A $0.3B capacitor maker at 0.3x sales whose AI-server sales are surging and whose prices just went up: the closest thing to SanDisk's set-up before its run left in Japan.
Needs The latest capacitor price hikes have to stick at ~30% while AI-server power supplies fill the new large-can capacity, taking sales to ~¥230B at a 12% net margin, which makes it a ~$3.1B company, about 3x Nichicon today on similar sales.
Driver Sales of about ¥230B in three years vs ¥131.8B last fiscal year: the latest aluminium-capacitor price increases (~30% cumulative in the bull) on top of AI-server PSU volumes filling the new large-can capacity; last quarter’s sales were already ¥38.0B, a ¥152B run-rate.
Risk Aluminium foil and electrolyte costs outrun the price hikes again, as they did last fiscal year, and the net-debt-heavy balance sheet makes the leverage cut both ways.
How the number is built Bull: revenue $912M TTM → $1.5B in year 3 at 12% net margin = $176M earnings; × 20x P/E, +15% shares = $3.1B market cap vs $297M today = 10.3x. Base: revenue $912M TTM → $1.1B in year 3 at 5% net margin = $52M earnings; × 14x P/E, +15% shares = $639M market cap vs $297M today = 2.2x.
The product link holds. The 30 percent price hike does not. Japanese posts call Nippon Chemi-Con a cheap AI-server capacitor name that rose 4.5 percent with the parts bounce, and the same writers say a US rally is not new orders.
A $0.7B boiler maker with a $2.4B AI-campus power order and a $14B pipeline: with gas turbines sold out to 2030, steam is the bypass.
Needs B&W has to turn its $14B pipeline into roughly two more GW-scale orders like the $2.4B Applied Digital contract and deliver them at an 8% net margin, which makes it a ~$7B company, about 1.5x Argan today.
Driver 2029 revenue: ~$0.9B legacy boilers/services plus ~1.5 GW a year of data-centre gas-fired steam power islands at ~$2.0B per GW, the price of the $2.4B, 1.2 GW (4 x 300 MW) Applied Digital/Base Electron order; drawn from a pipeline above $14B (4-6 GW).
Risk Fixed-price project losses: B&W has a long record of cost overruns, and one bad 300 MW island could wipe out the equity of a company that still carries debt and pension liabilities.
How the number is built Bull: revenue $834M TTM → $4.0B in year 3 at 8% net margin = $320M earnings; × 25x P/E, +15% shares = $7.0B market cap vs $676M today = 10.3x. Base: revenue $834M TTM → $1.9B in year 3 at 4% net margin = $76M earnings; × 15x P/E, +10% shares = $1.0B market cap vs $676M today = 1.5x.
The pipeline holds. The conversion rate does not. Posts cite a backlog above $2 billion, a pipeline above $14 billion and data-center FastPower, but also a stock that already multiplied and then fell about 66 percent off its high. The Applied Digital order is not what people are confirming.
Down 70% with a signed $865M, 10-year AI lease already billing: energised power is the scarcest input in AI, and this campus can grow 7x.
Needs Duke Energy has to deliver the ~200 MW WhiteFiber is evaluating at NC-1 and it has to be leased at the Nscale rate; only the 40 MW is confirmed today, and the bull makes it a ~$3.8B company, about Cipher Mining's size now.
Driver 200 MW of critical IT leased at NC-1 at the Nscale contract rate ($865M over 10 years for 40 MW = $2.16M per MW per year) = $432M, plus ~$100M of cloud; needs Duke Energy to deliver the ~200 MW extra WhiteFiber is evaluating (site up to ~300 MW gross).
Risk Financing: the build-out needs about $2B of debt or equity, the proposed secured financing is not yet in place, and dilution could be far worse than the 30% assumed.
How the number is built Bull: revenue $93M TTM → $530M in year 3; × 13x EV/sales, +30% shares = $3.8B market cap vs $427M today = 8.8x. Base: revenue $93M TTM → $220M in year 3; × 7x EV/sales, +15% shares = $817M market cap vs $427M today = 1.9x.
The existing contract holds. The extra 200 MW does not. Posts say WhiteFiber is down about 78 percent, with 40 MW under the $865 million Nscale lease and a roadmap cut to 70 MW by year-end, not a Duke expansion.
A loss-making power-chip maker at 1x sales with its own fab and a seat on Nvidia's 800 V rack: fill the fab and the margin swing is SanDisk-shaped.
