Zero One Investment Research Daily Intelligence Brief, July 27, 2026

NVIDIA weighs a US$250bn guarantee for OpenAI; Korea signs US$950bn of chip deals

Top Market Signals

NVIDIA would underwrite OpenAI's landlord, not its chips

NVIDIA (NVDA US) is in talks to provide roughly US$250bn of financing guarantees to OpenAI, according to a Wall Street Journal report relayed by Reuters and CNA. The guarantee would let OpenAI lease a 10GW campus in southern Ohio being developed by SoftBank's energy arm, a project whose total cost including chips could pass US$500bn.

The structure is the part worth reading closely. People familiar with the talks told the paper the US$250bn covers the data center lease payments and the construction debt behind them, and specifically does not cover the NVIDIA chips that will run inside. NVIDIA is separately in discussions to finance OpenAI's chip purchases, a facility that could reach US$350bn, on top of the US$30bn it has already invested in the company. Nothing is signed, and the talks could still collapse.

Power is the reason the project sits where it does. The electricity allocation is controlled by the US government and funded solely by Japan under a recently agreed trade deal, with Commerce Secretary Howard Lutnick personally involved in the allocation decision. OpenAI has held weeks of talks over the Ohio lease and is the most committed bidder, though Anthropic, Microsoft (MSFT US) and Google have all met Lutnick about the same capacity in recent weeks.

Our 23 July 2026 Blade flagged AMD committing up to US$5bn to Anthropic in an arrangement that routes back into AMD server purchases. This is the same mechanism two orders of magnitude larger, and it is now drawing named opposition: Bloomberg reported NVIDIA is working on a fresh round of infrastructure deals worth more than US$750bn, and both Jim Chanos and Michael Burry have publicly described the OpenAI structure as NVIDIA financing its own chip sales. The guarantee carves the chips out of the collateral, which is the answer to that charge, but it also means NVIDIA carries the property risk while the chip revenue stays on a separate contract.

Linked stocks: NVDA US, MSFT US, 2330 TT

Sources: 輝達擬提供2500億美元擔保　助OpenAI建資料中心, Bloomberg, 傳輝達提供2500億美元擔保 支持OpenAI資料中心

US$950bn of Korean chip supply goes under contract

Samsung Electronics (005930 KS) and SK Group signed about US$950bn of AI chip agreements with US technology buyers over the weekend, the largest single tranche of forward memory and foundry commitments the industry has seen. Korean shares priced it on Monday, with the KOSPI closing up nearly 1% and SK Hynix (000660 KS) rebounding more than 3%.

SK Group and NVIDIA account for the larger half. The two signed letters of intent covering a partnership described as worth more than US$500bn, spanning three parts: SK Telecom will build a 2GW NVIDIA Vera Rubin DSX AI factory with the first data center entering service in 2027, SK Hynix enters a long-term arrangement to secure and co-develop next-generation AI memory including HBM (high-bandwidth memory), and the group will offer sovereign AI services. Because these are letters of intent, the financial split across SK Telecom, SK Hynix and the wider group was not disclosed.

Samsung's US$200bn memorandum with Broadcom (AVGO US) is the one that matters most for Taiwan. Signed on 24 July 2026, it covers HBM4 supply and 2nm manufacturing through 2030, and it is structured as turnkey work: foundry and advanced packaging bundled with the same vendor rather than split across suppliers. That bundling is the specific thing Samsung has been unable to sell against TSMC (2330 TT), and Broadcom is the largest merchant designer of custom AI accelerators.

The read-through is that memory and leading-edge capacity through 2030 is being contracted rather than bought on the spot market, which suits the incumbents and squeezes anyone arriving late. It also gives Samsung's foundry arm its first anchor customer of a size that can fund the node roadmap, arriving a week after the union survey flagged in our 21 July 2026 Blade found 81.5% of 8,297 foundry employees intending to leave.

