This deep dive covers back-end semiconductor equipment. The sector has already run hard. The obvious question writes itself.

Is the party already over?

We also saw 30–40% single-day moves in some names over the past two days. The timing may feel awkward. But we wanted to check whether anything still looks interesting — and whether the names that have gotten all the attention are actually fairly priced.

Is the Party Already Over?

Back-end equipment stocks have been rising steadily for one to two years. At first glance, the chart suggests it might be too late.

We decided not to take that at face value. Three questions had to be answered first.

Is the memory shortage real? Will capital spending into back-end continue through 2027? And does the cycle hold together beyond 2028?

On the memory shortage: We addressed this a month ago in our memory check-up piece. SK Hynix and Samsung both said explicitly in their IR communications that this cycle is structurally different from past ones. Samsung noted that 2027 demand is already being booked now.

One more data point: DDR4 spot prices are up 1,360% from their 2024 lows. AI data centers are absorbing roughly 70% of global DRAM demand. New fabs take four to five years to build. The 2026–2027 window is structurally incapable of clearing the shortage. Answer: real.

On 2027 CAPEX: TSMC’s 2026 capex is approximately $56 billion — an all-time high. CoWoS capacity is set to quadruple from end-2024 levels. NVIDIA has locked in roughly 60% of 2026 capacity and 50% of 2027. The 10-percentage-point decline isn’t demand softening — it’s AMD, Broadcom, and Google taking their own slices with custom AI chips.

The clearest signal: TSMC outsourced advanced packaging for the first time. Approximately 240,000–270,000 wafers, with 180,000–190,000 going to Amkor. When TSMC runs out of room, the cycle still has legs. Answer: accelerating.

On post-2028: Back-end equipment typically runs six to nine months from order to revenue recognition. If 2028 and beyond were empty, the search for new positions would be less meaningful. The HBM roadmap looks like this: HBM4 in 2026, HBM4E in 2027–2028, HBM5 in 2028–2030, HBM5E around 2031. Each generation requires new equipment cycles. Answer: no cliff.

All three questions passed. But the process surfaced an unexpected clue.

Beyond HBM5, stacking 20-plus layers changes the bonding method entirely. TC bonders — where Hanmi Semiconductor leads — remain the primary tool through HBM4E. The extension came from JEDEC relaxing height limits, giving the existing approach another generation. But beyond that, hybrid bonding takes over: copper-to-copper, no bumps. BESI and ASMPT are better positioned there.

The signals are already in the order data. SK Hynix placed its first mass-production hybrid bonding order in March 2026 — a system jointly developed by AMAT and BESI. Samsung has decided to introduce hybrid bonding starting with HBM4, running parallel qualification with SEMES and BESI equipment.

Hanmi isn’t standing still. It’s investing over ₩50 billion in a dedicated hybrid bonding line at its Incheon Plant 7, targeting a 2027 ramp, and has received a development order from SK Hynix on the technology.

The center of gravity is shifting — but it’s not a displacement story. The bonding pie is getting bigger. The seats inside it are being rearranged.

The party isn’t over. It’s moving to a different room.

What Is Actually Happening in Back-End?

Back-end works differently from front-end.

Our earlier piece on front-end materials concluded: the same fab wears out faster. As processes become more precise, consumables get used up more quickly — recurring revenue on every wafer.

Back-end is different. A single piece of equipment costs tens to hundreds of millions of won. It doesn’t wear out daily. You need to buy more units to add capacity. The right frame here is:

The same chip stays longer.

HBM stacking requires bonding at every layer — 12 times for a 12-layer stack, 16 for a 16-layer stack. At each step, there’s a KGSD check (Known Good Stacked Die): confirming that everything stacked so far is functioning correctly. One bad layer makes everything above it worthless. As bonding repeats, so does verification.

Two engines are driving this.

First, HBM stacking itself. More layers, more bonding cycles, more test insertions per die. The throughput ceiling on bonding equipment becomes the ceiling on HBM output.

Second, AI accelerator packaging. NVIDIA’s CoWoS (Chip on Wafer on Substrate) process bonds GPU dies with HBM into a single package. Add SoIC (System on Integrated Chips) on top, and the number of test insertions per advanced AI chip climbs to 10–20. That’s the number Advantest’s CEO cited directly.

Three pieces of evidence that this is translating into real orders: CoWoS capacity is expanding roughly 4x from end-2024 levels, NVIDIA has locked in the majority of capacity through long-term agreements, and TSMC has opened up outsourcing for the first time. Orders placed today turn into revenue in the second half of 2026 through the first half of 2027. These aren’t stocks that have already priced everything in — they’re companies just beginning to recognize revenue.

Mapping the Back-End Value Chain

Back-end is the sequence of steps between a wafer entering the line and a finished chip coming out. Cut the wafer into dies, stack and bond the dies, heat-treat, surface-flatten, and test electrically.

We mapped nine categories across the process flow. The ★ marks the dominant player in each category.

Dicing — cutting wafers into individual dies. Disco (6146 JP) ★ dominates globally. Iotech (039030 KS) has domestic exposure.

TC Bonding — thermal compression bonding, stacking dies with heat and pressure. Primary method through HBM4E. Hanmi Semiconductor (042700 KS) ★ leads domestically. ASMPT (0522 HK) competes globally.

Hybrid Bonding — copper-to-copper direct bonding, no bumps. Primary method from HBM5 onward. BESI (BESI NA) ★ leads globally. ASMPT competes. Hanmi is targeting entry by 2027.

Annealing — post-bonding heat treatment to stabilize the bond interface. DIT (110990 KS) ★ has co-developed with SK Hynix since 2019. Iotech also has exposure.

Descum / CMP — surface planarization before and after bonding. PSK Holdings (031980 KS) ★ is the primary domestic player. AMAT competes globally.

Memory Testers — electrical inspection of HBM and DRAM. Advantest (6857 JP) ★ dominates.

Logic Testers — electrical inspection of GPUs, CPUs, and ASICs. Teradyne (TER US) ★ leads.

Handlers, Sockets, and Pins — interface components between chips and testers. Techwing (089030 KS) ★ makes handlers and cube probers. ISC (095340 KS) ★ makes test sockets. LEENO Industrial (058470 KS) ★ makes test pins.

OSAT — outsourced semiconductor assembly and test. Amkor (AMKR US) ★ is the largest US-headquartered player. ASE competes.

OSAT sits separately — it’s not a single process step but a full outsourcing model for the entire back-end sequence.

A few things stand out from this map.

Bonding splits into two rows. TC bonding is Hanmi’s territory. Hybrid bonding is BESI’s. The center-of-gravity shift from the first chapter shows up right here in the table.

Testing is deeper than it looks. Three of the nine rows are test-related. As inspection frequency increases, every player in this chain benefits together.

Annealing and descum — DIT and PSK Holdings — sit outside bonding and testing. But they’re required regardless of which bonding method wins. More on that shortly.

Row 9 is where TSMC’s outsourced wafers land. Roughly 70% of that volume goes to a single company: Amkor.

Bonding — The Center of Gravity Is Moving

Bonding is the heaviest position in back-end. Every layer added requires one bonding step. The throughput ceiling of bonding equipment is the production ceiling for HBM.

In Q1 2026, Hanmi Semiconductor and BESI moved in opposite directions within the same quarter. Hanmi’s revenue fell 88% quarter-on-quarter. BESI posted 28% year-on-year revenue growth and guided for 30–40% growth in Q2. AMAT — the world’s largest equipment company — acquired approximately 9% of BESI during the same period. The direction became clear.

The underlying story is a shift in bonding method. TC bonding inserts solder bumps between dies to make the connection. Stack beyond 20 layers and even the bump space becomes a constraint. Hybrid bonding eliminates the bumps entirely — copper bonded directly to copper. Thinner stack, faster signal, better thermal dissipation.

