Zero One Daily Intelligence Brief — April 23, 2026

Intel Beat and TSMC Roadmap Reset the AI Silicon Balance of Power

Top Market Signals

Intel's AI payoff lands with a TeraFab twist

Intel (INTC US) delivered a strong Q1 revenue forecast beat and signaled that the chipmaker is finally capturing share of the AI infrastructure build-out, while CEO Lip-Bu Tan paired the results with a pointed endorsement of Elon Musk's TeraFab chip manufacturing project.

Intel reported Q1 2026 revenue of $13.6bn, with GAAP loss per share of $(0.73) and non-GAAP EPS of $0.29. Intel Foundry revenue climbed 16% YoY as the company expanded manufacturing partnerships, and management said 18A and 14A process progress is running ahead of expectations. The forecast beat drove shares sharply higher in after-hours trade and lifted CPU peers AMD (AMD US) and Arm Holdings (ARM US) with it; ARM hit a new all-time high after Intel's print, extending a 26% one-week rally.

The more strategic disclosure was Tan's remarks on the Musk-led TeraFab project, where the Intel CEO said he could "think of no better partner than Elon Musk" to explore unconventional ways to improve chip manufacturing and drive down cost. Musk is separately planning a SpaceX-Tesla joint venture around the TeraFab concept that analysts are already framing as a potential disruption vector for Nvidia (NVDA US) and the broader AI silicon incumbents. Intel's participation positions 18A and 14A as the candidate nodes for that capacity, and fits the company's narrative that an external anchor customer is materializing. Tesla (TSLA US) reinforced the ambition on the same day, announcing plans to spend an additional $25bn this year to support Musk's AI roadmap.

For the AI server read-through, Bloomberg analysis highlighted that Intel's data center CPU demand is beating expectations, suggesting general-purpose server refresh is running alongside the accelerator cycle rather than being crowded out. That is consistent with Taiwan ODM commentary that ASIC volume growth pulls general-purpose server co-shipments, and it supports the thesis that the AI capex wave is broadening beyond pure GPU content.

Linked stocks: INTC US, AMD US, ARM US, NVDA US, TSLA US Sources: Bloomberg, Tom's Hardware, Google News, Google News, Bloomberg

TSMC unveils A13 and A12, defers High-NA EUV, and breaks ground on Arizona packaging

TSMC (2330 TT / TSM US) used its 2026 North America Technology Symposium to unveil two new process nodes beyond A14, lock in a 2029 timeline for its Arizona advanced packaging fab, and disclose that it will skip ASML's High-NA EUV before 2029 on cost grounds.

At the forum, TSMC revealed the A13 node as a nanosheet-architecture refinement emphasizing density, performance, and power efficiency for AI and HPC workloads, with design-technology co-optimization (DTCO) central to the pitch. The company paired that with A12, a strengthened variant of A14 incorporating backside power delivery; both A13 and A12 are targeted for 2029 production. The roadmap effectively gives TSMC a series of optimized 2nm-and-below nodes stretching through the decade and widens its gap over Samsung (005930 KS) and Intel on stated cadence.

Advanced packaging took equal billing. Zhang Xiaoqiang, senior vice president of global business and deputy co-COO, said TSMC will bring an Arizona chip packaging fab online by 2029, the company's first packaging site outside Taiwan. CoWoS capacity is now running at 5.5x reticle size, with a 14x reticle-size version planned for 2028. The Business Journals confirmed the Phoenix ground-breaking. The readthrough for Taiwanese suppliers is direct: equipment names including Scientech, GPTC and consumables peers are positioned as first-round beneficiaries as CoWoS and 3D stacking move from "supporting" to "system core" roles in the AI compute stack. The Cadence (CDNS US) collaboration disclosed alongside the event targets next-gen AI silicon design on TSMC nodes.

The most commercially significant deferral was on High-NA EUV. Zhang said TSMC has no plan to deploy ASML's (ASML NA) latest High-NA EUV tools before 2029, citing cost: a single machine retails for over €350mn, roughly double current-generation EUV. TSMC will instead deepen exploitation of existing EUV, reframing the advanced-process race from an "equipment arms race" to one of process integration and capital efficiency. For ASML, this confirms near-term High-NA demand will lean on memory and Intel, not TSMC's 2nm / A-series ramp. For the Taiwan supply chain, the message is that capex intensity is being managed, which supports the duration of the current upcycle.

Linked stocks: 2330 TT, TSM US, ASML NA, CDNS US, 005930 KS, INTC US Sources: 台積美先進封裝廠2029啟用, 台積電暫緩導入High-NA EUV, 台積A13製程出鞘 台鏈熱起來, Google News, Google News

Texas Instruments surges 19%, signaling analog cycle inflection

Texas Instruments (TXN US) closed 19.4% higher in its best single-day move since 2000 after Q1 results showed both AI data center and industrial analog demand accelerating, with management opening the door to a further round of price hikes in the second half.