Needs AI-server power stages and 800 V parts have to take revenue from ~$680M to ~$1.35B within three years and refill the Oregon fab so margins get back to their last peak, which makes it a ~$5.4B company, under a tenth of Monolithic Power.
Driver Revenue roughly double TTM within three years: the company guided AI and server revenue up more than 60% quarter on quarter for the latest quarter, power ICs hit a record $72.7M quarter, and it supports Nvidia's 800 V DC rack; the extra volume runs through its own under-used Oregon fab.
Risk AI servers are still under a fifth of sales; if Nvidia's power-stage share goes to Monolithic, Infineon or Vishay, the fab stays half-empty and losses continue.
How the number is built Bull: revenue $679M TTM → $1.4B in year 3 at 14% net margin = $189M earnings; × 30x P/E, +5% shares = $5.4B market cap vs $741M today = 7.3x. Base: revenue $679M TTM → $850M in year 3 at 6% net margin = $51M earnings; × 20x P/E, +3% shares = $990M market cap vs $741M today = 1.3x.
The 800-volt option holds. The earnings path does not. Posts say Alpha and Omega was named for Nvidia's 800V rack but AI servers are still under a fifth of sales, revenue is about $680 million, and the company is still loss-making.
The purest small-cap version of SanDisk's price leverage: quarterly revenue up 59% in one quarter on price alone, and still valued at about 2x run-rate earnings.
Needs Legacy DRAM prices have to stay near today's records into 2028 as the majors keep exiting DDR3/DDR4, holding a ~45% net margin on ~$3.6B of revenue, which makes it a ~$16B company, two-thirds of Winbond today.
Driver 2028 revenue a third above the latest quarterly run-rate: Q3 revenue was about NT$22.2B (+59% q/q on price, not volume), a ~$2.7B annual pace, with record EPS of NT$36.79; legacy DDR3/DDR4 prices stay high while Samsung, SK Hynix and Micron move capacity to HBM.
Risk It is fabless and buys wafers, so a legacy-DRAM price roll-over hits it twice (price and wafer cost), and the market is already pricing it at about 2x run-rate earnings for exactly that reason.
How the number is built Bull: revenue $1.5B TTM → $3.6B in year 2 at 45% net margin = $1.6B earnings; × 10x P/E = $16.2B market cap vs $2.4B today = 6.7x. Base: revenue $1.5B TTM → $1.8B in year 2 at 25% net margin = $450M earnings; × 8x P/E = $3.6B market cap vs $2.4B today = 1.5x.
The quarter holds. A further 6x does not. Posts say Elite Semiconductor just printed record profit, with EPS of NT$36.79 and revenue up about fivefold, and the same tape is already selling the stock with the other memory names.
Trades at about 2.6x its first-half earnings run-rate: the SanDisk price effect at module-maker scale, valued as if it ends tomorrow.
Needs The memory shortage has to last through 2028 so Team Group keeps a ~25% net margin on ~$2B of revenue, which makes it a ~$5.5B company, about 1.5x ADATA today.
Driver 2028 revenue about double the first-half run-rate (H1 revenue NT$16.6B, EPS NT$53.44, a record): module prices stay at shortage levels and industrial/B2B SSD volumes grow ~50%.
Risk Module makers earn inventory gains on the way up and inventory losses on the way down; if contract prices roll over in 2027 the earnings vanish faster than SanDisk's.
How the number is built Bull: revenue $838M TTM → $2.0B in year 2 at 25% net margin = $500M earnings; × 11x P/E = $5.5B market cap vs $834M today = 6.6x. Base: revenue $838M TTM → $900M in year 2 at 8% net margin = $72M earnings; × 10x P/E = $720M market cap vs $834M today = 0.9x.
Does not hold up as a SanDisk. Team Group's own posts are product awards, and the Taiwan tape only notes the stock staying green on a weak memory day. The shortage talk is Nanya, Winbond and Phison.
A $0.3B copper-foil maker at under 1x sales, listed among the AI-server HVLP suppliers, with consensus already expecting earnings to more than double.
Needs Fukuda has to win real HVLP foil volume as the leaders sell out, doubling sales at Mitsui Kinzoku's margin, which makes it a ~$2.0B company, a quarter of Mitsui Kinzoku today.
Driver Sales roughly double as HVLP copper foil for AI-server boards and copper powder for capacitor electrodes overflow from the market leaders; consensus already prices earnings up ~2.5x (forward P/E 11.7 vs 29 trailing).
Risk The AI copper-foil money is going to Mitsui Kinzoku and JX; if Fukuda stays a second-source supplier the doubling never comes.