TC bonding remains the primary method through HBM4 and HBM4E. The full transition to hybrid bonding is an HBM5 story — 2028 and beyond.

Hanmi isn’t a finished company. JEDEC’s height relaxation gave TC bonding another generation of runway. Hanmi has also introduced next-generation TC bonders targeting HBM5 and HBM6. And it’s building a hybrid bonding line of its own, targeting a 2027 ramp — while holding a development order from SK Hynix on the technology. It’s a dual-track strategy: protect the present, plant a foot in the future.

That said, the market already knows the TC bonding story. A ₩30 trillion market cap is the result of that knowledge. And the HBM5 transition risk is just starting to surface. Known strength and emerging concern coexist in the same ticker.

A note on annealing: DIT has co-developed laser annealing with SK Hynix since 2019. We initially considered it a hidden hybrid bonding beneficiary. On closer review, two things gave us pause.

DIT’s laser annealing is weighted more toward post-ion-implant wafer recovery in next-generation DRAM than toward hybrid bonding void elimination. And the thermal treatment step inside hybrid bonding appears to be converging toward AMAT’s integrated system. The direct hybrid bonding narrative for DIT was overstated.

More significantly: DIT’s core laser annealing technology is currently subject to criminal proceedings. Prosecutors conducted a search and seizure in 2024 on alleged trade secret misappropriation. Indictments followed in 2025, and the trial is ongoing. The outcome could affect the foundation of the growth thesis.

Iotech also has annealing exposure, but its back-end growth axis sits more in laser dicing and debonding. It has begun supplying laser annealing to Samsung lines and debonding tools to TSMC — tracking the dicing-and-debonding path rather than annealing.

Testing — A Real Bottleneck, but the Market Already Knows

Chips are expensive, so they get tested repeatedly. HBM is checked at every stacking step and again after full assembly. An AI accelerator package with multiple HBM units attached — each worth tens of thousands of dollars — cannot afford to discover a defect post-packaging. Ten to twenty test insertions per advanced AI chip. More inspections means more equipment, more sockets, more pins.

Two companies divide the tester market. Advantest (6857 JP) covers memory. Teradyne (TER US) covers logic. Since HBM is memory, Advantest is the more relevant name for the thesis we’re tracing here. The market has already reached the same conclusion.

In the gap between the testers and the chips, Korean companies hold strong positions.

Techwing (089030 KS) makes handlers and cube probers. The cube prober emerged as the critical inspection tool for stacked HBM dies, and Techwing’s stock is up nearly 14x from its lows on that recognition.

ISC (095340 KS) specializes in rubber test sockets. The HBM angle is understood. What the market has not fully priced is the potential extension into next-generation ASIC sockets and photonics testing.

LEENO Industrial (058470 KS) makes test pins. Its strength is stability across a diversified customer and chip base rather than high-beta exposure to any single theme.

Testing has a genuine bottleneck. But it’s a bottleneck the market has already recognized. Advantest is firmly framed as an HBM beneficiary. Techwing has already run 14x. Incremental upside now depends on new triggers or consensus upgrades.

Beyond Bonding and Testing — Where the Tollgates Sit

We flipped the question. Regardless of whether TC bonding or hybrid bonding wins, regardless of which technology becomes dominant — what positions does every participant have to pass through?

PSK Holdings (031980 KS) is one of those positions. Descum is the process that planarizes and cleans the chip surface before and after bonding. The surface has to be clean and flat for bonding to work. That requirement doesn’t disappear with the bonding method. If anything, hybrid bonding — which demands more precise surface contact — makes surface condition more critical, not less. As CoWoS-type advanced packaging scales up, descum demand scales with it. This was our conclusion in the glass substrate deep dive as well — the path led to TSMC and PSK Holdings.

Amkor (AMKR US) is the other tollgate. TSMC outsourced advanced packaging volume for the first time, and roughly 70% of that went to Amkor. TSMC keeps the high-value processes in-house and routes assembly and test to external partners. As long as CoWoS volume grows, the flow to Amkor grows with it. Amkor is also building advanced packaging capacity in Arizona alongside TSMC, positioning itself for the onshoring trend in AI chip supply chains.

Pricing the Bottlenecks

We walked through the bottlenecks process by process. Now we check how much of those bottlenecks are already priced in.

This valuation is a coordinate tool — it shows where the current share price sits relative to scenarios, not when to buy or sell. For companies with clear earnings visibility, think of it as an early preview of individual target prices in future reports. For companies with financial risk, dilution has already been factored in, so numbers may look more conservative than sell-side consensus. That’s intentional.

All companies were classified by business character and layer position into four types, with fair value midpoints and upper bounds derived accordingly. The midpoint reflects a conservative scenario; the upper bound reflects the most aggressive scenario that remains internally consistent.

All prices as of the time of writing. Sorted by distance from the midpoint of the fair value range and upper bound.

Read more

On June 12, a new ticker appears on Nasdaq. SPCX. SpaceX, the most valuable private company in human history, enters the public market at a reported valuation of roughly $1.75 trillion.

In the weeks leading up to the listing, capital flooded into anything adjacent to space. Satellite and rocket-adjacent names ran anywhere from double to several multiples in the span of a year. Then, right around peak excitement, the broader market took a breath. Nasdaq dropped more than 4% in a single session in early June, led by AI and semiconductor names, and space stocks got caught in the same wave. This looked less like a sector-specific problem and more like an overextended market pausing.

When that kind of noise hits, it is worth slowing down. Overpriced assets and genuinely overlooked ones tend to get lumped together in a selloff. I wanted to walk through this landscape with a farmer’s eye: while everyone was looking up at the rockets, was there still a field worth planting in?

Reading the IPO for What It Actually Is

When you open SpaceX’s filing, the revenue breakdown is the first thing worth pausing on. Of $18.7 billion in 2025 revenue, launch services accounted for $4.1 billion, roughly 8% of the total. Connectivity (Starlink) generated $11.4 billion, or 61%. The company we have been calling a “rocket company” derives most of its revenue from satellite internet.

That framing matters for how you read this listing. Read the revenue mix at face value, and this looks less like a launch vehicle IPO and more like a satellite internet infrastructure IPO. The interpretation is worth considering.

SpaceX’s center of gravity appears to be shifting from launch toward connectivity — toward what happens after the satellites are in orbit. Launch increasingly looks like an internal logistics function: a way to put Starlink satellites up faster and cheaper. In 2025, SpaceX conducted 165 launches, of which only 43 served external customers. The other three-quarters were for its own constellation. Launch frequency is growing explosively, but an increasing share of that cadence is pointed at one destination: Starlink.

There are ways to gain exposure to SpaceX directly. Several funds have been accumulating pre-IPO shares, each with different exposure levels and structures. Here is a quick comparison:

• BPTIX (mutual fund): direct equity stake, approximately 32% exposure as of late April, though large inflows through May have diluted that share considerably.

• XOVR (ETF): indirect exposure via SPV, approximately 23%, the highest single-name SpaceX exposure among publicly traded ETFs.

• RONB (ETF): direct equity stake, was above 20% at peak, but inflows have diluted the position over time. Annual fee of 1.0%.

• DXYZ (closed-end fund): indirect via SPV, largest single holding at roughly 10%. Trades at a meaningful premium or discount to NAV depending on market sentiment.

• NASA / UFO / ARKX (ETFs): diversified space economy exposure, low single-digit SpaceX allocation. UFO and ARKX are expected to add SPCX after the listing.

There is a structural trap in most of these vehicles worth noting. Private shares cannot be accumulated quickly. As capital pours into popular funds, SpaceX’s weighting actually declines, because the denominator grows faster than the numerator. You buy in for the exposure and then watch that exposure shrink. It is a genuinely odd dynamic.

Two additional headwinds apply across the board. The scarcity premium that came with being a private company evaporates the moment the stock lists. And regardless of which vehicle you use, you are ultimately paying the $1.75 trillion headline valuation.

