Zero One Daily Intelligence Brief — May 28, 2026

Every Major Memory Maker Invests in Anthropic; TSMC Huawei Threat? Jensen Says Over-Rated

Top Market Signals

Anthropic tops OpenAI in valuation as every major memory chip maker joins US$65bn raise

Anthropic raised US$65bn in a Series H round at a US$965bn post-money valuation, surpassing OpenAI as the world's most valuable AI startup, with Samsung Electronics, SK Hynix, and Micron joining as strategic infrastructure partners.

The round was led by Altimeter Capital, Dragoneer, Greenoaks, Sequoia, Capital Group, Coatue, and D1 Capital Partners, with additional institutional participation from Baillie Gifford, Blackstone, Brookfield, DST Global, DE Shaw, and Fidelity. Roughly US$15bn of the total represents hyperscaler pre-commitments, including the US$5bn Amazon investment announced in April. The company indicated this round is likely its final fundraise before an IPO.

Separately, Apollo Global Management and Blackstone are working to bring investors into a roughly US$36bn debt financing facility to purchase Google-sourced chips for Anthropic's compute buildout. The deal structure routes procurement through a third-party vehicle to remove it from Anthropic's balance sheet. Korean media also reported that Samsung Foundry is the leading candidate to win orders for a custom "Claude AI Chip" following similar decisions by Tesla and NVIDIA to use Samsung's advanced process nodes.

The strategic participation of Samsung, SK Hynix, and Micron is the significant supply-chain read-through. Together these three companies supply effectively all global HBM — a reminder that AI is not just the most important driver of GPU and server demand but also of memory demand: each HBM stack consumes multiples of the fab capacity of a conventional DRAM module, making AI the most memory-fab-intensive computing generation by a wide margin. AI model companies are now structuring capital raises that pull memory suppliers directly into the ownership cap table, binding them commercially to the GPU compute buildout they already serve. The pattern is the same one that ran through NVIDIA's equity stake program in suppliers: customers securing supply through capital rather than just contracts.

Linked stocks: NVDA US, 005930 KS, 000660 KS, MU US

Sources: Anthropic 新融資獲 650 億美元，三星、SK 海力士及美光策略參與, Bloomberg, Google News

Samsung trying to take memory leadership from Micron and SK Hynix as it ships world's first HBM4E 12-High samples to global customers

Samsung Electronics (005930 KS) shipped the world's first HBM4E (high-bandwidth memory 4E) 12-High samples to global customers on 28 May 2026, a full generation ahead of the HBM4 products SK Hynix currently leads in production.

HBM4E is the next-generation specification after HBM4, adding another layer to the 12-High stack and delivering higher bandwidth density per package. Samsung's announcement follows SK Hynix's dominance of the HBM4 supply chain for NVIDIA's Blackwell and Vera Rubin platforms, and Korean media framed the news pointedly: "Is SK Hynix watching?" The sampling milestone puts Samsung roughly six to nine months ahead of commercial HBM4E shipments, depending on customer qualification cycles.

The competitive significance is timing. SK Hynix has locked in preferred supplier status on HBM4 with NVIDIA and major hyperscalers, with gross margins reportedly running near 72% on AI memory. Samsung, which spent much of 2025 struggling with HBM yield issues, appears to have closed the gap at the HBM4E generation. HBM4E qualification at NVIDIA would represent Samsung's first re-entry into the top-tier AI memory supply chain since losing ground to SK Hynix in 2024.

For memory pricing broadly, sampling of HBM4E signals that high-end HBM capacity will remain constrained through at least mid-2027 while both Samsung and SK Hynix qualify the next generation alongside shipping current-gen volumes.

Linked stocks: 005930 KS, 000660 KS, NVDA US, MU US

Sources: Co, Google News

Jensen Huang holds Taiwan's 5th "Trillion Dollar Banquet"; Says Huawei threat to TSMC over-rated

NVIDIA (NVDA US) CEO Jensen Huang hosted his 5th "Trillion Dollar Banquet" (兆元宴) in Taipei on 28 May 2026, convening Taiwan's semiconductor and AI supply chain leadership at a single dinner table, while publicly dismissing Huawei's newly announced "Tau (τ) Scaling Law" as no threat to TSMC's position.

