D imaging tool — sole-qualified at HBM leaders
Onto's 3D imaging tools are the only ones qualified at two of the three HBM makers . As HBM stacks get taller (HBM3 → HBM4 → HBM4E), inspection time goes up and Onto's revenue scales linearly. Direct beneficiary of the memory supercycle.
Highest-beta NAND/SSD play for AI
SanDisk makes NAND flash and enterprise SSDs — the medium-fast storage layer right next to the GPU. AI training generates massive intermediate data ('checkpoints') that gets dumped to SSDs constantly. Highest-beta way to play the AI storage cycle.
HBM compression molding — sleeper monopoly
Tiny Japanese company with a functional monopoly on the molding step that seals every HBM stack . Almost nobody outside the industry knows this name. Pure-play on HBM volume — every memory stack the industry ships pays a small toll here. OTC.
Silent picks-and-shovels for every fab
Entegris supplies the ultra-pure chemicals, gases, and filtration systems every chip fab needs — including the HBM lines. Boring, defensive, and impossible to replicate without years of qualification. Quiet beneficiary of the entire fab buildout.
Modern AI chips aren't single chips — they're stacks of chips fused into one package.
To make AI chips faster, manufacturers stopped just shrinking transistors (that got too hard). Instead, they stack multiple chips on top of each other and fuse them with copper threads finer than a human hair. The machines that do that fusion (hybrid bonders) come from essentially one Dutch company . Around it sits a small group of specialty equipment names — wafer-level burn-in (Aehr), thermo-compression bonding (Kulicke & Soffa), 1.6T linear pluggable optics (AAOI). These are the 'outlier' picks: niche but irreplaceable.
TSMC can have all the leading-edge silicon they want. If they can't get bonders, they can't ship NVIDIA's next chip.
The hybrid bonder duopoly — fusing chips at the nanometer
Dutch company that makes the machines that fuse two chips together with nanometer accuracy (called 'hybrid bonders'). Required to build the most advanced AI chips. Q1 2026 orders €269.7M — a record. Applied Materials owns 9% of them and may try to buy them outright. ADR is on the pink sheets, which is why most American investors have never bought it.
1.6T LPO — single hyperscaler order play
AAOI is positioned to be one of the very few suppliers of next-gen 1.6T 'linear pluggable optics' (LPO) — a cheaper, lower-power alternative to traditional transceivers. A single $200M+ hyperscaler order would re-rate the stock. Lumpy, binary, but huge optionality.
Wafer-level burn-in monopoly
Aehr makes the only volume machines that 'burn in' chips at the wafer level — running them hot to weed out defective ones before they're cut and packaged. Critical for SiC power chips and increasingly for high-reliability AI silicon. Tiny float, niche monopoly.
TCB transition + wire-bond cash cow
Kulicke & Soffa runs the legacy wire-bond business that funds their next-gen 'TCB' (thermo-compression bonding) machines — the alternative path to BESI's hybrid bonders for connecting stacked chips. Either TCB takes share or it doesn't, but the wire-bond cash flow underwrites the bet.
Inside an AI computer, data moves between chips on beams of light.
Modern AI uses so many chips at once that copper wires can't carry the data fast enough. So everything inside an AI data center is now connected with tiny lasers and fiber-optic cables . Every laser needs a special crystal called indium phosphide. There are roughly two companies outside China that make those lasers (Lumentum and Coherent). NVIDIA put $4 billion into them in March 2026. The next inflection — 'co-packaged optics' (CPO) — moves the laser right inside the chip package, and a fresh wave of names (Sivers, POET, Fabrinet, Ciena) sits in that stream.
You can build a million GPUs, but if you can't connect them with light, you have a million paperweights.
Vertically integrated InP + 1.6T/3.2T + CPO stack
Coherent makes the lasers, packages them into transceivers, and is one of the few suppliers building the next-generation 'co-packaged optics' (CPO) systems — where the laser sits right next to the chip. NVIDIA gave them $2 billion in March 2026 to lock in supply. They own the entire stack from indium phosphide crystal to finished transceiver.
EML laser supply — the structural shortage
The biggest non-Chinese maker of the tiny lasers (called 'EMLs') that go inside fiber-optic cables . The market is structurally 25-30% undersupplied. Their 'OCS' (optical circuit switching) backlog is over $400M. NVIDIA invested $2B in March 2026 to lock them in for 32 months.