TI's Q1 print surprised on demand breadth. The company flagged that AI data center-linked analog demand surged, while industrial applications showed a clear recovery trend. Crucially, TI said current industrial demand is still well below the prior peak, implying the recovery has further to run. Management signaled it does not rule out another price increase in H2. TI's DLP and embedded processing senior vice president Amichai Ron said in a separate briefing that edge AI opportunities extend beyond robotics, arguing that existing product lines can also be repositioned for AI use cases, a broader monetization angle than investors have priced.

The stock's reaction rippled through the analog complex. On Semi (ON US) and Analog Devices (ADI US) were noted as heading for record highs on the TI print, and TI's strength is a read-across to the wider industrial semi cycle that has been languishing for six quarters. TI is simultaneously the AI data center analog beneficiary (power management, signal chain) and the broadest proxy for industrial end markets, so a beat of this magnitude carries dual signal value.

Linked stocks: TXN US, ON US, ADI US Sources: Google News, 資料中心、工控類比晶片暢旺 德儀喊出不排除下半年再漲價, 德儀嵌入式資深副總：邊緣AI商機不只機器人 既有產品也能轉型, Google News

Google's 8th-gen TPU split shifts ASIC server share toward Taiwan

Google Cloud Next 26 introduced a bifurcated 8th-generation TPU, with TPU 8t optimized for training and TPU 8i optimized for low-latency inference, and Taiwan system assemblers are quietly increasing their share of the build at Celestica's expense.

The dual-chip TPU architecture is a first for Google, co-developed with DeepMind, and both variants are expected to ship late in 2026. Per DIGITIMES Research, ASIC server shipments are projected to grow 64.2% in 2026, outpacing GPU server growth of 43.8%, and Google's TPU is expected to hold roughly 46% share of the ASIC accelerator market, with the majority still TPU v7 (Ironwood). The supply-chain readthrough is where today's signal lands: Celestica (CLS US) remains the primary TPU server assembler, but Hon Hai (2317 TT) now handles L10-L11 integration, and Inventec's (2356 TT) share of CPU motherboard production has reportedly risen from around 10% to 30%, eroding Celestica's historical lock. IC design is also broadening beyond Broadcom (AVGO US) with MediaTek (2454 TT) now participating.

Volume ramp is anchored in 2H26 for assemblers (Celestica, Inventec, Hon Hai), with thermal and power component suppliers Auras (3017 TT) and Cooler Master stepping up production at the end of Q2. The Rambus (RMBS US) SOCAMM2 server module chipset announcement sits adjacent: AI training and inference systems are demanding new memory subsystem topologies, expanding beyond traditional DRAM form factors. This broader ASIC ramp also pulls general-purpose server volumes, which benefits Quanta (2382 TT) as Google's general-purpose partner.

Linked stocks: GOOGL US, CLS US, 2317 TT, 2356 TT, 2454 TT, AVGO US, 3017 TT, RMBS US, 2382 TT Sources: Google第8代TPU衝刺ASIC戰力 台灣系統廠代工比重急增, Google TPU分拆策略優化AI算力 Gemini共同設計、最大化能效, 科技1分鐘：Google第八代TPU 8t、TPU 8i, Google News

MATCH Act takes aim at CXMT and YMTC, tightening memory supply

The US House Foreign Affairs Committee passed 20 bipartisan export-control bills including the MATCH Act, which specifically targets equipment sales into China memory makers ChangXin (CXMT) and Yangtze Memory (YMTC) after direct lobbying from Micron (MU US).

The MATCH Act would plug gaps in US chip equipment restrictions, including broader coverage of DUV lithography tools, and would pressure allies to match US stringency. Reuters reporting cited by United Daily News noted that Micron CEO Sanjay Mehrotra personally participated in the drafting process and held a closed-door roundtable with committee members, framing CXMT and YMTC as a national security issue to prevent China memory from replicating the dominance trajectory of Chinese solar.

The industry context makes the timing consequential. YMTC's Q1 revenue exceeded RMB 20bn with YoY growth more than doubling, and its NAND global market share has now passed 10%. YMTC runs two existing fabs with combined capacity of 200k wafers per month, is completing a new fab this year, and plans two additional fabs each targeting up to 100k wafers per month. CXMT is focused on DRAM and unveiled DDR5 and LPDDR5X product lines in November 2025. If MATCH-level controls land, Taiwan memory makers Nanya Technology (2408 TT) and Winbond (2344 TT) are directly positioned to benefit from reduced red-supply-chain expansion risk. The bills still require House floor votes; the most likely path to enactment is attachment to the annual defense authorization bill.

Separately, HP (HPQ US) chief commercial officer Dave McQuarrie warned that memory pricing pressure driven by AI data center demand could persist multiple quarters and may not ease until after 2027, and HP is using bundling and spec flexibility to cushion PC pricing impact. SK Hynix (000660 KS) posted record quarterly profit on AI-driven demand, and Samsung (005930 KS) and SK Hynix are positioned for a close Q1 memory operating margin race per thelec.net, with Korea Investment & Securities raising its SK Hynix price target on superior profitability.