If the core position is already expensive and the exposure is diluting or trading at a premium, the more interesting question is: who supplies the field? What grows whether SpaceX wins or whether its competitors do?

The Two Fields: A Value Chain Map

The value chain SpaceX is reshaping splits into two broad areas. The first is the launch and materials field, where demand grows with every additional rocket that flies. The second is the connectivity and terminal field, where demand grows as more satellites go up and more people need to receive that signal.

The Launch and Materials Field

Rockets and satellites either get used once or need to be continuously replenished. Every additional launch means more demand for specialty metals and high-frequency components. Think of it as fertilizer and equipment: the more often you plant, the more you use.

The critical question here is whether a company’s exposure is confirmed by contract or merely expected. There is a meaningful difference between a company that appears in SpaceX’s public filings as a supplier and one that is simply assumed to benefit.

Four names are worth tracking in this field:

• HVM Co., Ltd. (295310 KS) earns three stars. This is the closest thing Korea has to a verified direct supplier of superalloy castings for SpaceX engine components. HVM has internalized a complex vacuum melting process and received Tier 1 flight materials supplier certification in March 2026, bypassing distributors entirely. Space-related revenue has grown from 3% to 56% of the total in three years. The underlying business is as confirmed as it gets in this value chain.

• Filtronic (FTC LN) also earns three stars. A UK-based maker of high-power GaN radio frequency amplifiers for satellite applications. SpaceX holds an equity stake in the company, which says something about the depth of the relationship.

• Spear (347700 KS) earns two stars. A superalloy supplier with a widely cited nominal 10-year supply agreement. However, reading the contract more carefully, the headline figures reflect demand forecasts rather than guaranteed volumes, and operating margins have recently collapsed to the low single digits. Same “SpaceX alloy” label as HVM, meaningfully different contract quality.

• Wistron NeWeb (6285 TT) earns two stars. The primary contract manufacturer for Starlink user terminals. Hardware exposure, not software.

The Connectivity and Terminal Field

Even a fully deployed satellite constellation is useless without antennas and chips to receive the signal. This is where SpaceX’s center of gravity has already shifted.

The largest chokepoint in this field, smartphone chips, is already controlled by Qualcomm and MediaTek. But a careful look at those companies makes clear that satellite is a feature, not a driver. What actually moves their stocks is AI data center demand. Satellite NTN capability is embedded in their chips as a by-product of the broader cellular roadmap.

The more interesting exposure sits where the giants have not bothered to go directly: antenna systems and ground stations, and emerging satellite-direct chipsets.

• Intellian Technologies (189300 KS) earns two stars. Intellian does not supply Starlink. It supplies Starlink’s competitors: OneWeb, Telesat, Amazon Kuiper, AST. Low-earth orbit revenue now exceeds 64% of the total, and the customer base covers essentially every non-Starlink LEO operator. That neutrality is a structural hedge: Intellian wins as long as any non-Starlink operator succeeds. The flip side is that revenue visibility is low, because all the variables sit on the customer side. Q1 operating margin dropped to 0.9%, which is either a reflection of US factory ramp costs or a structural problem. Q2 recovery will be the deciding data point.

• ConTec (451760 KS) earns two stars. A ground station services provider, operator-neutral, now profitable. The caveat is structural: ConTec consolidates a subsidiary it only partially owns, which means consolidated profit figures overstate what flows to equity holders. The discount in the stock may be less about the business and more about the market correctly reading that accounting structure.

• GCT Semiconductor (GCTS US) earns one star. The most early-stage name in the group. GCT is targeting the satellite-direct-to-device chipset market, and confirmed production orders have been announced. First shipments are expected in the second half of 2026. The company has disclosed going-concern risk and trades with thin liquidity. The moat exists in concept; the toll booth has not yet opened.

• Satrec Initiative (099320 KS) earns two stars. Satellite manufacturing and earth observation data. Solid business, but the investment thesis is more about Korean government and export contracts than the SpaceX value chain directly.

• AST SpaceMobile (ASTS US) earns two stars. The pure-play direct-to-device satellite communications name. Full bull thesis is well understood by the market.

• Qualcomm / MediaTek / Semtech are included as reference points. Their NTN chip capabilities are real, but space is not the earnings driver. The data center narrative is.

(Star ratings reflect the author’s qualitative assessment of contract confirmation, competitive moat, and technology substitutability. Tickers follow Bloomberg convention.)

The pattern visible across both fields: three-star names tend to be the ones the market already knows about. The less-covered opportunities sit one level below, among names where the contract is just being confirmed or where operator-neutral positioning has not yet been fully appreciated.

Where the Real Bottleneck Is

Read more

In our last piece, we tracked how Korean front-end equipment names closed the valuation gap with their global peers. Jusung Engineering surged over 80% in three months. The thesis played out. That trade has moved from discovery into management territory.

So where does the next opportunity lie?

In March, at Nvidia’s GTC 2026, SK Group Chairman Chey Tae-won made a statement worth dwelling on. “The supply shortage stems from a wafer shortage,” he said. “Securing more wafers takes at least four to five years. We expect industry-wide supply shortages of more than 20% to persist through 2030. That said, we don’t want chip prices to spike from the shortage either.”

The important part isn’t that the shortage continues. It’s that there is no quick fix. Building cleanrooms, securing wafers, installing equipment, ramping yields — all of that takes four to five years, no matter how aggressively Big Tech accelerates CAPEX. The physical timeline cannot be compressed.

So semiconductor manufacturers have made a choice: rather than racing to build new fabs, they are squeezing more out of the fabs they already have. More advanced process nodes. Higher yields. Better chips from the same wafer. This is the process upgrade cycle now underway.

That’s the starting point for this piece. In a world of full-capacity fabs and generational node transitions, there is a layer of the supply chain that benefits more quietly than equipment makers — but arguably more reliably. Materials and consumables.

Equipment sells once. Materials and consumables are consumed every quarter, for as long as the fab is running. And as process sophistication increases, they are consumed faster.

1. Two Engines Driving Demand

To build the case that materials and consumables demand will grow meaningfully over the next year or more, it helps to start with the macro picture.

The first engine is volume.

Global 300mm silicon wafer shipments rose 13.1% year-over-year in Q1 2026 to 3,275 MSI (SEMI, Q1 2026). Demand for 300mm wafers is particularly strong, driven by AI data center investment. SEMI projects global 300mm fab equipment spending to reach $133 billion in 2026, up 18% from the prior year. More lines running, more area covered — and materials and consumables scale proportionally.

This engine is cycle-dependent. If memory prices roll over or Big Tech pulls back on CAPEX, it softens.

The second engine is process sophistication — and this is the one that matters.

According to Bank of America estimates on SK Hynix’s 2026 capex, roughly 40% of total spending goes toward upgrading existing lines to the latest process nodes. The remaining 60% is split between infrastructure (30%) and net new capacity additions (30%).

That 40% process conversion figure is significant. It increases materials and consumables consumption without adding a single wafer to total output.

The numbers are striking. In DRAM, the number of ALD (atomic layer deposition) steps per wafer has grown sharply with each generation — from 2 to 3 steps in the 1x/1y era, to 8 to 9 steps in 1a/1b, to more than 10 steps in 1c (SK Securities Research). Every additional ALD step means more precursor chemicals consumed, and faster wear on the quartz parts and SiC focus rings inside the chamber.

The same dynamic plays out in HBM. HBM4 requires 33% more wafers per unit than HBM3E, as I/O bandwidth doubles and foundry process steps are added. In NAND, each successive layer stack increases plasma energy intensity, shortening the replacement cycle for focus rings and other consumables.

This engine is structural. It holds as long as process miniaturization continues.