How the number is built Bull: revenue $271M TTM → $560M in year 3 at 14% net margin = $78M earnings; × 25x P/E = $2.0B market cap vs $312M today = 6.3x. Base: revenue $271M TTM → $330M in year 3 at 7% net margin = $23M earnings; × 15x P/E = $346M market cap vs $312M today = 1.1x.
Does not hold up. AI copper-foil posts are about JX Metals and Mitsui, not Fukuda Metal Foil. There is no investor thread on 5821.
United States · 1-yr +92% · Peers: packaging & back-end tools
Market cap $4.1B → $12.3B at 3x
3xin 3 yearsConviction 7/10
TLDR For Humans
A $4bn HBM bonder maker at half the multiple of back-end peers, with EPS up 19x off the trough.
Makes the wire bonders and thermocompression bonders that attach dies and stack HBM layers. Its TCB tools are a gating step for the taller HBM4 stacks, and its fluxless TCB is aimed at the next-generation bonding market.
Why it’s underrated Almost no dedicated posts on X - the bonder conversation is BESI versus ASMPT, and Kulicke is forgotten despite sitting in the same oligopoly. Revenue +44% and EPS +1,888% off the depressed base (margin recovery), yet it trades at 13.4x forward and 3.8x EV/sales - roughly half the multiple of the back-end tool peer group. A $4bn company selling picks and shovels to the HBM build-out.
Auto/industrial bonding recovery adds a second engine
Risks
Equipment orders are lumpy and cyclical
BESI and ASMPT are bigger and better-resourced rivals
Chinese bonder competitors undercut on price
Margin recovery is young - the EPS base is still small
How the number is built ~25%/yr EPS as the TCB/HBM4 bonder cycle builds (rev +44% TTM; EPS +1,888% is off a tiny base) × the average of a P/E re-rating 13.4x→22x (22x cap) and an EV/S re-rating 3.8x→6.9x (capped); ≈3.4x, rounded down.
Kulicke & Soffa has almost no dedicated posts. When bonders come up, the argument is BE Semiconductor versus ASMPT, not Kulicke.
The world's #1 cable maker for the power bottleneck at 1.7x EV/sales — the AI trade ran everywhere except here.
The world's largest cable maker: high-voltage and HVDC grid cable, submarine interconnects, and the optical and low-voltage cabling that wires data centres. When a gigawatt campus needs power, or a grid needs an interconnect, the order lands here or at one of two European rivals. Post the Encore Wire deal it is also a leading US building-wire supplier.
Why it’s underrated It owns the tightest physical bottleneck in the AI power chain - multi-year lead times on HV/HVDC cable - yet trades at 1.7x EV/sales and ~19x forward earnings while US power names at 3-7x EV/sales get the AI tape. X barely names it (the conversation stops at GE Vernova, Eaton, Vertiv). 1-year move of +17% is nothing next to the ~3x the power trade put up elsewhere.
Record grid and data-centre order intake flowing into 2027-28 revenue
HVDC interconnects (Germany, Italy, North Sea) and offshore-wind awards are multi-year, fixed-margin contracts
US data-centre build-out pulling through Encore Wire volumes
European grid capex plans (REPowerEU successors) keep accelerating
Risks
Project execution and raw-material (copper, aluminium) price swings
Competition from NKT and Nexans compresses tender margins
Europe capex timing slips if power demand disappoints
Already +212% over three years, so part of the re-rating has happened
How the number is built ~19%/yr EPS on the grid-cable supercycle (rev +12.5% TTM × operating leverage; TTM EPS +81% includes one-offs) × the average of a P/E re-rating 19.4x→21.1x and an EV/S re-rating 1.7x→2.6x; ≈3.1x, rounded down.
Prysmian is in the same gap. Data-center cable capex is being discussed, and Prysmian itself is barely named.
Japan · 1-yr +56% · Peers: power equipment & cable
Market cap $36.8B → $110B at 3x
3xin 5 yearsConviction 7/10
TLDR For Humans
Sold off with the Japanese wire trio on no bad news; 1.2x EV/sales for the cables every data centre and grid needs.
One of Japan's big three wire-and-cable groups: high-voltage power cable for grids and data centres, optical fibre and cable, and the world's largest wire-harness business for cars. Every AI data centre and every grid upgrade needs its cables somewhere in the chain.
Why it’s underrated 1.2x EV/sales for a company at the centre of the power bottleneck. The whole Japanese wire trio broke its 75-day average in the AI-power sector unwind with no company-specific bad news, and the bounce has been weak - while the demand story (data-centre power, optical fibre) did not change. EPS +103% TTM, dividend 2.1%, and it has not been chased by the US timeline at all.
