Haelangdal
Founder Analyst
AI infra · macro · mega-theme research
Recent Reports
- Company·Jun 3, 2026·18 min
Marvell Technology — From Compute to Connectivity, the Optical DSP Ruler of AI Infrastructure
Marvell Technology is not a single-stock bet but the most direct exposure to the structural shift in which AI infrastructure's center of gravity moves from 'compute' to 'the connectivity that links compute.' As AI models balloon, the bottleneck moves from the chip itself to the connectivity linking chip to chip and rack to rack, and this shift grows custom ASICs and optical interconnects faster than GPUs. Marvell holds the rare combination of being the overwhelming No. 1 in the optical DSP merchant market — the brain of the optical module — with a 60-70% share, plus No. 2 in custom silicon. In June 2026, when Jensen Huang called it 'the next trillion-dollar company,' the stock jumped 32.5% in a day. But much of that growth is already in the price. What remains to be verified is whether the guided growth path materializes in quarterly results.
- Company·Jun 1, 2026·16 min
SanDisk — The Purest Beta of NAND Memory Re-Rated in the AI Era
SanDisk is not a single-stock bet but the purest exposure to the NAND flash memory cycle being re-rated in the AI inference era. Past NAND booms were driven by smartphones and PCs; this cycle is led by AI-inference enterprise SSD demand and a structural HDD shortage. Demand is firm while suppliers froze capacity, forming a seller's market where prices jump each quarter, and the biggest beneficiary is the pure play that does only NAND without DRAM. SanDisk's market cap swelled roughly 43-fold after its February 2025 spinoff from Western Digital. But much of that picture is already in the price. From here, returns hinge not on story but on which point of the cycle you enter.
- Research·May 30, 2026·14 min
Weekly Market Review: Week 22 (May Week 5)
May Week 5 (5/25-5/29) saw AI infrastructure stocks prove pricing power twice in one week. On 5/26, Snowflake jumped +36.5% on quarterly earnings — its largest single-day gain ever. On 5/29 AMC, DELL reported a Q1 FY27 big beat (revenue $43.8B, EPS $4.86, AI server $16.1B +757% YoY) and surged +15% after-hours. DELL also raised FY27 guidance to $165–169B revenue and $60B AI server (from $50B). Both moves were triggered by accelerating AI data and AI server revenue visibility, signaling that the first-stage AI infrastructure revenue has arrived in markets. KOSPI rallied +3.55% (+290.86p) on 5/29 to a record-high 8,476.15 — driven by Samsung shipping the first 7th-gen HBM samples, an imminent Jensen Huang Korea visit, and Iran-Hormuz deal optimism. US S&P 7,580, NASDAQ 26,972, and Dow 51,032 all closed at record highs, with VIX at 15.32. However, April PCE released 5/29 raised inflation reignition concerns — headline 3.8% hit a 3-year high while core 3.3% matched consensus. Key variables next week: 6/4 Broadcom Q2, 6/5 Lululemon, DELL pricing follow-through into 5/30, Jensen Huang visit, and KOSPI 8000+ stability.
- Company·May 27, 2026·34 min
Samsung Electro-Mechanics — The Top Korean Component Pick the Data-Center BOM Points To
Samsung Electro-Mechanics is the only Korean company exposed to all three items exploding in the data-center Bill of Materials (BOM) — MLCC, ABF substrate, and glass substrate. On the VR200 NVL72 rack BOM, after memory (+435%) came PCB (+233%), MLCC (+182%), and ABF substrate (+82%), and these three converge on one company. The benefit already shows in earnings — Q1 2026 topped KRW 3 trillion in quarterly revenue for the first time. But at ~73x 2027 estimated earnings, 3.6x global peers (~20x), being a good company and being a good price now are two different things.
- Company·May 26, 2026·15 min
Bloom Energy — The Fastest Fix for the AI Power Bottleneck
The real bottleneck for AI data centers is not chips but electricity, and Bloom Energy is the company that fills that power gap with SOFCs (Solid Oxide Fuel Cells) in 6~9 months. In the first quarter of 2026, revenue rose +130% YoY and GAAP operating income swung to a profit, crossing from loss-making growth to profitable growth. Anchor contracts — Oracle 2.8GW, AEP 1GW, Brookfield $5 billion, Nebius about $2.6 billion — prove this demand exists not as an estimate but as contracts. But with market cap up roughly 22x in 12 months and P/S above 30x, it is already an expensive zone, so the investment decision is not the target price but whether the logic holds.
