Zero One Daily Intelligence Brief — May 6, 2026

Apple Tests Foundry Diversification, Lifting Intel to Record as AMD Triples Server CPU Outlook

Top Market Signals

Apple tours Intel and Samsung fabs; Intel breaks $100 to record on the news

Apple is reportedly evaluating chip-manufacturing alternatives beyond TSMC, with the company touring Samsung's Texas fab and engaging Intel on potential foundry work. Apple is TSMC's largest customer at over 20% of revenue, and a long-running exclusive supply relationship is now being tested. Intel jumped nearly 13% on the day, breaking $100 to a record high. Micron rose more than 10% intraday, AMD and Super Micro also surged, and TSMC's ADR fell about 2%.

The strategic read is that AI-driven advanced-process tightness has pushed Apple to actively shop for second-source capacity. Apple's volume isn't going to TSMC's bleeding edge (N3 / N2 family for iPhone) overnight — Intel 18A and Samsung's Foundry US ramp are not yet at the capability or yield bar to take the hero socket — but the signaling effect matters. Apple has historically used the threat of multi-sourcing to negotiate pricing with TSMC, and the foundry market reads any motion as evidence that the AI capex cycle is creating room for non-TSMC alternatives even in the most demanding logic accounts. For Intel, the more concrete near-term path is high-performance compute and AI accelerator packaging — the EMIB-T story we covered May 4 — and any production volume from a customer of Apple's caliber would compound the rerating that's already happened. The market also extended the move to memory: with AI capex pull running across the stack, Micron's break above $635 is part of the same positioning.

Linked stocks: AAPL US, INTC US, 005930 KS, 2330 TT, MU US, AMD US Sources: 第二來源啟動 蘋果晶圓代工再+1, 美股收盤》英特爾飆近13％

AMD Q1 data center revenue +57% YoY; server CPU TAM tripled to USD 120 billion

AMD reported Q1 2026 results that beat expectations, sending the stock up roughly 12% after-hours. Data center segment revenue grew 57% YoY on AI accelerator and EPYC server CPU demand, and CEO Lisa Su raised AMD's server CPU TAM forecast to approximately USD 120 billion — described as a tripling of the prior estimate — citing agentic AI as the demand engine that's redrawing CPU consumption per inference workload. The EPYC Verano roadmap was reaffirmed, and forward guidance was framed as growth-positive into Q2 and full-year 2026.

The TAM revision is the operative signal. AMD has been gaining server CPU share against Intel for six consecutive quarters, and the +57% data center print is now the run-rate that supports that share trajectory. The agentic-AI-rewrites-CPU-demand argument is consistent with what hyperscalers have been disclosing on their own calls — that inference workloads pull 2-3x the CPU time per dollar of GPU spend versus training-only workloads — and is the reason Intel's data center recovery is happening simultaneously rather than zero-sum. For the Taiwan supply chain, this confirms the AI-server pull-through that Hon Hai (next signal) and Quanta have been reporting. The upside risk to TSMC remains intact even on a day when the Apple foundry-diversification story sent its ADR down: AMD CPUs and GPUs are TSMC-fabbed, and the TAM expansion is a TSMC volume story regardless of whose logo is on the chip.

Linked stocks: AMD US, INTC US, 2330 TT, NVDA US, 2317 TT Sources: 資料中心業務年營收大漲 57%，AMD第一季財報亮眼

Global Unichip April revenue +157% YoY on Google CPU and CSP AI ASIC orders

Global Unichip Corp (GUC, 3443 TT), TSMC's design-service subsidiary, posted April revenue of NT$4.29 billion, up 156.9% YoY and 11.2% MoM — its second-highest single-month figure ever. Year-to-date April revenue is NT$15.74 billion, +81% YoY. The company has confirmed wins on Google CPU silicon and AI ASIC sockets at multiple cloud service providers, with shipments now ramping into volume.

For positioning, GUC is the cleanest pure-play on the AI-ASIC-as-foundry-volume thesis: TSMC owns it, the customer roster is hyperscaler-grade, and the design-service revenue line lags wafer ramps by one to two quarters — meaning April +157% YoY tracks the actual silicon entering production through 2H26. The Google CPU win is also the second confirmation in 10 days (after MediaTek's TPU socket, covered last week) that hyperscaler custom silicon is sourcing at multiple Taiwan design-service vendors rather than concentrating with one. Volume to GUC compounds with TSMC's CoWoS allocation: each ASIC tape-out at GUC translates directly into TSMC wafer starts plus advanced-packaging slots.