Data-centre and grid cable orders in Japan and the US
Optical-fibre capacity expansion (read-across: Furukawa's India build-out)
Sector bounce after the AI-power unwind reverses
Auto harness margin recovery as EV costs normalise
Risks
Conglomerate complexity: autos, telecoms and cables in one P&L
Auto harness exposure to EV production slowdowns
Yen swings move the story in USD
Cable pricing power thinner than for transformers and switchgear
How the number is built ~19%/yr EPS on grid + fibre capex with margin recovery (rev +12.3% TTM) × the average of a P/E re-rating 15.0x→21.1x and an EV/S re-rating 1.2x→2.1x (capped); ≈3.8x, rounded down.
Sumitomo Electric is traded as one of three Japanese wire names with Fujikura and Furukawa. Local traders say the whole group broke its 75-day average on a sector unwind in AI and data-center power, with no stock-specific bad news, and that the bounce has been weak.
The cheapest stock in the whole screen at 0.75x EV/sales: high-voltage cable, full order book, 2.4% dividend.
France's high-voltage cable champion: grid connections, offshore-wind export cables and data-centre power links. The electrification of everything - including the AI build-out - runs through HV cable, and Nexans is one of three European firms that can make it at scale.
Why it’s underrated The cheapest stock in the entire screen at 0.75x EV/sales, flat over the past year (+2%) while the power trade ran. 13.8x forward earnings and a 2.4% dividend. TTM EPS is ugly (-77%) on one-offs and project provisions, which is exactly why the market has left it behind - the order book, not the trailing P&L, is what the grid build-out pays.
TTM earnings are genuinely poor - the turn must show up
Prysmian and NKT compete on every large tender
French energy policy shifts could slow grid capex
How the number is built ~15%/yr EPS off a normalized base (TTM EPS hit by one-offs) × the average of a P/E re-rating 13.8x→21.1x and an EV/S re-rating 0.75x→1.35x (capped); ≈3.3x, rounded down.
Nexans is quiet on X. US power lists name GE Vernova, Eaton and Vertiv. The European high-voltage cable makers are not in that conversation.
Advanced packaging is the gating step for every AI chip; Amkor is up just 20% in a year and half the multiple of the tool makers it feeds.
One of the two big independent outsourced assembly-and-test (OSAT) houses. It packages the accelerators, HBM stacks and chipsets that go into AI servers, and is building an advanced-packaging campus in Arizona next to TSMC for US-based AI silicon.
Why it’s underrated Advanced packaging is the gating step for every AI accelerator, but Amkor has barely participated: +20% in a year, 14.6x forward and 1.35x EV/sales - roughly half the multiple of the back-end tool makers it feeds. Per this week's posts, customers are prepaying for AI packaging capacity while the market fixates on the negative free cash flow of the build-out. Computing revenue is up about a quarter year-on-year.
Arizona advanced-packaging campus ramps (TSMC-adjacent, US CHIPS-funded)
Customer prepayments for AI packaging capacity
HBM and 2.5D packaging attach rates rising with every accelerator generation
Apple/Nvidia/TSMC US packaging demand
Risks
Negative free cash flow while the build-out is funded
OSAT is a structurally lower-margin business than tool makers
Capital intensity rising; utilisation is cyclical
Chinese OSATs (JCET, Tongfu) compete on price
How the number is built ~24%/yr EPS on the AI-packaging ramp (rev +15.7% TTM × leverage; EPS +99.9% is base-effect) × the average of an OSAT-peer P/E re-rating 14.6x→17.5x (ASE/JCET territory, not the 25.7x tool-maker median) and a flat EV/S 1.4x (ASE ~1.4x); ≈2.1x, rounded down.
Amkor is active this week and split. One camp says the quarter missed and free cash flow went negative because of the packaging build-out. Another points at computing revenue up about a quarter from last year, customers prepaying for AI packaging, and a forward multiple still in the mid-teens. The Arizona advanced-packaging campus is the corporate post everyone quotes.
Phison's CEO called the 2027 NAND shortage; the market quotes him to buy SanDisk while Phison itself trades at 5x forward.
Taiwan's NAND controller and module house: it designs the controllers and builds the enterprise SSDs that AI servers and data centres buy. Its chief executive has become the shortage oracle of this cycle - his August call was that 2027 NAND constraints will be worse than 2026 because a new fab takes up to four years.