- Research·May 26, 2026·15 min
Weekly Market Review: May Week 4
May Week 4 (5/18-5/22) saw NVIDIA's Q1 FY27 full beat (revenue $81.6B, EPS $1.87, Q2 guidance $91B) combine with Samsung's dramatic strike settlement to send KOSPI surging +6.55% (+472.46p) on 5/21, triggering a buy-side circuit breaker. U.S. markets extended their rally with S&P 500 posting its 8th consecutive weekly gain, closing 5/22 at 7,473.47 (+0.37%), while Dow hit an all-time high of 50,579.70. The week's pivot came on 5/20 (Wed) with NVIDIA earnings — 5/18-19 KOSPI had plunged -3.25% under pressure from 30-year treasury yields hitting 5.19% (19-year high), but NVIDIA's beat and Samsung's strike resolution simultaneously hitting on 5/21 reversed everything. AI supercycle fundamentals overwhelmed rate and labor dual risks. Next week: 5/29 PCE inflation data and US-China trade negotiation progress are key variables.
- Editorial·May 23, 2026·22 min
The Place Memory Took — Vera Rubin Rewrites the Data Center BOM
Morgan Stanley's estimate for the Vera Rubin VR200 NVL72 rack puts memory at 25.6% of a $7.8M total — 2.7x wider than GB300's 9.4%. Combining HBM4 (20.7 TB) with LPDDR5X (54 TB), the "HBM 15.5%" proposition evolves into "memory 25.6%." NVIDIA's Q1 FY27 triple beat (5/20) and KOSPI's record +8.42% session on 5/21 validated the proposition in real time. But filling the expanded ceiling depends on VR200 ramp timing and HBM4 / LPDDR5X pricing — potential, not guarantee.
- Company·May 22, 2026·22 min
PSK Holdings — The Small Window of Descum Is Opening
PSK Holdings is the only Korean equipment maker breaking Japan ULVAC's 30-year Descum monopoly in HBM and CoWoS back-end processing — and the only Korean supplier to all three memory players AND TSMC simultaneously. Shinyoung Securities' "Reflow demand per CoWoS capacity = 6× HBM" data has been under-cited by the broader market. TSMC CoWoS capacity expands from 40K/month in 2025 to 90K/month in 2026 (2.25×) — and this 6× multiplier lands in P&L during that window. But the company itself guided 2026–2030 operating margin to ≥20%, a ~15pp gap to 2025's 35.3%. The answer is "Yes, but."
- Company·May 20, 2026·28 min
NVIDIA Earnings D-1: Six Bullish Theses for the Next 24 Hours
This NVIDIA earnings print is the decisive variable that determines whether the Big Four's roughly $700B 2026 AI infrastructure CapEx guide is actually converting into NVIDIA revenue. The next-generation Vera Rubin GPU is heading into volume production in H2 2026, and the Korean supply chain rides four parallel beneficiary lines — memory, optics, power, and packaging. Consensus revenue $78.8B, Q2 guide consensus $86B, TTM P/E 41.28x, forward NTM PEG ~0.4 — the cheapest among megacaps. Q2 guide above $87B + 75% gross margin + Vera Rubin H2 timing reconfirmation — if all three converge, additional upside trigger activates.
- Research·May 19, 2026·24 min
MACRO 101 — Why Standing Still Costs You Money
US federal debt has reached $39 trillion and cannot be repaid directly. The core hypothesis of this report is that the Trump administration's policy package is building a Financial Repression 2.0 system that slowly liquidates someone's real assets to reduce the debt burden. On May 18, 2026 the US 30-year Treasury yield rose to 5.13%, roughly 5x its July 2020 level of around 1%. Japan's 30-year yield hit a record 4.19% — the highest since the bond's 1999 launch. Five categories of Treasury buyers — foreign central banks, China, Japan, the Fed, and US households (the latter exposed by the Fed itself as misclassified hedge-fund basis trades worth $1.4T) — have all stepped back. The GENIUS Act forces stablecoin issuers to become forced buyers of short-dated Treasuries, and the SEC's May 18 tokenization announcement expands stablecoin demand. Tether posted $10B profit in Q1–Q3 2025 and now holds $135B in Treasuries, ranking alongside Saudi Arabia and Korea as a top global holder. The same Financial Repression mechanism that took US debt/GDP from 119% in 1945 to 32% in 1980 is being reassembled through the GENIUS Act, SEC tokenization, tariffs, and AI capex policy. The bill is paid by four cohorts: stablecoin holders, bank depositors, long-duration Treasury holders, and unhedged foreign dollar holders. This report walks through seven chapters designed for first-time students of macro, finance, and crypto — the question is simply 'why is this happening now?'