Optical contract manufacturer — the assembly floor
Thai-based factory that assembles the finished optical transceivers for Lumentum, Coherent, and most of the industry. Doesn't matter who wins the laser war — Fabrinet builds it. Q2 FY26 revenue hit $1.13B, up 38% YoY.
WaveLogic 6 — 1.6T coherent optics for cloud
Ciena's WaveLogic 6 Extreme is one of the only 1.6T coherent optics platforms shipping . Cloud-direct revenue up 76% YoY. As hyperscalers connect AI campuses to each other across cities, this is the layer that does it.
Independent CW laser supplier — the overflow choke
Tiny Swedish company that makes the continuous-wave lasers used in silicon photonics modules . Sits as the 'overflow' supplier when AMD, Meta, Ayar Labs, POET, and Lightmatter all need lasers and the big two (Lumentum, Coherent) are sold out. OTC.
CPO engine outlier — small-cap optionality
Small Canadian company building 'optical interposer' chiplets — basically a way to bolt photonic modules directly onto custom AI chips . Pre-meaningful-revenue. If their chiplet platform gets a single major hyperscaler design win, the stock could double or triple. Lottery ticket with real engineering.
Specialty glass and optical fiber
Corning makes the actual fiber-optic cable that physically connects every AI data center — plus advanced display glass. Already in your Berserker portfolio at +101% from buy. Every meter of new AI fiber buildout means Corning fiber. Pricing power is improving as supply tightens. Manage via your existing 50-day moving-average rules.
Even with light pipes, you still need wires, switches, and sockets. A lot of them.
An AI data center has tens of thousands of physical connections — between GPUs, between racks, between rows. Each one needs a connector, and the long-haul ones need 'retimers' to clean up the signal. Astera Labs owns the retimer market. Credo owns the AI cable market. Arista and Cisco own the Ethernet-switch market. And Amphenol sits underneath all of it making the physical sockets that everything plugs into. Five companies tax every AI rack ever shipped.
Hyperscalers can buy GPUs from anyone. The fabric that connects them is bought from a much smaller list.
PCIe Gen 6 retimer monopoly + Scorpio scale-up fabric
Astera Labs makes the 'glue chips' that connect GPUs to each other and to memory inside an AI rack — including PCIe retimers (which boost signal strength so the chips can actually talk) and the new Scorpio fabric switches. Q1 2026 sales were $308M, up 91% YoY. Their 320-lane Scorpio X-Series is shipping to hyperscalers now.
ZeroFlap AECs — the de facto cabling standard
Credo's 'Active Electrical Cables' (AECs) are the cables hyperscalers are standardizing on for AI server-to-server links . 'ZeroFlap' branding promises no link drops at AI workloads. Quietly winning the AI cabling war.
Only Ethernet-AI fabric vendor at GB200 scale
Arista makes the Ethernet switches that connect every AI rack to every other AI rack . Meta, Microsoft, and Oracle all use them. They're the only vendor proven at NVIDIA GB200 deployment scale. Boring, dominant, prints money.
Connectors for every AI server — the everyone-needs-it tax
Amphenol makes the physical connectors and sockets in every AI server, NVLink rack, and CPO module . Every cable, every backplane, every rack pays a small Amphenol toll. The most boring and most reliable infrastructure name on the page.
Acacia + 1.6T silicon photonics dark horse
Often dismissed as a legacy networking company, but Cisco's Acacia subsidiary makes silicon-photonic transceivers at 1.6T . Trades cheap relative to Arista. Dark-horse beneficiary if CPO adoption pulls Acacia into the spotlight.
An AI data center now consumes as much electricity as a small city — and the grid wasn't built for it.
A modern AI training campus needs 1 gigawatt of electricity — about as much as 750,000 homes. The U.S. power grid was not designed for this. So hyperscalers are buying their own power plants (gas turbines from GE Vernova and Siemens Energy, nuclear PPAs from Talen/Constellation/Vistra, uranium fuel from Cameco, eventually small modular reactors from NuScale), their own switchgear (Powell, Eaton, Hubbell), their own cooling (Vertiv, Modine, nVent), their own backup generators (Generac, Cummins, Bloom), and hiring contractors (Quanta, EMCOR, Comfort Systems) to install it all. This is the largest chokepoint by company count — 22 names — because the build-out is enormous, physical, and slow.
Money cannot turn into electrons faster than the turbines can be cast and shipped. The bottleneck is physical, not financial.