Why it’s underrated 5.2x forward earnings and 3.0x EV/sales for one of the two controller duopolists in the tightest part of memory. The stock itself is quiet - its CEO's shortage line is quoted to argue for SanDisk and Kioxia instead. Revenue +157% and EPS +841% TTM, yet it trades like a mid-cycle parts maker rather than the AI-storage toll booth it has become.
NAND supply-demand gap widening into 2027 (management's own forecast)
Enterprise SSD attach to AI servers keeps compounding
In-house controller + module margin expansion with NAND prices
Retail module inventory restocking after the drawdown
Risks
Peak-cycle earnings: NAND inventory gains reverse hard in a downturn
Customer competition: NAND makers increasingly build their own controllers
Taiwan geopolitical risk premium
TWO listing: thinner liquidity than TSE names
How the number is built Peak-cycle math, assumes the shortage extends as management calls it (2027 worse than 2026): EPS +20%/yr for 3 years from peak, then an exit at 10x on 70% of peak EPS (mid-cycle haircut) vs today's 5.2x; ≈2.3x, rounded down. With EPS flat from peak the same math gives ≈1.4x.
Phison's own stock is quiet. Its chief executive's line from August — that 2027 NAND constraints will be worse than 2026 because a new plant takes up to four years — is still being used, mostly to argue for SanDisk and Kioxia.
The NAND-controller duopolist the shortage conversation forgot: +71% revenue growth, Strong Buy, nobody talking about it.
Fabless designer of NAND flash controllers - the silicon between the flash and the server, phone or SSD - plus enterprise SSD controller platforms for AI servers. Nearly every non-Samsung flash module ecosystem uses its controllers, and it sells into eMMC/UFS for AI smartphones.
Why it’s underrated The capital-light way to own the memory shortage: it rides NAND bit growth without owning wafers or inventory. Revenue +71% and EPS +268% TTM, yet the ticker barely appears on X - the shortage conversation jumps from Phison's CEO straight to SanDisk and Kioxia. Its 14x forward P/E looks rich next to Phison's 5x, but Phison's earnings embed NAND inventory gains at peak prices; SIMO's are cleaner, royalty-style controller economics. Consensus is Strong Buy with a target ~78% above spot.
NAND controller shortage into 2027 (Phison CEO: supply constraints worsen next year)
Enterprise/AI SSD controller ramps at hyperscalers
eMMC/UFS attach to the AI-phone replacement cycle
PC/NB SSD attach rates still rising
Risks
Customer concentration: the module makers and NAND vendors are few
Chinese controller rivals (Maxio) push domestic substitution
A NAND price downturn hits both volumes and royalties
Small cap (~$7bn): liquidity and volatility
How the number is built ~25%/yr EPS on the NAND-controller + enterprise-SSD ramp × the average of a P/E re-rating 14.5x→22x (22x cap; Phison's 5x is peak-distorted, ignored) and an EV/S de-rating 5.2x→3.0x (peer); ≈2.0x, rounded down.
Silicon Motion barely appears. The NAND-controller conversation quotes Phison's chief executive on a shortage into 2027, and then the money discussion jumps to SanDisk, Kioxia and Micron.
United States · 1-yr -19% · Coherent, Lumentum, Corning
Market cap $12.3B → $24.6B at 2x
2xin 3 yearsConviction 7/10
TLDR For Humans
The optics bottleneck's manufacturer: down 19% in a year, +36% revenue, 18.8x forward against ~26x peers.
The largest independent contract manufacturer of optical modules and components: it builds the transceivers, amplifiers and photonic assemblies that Nvidia, Ciena, Infinera and the hyperscalers ship into AI clusters. When optics is the bottleneck, Fabrinet is the hands that assemble it.
Why it’s underrated Optics is the tightest bottleneck after memory, but the crowd chased the chip names: Fabrinet is DOWN 19% over a year while revenue grew 36% and EPS 42%. 18.8x forward earnings and 2.5x EV/sales against ~26x and ~6.6x for the optics peer group, net cash, and it reports fiscal Q1 soon.
New Thailand campus adding capacity for AI photonics
Next earnings as the near-term catalyst
Risks
Customer concentration: a handful of programmes drive the AI ramp
Optics orders can air-pocket between product generations
Component shortages (lasers, wafers) can cap output
US-Thailand manufacturing footprint carries tariff and logistics risk
How the number is built ~25%/yr EPS on the AI transceiver ramp (rev +36% TTM, EPS +42%) × the average of a P/E re-rating 18.8x→22x (vs ~26x optics peers; 22x cap) and an EV/S re-rating 2.5x→4.4x (capped); ≈2.9x, rounded down.