To put a rough size on it: the global fab equipment market is roughly $133 billion. Memory accounts for approximately 36%, or around $48 billion. Applying the 40% process conversion share to that memory figure implies roughly $19 billion in structural consumables-linked demand — the kind that repeats every quarter regardless of whether new equipment orders are placed. It doesn’t show up clearly in traditional CAPEX analysis, which is precisely why it tends to be underappreciated.

Both engines are running simultaneously right now. That makes this an unusually clear moment to look at the materials and consumables layer.

2. The Value Chain Map

Process sophistication increases consumables demand differently at each step of the fab flow. Here is how demand builds, following the wafer through the process sequence.

Read more

Lately I’ve been telling a slightly different kind of story here. Tracking the big macro flows — semiconductor supply chains, energy transitions, the kind of themes where trillions in government capital move markets — that’s the core of what this channel does. But sometimes, separate from those large structural narratives, there’s a company you’ve been quietly watching for a long time, and one day something clicks into place. Today is one of those days.

Let me tell you about a company called $GCTS.

I’ve been following this company since last autumn. It’s a Nasdaq-listed fabless semiconductor company founded by CEO Kyungho Lee. The fact that a Korean entrepreneur built and listed a semiconductor company in the United States was enough to catch my attention. But once I understood the relationship between $GCTS and Anapass (KRX: 123860, KRW), I started tracking it far more closely.

Anapass is a fabless company that supplies OLED display driver chips to Samsung Display. Korean investors will likely recognize the name. What’s less well known is that Anapass has been lending hundreds of billions of Korean won to GCTS over the years, effectively serving as its financial lifeline. The two companies are legally separate, but financially they are deeply intertwined.

What I was waiting for was simple: GCTS’s 5G chipsets actually selling. If real shipments materialized, GCTS would rise on its own merits — but more importantly, the discount weighing on Anapass, caused entirely by its GCTS exposure, would begin to lift.

While waiting for that news, I looked more closely at GCTS’s peer group. And in doing so, I stumbled onto something unexpectedly strange.

That’s the story I want to tell today, from the beginning.

Read more

On May 14th, President Trump is heading to Beijing. It’s the first visit by a sitting U.S. president to China in nearly eight years — since November 2017. The signal is clear: both sides want to build on the partial rapprochement reached at the Busan summit and take things a step further.

Every investor’s mind naturally drifts to the same question: which themes and companies stand to benefit?

But that question has a trap built into it. Nobody knows whether the talks will be delayed again, or what depth of agreement will actually emerge. Tariff rollbacks could accelerate. A permanent lifting of rare earth export controls might be announced. Or the Iran variable could muddy the waters enough that the whole thing amounts to little more than a directional statement.

Predicting the outcome of these negotiations is no different from trying to forecast the weather. A farmer doesn’t predict the weather. A farmer picks seeds that grow no matter what the weather does. That’s what this piece is about — finding the fields that keep growing, whether the talks end in détente or escalation.

[1] Where We Are Now: The Current Coordinates of the U.S.-China Conflict

The arc from early 2025 to today breaks into three phases.

The first was collision. Shortly after Trump’s second inauguration, the U.S. began ratcheting up tariffs on Chinese goods. China retaliated in kind. U.S. tariffs on Chinese imports peaked at 145%; China’s tariffs on U.S. goods hit 125%.

The second was a ceasefire. At the Geneva talks in May 2025, both sides agreed to a 90-day tariff reduction. U.S. tariffs dropped from 145% to around 30%; China’s from 125% to roughly 10%.

The third was partial resolution. The Busan APEC summit in October 2025 brought the two leaders face-to-face for the first time in over six years. They agreed on further tariff cuts, a one-year suspension of China’s rare earth export controls, and a resumption of Chinese purchases of U.S. soybeans.

That’s the surface story. But beneath it, something more structural has been happening. While tariffs were coming down, China was quietly sharpening its export controls on gallium, germanium, antimony, graphite, tungsten, and LFP battery technology. The real battlefield has shifted from tariffs to resource weaponization.

And there’s a hard deadline attached to all of this. The export control suspensions agreed at Busan expire on November 27, 2026. Trump’s Beijing visit and the U.S. midterm elections fall in the same month. The timing is not coincidental. I’m using this trip as a trigger to map the entire strategic minerals supply chain.

[2] What’s Actually on the Negotiating Table

Tariffs have already been substantially adjusted. The real bargaining chips now are commodity-level export controls. Here’s how China’s leverage breaks down across key materials, scored by importance (how structurally critical the material is) and urgency (how close to the deadline or how fast the situation can escalate) — each on a scale of 1 to 5.

Gallium (GaN/GaAs feedstock): China controls 80% of global production. Controls began August 2023, suspension expires November 27, 2026. Importance: 5. Urgency: 5. Agreement odds: moderate.

Graphite (battery anode material): China refines 80%+ of global supply. Controls began October 2023, suspension expires November 27, 2026. Importance: 5. Urgency: 5. Agreement odds: moderate.

AI semiconductor export controls (H200 and above): Controls have been in place since 2022, currently under conditional easing discussions. Importance: 5. Urgency: 4. Agreement odds: conditional.

Germanium (optical communications, semiconductors): China controls 60% of production. Controls began August 2023, suspension expires November 27, 2026. Importance: 4. Urgency: 5. Agreement odds: moderate.

Tungsten (semiconductor interconnects): China controls 80% of production. License system introduced February 2025, ongoing. Importance: 4. Urgency: 4. Agreement odds: low.

Antimony (defense, batteries): China refines 80% of global supply. Controls began September 2024, suspension expires November 27, 2026. Importance: 3. Urgency: 5. Agreement odds: moderate.

LFP battery technology: Manufacturing process controls introduced in 2025, currently under negotiation. Importance: 3. Urgency: 3. Agreement odds: low.

Agricultural goods (soybeans, grains): Busan partial agreement already in place, expanded purchases under discussion. Importance: 2. Urgency: 3. Agreement odds: high.

Aircraft (Boeing orders): Suspended, currently in discussion. Importance: 2. Urgency: 3. Agreement odds: high.

The asymmetry in this table tells the real story. The items with the highest strategic importance and urgency — gallium, graphite, AI chips — are the hardest to resolve. The items most likely to yield quick agreement — agriculture and aircraft — are the least strategically significant.

That asymmetry reveals the actual structure of the trade. Agricultural goods, aircraft orders, and relaxation of strategic mineral controls are the cards the U.S. wants to take home. Additional tariff cuts and access to AI semiconductors are what China wants in return. Even China’s willingness to extend the export control suspensions isn’t purely a concession — the fact that a suspension needs to be extended at all is itself a demonstration of China’s leverage. This exchange is the core deal structure of the Beijing summit.

[3] The Question That Matters More Than the Outcome

After looking at the negotiating table, the instinct is to think: “If gallium gets decontrolled, buy GaN plays. If graphite tightens, buy de-China graphite companies.” That’s not wrong. But as I said earlier, trying to profit by predicting negotiation outcomes is structurally the same as weather betting.

Let me reframe the question.

“Rather than asking which direction material prices will move, ask whether the structural demand for what those materials make is growing.”

Whether gallium gets decontrolled or tightened, the fact that 5G base stations and AI data centers need GaN semiconductors doesn’t change. With that lens, I’m grouping companies into three categories.

• Group A: Structural beneficiaries regardless of negotiation direction. These companies win on easing (lower input costs) and on tightening (higher pricing power as alternative suppliers). End demand is the constant.

• Group B: Direct beneficiaries if talks ease. These companies get a direct revenue or margin tailwind from a favorable outcome.

• Group C: Thesis strengthens if talks deteriorate. These are de-China supply chain plays — compelling on escalation, diluted on détente.

Seeds that grow in rain or drought. Seeds planted when rain is coming. Seeds built for drought. Which you choose depends entirely on how confident you are about forecasting the weather.