Heavy-frame gas turbines — the AI power backbone
Maker of the giant H-class gas turbines that power new AI data center plants . $2.4 billion in data-center electrification orders in Q1 2026 alone — more than all of 2025 combined. Backlog $163 billion and growing. Order book sold out into 2030.
Third leg of the gas turbine triopoly
Siemens Energy is the third leg of the gas-turbine triopoly alongside GE Vernova and Mitsubishi. Almost doubled gas turbine sales (100→194 units) in fiscal 2025. 60% of 2025 gas turbine orders are tied to data centers. Partnered with Oklo on small-modular-reactor steam turbines. European ADR — liquidity is less than U.S. names. Wind subsidiary still bleeding cash, which is why the stock isn't up as much as GEV.
Liquid cooling pure-play — $15B backlog
Vertiv makes the uninterruptible power supplies, power distribution, and especially the liquid cooling systems for AI server racks. Modern AI racks pull 50-100 kilowatts (10x older racks) and require liquid cooling. $15 billion backlog. Direct NVIDIA GB200/300 partner.
1,920 MW AWS PPA through 2042
Talen owns the Susquehanna nuclear plant in Pennsylvania and signed a long-term power contract with Amazon for 1,920 MW . That deal locks in cash flow for 18 years. Direct power-to-data-center deal, no grid middleman.
Three Mile Island restart for Microsoft
Constellation owns the largest U.S. nuclear fleet , including the famous Three Mile Island plant. Microsoft signed a 20-year deal to restart Unit 1 and buy all the power. Meta then signed a similar deal at the Clinton plant. Nuclear is the only carbon-free 24/7 power.
Comanche Peak 1,200 MW nuclear PPA
Vistra is an independent power producer with a heavy nuclear fleet in Texas. Behind-the-meter nuclear deals lock in 20-year cash flows at premium prices. Meta is the biggest customer.
Uranium mining — the fuel for the nuclear thesis
If Microsoft, Meta, and Amazon are all signing 20-year nuclear PPAs, somebody has to dig up the uranium that fuels those reactors . Cameco is one of the world's largest. Spot uranium prices have already tripled in two years. Hyperscaler nuclear deals push utilities into long-term uranium contracts at higher prices — Cameco is the cleanest play. Spot prices are volatile, and Kazatomprom production levels remain a wildcard.
Small modular reactor designer
NuScale designs small modular reactors (SMRs) — bite-sized nuclear plants you could drop next to a data center campus . The dream is one hyperscaler picks them as their SMR partner, and the stock 5x's. The reality is no revenue from operating reactors yet, the first commercial deal collapsed in 2023, and dilution risk is real. Pure lottery ticket on the SMR adoption curve.
Switchgear/transformer pure-play
Texas-based maker of custom electrical 'switchgear' — the prefab electrical buildings dropped onto data center sites. About half of all U.S. 2026 data center builds are reportedly delayed by power gear shortages. Powell's order book reflects it.
HTS wires + DPA national-emergency play
AMSC makes 'high-temperature superconducting' wires and power electronics that move enormous current through small cables. Defense Production Act priority designation. Niche, but the only listed pure-play in the category.
Solid-oxide fuel cells — on-site dispatchable power
Bloom makes solid-oxide fuel cells that generate electricity on-site from natural gas . When data centers can't get a grid connection (and many can't until 2028+), Bloom's modules are the workaround. Already powering hyperscaler campuses today.
Data-center cooling — FY28 target >$2B
Modine is a thermal-management company that pivoted hard into AI data-center cooling . FY25 data-center revenue was $644M, up 119%. Management targets >$2B by FY28. Cleanest mid-cap cooling pure-play after Vertiv.
Datacenter racks/enclosures — Siemens + NVIDIA reference
nVent makes the specialized racks and enclosures that house AI servers , plus liquid cooling distribution. Q3 2025 datacenter orders +270%. Reference architecture partner with Siemens and NVIDIA.
Largest army of high-voltage line workers
Quanta owns the largest private workforce of high-voltage transmission linemen on the continent . You can't connect a data center to power without them. Record $48.5B backlog. CEO calls the addressable opportunity $2.4 trillion through 2030.
Mission-critical electrical/mechanical construction
EMCOR is the largest specialty mechanical/electrical contractor in the U.S. — the people who actually install the wiring, plumbing, and HVAC in data centers. Record remaining performance obligations of $15.8B.
MEP/HVAC contractor — backlog doubled to $12.45B
Comfort Systems is the second-largest mechanical/electrical/plumbing contractor . Backlog has doubled in the last year to $12.45B. Stock has been one of the quietest 10-baggers of the AI era.