Attendees included TSMC chairman Wei Che-chia, Foxconn chairman Young Liu, Quanta chairman Lin Bai-li, ASUS chairman Shih Chung-Hwa, Pegatron chairman Tong Tze-Hsien, Wistron chairman Lin Hsien-ming, and MediaTek CEO Tsai Li-hsing. The dinner took place the week before COMPUTEX 2026 opens on 2 June. Taiwan's total stock market capitalization has doubled to approximately NT$155 trillion since Jensen Huang held the first banquet in 2023.

On Huawei's Tau Scaling Law, which claimed die-stacking techniques could bypass leading-edge process constraints and reach 1.4nm-equivalent transistor density by 2031, Jensen Huang said: "This is a breakthrough for Huawei, but it is not a threat to TSMC. TSMC and Taiwan have had this technology for nearly 10 years." He specified that hybrid bonding and 3D packaging, the core techniques behind Tau, are already embedded in TSMC's CoWoS (chip-on-wafer-on-substrate) and SoIC (system-on-integrated-chips) roadmaps.

Jensen Huang flagged energy as the constraint beyond chips. He said NVIDIA's annual procurement from Taiwan's supply chain has risen from roughly US$10-15bn several years ago to US$100-150bn today. He also stated that AI will become 10 to 100 times larger than the prior chip-and-systems market, but that Taiwan needs significantly more energy to support both manufacturing scale-up and the domestic AI data centers required to capture that value.

Linked stocks: NVDA US, 2330 TT, 2317 TT, 2382 TT

Sources: 談華為半導體新突破　黃仁勳：台積電和台灣領先10年, 黃仁勳示警台灣 AI 下一關卡，不是晶片、是電力, 點評華為「韜定律」黃仁勳：威脅不到台積電！

Wiwynn warns AI data center bottleneck widens from memory to networking chips, easing only by 2028

Wiwynn (6669 TT), one of NVIDIA's largest AI server manufacturers, warned Bloomberg on 28 May 2026 that supply shortages in AI data center buildout have expanded well beyond memory chips to include networking chips and other critical components, pushing hardware prices to record levels.

Wiwynn chairman Hong Li-ning told Bloomberg that with Meta, Microsoft, and other hyperscalers accelerating capital expenditure, hardware demand across the full data center stack will remain strong for three to five years. "The items in short supply change somewhat each year; we may only start to see supply tightness ease in late 2027 or 2028," she said.

The company currently sources 80% of its revenue from US customers. Wiwynn's first Texas factory, in El Paso, is now operational, and the company plans to build three additional US factories over the next two years. Wiwynn's full-year revenue is expected to grow nearly 30% this year as inference demand from cloud service providers drives strong orders on standard servers alongside AI configurations.

The statement extends the supply tightness narrative flagged in prior Blades beyond memory. Our 27 Apr 2026 Blade noted Nanya Technology's qualification for NVIDIA's Vera Rubin LPDDR supply chain as evidence of memory tightness spreading to new vendors. Wiwynn's Bloomberg warning covers the next layer: networking chips, power components, and broader server hardware, where lead times are stretching independently of the memory cycle.

Linked stocks: 6669 TT, NVDA US, 2317 TT, MSFT US

Sources: 不只記憶體！輝達伺服器製造商緯穎：AI瓶頸2028才見緩解, 緯穎示警AI供應鏈瓶頸擴大 短缺恐延續至2028年

Marvell FY27 Q1 beats on AI custom chips; FY28 revenue guide raised 10% to US$16.5bn

Marvell Technology (MRVL US), an NVIDIA portfolio company and major TSMC customer for 3nm and 2nm AI silicon, reported FY27 first-quarter revenue of US$2.42bn (+28% YoY), beating consensus, and raised its FY28 full-year revenue outlook by 10% to approximately US$16.5bn.

For the current quarter (FY27 Q2, ending July 2026), Marvell guided revenue to approximately US$2.7bn with non-GAAP EPS of US$0.93, both above analyst expectations. CEO Matt Murphy said custom-chip-related order intake is "extremely strong" across all major US hyperscalers. Marvell has active custom silicon programs with every major cloud service provider.