Fabrinet is on oversold-infrastructure screens and on lists of AI names under 20x forward earnings. There is little fresh company-specific argument beyond 'optics manufacturer, sold off with the rest of the complex.'
Japan · 1-yr +9% · Peers: packaging & back-end tools
Market cap $1.1B → $2.2B at 2x
2xin 3 yearsConviction 7/10
TLDR For Humans
A $1.1bn near-monopoly in HBM moulding equipment that has barely moved while the whole sector re-rated.
A Kyoto-based maker of compression-moulding equipment with near-total share in the moulding steps of advanced packaging - the encapsulation of HBM stacks, panel-level fan-out and chip-first packaging. Every HBM module that ships passes through tooling like Towa's.
Why it’s underrated A $1.1bn company with a near-monopoly in a mandatory process step that has barely moved (+9% in a year, +31% in three) while the Japanese back-end basket (Ibiden, Disco, Advantest) re-rated multiples of that. X has no thread on it at all. Revenue +29% TTM with 16.4x forward earnings.
HBM stack heights rising -> more and newer moulding steps
Panel-level packaging transitions to volume production
Order book visibility from memory makers' 2027-28 fab plans
Compression moulding share gains in power and automotive chips
Risks
Tiny market cap: orders can gap the stock
Customer capex timing is everything in tool land
Single-product-line concentration
Thin liquidity on the Tokyo exchange
How the number is built ~18%/yr EPS as HBM stack heights rise (rev +29% TTM, TTM EPS +14.8% in a near-monopoly step) × the average of a P/E re-rating 16.4x→22x and an EV/S re-rating 2.3x→4.2x (capped); ≈2.6x, rounded down.
Towa has no real thread. Packaging-equipment attention is on BE Semiconductor, ASMPT and, historically, Hanmi.
Hong Kong · 1-yr +36% · Peers: packaging & back-end tools
Market cap $6.3B → $12.6B at 2x
2xin 3 yearsConviction 6/10
TLDR For Humans
If slower HBM4 speeds favour the existing TCB tools, ASMPT wins over BESI — and BESI is the one that just ran.
The leader in thermocompression bonding (TCB) - the incumbent tool that stacks HBM dies and bonds 2.5D/3D packages - plus a large SMT (surface-mount) equipment business. Its TCB tools are installed across the memory makers and OSATs.
Why it’s underrated The relative-value case of the week: if Nvidia's slower HBM4 pin speed delays hybrid bonding, the existing TCB toolset wins, which favours ASMPT over BESI - and BESI is the one that just rallied 11% in a day. ASMPT trades at 2.7x EV/sales and 22x forward against that backdrop, with revenue +32% TTM as packaging capex returns.
HBM4 spec shift keeps TCB the incumbent bonding tool
Memory makers' advanced-packaging capex into 2027
SMT recovery as electronics manufacturing restocks
China advanced-packaging demand (domestic HBM efforts)
Risks
Hybrid bonding eventually displaces TCB - timing is the fight
TTM earnings still trough-level (no EPS growth to show yet)
Hong Kong listing discount and China exposure
BESI is executing strongly and taking share of the narrative
How the number is built ~20%/yr EPS as TCB orders convert (rev +32% TTM off trough earnings) × the average of a flat P/E (already at the 22x cap) and an EV/S re-rating 2.7x→4.9x (capped); ≈2.4x, rounded down.
ASMPT's main mention is that relative-value note: a possible winner against BE Semiconductor if hybrid bonding slips. Otherwise it is quieter than BESI's move.
Taiwan · 1-yr +38% · Asia Vital Components, Modine, nVent
Market cap $4.1B → $8.2B at 2x
2xin 3 yearsConviction 6/10
TLDR For Humans
The cooling trade's quieter Taiwan play: +71% revenue, 20.9x forward, in line with peers.
Taiwan's leading maker of vapour chambers, heat pipes and cold-plate modules for the 100kW+ AI racks, plus fans and blowers. It ships the thermal hardware that lets Nvidia-scale servers run liquid-cooled.
Why it’s underrated The cooling trade bid Vertiv and AVC to the moon; Auras has the same AI-server thermal exposure at 20.9x forward and 4.5x EV/sales - roughly in line with peers - with revenue +71% and EPS +155% TTM. It has run +38% in a year, so this is a relative-value pick in the hottest bottleneck, not an undiscovered one.