Linked stocks: 3443 TT, 2330 TT, GOOG US, 2454 TT Sources: 4月營收速報》創意創單月次高 年增156%

Hon Hai April revenue at record on AI server pull; Q2 "not the slow season"

Hon Hai Precision (2317 TT, Foxconn) reported April revenue of NT$832.1 billion, up 29.7% YoY and 3.5% MoM — an all-time high for the month. Year-to-date April is NT$2.96 trillion, +29.7% YoY, also a record. Management said Q2 is holding "stable growth" despite Q2 being the seasonally slow quarter for ICT manufacturing, citing the AI server pull-through as the offsetting driver.

Hon Hai's monthly print is the highest-frequency signal on AI server build velocity at the OEM scale. Cloud and Networking — which contains the AI-server line at scale — has been the segment driving the mix shift, and a +29.7% YoY in April with no slowdown into Q2 implies the GB200 / GB300 production ramps that customers (Microsoft, Meta, Google, Oracle) have been disclosing on their calls are now translating into hardware shipments through the contract manufacturers. The "淡季不淡" (slow season not slow) framing is unambiguous Q2 guidance: order books for May and June are at or above April's level. For the AI-server bill of materials read-through — substrates, networking, power, thermal, optical — Hon Hai's April is the ground-truth confirmation that the supply chain isn't yet rolling over.

Linked stocks: 2317 TT, NVDA US, 2330 TT, 3017 TT, 2382 TT Sources: 4月營收速報》鴻海創同期新高 Q2淡季不淡

Vanguard pivots Singapore JV fab into CoWoS interposer service

Vanguard International Semiconductor (5347 TT) used its earnings call to disclose that its Singapore VSMC 12-inch fab JV is being repositioned, with customer "consigned equipment plus long-term capacity commitment" arrangements converting the fab from generic foundry into a silicon interposer manufacturing service for CoWoS advanced packaging. Total fab investment was reduced from USD 7.8 billion as the product mix simplifies and process complexity drops with the interposer focus. Management guided Q2 positively on the back of the new structure plus tier-2 foundry pricing recovery.

The strategic read is that interposer wafer capacity is now scarce enough that a hyperscaler-tier CoWoS customer is willing to fund dedicated fab capacity — putting equipment on consignment at Vanguard rather than waiting for TSMC to ramp its own interposer lines. This is functionally the same operating model that Samsung and SK Hynix offer their HBM customers (long-term agreements with consigned cap), now showing up in interposer wafers. Once the Vanguard line is qualified, it adds non-TSMC capacity to the AI-package supply pool — relevant for hyperscalers managing CoWoS allocation risk through 2027. The customer name was not disclosed, but the structure (consigned equipment plus LTA) typically points to a single large counterparty.

Linked stocks: 5347 TT, 2330 TT, 4958 TT, 3037 TT Sources: 打入CoWoS供應鏈 世界先進 看旺第2季

Hyperscaler capex now ~USD 725 billion; memory tightness extends through 2028

The four major cloud service providers have collectively raised aggregate capex guidance to approximately USD 725 billion, per the latest disclosed roadmaps, and memory contract structure is shifting toward 3-to-5-year long-term agreements — a step beyond the typical six-month negotiation rhythm. Winbond's general manager called memory "a strategic resource in the AI cycle" on his investor call. Industry forecasts now extend the structural memory shortage to 2028, deferring the previously expected 2027 supply normalization.

The implication for memory pricing power is straightforward: the LTA shift transfers price-discovery power from buyers to sellers across the cycle's expected duration. Suppliers signing 3-5 year commitments at AI-cycle ASPs are locking in margins through what would historically be the down-leg of the next cycle. For hyperscalers, the trade is supply security against price ceiling — and the willingness to sign suggests internal demand models put DRAM and HBM bit-growth requirements above what fab-equipment installation rates can deliver through 2027. This rerates the Korean memory duo (Samsung Memory, SK Hynix), Micron, and the Taiwan specialty DRAM names (Nanya Tech, Winbond, Macronix) on a longer-duration multiple.

