Monday, August 03, 2026

Good morning. The overnight tape is all about the supply-demand squeeze in AI silicon: TSMC PULLING 3NM TO 180K WSM BY EARLY Q4 AND 2NM TO 100K WPM BY YEAR-END — tightest leading-edge supply backdrop in years. SK HYNIX RECORD PROFIT MISSED, SHARES SLUMP 10% — memory drags while semis and hyperscalers lead. Apple's MacBook Air shortage confirms the silicon crunch is real, with Apple steering customers to the Pro. Earnings reactions: GOOGL leads this cycle — GCP growth accelerating from 63% to 82% y/y, margins up 3pts sequentially. AWS reaccelerated to 37% (highest since Covid); Azure guides to 45% in CQ3. The hyperscalers get paid. Asia: Nomura just jacked up Samsung OP forecasts to KRW 635tn next year and KRW 774tn the following, with shareholder return yields hitting 22.3% — if credible, Samsung is deeply undervalued. Hynix's 10% slump keeps memory sentiment mixed. Frame the day around: (1) value shifts from the model layer to the infrastructure layer — OpenAI and Anthropic commoditize, hyperscalers are the toll booth. (2) TSMC's pull-ins make advanced node allocation the scarce resource — NVDA, AVGO, AMD all fighting for the same 3nm/2nm capacity. (3) Enterprise AI hits the tipping point in 3-4 quarters — Microsoft's 30M Copilot seats vs 450M M365 paid seats shows the runway. (4) Memory bifurcates: HBM demand robust, but expectations overran; MU trades at a premium to Hynix/Samsung despite the same Taiwan Strait risk as TSM — that mispricing resolves. We'll hit up GOOGL, TSM, and MU first, then get to the power and neocloud complex (TLN, VST, CORZ).

1. Supplementary Coverage

AMZN

AWS reacceleration is the cleanest cloud validation this cycle. Growth jumped 28% → 37% y/y in Q2, highest since Covid, with normalized op margins +1pt sequentially. Market paid with a 15% rally. Read-through: hyperscaler AI spend converts to revenue, not just capex. Capex raised 10% to $220B for 2026; contracted backlog sits at $496B, up from $364B. Debate shifts from AI demand to FCF durability.

MSFT

Azure lit the +16% print. Growth accelerated 39% → 43% y/y, CQ3 guide at 45%. That's the proof AI spend monetizes inside Microsoft's own cloud. Co-Pilot: ~450M M365 paid seats, but only ~30M carry Co-Pilot. If attach moves, multi-year annuity. Contracted backlog $678B, but only ~30% of RPO converts in 12 months. Bullish demand, bearish near-term conversion.

GOOGL

GCP is the strongest cloud print of the cycle. Growth accelerated 63% → 82% y/y, op margins +3pt sequentially. The 7% post-earnings drop was forced-sale tape, not fundamentals. TPU demand projected at 15M units by 2028 vs NVDA's ~13M/year. Google moved packaging to Intel EMIB because NVDA locked up TSMC CoWoS. Vertical integration is real and accelerating. DeepMind wants to be the Android of robotics — optionality, not near-term revenue.

TSM

TSMC is pulling capacity forward because demand is real. 3nm to 180k wpm by early Q4 (not end-26); 2nm to 100k wpm by year-end. Strongest leading indicator for AI silicon. NVDA pre-booked >50% of CoWoS for three years, forcing Google to Intel EMIB. NVDA, AMD, AVGO all vying for 3nm/2nm allocation. Foundry capacity effectively sold out.

NVDA

NVDA still owns the advanced-node queue. Pre-booked >50% of TSMC CoWoS for three years; TSM pulling 3nm/2nm forward. HBM downgrade risk from 12-Hi to 8-Hi is the swing factor — checks over two weeks suggest customers are considering it. At 13x GAAP earnings, cheapest multiple in a decade; Gerstner wants capital return raised to 70-75% of FCF. OpenAI talking up to a $250B backstop around 10GW Ohio data center debt — unconfirmed, but if it closes, decade-level visibility.

AVGO

AVGO is in the allocation scramble with NVDA and AMD as TSM pulls 3nm/2nm forward. Supply constraint is a tailwind for whoever holds allocation. MediaTek's $5B ASIC push targets 15-20% of an $80B custom AI chip market by 2027. AVGO still owns the high-complexity segment, but the competitive set is expanding.

AMD

MI300X demand is real and organic. A neocloud reported 113 new MI300X customers in 90 days with zero marketing spend. That's a velocity stat, not a backlog number. AMD is placing a >$14B bet with Core Scientific for 530MW across five sites, scaling to 2.5GW — chip seller to infrastructure enabler. Also raising GPU/memory bundle prices ~10% to board partners. Defends margins, passes cost down.