Linked stocks: 005930 KS, 000660 KS, AVGO US, 2330 TT

Sources: Skhynix, Cnbc, Samsung

NVIDIA trims Vera Rubin memory as prices bite

NVIDIA has begun redesigning the Vera Rubin NVL72 rack to carry less memory, according to a GF Securities research note reported by TechNews and Liberty Times. The brokerage estimates the changes cut the platform's relevant costs by close to half, and they are a direct response to memory pricing rather than to any performance decision.

The specific cuts are on the cheap memory, not the expensive kind. SOCAMM modules (small compressed attached memory modules, the compact DRAM packages that sit next to the processor) drop from a planned 192GB to 96GB. Total memory attached to the Vera CPU falls from a planned 54TB to 55TB down to 28TB. The 20.7TB of HBM4 per rack that does the actual AI work is untouched.

The arithmetic behind the decision is unusually explicit. GF Securities calculates that without the changes, memory would cost US$2.1m per VR200 system and account for 29% of the bill of materials, against a 20% ceiling NVIDIA works to. Trimming the LPDDR5X content takes that line from an estimated US$1.2m to US$586k, and to US$293k if the capacity is cut to a quarter. Bernstein separately puts the cost of a single NVL72 Rubin rack at up to US$9.1m, above an earlier US$7.8m estimate that used pre-inflation memory prices, and models HBM4 reaching US$53 per GB by 2027.

For the memory makers this cuts two ways. NVIDIA locked in long-term contracts before the shortage, so its own supply is secure, but a platform vendor engineering memory content downward is the first hard evidence that current DRAM pricing has passed the point where the largest buyer will simply absorb it. The pressure lands on commodity LPDDR content rather than HBM, which keeps the high-margin business intact and pushes the adjustment onto conventional DRAM lines.

Linked stocks: NVDA US, 000660 KS, 2408 TT, MU US

Sources: TechNews, 貴到撐不住！輝達砍半Vera Rubin記憶體容量

CXMT debuts as mainland China's most valuable stock

ChangXin Memory Technologies closed its first day on Shanghai's STAR Market up 466%, reaching a market value of about RMB 3.3tn (US$487bn) and passing every other company listed on the mainland. The Hefei DRAM maker raised RMB 57.92bn (US$8.6bn) at RMB 8.66 per share, the largest initial public offering in Asia this year and the second largest in STAR Market history behind SMIC.

Turnover on the debut exceeded RMB 140bn. CXMT was founded in 2016 by chairman Zhu Yiming and is now the fourth largest DRAM manufacturer in the world, having moved from dependence on state subsidy and sustained losses to genuine pricing power as AI data center construction turned memory into the scarcest input in the chain.

Pricing is where the story turns uncomfortable for the incumbents. DigiTimes reports that CXMT and NAND maker YMTC have begun quoting above Samsung on some parts, an inversion of the historical order in which Chinese memory competed on price alone. Chinese PC brands are reported to have booked CXMT capacity through the end of 2027.

The constraint has not gone away. CXMT still faces equipment export controls that cap how far it can push process nodes, and its filings did not spell out HBM investment. What has changed is the funding position: a company that was a subsidy case two years ago now has US$8.6bn of fresh equity and a valuation that lets it pay for talent and tools at market rates.

Linked stocks: 005930 KS, 000660 KS, MU US, 2408 TT

Sources: Cnbc, Scmp, 硬槓華為、報價超車三星　中國記憶體雙雄翻轉產業秩序

Korea's chip exports are rotating toward Taiwan

The Bank of Korea published a production and supply chain map of the country's main manufacturing industries on 27 July 2026, and the semiconductor trade data in it is the clearest official measure yet of how the AI buildout has rewired Asian chip flows.

Taiwan took 9.0% of Korea's semiconductor exports in 2022, its fourth largest market. Last year that share reached 19.9%, moving Taiwan to second. China remains the largest single destination but its share fell from 53.1% to 40.3% over the same period.

The central bank attributes the shift to a change in what is being bought. Jung Sung-yeop, who heads regional research support at the bank's research department, said demand has moved from CPUs and conventional DRAM toward GPUs and HBM, and that this has made the global AI semiconductor supply chain anchored on NVIDIA, TSMC, Samsung Electronics and SK Hynix more firmly established.