- Editorial·May 17, 2026·20 min
AI Data Center BOM Reach Rate
Decompose the BOM (Bill of Materials) of an AI data center and the theoretical revenue ceiling each industry must reach reveals itself. Inside NVIDIA's NVL72 at USD 3.1M sit HBM USD 480K, CoWoS packaging USD 180K, optical USD 200K, cooling USD 50K — NVIDIA 61% · HBM 15.5% · TSMC 5.8% · optical 6.5% · cooling 1.6%. Scale to 1GW DC capex USD 38B and industry attribution becomes NVIDIA USD 18B · HBM trio USD 7.65B · TSMC USD 4B · optical USD 3.8B · power USD 3.5B. Of the USD 2,660B room from 70GW × USD 38B over 2026~2028, NVIDIA has recognized 20.2%, HBM 14.9%, optical 15.8%, power 10.2%. By BOM accounting identity no component can exceed its BOM share, so as NVIDIA moves toward 100% of the room HBM, optical, and power must follow at their shares. Part 3 of the AI Data Center Deep Dive series — Part 1 (the infrastructure split) and Part 2 (agentic penetration) explained the drivers of this capex; Part 3 reverse-engineers which industries the same capex flows into.
- Editorial·May 16, 2026·18 min
Reverse-engineering Agentic AI Penetration
Claude holds 6% of generative-AI web traffic — about 1/16 of ChatGPT — yet annualized run-rate revenue of $30B has overtaken OpenAI. Two species of users sit behind those numbers — the chatbot user the traffic gauge sees, and the token-heavy agent user the ARR gauge bills. This is the sequel to AI Data Center's Split (Part 1): it crosses Google token throughput, Similarweb web traffic, and OpenAI / Anthropic ARR to reverse-engineer the depth of agent penetration. 134x token growth, an 11.3x per-user ARR gap, and Claude Code's single-product ARR multiplier (to $2.5B in 14 months) anchor the evidence. Agent revenue penetration is around 36% for the two-company combined figure, against under 5% of traffic — the revenue curve follows tokens, not users. First-order beneficiaries are the TPU value chain (Alphabet, Broadcom, Amazon) and memory HBM (SK Hynix, Samsung, Micron); OpenAI exposure flows through Microsoft and Oracle.
- Research·May 16, 2026·15 min
Weekly Market Review: Week 20 (May Week 3)
Week 20 (May 10-16) was the week the KOSPI hit 8,000 for the first time intraday — then crashed -6.12% the same day. After closing at a record 7,981.41 on May 14 — the fourth straight all-time high — the KOSPI touched 8,046.78 in early trading on May 15, only to be hit by a record KRW 5.62 trillion of foreign selling that drove the index down 488.21 points; Samsung dropped -8.61% and SK Hynix -7.66%, and a sell-side circuit breaker (KOSPI 200 futures -5% in under a minute) triggered, erasing the week's entire gains in four hours. AI super-cycle fundamentals were solid — SK Hynix's forward PER overtook Samsung's for the first time, and KB Securities raised its 2026 KOSPI target from 7,500 to 10,500 (+40%). But the KOSPI's 12M forward PER reached 30x — sharply higher than the S&P 500's 22x — and Samsung plus SK Hynix together crossed 50% of KOSPI 200 weight for the first time, surfacing single-industry concentration risk as the multiple ceiling. Five factors converged into the May 15 crash — Trump's hawkish line on Iran, US 10-year yields above 4.5%, USDKRW touching 1,500, Samsung's May 21 general strike imminent, and overheated valuations triggering profit-taking. The S&P 500 hit a record 7,501.24 on May 14 alongside NASDAQ 26,635.22, then both fell on May 15 (-1.24% / -1.54%); VIX moved from 17.26 to 18.43 to close the week.
- Research·May 13, 2026·18 min
When GPUs Become an Asset: AI Infrastructure's Capital Structure Gets Rewritten
On May 12, 2026, CME announced GPU rental futures — not a routine product release but an inflection point that rewrites AI infrastructure's capital structure. Cash-settled futures on Silicon Data indexes are expected to launch in H2 2026, likely compressing the seven-year evolution path Bitcoin futures took. Neoclouds transform from landlords into spread traders, and banks, for the first time, will value GPU residuals objectively.
- Note·May 12, 2026·3 min
Surely, the throne of this new era goes to them — doesn't it?