MU

Memory is in a structural shortage, not a cyclical uptick. DRAM and NAND inventory at three weeks; shortage persists through 2027. MU is the purest US-listed exposure. Taiwan Strait risk is real but unpriced — HBM wafers don't come from Idaho. NVDA's HBM downgrade risk (12-Hi → 8-Hi) is the key downside swing. A credible bearish view is forming; watch for an HBM oversupply thesis in 2027.

SNDK

SNDK stands in the middle of the NAND shortage. DRAM/NAND inventories at three weeks; supply constraint runs through 2027. Memory cost inflation spills into consumer devices and AI servers. Tighter NAND pricing is the setup.

LITE

InP shortage is the next bottleneck trade. Lumentum's CEO says it "will get worse than memory." AXT and JX Advanced Metals are locking up substrate capacity with deposits and multi-year expansions. LITE has pricing power if it owns the InP supply chain. Active copper PAM6 may extend copper's life ~5 years — a headwind, not a killer. LITE flagged in the GAI infrastructure volatility basket; don't confuse sentiment beta with fundamental deterioration.

AXTI

AXTI just locked up the InP substrate bottleneck with an $87M deposit through 2031. Six-year capacity commitment on the critical material for optical transceivers. Lumentum's CEO says InP shortage gets worse than memory. AXTI is the upstream leverage.

AAPL

MacBook Air shortages are the first consumer-visible casualty of the AI memory crunch. Apple is steering customers to MacBook Pro, adding "subject to availability" warnings. Mix tailwind, volume tax. Apple is morphing into the world's biggest subscription business — $2,000+ foldable at $60-70/month. Health is the next wearable wedge. Long-dated optionality.

TLN

Power-side AI beneficiary as datacenter power constraints move up the supply chain. No specifics in the window. Positioning memo, not a fundamental update.

VST

Same as TLN — power-side AI beneficiary. Datacenter power constraints moving up the supply chain. No specifics. Positioning memo.

CORZ

Core Scientific just became AMD's infrastructure arm. >$14B, 530MW deal across five sites scaling to 2.5GW. Massive contracted revenue base. Validates the pivot from Bitcoin mining to AI hosting.

INTC

Intel picked up Google as an EMIB packaging customer because NVDA locked up TSMC CoWoS. Foundry-services win, not a core CPU turnaround. Proves advanced packaging has a role in the AI supply chain. Question is whether it scales beyond packaging.

BABA

Alibaba is a channel for NVDA GPUs into China's AI labs. Moonshot reportedly accessed ~20,000 NVDA GPUs through Alibaba. Critical AI compute intermediary despite export controls. Small relative to US scale, but the route is proven.

QCOM

Qualcomm is eating AI cost inflation across memory, wafer, packaging, testing. Double-digit price hikes. Mix of pricing power and volume risk. China's CXMT nearing final LPDDR6 development — another reason the memory crunch isn't one-dimensional.

NBIS

Nebius calls its pivot capital-light, but the cash flow says otherwise. 89% of revenue prepaid — mostly Meta's $12B deal — yet projected FCF is negative ~$20B from 2026-2028. Prepaid revenue gives visibility; the buildout still burns capital. "Capital-light" is a stretch.

RDDT

Reddit is compounding. Q2 revenue hit $805M, ad revenue $762M, up 64% y/y. High-growth ad business with AI data licensing optionality. Market should pay for growth this clean.

GM

GM bought back $2B in Q2 — 25M shares. Spending on buybacks instead of chasing EV losses. Signal they see the stock undervalued. Supports per-share metrics even if volumes stagnate.

NFLX

Netflix is chasing YouTube creators, including 'Hot Ones,' because it finally sees the threat. Creator economics far cheaper than Hollywood production — could expand margins. Near-term hurdle is deal negotiation; long-term risk is diluting the premium brand.

SNOW

SNOW sits in the accelerant path of agentic AI. Needham sees demand for all data management systems accelerating, not just vector DBs. Agents need governed, structured enterprise data — Snowflake's turf. Question is whether consumption-based revenue converts this into reacceleration.

MDB

MDB is in the agentic AI data management accelerant path. Demand accelerating across Postgres, Databricks, Snowflake, MDB. Agents need flexible, operational data stores. MongoDB is the developer default. Market underweights this broad pull.

ASML

ASML's service revenue growth is the leading indicator the market is ignoring. Combined ASML/Lam/TEL/KLA service revenue grew 33% y/y in June quarter vs 20% for equipment. Utilization is inflecting before WFE orders. Service cash flow funds the dividend — downside support.