Linked stocks: MU US, 2408 TT, 2344 TT, HPQ US, 000660 KS, 005930 KS Sources: 美光出招 圍堵記憶體紅鏈 南亞科、華邦等將受惠, 惠普警告記憶體緊縮恐再延多季 以搭售與規格彈性緩PC漲價衝擊, Google News, Google News

GPT-5.5 lands as Meta cuts 10% and Anthropic valuation tops $1 trillion

OpenAI launched GPT-5.5 while Meta (META US) prepared to lay off roughly 8,000 employees (about 10% of staff), Microsoft (MSFT US) readied its first-ever voluntary departure program, and Anthropic's implied secondary-market valuation overtook OpenAI at roughly $1 trillion.

OpenAI president Greg Brockman called GPT-5.5 a step toward a unified "super app" combining ChatGPT, Codex, and an AI browser, and chief scientist Jakub Pachocki noted significant short-term improvements with extremely significant improvements expected in the medium term, while characterizing the last two years as surprisingly slow. OpenAI published benchmark data showing GPT-5.5 outperforms Google Gemini 3.1 Pro and Anthropic Claude Opus 4.5 across multiple tests. The model release is paired with new consumer-app connectors (Spotify, Uber, Instacart, TurboTax) pushing Claude and ChatGPT toward daily-utility positioning.

The valuation backdrop: Anthropic's "Mythos" AI model has pushed the company's secondary-market implied valuation to around $1tn, overtaking OpenAI in the market-cap race per New York Post reporting; Bloomberg separately notes Mythos is driving fear about AI deployment speed even as it potentially benefits Anthropic commercially, with potential valuations approaching $800bn on other framings. Jeff Bezos's Project Prometheus physical-AI lab closed a $10bn round at a $38bn valuation.

On the labor side, Meta's 8,000-job cut and Microsoft's first-ever voluntary exit program are happening alongside AI capex scaling, not despite it. SoftBank (9984 JT) secured a $10bn loan backed by its OpenAI stake, and is converting part of an Osaka factory into one of Japan's biggest large-scale battery lines to power its own AI data centers. For investors, the signal is that hyperscaler labor restructuring and LLM model churn are now running at roughly one-quarter tempo, with Anthropic, OpenAI, Google, and Meta each refreshing within the last three months.

Linked stocks: MSFT US, META US, GOOGL US, 9984 JT Sources: TechCrunch, Google News, The Verge, AI投資與成長放緩夾擊 微軟推優退方案調整體質, Bloomberg, Bloomberg

先聲 First Word — Exclusives from Chinese-Language Sources

SMIC pivots seriously into advanced packaging

SMIC is quietly shifting strategic focus from pure front-end to back-end advanced packaging, setting up a fully-owned RMB 3.1bn subsidiary, Shanghai Xin Sanwei Semiconductor, in Shanghai Lingang Free Trade Zone, and is aggressively recruiting an advanced packaging team. The move positions China's largest foundry to compete with ASE (3711 TT) and TSMC InFO/CoWoS at the back-end and absorb more of the value chain domestically, a material negative for Taiwan OSAT share over the 2027E-2028E horizon if execution lands. (24 Apr 2026)

Source: 中芯國際重返封測「玩真的」 傳先進封裝團隊大舉招兵買馬

Taiwan packaging equipment maker turmoil threatens CoWoS and CoPoS order redistribution

A Taiwanese packaging equipment supplier is in the middle of sharp senior-management upheaval, and supply-chain reporting indicates TSMC CoWoS, panel-level packaging (CoPoS), and ASE group equipment orders could be reshuffled. Scientech / Hong Su Technology (弘塑), which rode the TSMC advanced packaging ramp since 2023, was flagged as the reference name in the coverage. Order redistribution at this stage of the CoWoS buildout is not a minor supplier footnote; it bears on 2027E capacity readiness. (24 Apr 2026)

Source: 台系設備廠宮鬥疑雲難解 台積CoPoS設備訂單恐重分配

Shenzhen stands up China's first 10,000-card domestic AI compute cluster

Shenzhen's municipal government, with Huawei, Phytium, and YMTC, unveiled China's first 10,000-card-scale domestically produced AI compute cluster at Shenzhen Bay Technology Ecology Park, positioned to break foreign monopoly and progress toward self-sufficient compute. Separately, China's MIIT disclosed that nationwide intelligent compute capacity reached 1,882 EFLOPS as of end-March, a figure some experts argue implies a "dark pool" of compute orders of magnitude larger than publicly reported. Together these point to a sharper trajectory for the domestic alternative stack than Western counts assume. (24 Apr 2026)

Source: 深圳建成中國首個萬卡級國產算力叢集 打破海外壟斷、邁向中國自主算力

Packaging price hikes now outpace foundry hikes for chipmakers

DIGITIMES reports that over the past several weeks, chip vendors from majors to mid-cap IC designers have issued price-increase notices or reopened pricing negotiations with customers, and for many IC designers packaging is now the single largest cost-pressure source, rising faster than foundry. This inverts the standard narrative that foundry is the marginal cost driver. If sustained, it meaningfully re-rates OSAT pricing power versus foundry and pressures fabless gross margins into 2H26. (24 Apr 2026)

Source: 封測漲價比晶圓代工還多 晶片廠面臨最大成本壓力源