Hyperscaler capex hit a record $133B in Q1 but the depreciation machine is just getting started — $430B+ annual charges loom by 2030. China's domestic AI chip substitution is accelerating as NVIDIA's H200 gets blocked at customs. Memory market dynamics shifting as commodity DRAM margins eclipse HBM profitability.
Big Tech spent $133B in combined capex in Q1 2026, up 70% year-over-year, with depreciation already hitting $41.6B in the quarter. Server equipment depreciates over 5-6 years, meaning today's record spend locks in a structural earnings drag through the end of the decade. Consensus now pegs annual depreciation for MSFT, AMZN, META, and GOOGL exceeding $430B within five years — against just $372B in combined net income last year. The useful-life extension lever (from ~4 years to 5-6) has already been pulled. At 40% capital intensity, the market may not reward the treadmill forever.
↳ Alphabet's Q1 results validate the bulls. Market cap hit $4.65T — just shy of NVIDIA's $4.85T — with the single-day market cap increase setting a record. Cloud and AI services demand was the clearest signal yet that aggressive AI investment is generating returns.
Huawei is on track to capture the largest share of China's AI chip market this year. Revenue is expected to reach approximately $12B, up from $7.5B in 2025 — a 60%+ jump. The bulk of orders are for the 950PR chip, in mass production since March, with an upgraded 950DT planned for Q4. Beijing has directed Chinese tech firms to support domestic suppliers and restrict NVIDIA usage to overseas operations. Simultaneously, U.S. regulators require all NVIDIA chips ordered by Chinese customers to remain in-country — and H200 chips are reportedly unable to clear customs. Morgan Stanley projects China's AI chip market reaches $67B by 2030, with 86% supplied domestically. This year's domestic market: ~$21B.
↳ SMIC has reportedly produced early 3nm test wafers as of February 2026, suggesting the domestic process technology roadmap is advancing despite sanctions. Yield and HVM timelines remain opaque.
Korea's preliminary April semiconductor export data is explosive. DRAM: $9.25B (+15% MoM, +83% QoQ). DRAM modules: $6.14B (-17% MoM, +70% QoQ). SSD: $3.84B (+20% MoM, +181% QoQ). MCP: $8.16B (-5% MoM, +67% QoQ). NAND was the lone soft spot at $1.67B (-34% MoM, +17% QoQ). The critical signal for positioning: profitability in commodity memory has surpassed HBM, giving the Memory Big 3 stronger incentive to allocate capacity toward commodity DRAM/NAND rather than HBM. Near-term HBM shipment mix may disappoint relative to bull expectations.
↳ Epitaxial wafer supply-demand is tightening faster than expected. Leading-edge logic wafer demand (7nm and below) projected to inflect next year, reaching ~1M wpm by CY28 — roughly 10% of total 300mm equivalent demand. Wafer ASPs are turning after years of AI being a rounding error. GlobalWafers, SUMCO, Shin-Etsu, and Siltronic stand to benefit.
Former TSMC packaging VP Douglas Yu has joined MediaTek, a move insiders expect will substantially de-risk MediaTek's first adoption of EMIB-T advanced packaging. Industry source: "TSMC assumed customers would not leave, but they did." TSMC is reportedly rushing to fill the packaging gap between CoWoS and next-gen solutions. Separately, low TSMC SoIC yields are forcing the industry to lower near-term CPO shipment forecasts for 2026. Teradyne acknowledged Advantest's capabilities in CPO testing during its Q1 call — a surprise admission that the tester duopoly is more balanced than consensus assumed.
↳ Goldman has a TP5000 on MediaTek, flagging 90% upside driven by next-gen AI ASIC projects with ASPs multiple times higher than the current generation. MediaTek is now engaging in compute die design in addition to I/O dies, making the franchise gross-margin accretive.
The GPU cloud market is supply-constrained. Top AMD bare metal providers — Vultr, DigitalOcean, TW — are all sold out on MI355x inventory. H100 pricing at $3.39/hr vs. MI300x at $1.99/hr represents a ~70% discount, suggesting AMD is aggressively buying inference share. On the infrastructure side, Applied Digital signed a 15-year, 300MW lease with a second U.S. investment-grade hyperscaler at its Delta Forge 1 campus, adding ~$7.5B in contracted value. Total contracted revenue now exceeds $23B, with over 50% backed by IG tenants. Initial operations expected mid-2027.
Apple's Q2 FY2026 call revealed a key supply chain signal: memory is not the bottleneck — leading-edge SoC capacity at TSMC is. Apple can secure enough memory but cannot ship enough products due to insufficient TSMC leading-edge node allocation. Memory remains a cost variable pressuring margins, but the unit ceiling is set by fab capacity. Implications: TSMC's pricing power on leading-edge nodes stays elevated, and Apple-exposed memory suppliers face margin headwinds without volume relief.
MPWR's Q1 call confirms AI demand broadening from GPUs into CPU servers, optical modules, switches, storage, DDR5, and higher-voltage power conversion. Aixtron posted €59M Q1 revenue and €171M in orders, management sizing the optoelectronics opportunity at 80-100 G10 tools per year — a €300-400M annual TAM with multi-tool orders ramping. AXT reports indium phosphide substrate demand accelerating faster than qualified supply, with customer procurement shifting toward multi-year allocation security. Cohu is pivoting its Eclipse handler toward HBM thermal test. Turing Award winner David Patterson is now calling High Bandwidth Flash (HBF) "highly likely to stand at the center of the next bottleneck." Teradyne beat Q1 at $1.28B revenue and 60.9% gross margin but guided Q2 to $1.15-1.25B (midpoint -6% QoQ), signaling a 2H26 transition.
Full analysis, implications, relative trade read-throughs, positioning reads and catalysts — delivered by email at 07:00 HKT daily.