LRCX

Lam is compounding through the service layer. Combined semicap service revenue grew 33% y/y vs 20% for equipment; Lam guides 10%+ long-term service growth. Service tracks utilization, and utilization is inflecting. Funds the dividend — a real floor.

KLAC

KLA has the best long-term service target in the group at 13-15%. Combined semicap service revenue up 33% y/y vs 20% for equipment. Process control intensity rising as fabs squeeze yields. Service cash flow funds the dividend.

AMAT

AMAT's service business is targeting mid-teens growth while combined semicap service revenue runs 33% y/y. Leading indicator for utilization and eventual WFE. Service funds the dividend and cushions the cyclical trough. Semicap beta with a service floor.

IBM

IBM's lithography roadmap targets EPE from 3.0nm to 1.2nm, moving from nanosheet to nanostack. R&D leadership, not commercial revenue. Positions IBM as a research partner in advanced semiconductors. No P&L impact for years.

IONQ

IONQ is buying SKWT to become a full-stack quantum company. Expands control over the stack but adds integration risk. Quantum revenue remains immaterial for years. Strategic optionality — market treats it as dilutive unless commercial contracts land.

EPAM

EPAM spiked +17.5% premarket on a large OpenAI upskilling push. Narrative move, not an order book signal. AI upskilling doesn't convert to revenue until clients actually buy AI delivery projects. If the order pipeline doesn't accelerate, the gap fills.

WULF

TeraWulf is re-rating on the AI power bottleneck. Anthropic-linked ~$19B order and power plant acquisition put independent power plus cooling at a premium. AI bottleneck shifted from silicon to grid. WULF owns the kind of asset scarcity flows into.

UPS

UPS is the anti-AI name in the FCF scrutiny basket. Bloomberg flagged UPS among companies where massive AI spending raises FCF questions — but UPS has no AI story. If rates stay high, UPS's cash-flow profile gets re-rated lower. Risk-off signal for FCF-poor legacy industrials.

VLO / MPC / PSX / DINO

Refined product and logistics assets are getting blown up — bull set-up for US independent refiners. Aggressors targeting downstream, not crude. Product supply disruption widens cracks. VLO owns the logistics advantage; MPC has the refining and marketing footprint; PSX's chemicals/logistics network adds leverage; DINO is the small-cap, highest-beta way to play it.

ORCL

ORCL's backlog is the biggest proof-of-demand artifact in tech. Contracted backlog reached $638B, up 363%, with $75B in customer prepayments and customer-supplied hardware. But expect only ~12% of RPO to convert in the next 12 months. The market is paying for 2028, not this year.

SSNLF

Nomura and Goldman just threw a massive gauntlet down on Samsung. Nomura sees operating profit going KRW 391tn → 635tn → 774tn over three years; Goldman added it to APAC Conviction List with a KRW 490,000 target, 87% upside. Implies roughly $1.25T in cumulative operating profit and shareholder-return yields up to 22.3%. Market remains skeptical — that's the opportunity. Profitability now depends almost entirely on the chip business — concentration risk, but memory/HBM leverage acts directly on EPS.

HXSCL

SK Hynix printed record profits and still dropped 10% because consensus was even higher. Expectations ran ahead of fundamentals. HBM pricing strength confirmed, but the market questions sustainability. Memory longs got the demand data they wanted; the marginal buyer wants 2027 visibility. New memory capacity takes 3-4 years to build — supply stays tight. But the 10% drop shows how fragile positioning is when good news is already in the price.

2. Street Color / Heard (unverified)

Hearing OpenAI regained the mandate from Anthropic. Six months ago Claude Code was "flying off the shelf"; Altman cut side hustles, reallocated compute, pivoted half the team to Codex. Codex launched because Claude Code was winning. "The yolo CEO reigns supreme once more."

Word is enterprise AI adoption crosses the tipping point in 3-4 quarters — token costs overwhelm incremental gross profit and models beat the median knowledge worker. Sooner for international.

AMD channel check: 113 new MI300X customers in 90 days, zero marketing spend. On-demand GPU demand is broadening beyond incumbents.

Hearing NVDA HBM downgrade risk from 12-Hi to 8-Hi — multiple checks over two weeks suggest customers are considering it. Single most important unconfirmed data point. If read as spec optimization, the pullback is a buy.

Word is NVDA in talks with OpenAI for up to a $250B backstop around 10GW Ohio data center debt. Unconfirmed. If it closes, decade-level demand visibility.

