Zero One Daily Intelligence Brief — May 4, 2026

Nvidia at Zero in China as Hunyuan Triples Weekly Inference Demand

Top Market Signals

Jensen Huang concedes Nvidia China share is zero, says policy "largely backfired"

Speaking to lawmakers at the Special Competitive Studies Project, Jensen Huang said Nvidia's China AI accelerator share is now zero, down from roughly 95% two years ago — and that the US export-control framework "probably does not make a lot of strategic sense" and has "already largely backfired."

The collapse follows the April 2025 BIS rule requiring a license even for H20 — the chip Nvidia had specifically engineered to remain compliant — which forced Nvidia to take a USD 4.5 billion charge in Q1 FY26 against H20 inventory and unfulfilled purchase obligations. A March 2026 US prosecution of a Supermicro co-founder also disrupted grey-market routing, with B300 server prices in China roughly doubling to about USD 1 million versus USD 550,000 in the US.

The strategic read is that the China data-center AI chip market has cleanly bifurcated, and the US side has lost it. Two years ago Nvidia was extracting effectively monopoly economics from Chinese hyperscalers; now those same buyers are sourcing Huawei Ascend, Cambricon, and increasingly domestic packaging — with the volume showing up in the Tencent Hunyuan and DeepSeek inference numbers (next signal). The base case is no longer "Nvidia recovers some share via a future H30/H40" — it is "Nvidia is structurally absent from China at hyperscaler scale, and the second-order question is whether China can supply enough silicon to keep up with its own demand growth." For Nvidia, the offset is global non-China demand which has more than absorbed the lost revenue; the longer-term cost is loss of optionality on a market that historically generated 20-25% of data-center revenue and acted as a margin pricing buffer. For the supply chain, this hardens the read that two parallel AI compute stacks now exist — one TSMC/SK Hynix/Samsung-anchored, one SMIC/CXMT/JCET-anchored — and that the Taiwan/Korea cluster's growth is no longer dependent on China being part of the customer base.

Linked stocks: NVDA US, AMD US, AVGO US, 2330 TT, 005930 KS, 000660 KS, 0981 HK Sources: Techloy

Chinese AI inference doubles US weekly token volume; Tencent Hunyuan Hy3 +799% WoW

OpenRouter data for the week of April 27 – May 3 shows Chinese AI models served 7.94 trillion tokens, more than double the US's 3.26 trillion — and the second time in two weeks that China has surpassed the US.

Chinese volume rose 81.7% week-over-week while US volume fell 34.6%. Tencent's Hunyuan Hy3 preview took the global #1 slot at 3.03 trillion tokens, +799% WoW. Moonshot's Kimi K2.6 was #2 at 1.28 trillion tokens (+15% WoW). DeepSeek-V4-Flash debuted at #9 with 0.70 trillion tokens (+344% WoW). DeepSeek priced V4-Flash input cache hits at CNY 0.02 per million tokens — roughly USD 0.0028 — and V4-Pro at CNY 0.025. Global token usage was 23.9 trillion for the week, +8.6% WoW.

The price-and-volume combination is the most aggressive Chinese inference push to date. Hunyuan Hy3 going from a standing start to 3 trillion tokens in a week is consistent with Tencent making it the default model in WeChat AI features at near-zero cost — a usage-driven flywheel that the US closed-model providers cannot match without conceding margin. The strategic implication for Western AI companies is that the price floor for inference is now being set by Chinese open-weights models running on domestic Huawei/SMIC silicon, distributed through hyperscaler-grade Chinese clouds. For Nvidia and the rest of the US AI compute stack, this is the demand-side counterpart to Huang's market-share concession: the inference workload is migrating to a stack they don't supply. For US foundation-model providers (OpenAI, Anthropic, Google), the read is that Chinese models are now pulling enterprise inference workloads on price even where capability gaps exist on the margin — and that pricing pressure feeds back into per-token economics globally.

Linked stocks: NVDA US, GOOG US, MSFT US, 0700 HK, 0981 HK, BIDU US Sources: 中國 AI 大模型周調用量再超美國，騰訊混元 Hy3 preview 登頂

MediaTek hires TSMC's CoWoS architect to anchor AI ASIC push

MediaTek (2454 TT) confirmed it has hired Yu Cheng-Hua, the TSMC backend R&D veteran widely credited as the chief technologist behind CoWoS, who retired from TSMC in 2025.

Yu joined TSMC in 1994 and spent his career on advanced packaging — he is described in the announcement as "an important driving force" of TSMC's packaging program. MediaTek says he will set its forward advanced-packaging roadmap and guide R&D investment, directly supporting the company's AI ASIC business, which MediaTek has guided to "billions of dollars in revenue by 2027."