[4] The Fields That Survive Either Outcome: Group A

The common thread across China’s weaponized materials: China controls the supply, but the demand for what those materials make grows independently of China. For Group A companies, the direction of negotiations doesn’t change the direction of the business. Easing stabilizes input costs; tightening strengthens their position as alternative suppliers.

Read more

Editor's note: This research was originally published on April 11, 2026, for Korean subscribers of Growth Wave. Since publication, the companies highlighted in this analysis have appreciated significantly. Our top pick, Sanill Electric (KRX: 062040), has returned +143% from the April 6 price. Samsung SDI (KRX: 006400) has returned +65%. Fluence Energy [$FLNC] is still searching for its fair value. We're sharing this piece in full as a window into how Growth Wave research works. All valuation figures and stock prices reflect April 6, 2026 levels.

For years, battery storage was treated as solar’s companion — a tool to smooth intermittency and stabilize the grid. But in 2024, the EIA made it official: battery storage now appears as an independent category in new capacity statistics, alongside solar, wind, and natural gas. Not a supplement. A power plant.

Do you see the black dot labeled Battery Storage — accounting for 24.3% of projected 2026 new US generating capacity?

In 2025, utility-scale battery storage added 15,775MW in the US. That’s +58.4% year-over-year. The 2026 plan calls for 24,268MW — another +56.7%. Combined with solar and wind, these three sources account for 99.2% of US net new generating capacity in 2026. Fossil fuels and nuclear: effectively zero net addition.

One thing worth clarifying upfront. The OBBBA (One Big Beautiful Bill), signed in 2025, significantly accelerated the ITC phase-out timeline for solar and wind — leading many market participants to conclude that clean energy incentives broadly had been cut. But standalone BESS is different. Battery storage systems were treated separately in the legislation, with full ITC maintained through 2033. Unlike solar and wind, BESS avoided the policy headwind. In fact, the solar and wind deadline rush is structurally pulling BESS demand up with it.

This deep dive sets solar aside for now and goes one layer deeper into BESS — Battery Energy Storage Systems. As ESS graduates from grid stabilization tool to independent power asset, we’ll map where capital has flowed first across the value chain, where it’s concentrating now, and where the market may not yet be pricing things correctly.

[1] Spring 2026: Where BESS Stands Today

What the data says

As noted above: 15,775MW added in 2025, 24,268MW planned for 2026. Over the same period, natural gas added just 4,152MW of new capacity. Batteries added 3.8x more than gas in a single year.

Per EIA STEO projections, cumulative US utility-scale battery storage capacity will reach 64.9GW by end of 2026 — nearly four times the ~17GW figure from early 2024, achieved in just three years.

Europe is moving in the same direction. The EU installed 27.1GWh of new battery storage in 2025, up 45% year-over-year. Utility-scale crossed 55% of the total for the first time. Germany alone has a backlog of 500GW in pending BESS grid connection applications.

And as noted: standalone BESS retains full ITC through 2033. The solar and wind deadline rush is pulling BESS demand higher alongside it.

What the field is saying

Beyond the numbers, signals from the ground point the same direction.

The most notable shift in the supply chain is in China’s ESS cell market. Major manufacturers are reportedly booked through Q1 2027, and spot cell prices have begun recovering from their 2025 lows. In Q1, lithium carbonate, copper, LiPF6, and 314Ah cell prices all rose in tandem across China’s ESS supply chain. What’s notable is that the industry appears to be accepting — not resisting — these price increases. That’s a signal that demand-side economics have reached a threshold of viability.

Similar signals are coming from upstream in the raw materials chain. The CEO of Ganfeng Lithium, the world’s largest lithium metal producer, publicly stated that an “AI data center-driven ESS supercycle has begun.” This is a supplier making the statement, which warrants some caution. But when a raw materials producer uses language this strong, it suggests demand visibility has reached a sufficient level to justify it.

Corporate actions confirm the same direction. Korea’s battery trio is investing approximately $20 billion to expand North American ESS capacity, targeting a 30% market share this year. LG Energy Solution’s ESS revenue is forecast to roughly triple year-over-year in 2026. LG Group Chairman Koo Kwang-mo personally visited the US Vertec facility and emphasized moving “beyond battery supply toward integrated solution capabilities” — a clear signal of intent to extend from manufacturing into services.

One paradox: How to read the Tesla miss

Tesla deployed 8.8GWh in Q1 2026 — down 38% from Q4 2025, and 39% below Wall Street consensus of 14.4GWh. YoY: -15.4%.

This should not be read as a demand signal for the BESS market. Global ESS installations grew +45% YoY over the same period. Tesla’s energy business posted five consecutive quarters of record gross profit through Q4 2025. Whether the Q1 deployment miss reflects a structural issue or a timing gap was a question for the April 22 earnings call.

[2] The BESS Value Chain Map

A battery storage system is not a single cell. Four distinct layers must stack in sequence — materials and components, system integration, grid connection infrastructure, and operating software — before electrons can flow in and out of the grid. As with our previous deep dives, we’ll analyze how capital flows through each layer, how quickly, and where the moats are being built.

L1. Materials, Cells & Components

Moat strength: ★★ (cells under commoditization pressure; select specialized components ★★★)

The physical foundation of BESS. LFP (lithium iron phosphate) chemistry has become the de facto standard for utility-scale ESS. Cell manufacturing appears to have entered a supply-constrained phase — but the same pattern that played out in solar modules could repeat: a wave of new capacity coming online eventually tips back into oversupply. Structural components, thermal management systems, pouch films, and cathode materials are growing alongside cell demand.

Two positions stand out in this layer. First, cell makers (LG Energy Solution, Samsung SDI, CATL) — direct beneficiaries of the ESS demand shift, but with EV businesses mixed in, reducing ESS-specific exposure. Second, ESS-dedicated component and materials companies (Hanjung N.C.S., Seojin System, Shinsung E&T, Yulchon Chemical) — with higher ESS revenue concentration and established lock-in structures with key customers. L&F sits between the two, with a dedicated LFP cathode factory under preparation.

On the supply side, non-Chinese ESS cell suppliers in the US market narrow to three: LG Energy Solution, Samsung SDI, and SK On. As FEOC regulations tighten and tariffs accelerate Chinese exclusion, demand concentration toward these three is a structural scenario worth tracking — less as a near-term bottleneck, more as a medium-term structural tailwind.

Hanjung N.C.S. (KRX: 107640) ESS Revenue: 77.3% · Concentration: ★★★★

Seojin System (KRX: 178320) ESS Revenue: 52% · Concentration: ★★★

Shinsung E&T (KRX: 036710) ESS Revenue: est. 60%+ · Concentration: ★★★

Yulchon Chemical (KRX: 008730) ESS Revenue: est. 50%+ · Concentration: ★★★

LG Energy Solution (KRX: 373220) ESS Revenue: 3x target 2026 · Concentration: ★★

Samsung SDI (KRX: 006400) ESS Revenue: Record Q4 2025 · Concentration: ★★

L&F (KRX: 066970) ESS Revenue: ESS-dedicated LFP · Concentration: ★★

CATL (HKEx: 3750) ESS Revenue: ~21% (H1 2025) · Concentration: ★★

L2. System Integration

Moat strength: ★★★ (track record-based contract credibility, software lock-in potential)

This layer procures cells, PCS (power conversion systems), BMS (battery management systems), and transformers — then integrates them into a functioning system, taking responsibility from design through commissioning and performance guarantees.

As cell prices fall and project volumes (Q) increase, contract flow is concentrating toward proven integrators.

Fluence is an ESS-dedicated integrator with a growing software-based recurring revenue stream. Tesla holds a strong North American position via Megapack, though Q1 2026 results introduced uncertainty. Eos Energy differentiates with aqueous zinc battery technology (non-lithium), but carries significant balance sheet risk. Sungrow’s ESS segment exceeded solar inverters for the first time in 2025, becoming its largest business unit, and is expanding rapidly in non-US markets.