Google TPU talk: 15M units projected by 2028 vs NVDA's ~13M/year. Packaging moved to Intel EMIB because NVDA locked up TSMC CoWoS. Vertical integration accelerating.

Hearing MediaTek's $5B ASIC push targets 15-20% of an $80B custom AI chip market by 2027, with Google TPU v9 in early 2028. AVGO still owns high-complexity, but the competitive set is expanding.

Lumentum CEO says InP shortage "will get worse than memory." AXT locking up substrate with $87M deposit through 2031. InP is the next bottleneck trade.

Apple supply chain: steering MacBook Air buyers to MacBook Pro amid memory/silicon crunch. "MacBook Air subject to availability" — something Apple has rarely, if ever, done. Consumer-visible casualty of the AI memory crunch.

Word is Apple Upgrade program positions $2,000+ foldable at $60-70/month. Path to next trillion via subscriptions, not units.

Co-Pilot math: ~450M M365 paid seats, only ~30M with Co-Pilot. Attach rate is the multi-year annuity. The debate is price/attach, not capability.

DeepMind wants to be the Android of robotics — brain-swapping-body architecture. No near-term revenue, but it's the optionality behind Google's AI stack.

Hearing model layer commoditizing; value shifting to infrastructure/cloud layer. "Ant and OAI are driving 90% of ARR." The most intelligent token is worth meaningfully more than the second.

Altman says a 1GW DC requires ~10,000 workers full-time for ~18 months. Labor intensity is the hidden constraint on the buildout.

Word is broad public neocloud analyses are "largely pointless" — every cloud serves its own niche. Proprietary diligence required.

Samsung: Nomura sees OP going KRW 391tn → 635tn → 774tn over three years. Goldman adds to APAC Conviction List, KRW 490,000 target, 87% upside. Implies shareholder-return yields up to 22.3%. Market still skeptical.

SK Hynix printed record profits, dropped 10%. Consensus was even higher. Good news already in the price.

A credible bearish memory view is forming: "I am not bullish on memory. More soon..." HBM oversupply in 2027 as capacity comes online is the emerging bear thesis. Bull case relies on LTA discipline.

Hearing memory cost inflation is the primary driver of hyperscaler capex increases. DRAM/NAND inventory at three weeks through 2027.

Qualcomm raising prices double-digit across memory, wafer, packaging, testing. Margin defense, but volume risk if price hikes cool handset demand.

China angle: CXMT nearing final LPDDR6 development. Domestic memory taking share is another supply-side twist.

Needham sees agentic AI accelerating demand for all data management systems — Snowflake, MongoDB, Databricks, Postgres. Not just vector DBs.

EPAM +17.5% premarket on OpenAI upskilling push. Narrative, not orders — yet. If the order pipeline doesn't accelerate, the gap fills.

WULF re-rating on Anthropic-linked ~$19B order and power plant acquisition. AI bottleneck shifted from silicon to grid.

Bloomberg flagged UPS in the FCF scrutiny basket — massive AI spending raises FCF questions for companies with no AI story. Risk-off signal for FCF-poor legacy industrials.

Refined product and logistics assets "getting blown up" — aggressors targeting downstream, not crude. Bull set-up for VLO, MPC, PSX, DINO.

Netflix chasing YouTube creators including 'Hot Ones.' Creator economics cheaper than Hollywood; margin expansion potential. Long-term risk is diluting the premium brand.

Apple hiring for health/wellness leadership across vision products — smart glasses and headsets as health devices. Healthcare monetization optionality underappreciated.

Semicap service revenue grew 33% y/y vs 20% for equipment (ASML/Lam/TEL/KLA combined). Utilization inflecting before WFE orders. Service funds dividends — downside support.

Gerstner wants NVDA capital return raised from 50% to 70-75% of FCF. At 13x GAAP earnings, the multiple is a floor not a ceiling.

IBM lithography roadmap: EPE 3.0nm → 1.2nm, nanosheet → nanostack. R&D leadership, not commercial revenue.

IONQ buying SKWT for full-stack quantum. Integration risk, strategic optionality.

Meta Business Agents passed 1M active businesses — no disclosed monetization. Market wants attach rates, not user counts.

Moonshot reportedly accessed ~20,000 NVDA GPUs through Alibaba. BABA as AI compute intermediary into China.

Oracle backlog $638B, up 363%, with $75B in prepayments and customer-supplied hardware. Only ~12% of RPO converts in 12 months. Market paying for 2028, not this year.

TSM pulling 3nm to 180k wpm by early Q4, 2nm to 100k wpm by year-end. NVDA, AMD, AVGO all vying for allocation. Foundry capacity effectively sold out.