The custom ASIC (application-specific integrated circuit) division is the high-growth driver: Marvell guided that segment revenue will grow more than 20% in FY27, double in FY28, and exceed US$10bn in FY29, up from the prior US$8bn target. The company also raised its FY27 full-year revenue outlook to approximately US$11.5bn, implying roughly 40% annual growth from FY26.

Marvell's read-through to TSMC is direct. The company processes all of its leading-edge AI silicon through TSMC's 3nm and 2nm nodes and relies on CoWoS packaging for high-bandwidth AI chip assemblies. The 10% upward revision to FY28 guidance means more advanced node wafer starts and more CoWoS throughput are required from TSMC two years out than were previously planned.

Linked stocks: MRVL US, 2330 TT, NVDA US, AVGO US

Sources: 邁威爾上調展望 反映AI需求暢旺 台積受惠, 邁威爾上季財報與展望報喜 協力廠華星光跟著旺

Synopsys raises FY26 guide; CEO says in-house chip design is now a hyperscaler structural trend

Synopsys (SNPS US), the dominant EDA (electronic design automation) software vendor, raised its FY26 revenue guide and reported Q2 revenue of US$2.28bn (+42% YoY) on 27 May 2026, as CEO Sassine Ghazi declared that hyperscaler in-house chip design has become a structural macro trend that Synopsys alone can monetize at the IP level.

Q2 results beat consensus revenue of US$2.25bn and non-GAAP EPS consensus of US$3.15. For Q3 (May-July 2026), Synopsys guided revenue of US$2.41-2.46bn, substantially above the US$2.21bn analyst estimate. Full-year FY26 revenue guidance was raised to US$9.63-9.71bn from the prior range of US$9.56-9.66bn.

Ghazi made the strategically significant statement during the earnings call: "On the IP side, the trend of hyperscalers designing their own data center chips is now established. Without Synopsys IP, you simply cannot build your own chip." He added that Synopsys is negotiating new agreements with major customers under which clients will pay per-chip royalties on devices manufactured using Synopsys technology, shifting the business model toward volume-linked revenue.

Elliott Investment Management, the activist fund, will receive a board seat as part of a separately announced cooperation agreement. Ghazi said he and Elliott have had "constructive" discussions on monetizing Synopsys' software assets and IP portfolio more aggressively.

Linked stocks: SNPS US, CDNS US, 2330 TT, MRVL US

Sources: 新思調高年度財測 CEO：自研資料中心晶片成趨勢

先聲 First Word — Exclusives from Chinese-Language Sources

Qualcomm deepens push into PCs, launching Snapdragon C for entry-level AI PCs

Qualcomm (QCOM US) launched the Snapdragon C platform on 28 May 2026, a new entry-level processor targeting students, families, and small businesses at a price point above US$300, with Acer, HP, and Lenovo as launch OEMs. The platform integrates an NPU for on-device AI inference, targeting a segment previously dominated by Intel's low-power Celeron and Atom lines. (28 May 2026) Source: 高通推入門款 Snapdragon C 晶片！鎖定平價 AI 筆電與學生市場

Vanguard International at full capacity, opening price increase renegotiations with customers

Vanguard International (5347 TT), Taiwan's leading mature-node foundry, reported at its 28 May 2026 annual shareholder meeting that capacity is fully loaded and it is reopening price discussions with customers amid rising utility, materials, and labor costs. In February 2026 industry sources had reported Vanguard hiked selected product prices by 15% from April. Chairman Fang Lue said the company expects 2H26 industry conditions to be better than 1H26. (28 May 2026) Source: 世界產能滿載 喊漲價 下半年產業景氣將優於上半年

Powertech Technology April profit +183% YoY; FOPLP and silicon photonics entering production

Powertech Technology (6239 TT), Taiwan's major IC assembly and test house, reported April 2026 net profit of NT$883m, +183% YoY and NT$0.95 per share. The company said FOPLP (fan-out panel-level packaging), silicon photonics, and CPO (co-packaged optics) technologies are entering initial production ramps this year, alongside pricing tailwinds from memory and logic chip price increases. Management expects quarterly revenue growth to accelerate throughout the year. (28 May 2026) Source: 力成自結4月獲利大增183% 每股賺0.95元