The hire is strategic on two axes. CoWoS supply has been the binding constraint on the entire AI accelerator market through 2024-2026, and MediaTek's ASIC franchise — which already counts Google TPU among its customers — depends on getting allocation from TSMC and on co-designing the package with the silicon. Bringing in the architect of the technology stack tightens that loop. It also reads as a specific response to Google's reported evaluation of going around MediaTek to procure TPU silicon directly, with Intel's EMIB-T (next signal) as a second-source candidate. If Google is pricing optionality, MediaTek is signaling it will defend the package-plus-design value proposition rather than concede the front-end. The signal also matters for talent flows: TSMC's backend R&D bench has been a near-monopoly on advanced-packaging IP, and Yu's move is the most senior public departure to a fabless customer to date.

Linked stocks: 2454 TT, 2330 TT, GOOG US, AVGO US Sources: 聯發科延攬台積電封裝老將強化AI ASIC與先進封裝布局

Intel EMIB-T reaches 90% yield as Google weighs direct TPU sourcing

Intel's EMIB-T advanced packaging has reportedly reached 90% yield, with Google simultaneously evaluating whether to procure TPUs directly from a foundry rather than via MediaTek.

The 90% number is the threshold at which a packaging process is "qualified" but still meaningfully short of the ~98% required for FCBGA-grade volume production — and the climb from 90% to 98% is harder than the climb to 90%, because residual defect modes are concentrated and architecturally rooted. Intel Foundry's pitch to Google would be a vertically integrated stack — front-end at 18A or 14A plus EMIB-T or Foveros-Direct backend — that bypasses TSMC's CoWoS allocation queue.

The strategic read is asymmetric. If EMIB-T continues to mature and Intel can demonstrate volume parity with CoWoS-S on production silicon, Google has a credible second-source for TPU packaging — which materially weakens MediaTek's pricing power on the next-gen TPU socket. If EMIB-T stalls between 90% and 98% (where most advanced-packaging programs historically have), TSMC and MediaTek retain the packaging franchise. Intel's stock action — +16% on the week — is partly pricing the optionality on this story plus the data-center CPU recovery already covered last week. For the AI ASIC supply chain, the variable to track is whether Google places any production TPU volume with Intel Foundry through 2026; that would be the demonstrated proof, not a yield report.

Linked stocks: INTC US, GOOG US, 2454 TT, 2330 TT Sources: Intel先進封裝技術EMIB良率傳達90%

Korea April chip exports +173% YoY confirm AI memory demand

Korea posted April exports above USD 80 billion for the second consecutive month, with semiconductor exports up 173% YoY — a record for the month.

Total exports grew 48% YoY. Computer exports — primarily SSDs into AI-server builds — surged 515.8% YoY. Automotive exports fell 5.5%. Korean trade data is the cleanest macro tape on the AI cycle because Samsung and SK Hynix together dominate global HBM and a meaningful share of NAND, and at this magnitude the chip line cannot be explained by anything other than HBM and high-density NAND volume plus ASP. The +515.8% computer print is the secondary tell: hyperscalers are buying SSD content in proportion to GPU shipments, not just the GPUs.

For positioning, the read is that Q2 momentum is firmly intact at the Korean export level — which historically leads earnings prints by 4-6 weeks for SK Hynix and Samsung Memory. It also reduces the probability that the recent global memory rally has run ahead of fundamentals: at +173% YoY the print is in line with, rather than ahead of, the price action in SK Hynix and Micron over the trailing six months.

Linked stocks: 005930 KS, 000660 KS, MU US, 2408 TT, 2337 TT Sources: 韓4月出口再度破800億美元 晶片出口飆173%

WIN-COM Q1 net profit +44.7% QoQ as H1 capacity expands 40%

WIN-COM Technology (6515 TT), the AI-server probe-card and test-interface specialist, posted Q1 revenue of NT$2.98 billion (+33.4% QoQ, +29.7% YoY), gross margin 43%, operating margin 26%, and EPS of NT$19.54 — all single-quarter records.

New capacity in Kaohsiung Renwu came online in April, contributing 30% of total capacity, and management guided H1 capacity expansion of 40%. The company sells into AI-server SoC and HBM testing — the high-frequency, high-speed testing socket that AI-server demand has been pulling on through 2025-2026.

The result is a clean confirmation that test-interface bottlenecks are still real. Probe cards historically scale with both wafer starts AND with chip complexity — and CoWoS-class accelerators with HBM4 require multiple test passes per packaged unit. The 40% H1 capacity step is large enough to suggest order books well into 2H26, and at 43% gross margin the franchise is not getting margin-compressed by capacity additions. WIN-COM and FormFactor (FORM US) are the two scaled probe-card vendors with AI-server exposure; this print is incremental positive for both.