Fluence Energy (FLNC) ESS Revenue: 100% · Concentration: ★★★★★

Eos Energy (EOSE) ESS Revenue: 100% · Concentration: ★★★★

Tesla / Megapack (TSLA) ESS Revenue: ~10–15% of total · Concentration: ★★

Sungrow (SZ: 300274) ESS Revenue: 41.8% (2025 annual) · Concentration: ★★★★

L3. Grid Connection Infrastructure

Moat strength: ★★★★ (custom manufacturing, long lead times, certification barriers)

This layer physically connects completed BESS systems to the power grid. Transformers and switchgear are the critical components. Even when cells are abundant, a project cannot go live if this layer is constrained. In our assessment, this is where the most severe physical bottleneck is currently forming.

Korea’s three major power equipment companies (HD Hyundai Electric, Hyosung Heavy Industries, LS Electric) specialize in high-voltage transmission transformers. ESS projects require a different product: specialty distribution transformers. The global field here narrows to three suppliers: Prolec GE (Mexico-US joint venture), TMEIC (Japan), and Sanill Electric (Korea).

Sanill Electric — one of our preferred names — stands out for three reasons. First, there are no domestic competitors. Smaller companies cannot clear the technical bar; larger companies find the per-unit price (~KRW 150M) too low to justify entry. Second, FEOC regulations and tariffs affect Sanill Electric less than its global peers. Prolec manufactures in Mexico; TMEIC is Japanese. Sanill’s ESS specialty transformers are delivered FOB per project — the tariff burden falls on the buyer, not the supplier. Third, Sanill maintains trading relationships of up to 26 years with global inverter companies including GE, TMEIC, and Siemens — making near-term displacement by new entrants structurally difficult.

As of 2025, ESS and renewables specialty transformers account for 64.1% of total revenue, with operating margins sustained at 36%.

Sanill Electric (KRX: 062040) ESS/Renewables Revenue: 64% · Concentration: ★★★★

Hyosung Heavy Industries (KRX: 005440) HV transformer primary; ESS expansion planned · Concentration: ★★

L4. EMS Software & Operations Optimization

Moat strength: ★★★ (data accumulation-driven network effects; still early stage)

The software layer that decides when the battery charges and when it discharges. As BESS asset scale grows, the value of optimization algorithms grows with it. The market may not yet be pricing this layer adequately.

Stem Inc. (STEM) ESS Revenue: 100% · Concentration: ★★★★★

Stem is an ESS operations software specialist, executing a software transition centered on its AI-based EMS platform, PowerTrack. Over 30GW of managed asset data is the core competitive advantage. The path to profitability is still in progress — worth keeping in mind.

[3] Bottleneck Analysis: Demand Is Strong, Execution Is Stuck

BESS demand is unambiguous. EIA data, corporate backlogs, and field orders all point the same direction. The friction is in execution.

L2 faces factory ramp-up timelines and tariff uncertainty. L3 faces physical supply constraints on transformers. L4 faces limited market awareness. Of these, L3 bottlenecks are likely to persist the longest — physical constraints don’t resolve through policy or intent alone.

L1 Materials & Cells Not currently a bottleneck. Medium-term structural tailwind. Resolution speed: Fast

L2 System Integration Factory ramp-up, tariff and policy uncertainty. Resolution speed: Medium

L3 Grid Connection Physical supply constraint. Resolution speed: Slow

L4 Software Early market awareness, pre-profitability. Resolution speed: Long-term

[4] Timeline: Now Through 2027

2025 — confirmed 15,775MW added in the US (+58.4% YoY). Year One of ESS as an independent power asset. OBBBA passes House — solar/wind ITC acceleration debate begins. Real driver behind 2025–2026 installation rush.

H1 2026 LG Energy Solution North America capacity reaches 60GWh. Samsung SDI SBB 2.0 US mass production begins. Seojin System and Hanjung N.C.S. US factories go live.

2026 FEOC regulations tighten. Chinese exclusion accelerates. US battery: 24,268MW planned. Cumulative 64.9GW projected.

2027 IRA tax credits begin phasing down for solar/wind. BESS and battery cells retain separate incentives.

The most events converge in 2026 — cell supply localization, Korean component factories going live in North America, FEOC tightening, and installation records, all in the same year. Solar and wind tax credits begin phasing from 2027, but BESS retains its incentives. Given that renewable economics are already viable without subsidies, the pace of installation is unlikely to slow materially. AI data center demand provides additional buffer.

[5] Valuation

How to read this table

Companies are grouped into four types based on their value chain position and business characteristics. Type A: mature growth; Type B: cyclical; Type C: transitional growth; Type D: early stage / turnaround.

The fair value midpoint represents a conservative scenario. The upper target represents the convergence point of an aggressive but logically defensible scenario.

These figures show you where the current price sits on the map — not a timing call. For companies with clear financials, think of these as an early look at the target prices that will appear in future reports. For companies with fragile balance sheets, dilution risk is pre-discounted, which is why numbers may come in below market consensus. That is intentional.

All prices as of April 6, 2026.

L1. Materials, Cells & Components

Yulchon Chemical: Established as the primary supplier of LFP battery pouch film for LG Energy Solution’s ESS line, currently in a sole-supplier structure. The Poseung new factory is ramping up supply volumes. Customer diversification efforts by LG ES warrant ongoing monitoring.

Samsung SDI: Signed a large-scale LFP battery supply agreement with a major US energy company in December 2025. The only non-Chinese company capable of supplying prismatic-format batteries in the US — a meaningful differentiator. H2 2026 US local LFP production launch and European shipment timing are the key variables for H2 results.

LG Energy Solution: North American local production capacity expansion to 60GWh is in progress. The structurally strongest non-Chinese position for large-scale LFP ESS production in North America. The quarter when ESS revenue first exceeds EV revenue is the next key milestone.

Hanjung N.C.S.: ESS revenue at 77.3% — the highest purity in this list. Official confirmation of Indiana factory operations is the next judgment trigger.

Shinsung E&T: Directly linked to LG Energy Solution’s North American ESS line. Direction is right. Current profitability metrics don’t support the price level. 2026 operating profit turning positive is the threshold condition.

Seojin System: ~KRW 2 trillion long-term supply agreement with SK On in place; Fluence partnership ongoing. Direction is clear. This valuation conservatively reflects ongoing CB conversion and net debt level. The narrative is real, but the balance sheet fragility sits underneath it.

L&F: LFP cathode dedicated factory and large-scale Samsung SDI ESS supply — direction is clear. But the current price appears to already reflect full realization of the optimistic scenario. Confirming Q3 LFP mass production proceeds on schedule is the prerequisite.

L2. System Integration

Fluence Energy: $5.5B backlog covers FY26 revenue guidance 1.6x — revenue visibility is solid. Current price is near fair value midpoint. Q1 FY26 gross margin came in at 4.9%, with H2 weighted toward 11–13%. Q2 gross margin recovery is the single most important indicator to watch.

Tesla / Megapack: Megapack is a core player in the BESS value chain. However, Tesla doesn’t lend itself to fundamentals-based valuation in this framework — a substantial portion of the current price reflects market conviction on robotaxi, FSD, and Optimus.

Eos Energy: Aqueous zinc battery technology with no lithium dependency, lower fire risk, and long-duration storage advantages — the technology direction is sound. But the balance sheet makes investment difficult. Warrants and convertibles outstanding represent potential dilution equivalent to 64% of current shares. Dilution risk dominates the price more than the technology does. Worth watching; not yet actionable.

L3. Grid Connection

L4. EMS Software

Stem Inc.: Current price sits ~14% above the fair value midpoint. ESS operations software specialist; 30GW+ of managed asset data is the core moat. Software-based recurring revenue is growing steadily, but net debt of approximately $267M structurally pressures equity value. Debt resolution pace needs confirmation before the narrative.