Switchgear + transformers + (now) cooling
Eaton makes everything between the grid and the chip rack — switchgear, transformers, busways. Just bought Boyd Thermal for $9.5B to add liquid cooling. Most diversified play in the entire power-grid chokepoint.
Utility transformers — the actual grid bottleneck
Hubbell makes the distribution transformers (the gray boxes on poles, but bigger) that step grid voltage down for buildings . Lead times are 2-4 years. Pricing power is at all-time highs. Boring, dominant, oversold relative to peers.
Last-inch power delivery for dense AI racks
Vicor makes the tiny power-conversion modules that sit right next to the GPU and feed it clean voltage. As GPU power density rises, this 'last-inch' delivery problem gets harder, and Vicor's specialty modules become more valuable.
Backup generation for data-center reliability
Generac makes backup generators for buildings . Pivoting from residential into the data-center market. Hyperscaler vendor approvals are moving forward. If they crack the approved-vendor list, this re-rates.
Backup diesel gensets — the proven fallback
Cummins is the largest U.S. maker of the backup diesel generator sets that data centers fire up when the grid goes down . Diesel gensets are the proven, reliable backup standard — every gigawatt of new data center load means a few hundred megawatts of backup gensets get ordered. Quiet, slow-and-steady winner. Truck-engine cyclicality is the bulk of the business though, so the AI tailwind only partially shows up in the stock.
Analog/PMIC tollbooth on every AI server board
TI makes the boring analog and power-management chips on every AI server board — voltage regulators, current sensors, signal conditioners. Mega-cap, slow-moving, but every server pays this toll.
Hyperscalers can't build fast enough alone. They rent.
@aleabitoreddit's edge is following the money trail. Hyperscalers don't just spend their $700B in capex internally — much of it flows outward to specialty companies. NeoClouds like Nebius and CoreWeave rent GPU capacity back to Microsoft, Meta, and Amazon on multi-year, multi-billion-dollar contracts. Iris Energy adds the power-and-land twist. Cipher Mining is the smaller-cap analog still mid-pivot from Bitcoin. Celestica builds the actual servers. And underneath all of it sits the storage layer — Seagate and Western Digital make the cheap mass HDDs that hold every byte of training data ever generated. Seven companies, two distinct sub-themes, one massive river of money.
Hyperscalers can't physically build their AI infrastructure fast enough alone, so they rent it. The renters have committed contracts visible in filings. The math is real.
European NeoCloud — Aleab's biggest position
Re-emerged from the wreckage of Russian-linked Yandex as a pure-play AI cloud company . Just signed $46 billion in long-term cloud deals, including a multi-year supply agreement with Meta. Raised $4.3 billion in convertibles in March 2026. Eigen AI utilization layer is unique IP.
Land + power + GPU financing — $9.7B MSFT deal
IREN flipped from a Bitcoin miner into a full AI cloud operator . $9.7B Microsoft contract is in hand. Their edge: they already own land + power infrastructure that takes new entrants 3-5 years to build.
Bitcoin miner pivoting to AI hosting
Cipher Mining is another Bitcoin miner trying to make the IREN pivot — using existing power infrastructure and land to host AI workloads. Smaller market cap means a single AI hosting contract moves the stock harder. Higher execution risk than IREN. Continued reliance on Bitcoin price for the legacy business until the AI conversion is real.
Largest pure-play GPU cloud — NVIDIA-aligned
CoreWeave is the biggest dedicated AI cloud provider . $66.8B revenue backlog. ~3.1 GW of contracted power. NVIDIA is a major investor. Trades at 7x sales — much cheaper than Nebius — despite being scaled. Stock down 13% on the day shows the volatility.
AI hardware manufacturing — FY26 guide ~$19B
Celestica is the contract manufacturer that physically assembles AI servers and networking gear for hyperscalers. FY26 guide of ~$19B reflects the buildout. The 'who makes what NVIDIA designs' play.
Mass-capacity HDDs — the AI cold-storage layer
Seagate makes the giant hard drives that store the petabytes of training data AI models learn from. SSDs are too expensive for cold storage. HDDs are still the cheapest dollar-per-terabyte by a wide margin. AI demand has finally inflected HDD pricing power up.
HDD duopoly + post-split focus
WDC is the second leg of the HDD duopoly with Seagate . Post-split, the company is now a pure HDD play (the flash business spun off). Same AI storage tailwind, slightly different risk profile.