Linked stocks: 6515 TT, FORM US, 3264 TT Sources: 穎崴Q1大賺近2個股本 上半年將新增4成產能, AI需求強　穎崴首季獲利創新高

Win Semi to raise Q2 prices as AI server optical chip demand accelerates

Win Semiconductors (2455 TT), Taiwan's GaAs foundry leader, reported Q1 net income of NT$195 million (+24.0% YoY) on revenue +20.9% YoY, and management plans to initiate Q2 cost pass-through with arsenide gallide product price hikes to support margin recovery.

Demand commentary points to AI server / optical data-center upgrades from 800G to 1.6T transceivers as the largest driver, with silicon-photonics integration "for faster data transmission and lower power consumption" being the architectural pull. CW (continuous-wave) laser transmitter demand is guided to grow "multiples" versus 2025. Adjacent ramps in mass production include AI glasses, automotive LiDAR, robotic vision, stratospheric drones, 5G base stations, and low-orbit satellites.

The cost pass-through is the operative signal. Taiwan GaAs foundries have been margin-squeezed for 18 months as customers absorbed input cost inflation; Win Semi raising prices in Q2 with customer acceptance implies the optical-transceiver supply chain has crossed into seller's-market territory just as the 800G-to-1.6T cycle accelerates. Read-throughs include Lumentum (LITE US) and Coherent (COHR US) on the laser side, Foxconn Interconnect Technology (6088 TT), Innolight (300308 CH), and Eoptolink for the transceiver builds, and AAOI (AAOI US) and FII for module assembly. The 1.6T transition is also closely tied to silicon-photonics co-packaged optics, which is the medium-term threat — but for the next 12-18 months pluggable 1.6T ramps faster than CPO penetrates volume.

Linked stocks: 2455 TT, LITE US, COHR US, AAOI US, 6088 TT Sources: 全新獲外資提高目標價 傳Q2啟動成本轉嫁

先聲 First Word — Exclusives from Chinese-Language Sources

Chiang Shang-yi takes Winbond independent director seat — first such role for the TSMC R&D legend

Chiang Shang-yi — the former TSMC R&D chief who is widely considered one of the most important technologists in TSMC's foundry history, and currently a Foxconn director — has been nominated as an independent director at Winbond Electronics (2344 TT), with the shareholder meeting scheduled May 29, 2026.

This is Chiang's first independent-director appointment. Winbond is Taiwan's specialty DRAM and NOR flash player and is developing CUBE, a 3D-stacking-with-TSV memory architecture targeting edge AI applications — a positioning that overlaps directly with Chiang's HBM and advanced-packaging expertise. The Jiao family is also expanding board representation, with Jiao Tze-yi (third generation) nominated as the institutional director rep for Walsin Lihwa. (May 4, 2026) Source: 蔣尚義將任華邦電獨立董事

Sintec (欣銓) Longtan AI ASIC test ramp begins Q3

Sintec Technology (3264 TT) used its Q1 2026 results call to confirm sequential revenue and profit growth quarter-by-quarter through 2026, with first-floor production at the new Longtan AI ASIC and wafer-test plant starting Q3 and the remaining six cleanroom floors phasing in over two years.

Q1 capex of NT$1.24 billion (+18% YoY) is going into advanced test equipment with 6-8 month lead times. IC design customers are 56.1% of revenue, IDM 42.7%, with comm chips the biggest growth contributor at 32.5% of Q1 revenue. Management cited "strong demand from US customers for AI ASIC testing" plus memory recovery and Taiwan telecom-chip spillover as the order-book drivers. (May 4, 2026) Source: 欣銓今年逐季成長 龍潭廠Q3起挹注

Richwave (4968) targets 35% gross margin recovery on 3x Wi-Fi 7 ramp

Richwave Technology (4968 TT), the Taiwan RF chip vendor, used its May 4 investor conference to reaffirm full-year double-digit revenue growth and a return to 35% gross margin, with H2 guided above H1.

Q1 gross margin was below target on supply-chain cost inflation; rather than raise prices, management is leaning on cost reduction, yield, materials substitution, and scale. The growth driver is Wi-Fi 7 — projected to ship over 3x YoY in 2026 — which carries higher margin than legacy SKUs. Wi-Fi 8 remains immaterial for the year. The setup is the standard Taiwan RF playbook: ride the new-spec ramp through the gross margin recovery, with mix shift carrying most of the lift. (May 4, 2026) Source: 立積法說會／看今年維持雙位數成長 下半年優於上半年、毛利率拚回35%