[6] Sanill Electric: The First Slot to Fill as BESS Scales

Across the value chain map and bottleneck analysis, one name keeps surfacing. Let’s go one layer deeper into Sanill Electric — whether the moat is real, whether the growth path is realistic, and what the current price is actually reflecting.

Is the moat real?

GE: 15-year relationship. TMEIC: 26-year relationship. What those numbers mean in practice: clearing Toshiba-Mitsubishi’s 200-point qualification checklist took five years. KEPCO’s checklist runs 30–50 items. Dozens of Korean companies have attempted US market entry in this segment. Sanill is the only one that succeeded.

The company’s IR materials explicitly state: “ESS/renewables specialty transformers: no domestic competitors.” Smaller companies cannot clear the technical bar; larger companies find the ~KRW 150M per-unit price point unattractive. The high-voltage transformer market — where a single unit runs KRW 50–100B — and the specialty transformer market are the same industry in name only.

Global competitors: Prolec GE (Mexico-US joint venture) and TMEIC (Japan). Neither is immune to US tariffs and geopolitical risk. Prolec manufactures in Mexico; TMEIC is Japanese. Sanill’s ESS specialty transformers are delivered FOB per project — the tariff burden falls on the buyer, effectively removing tariff exposure for Sanill.

The moat shows up in the margin data. FY2025 annual operating margin: 36.1%. Q4 2025 standalone: 38.7%. In manufacturing, this level of profitability is not achievable without pricing power.

How much can it grow?

FY2025: Revenue KRW 501.9B, operating profit KRW 181.6B. From KRW 64.8B in revenue in 2021 — 7.7x growth in four years. FY25 new orders: KRW 566B, up 38.5% YoY. Q4 2025 orders: KRW 176.7B — the strongest quarter on record.

Growth follows three phases.

Today: ESS/renewables specialty transformers (~KRW 150M per unit) account for 64.1% of revenue. As order sizes scale from 50 units to 100–200 unit batches, production efficiency rises and pricing leverage follows.

2027–2028: Large-capacity data center transformers (~KRW 3B per unit — 20x current ASP) enter the portfolio. Development initiated by direct customer request from Nextera Energy and others; currently producing 10–20 units.

2028–2029: 154kV ultra-high-voltage transformers (~KRW 15B per unit — 100x current ASP). Domestic railway orders already on the books; PG&E engagement in the US underway.

Q scales with capacity expansion. P scales structurally with product mix upgrade.

FY26 in numbers

ItemBasisEstimateFY26E RevenueCAPA KRW 600B × 97% utilizationKRW 580BFY26E OPMNormal range 36–39%, specialty mix expansion37%FY26E EPSIndependent estimate from revenue/OPMKRW 5,601

Q4 2025 was a record quarter. Annualized, that implies near-95% capacity utilization. H1 utilization of 84% reflected early-stage ramp of the second factory — by year end, it had already recovered. Management guided toward KRW 600B in FY26 revenue. We applied conservative order-backed estimates rather than taking guidance at face value.

OPM dipped in Q3 2025 on a one-time cost; Q4 confirmed normalization. FY26 should see incremental ASP improvement as specialty transformer mix expands.

FY26E EPS of KRW 5,601 is 3% above the November estimate of KRW 5,440. The EPS itself hasn’t changed much. What’s changed is the current price and the multiple justification.

Why is the stock still here?

An honest answer is required.

Chairman Park’s quality-first management philosophy means capacity expansion is deliberately gradual — no more than one year of orders accepted in advance, factory additions of roughly KRW 100B per year in sequence. This discipline is the source of the 26-year customer relationships and the 38% operating margin. But it also structurally limits near-term earnings surprise potential.

There’s a second reason the market applies a discount. Hyosung Heavy Industries has US local manufacturing — structurally insulated from tariff risk. Sanill produces entirely in Korea and exports. That’s where the valuation gap comes from: Hyosung at ~46x PER vs. Sanill at ~23.9x. A concrete decision on US local factory investment would be the trigger to close that gap.

As of last November, the stock had approached our target price, and we shifted to Hold. Since then, the price has declined while EPS estimates have risen. That’s why the view has changed.

One more thing worth saying: Sanill doesn’t need the discount to resolve before it can move. Chairman Park’s quality-first philosophy means backlogs accumulate steadily, capacity expands incrementally, and ASP rises structurally. This is a company where results compound even without a catalyst.

We view this as a name that can be held over the long term without discomfort.

What would change the view?

Two conditions:

First: a second confirmed direct order from a data center developer. One order of KRW 43.8B is confirmed. A second would justify applying a ~38x forward multiple, pushing the fair value midpoint above KRW 244,000. Orders for units priced in the billions are a signal that the ASP structure itself is changing.

Second: a US local factory investment decision. Expected after Q1 2026 results are confirmed. When this is announced, the valuation gap versus Hyosung narrows and tariff risk is removed simultaneously.

If both triggers arrive quickly, the multiple re-rates. If they’re delayed, the results don’t stop compounding. Either way, the long-term hold case remains intact.

One additional note: in March 2026, the second-largest shareholder sold a 10% stake via block deal. This has no connection to company fundamentals. If anything, the previously limited float expanding through this transaction could be read as a positive — broadening the institutional shareholder base.

Chairman Park has not sold a single share since listing. The solar business operated through subsidiary Sanill Partners has not yet generated revenue and should be monitored separately from the core business.

The farmer in Florence looks at the soil before planting the seed. This soil has already been tested. Now we’re waiting for rain.

Closing

For a long time, we looked at ESS as an attachment to solar — a buffer sitting beside the panels. But what we’ve been mapping together tells a different story. The EIA has officially classified battery storage as an independent power source. Global ESS installations grew 84.7% in 2025.

Demand is genuinely strong. Both the US and Europe point the same direction. Yet the major battery companies’ share prices have spent most of this period near historic lows — because the market still sees them as EV companies, while the companies themselves are in the middle of repositioning.

Capital’s gaze has only just begun to shift. LG Energy Solution, Samsung SDI, Ganfeng Lithium — the suppliers are publicly acknowledging the change. Analyst coverage is beginning to follow.

That’s why this deep dive set solar aside and went directly into BESS. Many of the companies in this value chain don’t show large upside in isolation — partly because they’re still rebuilding balance sheets weakened during the EV cycle. But when the market begins to reclassify these companies as ESS companies, the multiples will change. That reclassification happens slowly through accumulating quarterly results — and then suddenly.

We mapped where the bottlenecks are. We mapped where the moats are forming. Getting there one step ahead of other investors — that was the purpose of this deep dive.

Ciao. 🌱 — Growth Papa, Florence

All content is for informational purposes only and does not constitute investment advice. All investment decisions are the sole responsibility of the reader. Past performance of specific securities discussed in this research does not guarantee future results.

First question. What does optical networking connect, and is optical alone enough?

Inside an AI data center, thousands of servers need to exchange data with one another. Copper cables handled this job until AI cluster scales grew large enough to expose their limits — signal degrades over distance, and heat builds up. Optical fiber solves this by transmitting data as light. Optical networking solves the road problem: server to server.

But no matter how wide you build the road, the factory stops if the warehouse can’t deliver materials fast enough. A GPU needs to pull data from memory to run any computation. When that retrieval speed can’t keep pace with GPU processing speed, you have a memory wall.

Optical networking is a road problem. The memory wall is a warehouse problem. These are different bottlenecks.

Second question. What gets worse as AI shifts from training to inference?

Building an AI model involves two stages. Training runs in batches — there is room to breathe. Inference is different. Every time a user continues a conversation, the GPU must access the entire prior context in real time. This context data, called the KV cache, must sit in the fastest available memory: HBM, stacked directly on the GPU package.

