Big Tech burned through $133B in Q1 capex (+70% y/y), locking in depreciation drag through 2030. Alphabet's blowout quarter and $4.65T market cap suggest AI ROI is real — for some. Memory wafer ASPs are inflecting, commodity DRAM margins now top HBM, and Huawei is surging 60%+ in China's domestic AI chip market.
Big Tech spent $133B in Q1 capex, up 70% year-over-year. Combined depreciation hit $41.6B in the same quarter. Server equipment depreciates over 5-6 years, meaning today's record spend locks in an earnings headwind through 2030. Consensus estimates peg annual depreciation for MSFT, AMZN, META, and GOOGL exceeding $430B within five years — against just $372B of combined net income last year. The useful-life extension lever has already been pulled (from ~4 years to 5-6), so that accounting trick is exhausted. Some hyperscalers are now running at 40%+ capital intensity as a percentage of revenue. History is unkind to businesses sustaining that level without provable returns.
↳ The market will eventually demand proof that $133B/quarter generates incremental revenue, not just token volume. The depreciation treadmill is mathematically inescapable once servers are deployed. TSMC, for reference, can break even at just 25% fab utilization — the hyperscalers have no such structural cushion.
Alphabet surged to a $4.65 trillion market cap on a record single-day gain after blowout earnings showed strong cloud and AI demand. It now sits just below NVIDIA ($4.85T) for the world's most valuable company. The print is being read as proof that aggressive AI investment is generating returns. Goldman's veteran semiconductor analyst Covello is now explicitly recommending long hyperscalers, short semis — arguing compressed cloud valuations reflect excessive investor skepticism while chip stocks have already priced in the cycle. In Goldman's best-case scenario, hyperscalers demonstrate positive ROI, cloud multiples expand, and capex discipline then actually hurts semiconductor revenues.
↳ The contrarian risk: if ROI disappoints, both legs lose simultaneously — cloud multiples compress further AND capex cuts crush chip demand. But the relative-value call has real momentum after Alphabet's quarter.
April Korea semiconductor export data came in hot: DRAM hit $9.25B (+15% MoM, +83% QoQ), SSDs reached $3.84B (+20% MoM, +181% QoQ), NAND at $1.67B (-34% MoM, +17% QoQ). Epitaxial wafer supply-demand is tightening faster than expected. Leading-edge logic (7nm and below) wafer demand is expected to inflect next year, reaching ~1M wpm by CY28 — roughly 10% of total 300mm equivalent demand. GlobalWafers, SUMCO, Shin-Etsu, and Siltronics stand to benefit. Critical nuance: commodity memory profitability now exceeds HBM, giving the Memory Big 3 stronger incentive to prioritize commodity DRAM/NAND over HBM shipments in the near term. This shifts the mix dynamic the market has been pricing.
↳ Apple's Q2 FY26 call confirms memory is a cost variable pressuring margins, not a supply bottleneck. The real constraint is SoC / leading-edge node capacity at TSMC. This supports margin expansion for memory vendors while complicating the bull case for HBM-centric revenue stories.
Huawei will capture the largest share of China's AI chip market this year, with revenue surging at least 60% to approximately $12B (from $7.5B in 2025). The 950PR chip entered mass production in March; an upgraded 950DT is planned for Q4. Beijing is pressing Chinese tech firms to support domestic suppliers and restrict NVIDIA chip usage to overseas operations only. U.S. regulators are blocking H200 customs clearance in China. Morgan Stanley forecasts China's AI chip market reaches $67B by 2030, with 86% from domestic suppliers. Domestic market this year: approximately $21B.
Former TSMC packaging VP Douglas Yu has joined MediaTek. Insiders expect his arrival to significantly derisk MediaTek's first adoption of EMIB-T packaging. Industry source: "TSMC assumed customers would not leave, but they did." TSMC is now rushing to fill the packaging gap. Separately, low TSMC SoIC yields have led the industry to lower near-term CPO shipment forecasts for 2026. Goldman sees 90% upside for MediaTek (TP5000), with next-gen AI ASIC ASPs "multiple times higher" as the company moves into compute die design alongside I/O dies — a margin-accretive shift.
Applied Digital signed a 15-year lease with a second U.S. investment-grade hyperscaler at its Delta Forge 1 campus: 300 MW of critical IT load, ~$7.5B in contracted value, total contracted revenue now $23B+. Over 50% backed by IG tenants. Initial operations mid-2027. On the inference side, bare metal GPU supply is zero — top AMD MI355x providers (Vultr, DO, TW) are sold out. H100 pricing at $3.39 and MI300x at $1.99 at some platforms, a ~70% discount ratio. Price increases are being implemented across the board.
OpenAI appears to have cracked recursive self-improvement for Codex. GPT-5.5 performs comparably to Mythos on the UK AISI's cyber range eval for long-horizon agentic capability. Meanwhile, Anthropic's Opus 4.7 is reportedly underperforming Opus 4.6 — users cite degraded output and declining long-context success rates, attributed to compute scarcity. Inference throughput is emerging as the defining competitive variable. David Patterson (UC Berkeley, Turing Award) flags High Bandwidth Flash (HBF) as the likely next bottleneck after HBM, with demand surge incoming.
Full analysis, implications, trade read-throughs, positioning reads and catalysts — delivered by email at 07:00 HKT daily.
Full bull/bear breakdown, broker targets, and catalyst calendar — delivered by email at 07:00 HKT daily.