Cold-plate and vapour-chamber attach to next-gen AI racks
Liquid-cooling content per rack rising with power density
US/EU hyperscaler cooling retrofits
Taiwan supply-chain position (Nvidia-adjacent)
Risks
Taiwan thermal makers compete on price; margins can compress
Already +38% in a year - less re-rating headroom
Design wins are lumpy
Concentration in server OEMs
How the number is built ~25%/yr EPS as liquid cooling ramps (rev +71% TTM; EPS +155% is base-effect) × the average of a P/E re-rating 20.9x→22x and an EV/S re-rating 4.5x→4.8x (already near peers); ≈2.1x, rounded down.
South Korea · 1-yr +143% · Peers: SK Hynix, Micron, Kioxia
Market cap $1,187B → $2,375B at 2x
2xin 3 yearsConviction 6/10
TLDR For Humans
3.5x forward for the only company with DRAM and advanced logic under one roof — the HBM4 base-die winner.
The only company on earth with DRAM, NAND and advanced logic under one roof. That matters for HBM4: with Nvidia slowing the pin speed, the harder problem has moved to the base die - and Samsung can build both the memory and the logic. It also runs the foundry and the phone business that funds it.
Why it’s underrated 3.5x forward earnings for the company a desk note this week preferred over Micron and Hynix on the HBM4 base-die argument. Chinese cloud buyers are reportedly asking Samsung and Hynix for supply deals running as long as ten years. The market still prices the chaebol discount and the old HBM qualification lag, not the structural shortage the rest of the industry is now confirming.
HBM4 base-die advantage: only firm with DRAM + advanced logic
Long-duration supply deals from Chinese cloud buyers
Record memory margins flowing through consolidated earnings
Shareholder-return reform (buybacks) narrows the governance discount
Risks
History of HBM qualification slips vs SK Hynix
Foundry losses drag consolidated margins
Chaebol governance discount persists
KRW strength or memory downturn hits earnings in USD terms
How the number is built 3.5x fwd prices peak-cycle memory EPS; with EPS flat from peak (no growth credited from here) and an exit at 10x on 70% of peak (mid-cycle haircut): 0.7×10/3.5 ≈ 2.0x, rounded down.
Samsung is up hard with SK Hynix, and one detailed desk note prefers it to Micron and Hynix: after Nvidia slowed the HBM4 pin speed, the harder problem is the base die, and Samsung is the only firm with DRAM and advanced logic in one company. Separate posts say Chinese cloud buyers are asking Samsung and Hynix for supply deals that run as long as ten years.
Japan · 1-yr +90% · Shin-Etsu, Resonac, Elite Material
Market cap $3.5B → $7.0B at 2x
2xin 5 yearsConviction 7/10
TLDR For Humans
T-glass near-monopoly for AI chip substrates: EPS +249% TTM at 12x trailing, in a materials basket that forgot it.
The world's largest maker of T-glass (low-dielectric glass cloth) used in the advanced substrates and boards that package AI chips, plus glass-fibre, insulation and medical businesses. Every high-end AI accelerator board passes through glass cloth like Nittobo's.
Why it’s underrated A near-monopoly input to AI chip substrates that the market has barely noticed: the substrate makers it supplies (Ibiden, Unimicron) got the AI bid while Nittobo moved on earnings alone - EPS +249% TTM on rev +12% - and still trades at 12x trailing earnings against ~18x for materials peers. It has already moved (+90% in a year), so this is a multiple-catch-up case, not an undiscovered one.
Thin coverage: the ~23x fwd P/E vs 12x TTM reflects a stale consensus EPS estimate for this thinly covered name
Customer concentration in the substrate makers
Chinese glass-cloth makers entering low-end grades
Tiny cap (~$3.5bn) and thin liquidity
How the number is built ~17%/yr EPS as T-glass and glass-cloth capacity tightens (rev +12% TTM; EPS +249% is base-effect) × the average of a P/E re-rating 12.4x→18x on TTM earnings (the 22.9x fwd figure reflects a stale consensus estimate) and an EV/S de-rating 4.4x→3.7x (peer median); ≈2.6x, rounded down.
Nittobo has almost no posts. Glass cloth for AI substrates is a specialist topic the timeline has not picked up.
GE Vernova's turbine book doubled and its stock re-rated; the #2 gas-turbine maker is down 20% in a year.
Japan's heavy-industrial champion: gas turbines for firm power (the workhorse of the data-centre build-out), nuclear reactors (Japan's PWR vendor, with new-build and SMR work), defence (missiles, F-35 components) and rockets. It is the only company that sells every layer of firm power.