Linked stocks: 005930 KS, 000660 KS, MU US, 2408 TT, 2344 TT, 2337 TT Sources: 四大CSP資本支出上修至超過7千億美元　記憶體結構性供需缺口延至2028年

Samsung hybrid bonding decision targeted mid-2026; HBM4 lineup eyes 2027 Nvidia shipment

Dutch back-end equipment vendor Besi disclosed on its Q1 earnings call that Samsung is targeting a mid-2026 decision on adopting hybrid bonding for its HBM stack — the technology that replaces thermo-compression bonding with direct copper-to-copper interconnects and enables tighter pitch, lower thermal resistance, and higher stack heights. SK Hynix and Micron are on parallel paths. All three memory makers are aligning toward a 2027 HBM4 shipment timeline targeting Nvidia Vera Rubin and follow-on accelerators.

Hybrid bonding adoption is the gating variable for HBM4 economics. Without it, HBM4 stacks beyond 12-Hi face yield walls; with it, 16-Hi and 20-Hi become possible in production volumes. Samsung's decision matters disproportionately because Samsung has been a generation behind SK Hynix in HBM share, and a hybrid bonding adoption that lands in time for HBM4 ramps would close the Nvidia qualification gap. For equipment vendors — Besi, K&S, Hanmi Semiconductor (which is the TC bonder leader and would be partly displaced) — the mid-2026 disclosure is the timing trigger. Hanmi was up 28% on the week before this story, which the market is reading as both HBM-volume positive and hybrid-bonding-disruption manageable in the near term.

Linked stocks: 005930 KS, 000660 KS, MU US, NVDA US, 042700 KS Sources: Besi：三星混合鍵合估年中拍板　HBM三雄瞄準2027出貨NVIDIA

先聲 First Word — Exclusives from Chinese-Language Sources

Aspeed Q1 EPS +59% YoY on AI server BMC ramp; six straight months of record revenue

Aspeed Technology (5464 TT), the BMC controller specialist that's the largest single beneficiary of AI server build volume per server, posted Q1 2026 EPS of NT$37.41 — up 59.4% YoY and 18.0% QoQ — and April revenue of NT$1.28 billion (+81.6% YoY, +3.5% MoM). Year-to-date April revenue is NT$4.42 billion, +59.8% YoY. The company has now posted six consecutive months of record monthly revenue. The driver is straightforward: per-rack BMC content is rising from 80 chips per rack to 120 in the new AI server platforms, per Nuvoton's commentary on the same trend. (May 6, 2026) Source: 股王信驊 Q1每股大賺37.41元

Hon Jun limit-up after Jensen Huang flags "robotics era within 3 years"

Hon Jun Precision (2354 TT), the Foxconn Group chassis maker, hit limit-up at NT$58.50 on volume of 36,000+ lots after Nvidia CEO Jensen Huang said publicly that the robotics era will materialize within three years. Hon Jun has positioned robotics as one of three core growth pillars for 2026, and chairman Chen Kuo-pao reiterated the line in the company's investor materials. The Foxconn Group's broader robotics positioning — humanoid platforms, factory automation, and Nvidia Isaac integration — is now the speculative leg of the Foxconn complex rerating. (May 6, 2026) Source: 熱門股》黃仁勳點名機器人 鴻準飛天

Taiwan tier-2 foundries (Vanguard, UMC, Powerchip) all hike prices simultaneously — first time on record

In a rare alignment, all three Taiwan tier-2 foundries — Vanguard, UMC, and Powerchip — have successfully pushed through 2026 price increases simultaneously, per DigiTimes reporting. The mechanism is "silicon inflation" — material shortages in CCL, electronic-grade glass fiber, and substrate inputs combined with sustained AI-driven demand at TSMC's tier-1 capacity that's pushing overflow business into tier-2 nodes. Each tier-2 has a different leverage point: Vanguard via the new CoWoS interposer service (covered above), UMC via mature-node automotive and analog, Powerchip via DRAM and specialty memory. The collective pricing power is the structural read. (May 6, 2026) Source: 二線晶圓代工罕見全部成功喊漲　世界先進、聯電、力積電各擁長槍