This S&P 500 IT earnings season proved the seriousness of AI demand. It's a time to build long-term AI portfolios — not to chase short-term trades. Today's one-line note: the throne of the 2026 new era is approaching Nvidia's hour.
- Editorial·May 11, 2026·18 min
Is Samsung the Engine Behind a KOSPI 12,000?
This update recalculates the KOSPI 12,000 reverse-engineering framework using SK Hynix's May 11 close of KRW 1,900,000. SK Hynix market cap is now roughly ₩1,383T and its forward PER is about 6.8x, up from the 5.2x level cited in SK Securities' May 7 analysis. Samsung still trades around 6x, which strengthens the laggard re-rating argument. As of May 11, Samsung and SK Hynix together represent about 48.1% of KOSPI market cap. Samsung at 10x takes KOSPI toward 11,686; Samsung at 12x lifts the model output to roughly 13,087.
- Editorial·May 10, 2026·19 min
The Splitting of the AI Data Center
The AI data center has split into training and inference; both branches explode at once. Inference passed 33% in 2023 and crossed 55% in early 2026. McKinsey projects 70~80% by 2030. The split runs through location, power density, network, and components — backend fabric (NVIDIA, InfiniBand) and frontend network (Arista, Ethernet) are now distinct markets. Part 2 covers the optical backbone petabit era (Marvell, Coherent, Ciena), the GW-campus five-branch power split (gas turbine, existing nuclear, SMR, renewables, SOFC), and the diversification pattern of model companies (Anthropic three-stage, OpenAI full diversification). Part 3 maps category leaders globally — memory is the one place where Korea is the global number one. No portfolio weights. Read as an exposure matrix.
- Research·May 10, 2026·13 min
Weekly Market Review: May Week 2
May Week 2 (2026-05-03 to 2026-05-09) was the week when KOSPI 7,498, Samsung Electronics' $1T club entry, and AMD/ARM beats all pointed to the same message. Korea's memory super-cycle is no longer just a domestic valuation event; it is being exported into the US AI-infrastructure price list. KOSPI first broke 7,000 on May 6 at 7,384.56, then lifted to 7,498.00 on May 8. Samsung Electronics rose +16% on May 6 to KRW 270,000 and entered the $1T market-cap club. In the US, AMD's Data Center revenue +57% and ARM's own AI data-center CPU spread the semiconductor rerating. Lumentum's +90% revenue verified the optical bottleneck, while a stronger won and WTI below $100 kept risk appetite open.
- Editorial·May 7, 2026·14 min
The Memory Pickaxe of This AI Gold Rush, Revisited
The end of this AI gold rush cycle is not set by memory — the pickaxe of this rush. It is set by the miners, the hyperscalers. Memory prices climbing and inventories at multi-year lows are not peak signals. They are signals that the miners are still buying more pickaxes. Big Four 2026 CapEx +77% YoY to $725B; Jefferies estimates 92% of operating cash flow goes to CapEx, of which 28% to memory; multi-year LTAs with 10-30% prepayments and floor-price clauses. To time the cycle's end, watch token cost decay and big-tech bond spreads — not memory prices.
- Editorial·May 5, 2026·24 min
RAMpocalypse — What Happens After the Memory Big-3 Leave
AI data centers are devouring memory raw materials, and the memory super-cycle is unfolding as both an unambiguous golden age for industry, investors, and hyperscalers — and a clear regression for ordinary consumers — at the same time. Per Micron's official disclosure, 1GB of HBM3E requires roughly three times the wafer space of 1GB of DDR5; as the memory big-3 funnel capacity into HBM, general-purpose memory supply gets directly squeezed in a zero-sum structure. Counterpoint Research recorded the first-ever price inversion in memory market history — DDR4 (budget devices) at ~$2.10/Gb above server DDR5 ($1.50/Gb). Apple's base Mac mini jumped from $599 to $799 (+33%), Samsung Galaxy S26 and S26+ each rose by $100, and the 32GB DDR5 module spiked from $149 to $239 in two months (+60%). NVIDIA, for the first time since the 1995 NV1, will not release a new GPU line in 2026 (codename "Kicker," RTX 50 Super series indefinitely delayed); IDC forecasts the global PC market to contract -11.3%. At the same time, SK Hynix posted Q3 2025 operating income of KRW 11.4T (an all-time record), Samsung DS reported KRW 7T, and Micron's FY2026 CapEx of $20B is its largest ever — the industry is unambiguously at peak earnings. This report is built around social diagnosis rather than "buy SK Hynix because of the memory shortage," answering instead: why does this affect daily life, is it justified, and what comes next?
20 of 81 shown. Use search for full archive.