Because HBM is physically stacked on the GPU, capacity is constrained. The H100, today’s most capable GPU, tops out at 80GB of HBM. Running a state-of-the-art LLM through a long conversation pushes the KV cache past 150GB. The wall has already been breached.

HBM solved the first wall — bandwidth. In doing so, it revealed the second wall: capacity.

Third question. Are CXL and NVLink competing or complementary?

A natural objection arises: hasn’t NVIDIA already enabled memory sharing across GPUs through NVLink?

Looking closely, the two technologies solve different problems. NVLink accelerates GPU-to-GPU communication — making multiple GPUs behave as one. CXL expands memory capacity by building a memory pool outside the server, scaling into the terabyte range. They are not competing; they address different bottlenecks.

The more important distinction is this: NVLink only works within NVIDIA hardware. It does not connect to Google TPUs, Amazon Trainium, or AMD GPUs. As hyperscalers accelerate development of their own AI chips, the need for an open standard — one that doesn’t lock them into a single ecosystem — becomes structural. That is why Google, Amazon, Meta, and Microsoft have all joined the CXL Consortium and the UALink Consortium.

If optical networking emerged from the physical limits of copper, CXL is emerging from an industry-wide need to avoid ecosystem lock-in.

Fourth question. Can the memory wall narrative match optical networking?

Honestly, it is too early to say with confidence.

Optical networking had a real shortage. Demand was locked in at the hardware design stage, supply couldn’t keep up, lead times exploded, prices rose, and margins expanded.

CXL’s direction appears sound. Microsoft Azure deployed CXL memory to live cloud instances in November 2025 — not a demo, an actual deployment. But there is no shortage yet. The software stack hasn’t matured, and NVIDIA ecosystem customers can hold on without CXL for now. This looks closer to the early days of glass substrates than to optical networking. The right approach is to identify who the leaders will be before the inflection arrives.

Memory hierarchy map — the three walls appear in order

AI servers contain a memory hierarchy. Faster means smaller and more expensive; slower means larger and cheaper.

HBM sits directly beside the GPU — the fastest, but the most limited. A palm. DRAM and CXL form the next tier — slower than HBM but far greater in capacity. A countertop. CXL extends this countertop beyond the server itself. NVMe SSD comes next — slower still, but holds terabytes. A refrigerator. Storage is the slowest but practically unlimited. A warehouse.

The three walls appear in sequence across this hierarchy.

The first wall is bandwidth. HBM breaks it. This is the hottest segment right now. The second wall is capacity. CXL breaks it. This wall is just becoming visible. The third wall is the cost of data movement. Moving data between memory and compute consumes power. PIM — Processing-In-Memory — breaks this wall by bringing computation into the memory itself. Still in the standardization phase.

Just as HBM layers are physically stacked, this entire hierarchy represents the ordered sequence of technological evolution.

Technology sequence is investment timing

Looking at this moment honestly:

The first wall is at peak heat. SK Hynix, Samsung, and Micron are sold out on HBM through 2026. NVIDIA is cutting gaming GPU production by 30–40% to prioritize AI HBM allocation. Shortage is real and prices are rising.

The second wall is coming into view. Microsoft Azure’s November 2025 CXL deployment was real production. But there is no shortage yet. This is the direction-confirmation stage.

The third wall is still being standardized.

So what needs to happen for CXL to explode the way optical networking did? Five stages:

Stage 1 — First hyperscaler production deployment: Complete (MS Azure, Nov 2025)
Stage 2 — Deployment performance data published and adoption spreads: In progress
Stage 3 — CXL revenue appears in chipmaker earnings: Not yet
Stage 4 — Memory makers report CXL shipment volumes officially: Not yet
Stage 5 — CXL becomes standard server spec: Not yet

We are between Stage 1 and Stage 2. Optical networking exploded at Stages 3 and 4. The signals to watch: the moment Astera Labs’ Leo product becomes a headline item in quarterly earnings commentary, and the moment OpenEdge Technologies recognizes its first CXL royalty.

Fifth question. Who are the leaders in this cycle?

What struck me most in this research is that the current leaders are not the CXL companies.

The HBM leaders — SK Hynix, Samsung, and Micron — are generating revenue from HBM today while simultaneously preparing CXL modules. Astera Labs is already growing explosively on PCIe retimers while laying CXL on top as the next layer. The companies already making money in the field, while preparing the next layer — those are the leaders in this cycle.

First wall layer — HBM, active shortage

SK Hynix (000660) holds the number one position in HBM market share. It has supplied HBM3E to NVIDIA on a near-exclusive basis and leads the transition to HBM4. CXL module development is already underway, positioning it for the second wall as well. Q1 earnings are due today, April 23. The key risk is what comes after the HBM cycle peaks.

Samsung (005930) carries a perception of lagging in HBM, but the gap is closing quickly. Q1 memory division results significantly exceeded market expectations, and Samsung was first to mass-produce CXL DRAM modules. With DS, DX, SDC, and Harman all under one roof, the full picture requires a sum-of-the-parts view.

Micron (MU) holds the lowest HBM share of the three, but the trajectory is the same. Long classified as a commodity cycle stock and priced accordingly, that classification is now shifting. It has fully exited consumer memory and repositioned around AI data centers, pulling forward HBM4 mass production by one quarter. Its 2026 HBM allocation is fully contracted. The market may still be looking at yesterday’s Micron while the fundamentals are already somewhere else.

Second wall layer — CXL, early stage

Astera Labs (ALAB) is already growing explosively on AI server PCIe retimers — more than 32 units per NVIDIA server on average. Revenue grew 115% year over year in 2025. Its Leo CXL controller was the first to be deployed in production at Microsoft Azure. This is a company already generating revenue without CXL, now adding CXL as the next layer. That said, the current share price appears to have pulled forward a significant portion of that future growth. Concentration risk around a single major customer also warrants attention.

Rambus (RMBS) sells the blueprints for critical memory technologies — and the chips built from those blueprints. With 2,049 patents and low exposure to commodity cycles, it supplies IP across DDR5, HBM4, CXL, and PCIe 7.0. Steady compounding rather than sharp re-rating is the more likely profile. A useful counterweight to cycle risk within a portfolio.

Marvell (MRVL) is a familiar name from our optical networking work. Custom AI silicon and optical interconnects are the primary growth drivers today. The CXL switch is scheduled for sample delivery in the second half of 2026 — the next re-rating moment looks more like 2027–2028. Hyperscaler concentration remains a key risk.

OpenEdge Technologies (394280) is the only Korean IP company registered with the CXL Consortium. Its controller IP flows into memory makers as they ramp CXL products — a royalty structure that scales automatically as the market grows. The company is currently loss-making with a short runway. Financing risk is real: the royalty stream may not arrive before the balance sheet needs attention.

Between the second and third walls — storage controllers

Fadu (440110) designs SSD controllers for AI data centers — handling data that can’t fit in HBM or DRAM. It supplies AWS, Microsoft, and Meta, and grew revenue 112% year over year in 2025. The technology and customer base check out. Convertible bond maturities and potential additional financing create latent balance sheet risk.

Third wall layer — PIM, post-2028

Samsung and SK Hynix are jointly standardizing LPDDR6-PIM, which brings computation inside the memory itself. Still in standardization; mass production is not expected before 2028.

Signals that would shake this thesis

As long as NVIDIA continues strengthening NVLink, customers inside the NVIDIA ecosystem can hold on without CXL. If AI model efficiency improves faster than expected, the moment the second wall is actually felt could be pushed out. Memory cycle risk is also real — today’s shortage environment could eventually become oversupply as large capex programs come online. OpenEdge and Fadu face financing risk that could arrive before the thesis plays out. We are watching this one beat earlier than optical networking. Don’t forget that.

The walls are real, and the order in which they fall is already determined. Bandwidth first, then capacity, then the cost of moving data. Mapping which companies stand at which point in that sequence — that is where this theme begins.

Ciao. 🌱