Big Tech spent $133B in Q1 capex, up 70% y/y, locking in a depreciation treadmill that consensus sees hitting $430B annually by 2031—against $372B in combined net income last year. Huawei AI chip revenue surges to $12B as Beijing forces domestic substitution. TSMC SoIC yield issues are dragging CPO shipment forecasts lower across the industry.
Big Tech spent $133B in Q1 capex, up 70% y/y. Combined depreciation hit $41.6B in the quarter. Server equipment depreciates over 5-6 years. Consensus pegs annual depreciation for MSFT, AMZN, META, and GOOGL exceeding $430B within five years against $372B of combined net income last year. Useful lives already extended from ~4 to 5-6 years—that lever is spent.
Goldman's Covello is recommending long hyperscalers, short semis. The thesis: hyperscaler multiples have compressed on ROI skepticism while chip names have already been rewarded. If hyperscalers demonstrate positive ROI, they rerate higher—and capex stays elevated, eventually crowding semis if spending moderates.
↳ The risk to the bear case: this cycle looks more like the Model T than shale. Structural inference demand may keep the treadmill running far longer than depreciation bears expect.
Epitaxial wafer ASPs are turning. Leading-edge logic (7nm and below) wafer demand is inflecting, with models estimating ~1M wafers per month by CY28—roughly 10% of total 300mm equivalent demand. GlobalWafers, SUMCO, Shin-Etsu, and Siltronic are primary beneficiaries. Epi wafer supply-demand is tightening faster than expected.
April Korea export data came in hot: DRAM at $9.25B (+15% MoM, +83% QoQ), SSD at $3.84B (+20% MoM, +181% QoQ). The sequential jump is massive. But here's the tension: commodity memory profitability now exceeds HBM, incentivizing the Memory Big 3 to prioritize commodity DRAM/NAND over HBM shipments. Near-term HBM shipment mix could disappoint.
David Patterson (UC Berkeley, Turing Award) is flagging High Bandwidth Flash (HBF) as the next bottleneck after HBM—a surge in demand is coming.
↳ Apple's Q2 FY26 call confirms memory is a margin headwind, not a supply constraint. The real bottleneck is TSMC leading-edge SoC capacity. Mac mini and Mac Studio are allocation-limited, not demand-limited.
Huawei AI chip revenue is tracking ~$12B in 2026, up from $7.5B in 2025—a 60% increase. The 950PR entered mass production in March; a 950DT upgrade is planned for Q4. Morgan Stanley sees the China AI chip market reaching $67B by 2030, with 86% supplied domestically. This year, domestic vendors control ~$21B of the market.
Beijing is requiring Chinese tech firms to use domestic AI chips and limit Nvidia usage to overseas operations. US regulators reciprocating—Nvidia H200 chips reportedly cannot clear Chinese customs. Two-way restrictions accelerating decoupling.
SMIC has produced early 3nm test wafers (first reported Feb 2026). Still behind TSMC's edge, but closing faster than sanctions proponents expected.
↳ Nvidia's China addressable market is structurally shrinking. The question is whether domestic silicon can match performance at scale. The 950PR suggests the gap is narrowing meaningfully.
Low TSMC SoIC yields are forcing the industry to cut 2026 CPO shipment forecasts. A supply-side constraint, not demand. Optical networking revenue recognition may slip into late 2026 or 2027.
Teradyne acknowledged Advantest's capabilities in CPO testing during its Q1 call—a notable concession. Teradyne posted Q1 revenue of $1.28B with 60.9% gross margin, beating estimates, but guided Q2 to $1.15–1.25B (midpoint -6% QoQ). Transition period in 2H.
Former TSMC packaging VP Douglas Yu joined MediaTek, de-risking EMIB-T adoption. Industry source: "TSMC assumed customers would not leave, but they did." TSMC is rushing to fill the packaging gap between CoWoS and alternatives.
Bare metal GPU providers are sold out. Top AMD providers (Vultr, DO, TW) have zero availability on MI355x. One operator is raising prices on 8x bare metal boxes, citing zero market supply. Vultr moved massive MI355x volume. Most competitors aren't offering short-term contracts.
Pricing signal: DO listing H100 at $3.39/hr, MI300x at $1.99/hr—~70% discount on AMD silicon. The gap reflects AMD's share push and genuine demand elasticity at lower price points.
Applied Digital signed a 15-year lease with a second IG hyperscaler: 300MW critical IT load, ~$7.5B contracted value, total contracted revenue now $23B+. Over 50% backed by IG tenants. Operations mid-2027.
Alphabet posted a record single-day market cap increase after Q1 beat on cloud and AI demand. Market cap: $4.65T, within striking distance of Nvidia at $4.85T. The narrative shift: AI capex is generating returns, at least for GOOGL.
Covello's best-case scenario for long hyperscalers/short semis is playing out. But if capex stays at 40%+ capital intensity, the market may eventually stop rewarding the treadmill.
MediaTek (GS TP5000, 90% upside): Next-gen AI ASIC ASP is "multiple times higher" than current gen, with MediaTek engaging in compute die design alongside I/O dies. GM-accretive. Former TSMC packaging VP's arrival further de-risks execution.
Full analysis, implications, relative trade read-throughs, positioning reads and catalysts — delivered by email at 07:00 HKT daily.