Why it’s underrated GE Vernova's turbine book went from 50 GW to 116 GW and its stock re-rated - Mitsubishi Heavy, the number-two large gas-turbine maker, is DOWN 20% over the past year while the same order wave hits its Mitsubishi Power division. 2.2x EV/sales, EPS +59% TTM, plus a defence rearmament kicker the market is not paying for in the AI trade.
Order-book-to-sales conversion as turbine slots ship
Risks
28.7x forward is a premium - defense quality is priced in
Consolidated revenue growth is flat TTM (+0.5%)
Large project execution (aircraft, nuclear) carries tail risk
Yen appreciation against the USD earnings mix
How the number is built ~15%/yr EPS on turbine and defence order-book conversion (revenue flat TTM) × the average of a flat P/E (no re-rating credited — 28.7x fwd already exceeds the 21.1x peer median) and an EV/S re-rating 2.2x→2.6x; ≈2.2x, rounded down.
Mitsubishi Heavy is thin. Japanese defense-and-energy lists in this window name other heavy industrials, and the turbine shortage on English X is GE Vernova.
Humanoid leg joints are the scarce component; the load-bearing reducer monopoly trades at cyclical lows.
The precision-reducer maker whose cycloidal RV gears sit in the load-bearing joints - hips, knees, legs - of industrial robots and future humanoids. If humanoids scale, the high-torque joints are the scarce component, and Nabtesco is one of two qualified suppliers worldwide.
Why it’s underrated The humanoid trade bought Harmonic Drive to the moon, then got hit by Chinese price cuts - and Nabtesco never got the bid at all: +15% in a year, 1.5x EV/sales, 19x forward. Revenue is still down 8% TTM (industrial automation trough) so the optionality is free: you pay cyclical-low prices for the load-bearing joint monopoly.
Humanoid leg-joint volumes at the Chinese and US OEMs
Industrial-automation capex cycle turning up
China localization (domestic plants) shields margins
Aircraft and wind businesses add non-robot earnings
Risks
Chinese reducer price war spreads from harmonic to cycloidal gears
Humanoid volumes may stay small for years
Revenue still declining TTM - the trough may not be over
Big robots (the RV sweet spot) are the slowest part of the market
How the number is built ~15%/yr EPS as industrial automation recovers and humanoid reducers scale (rev -8% TTM trough) × the average of a P/E re-rating 19.4x→22x and an EV/S re-rating 1.5x→2.7x (capped); ≈3.0x, rounded down.
Nabtesco rides the same component map, as the cycloidal-reducer supplier for the load-bearing joints, with older posts citing a profit jump on robot orders. It gets less heat than Harmonic Drive, and the China-price thread hits the whole reducer complex.
SanDisk sold memory. When AI data centres ate the supply, prices jumped and its earnings exploded.
Each pick sells the next shortage
Power, chips, memory, cooling, optics and robotics: the parts the build-out needs next.
Small, cheap and ignored
Small companies with fixed costs, where a shortage turns into earnings fast, before the market notices.
Already ran, so we skipped them
S$SNDK SanDisk +639%SH$000660.KS SK Hynix +219%WD$WDC Western Digital +201%F$5803.T Fujikura +53%L$LITE Lumentum +347%
Also on the watchlist
KH$285A.T Kioxia HoldingsAV$3017.TW Asia Vital ComponentsTE$8035.T Tokyo ElectronN$NKT.CO NKTPG$POWERGRID.NS Power Grid Corp of IndiaV$VRT VertivE$6361.T EbaraAT$8271.TW Apacer TechnologyPS$PSIX Power Solutions InternationalI$IQE.L IQE
The agents that did the work
ClaudeScreened the market and built every revenue, earnings and valuation path.
DeepSeekPulled the first screen of prices and fundamentals worldwide.
GrokReaffirmed every pick on X: what people say and whether the story holds.
Bull case: the multiple if the shortage plays out, built from each company’s revenue and earnings path and the valuation it would need. Base case: if growth simply continues. 2x–3x projections: earnings growth × the average of a P/E and an EV/S re-rating, capped. Predicted market cap = today’s market cap × the projected multiple. Picks from 7 countries. This is research, not financial advice.
10X AI Stocks Crewlattice
10x$BWin 3 years
8.8x$WYFIin 3 years
Power
The product link holds. The 30 percent price hike does not. Japanese posts call Nippon Chemi-Con a cheap AI-server capacitor name that rose 4.5 percent with the parts bounce, and the same writers say a US rally is not new orders.