Read alongside the US$950bn of supply contracts signed this weekend, the trade data says the same thing from the other direction: Korean memory is increasingly shipped to Taiwan to be packaged with TSMC logic before it reaches a US customer, and the China-facing volume that once defined Korean memory exports is being displaced rather than merely growing more slowly. The Bank of Korea also noted that Korean semiconductor output has concentrated further at home, with the capital region rising from 74.5% of national production in 2014 to 82.3% in 2024.

Linked stocks: 2330 TT, 000660 KS, 005930 KS, 3711 TT

Sources: 韓銀：半導體對台出口增加，形成 AI 半導體供應鏈

Packaging shortage now gates NVIDIA's optical switches

Co-packaged optics, which moves the optical components onto the switch chip's own substrate instead of into pluggable modules, has reached volume production, TrendForce said on 27 July 2026. NVIDIA is shipping its new Spectrum-X switch to selected partners and Broadcom's 51.2T Bailly switch continues to ship in small volume. The research house named three supply constraints that will set how fast either can scale, and all three sit in Taiwan's part of the chain.

The first is the optical engine, which integrates lasers and modulators in one complex, yield-sensitive process with a thermal management problem attached. TrendForce says only a handful of suppliers can build them in volume. The second is silicon photonics wafer foundry and back-end assembly, where precision and reliability requirements make capacity slow to add, leaving little short-term flexibility. The third is advanced packaging: these switches lean heavily on TSMC's COUPE process, and AI accelerators and high-performance computing parts are competing for the same lines.

The product detail explains why the packaging line is the binding one. NVIDIA co-developed Spectrum-X with TSMC on COUPE, it handles 400 Tb/s, and capacity expands further in 2H26. Broadcom's Bailly was brought into production with Delta Electronics (2308 TT) and Micas Networks, cuts power consumption by up to 70% against traditional pluggable transceivers, and has been validated in deployment by Meta (META US).

Responses are already diverging. Some cloud operators are locking supply with long-term purchase agreements and investing directly in upstream silicon photonics firms, NVIDIA is deepening its TSMC relationship to integrate design through packaging, and Google is accelerating in-house optical components to reduce outside dependence. TrendForce expects vertical integration to become the norm and says firms controlling silicon photonics design, optical engine manufacturing and packaging together will lead when the market scales in 2027 to 2028.

Linked stocks: 2308 TT, 2330 TT, AVGO US, NVDA US

Sources: TechNews, 研調：拓展CPO交換器　瓶頸在光引擎良率、先進封裝, CPO 交換器正式跨入量產 輝達、博通小量出貨但三大瓶頸浮現

先聲 First Word: Exclusives from Chinese-Language Sources

Taipower sees semiconductor power demand doubling by 2035

Taipower estimates the semiconductor industry will draw about 52 billion kWh of electricity in Taiwan this year, rising to 80 billion kWh by 2030 and 110 billion kWh by 2035, with wafer fabrication the largest and fastest-growing block ahead of memory and advanced packaging and test. Chairman Tseng Wen-sheng compared a single data center to a shot put next to the bowling ball of a science park, noting two racks can exceed the Taipower headquarters building's own consumption, which puts a hard physical ceiling under every Taiwanese capacity announcement. (27 Jul 2026) Source: 台電估今年半導體用電量 520 億度，2035 年增至 1,100 億度

Server DRAM spot prices run 146% above contract

Spot prices for server DRAM have climbed to 146% above contract prices as AI investment pulls demand past supplier expectations, with sample material also reported tight, according to DigiTimes. A gap that wide is the mechanism by which the next contract round reprices, which is why the spread matters more to Samsung Electronics and SK Hynix earnings than the spot volume itself. (27 Jul 2026) Source: 樣品也吃緊　伺服器DRAM現貨價較合約價飆升146%

Low-orbit satellite parts lift Universal Microwave's margin

Universal Microwave Technology (3491 TT) reported unaudited 1H26 revenue of NT$1.92bn, up 70% YoY and a record for the period, with low-orbit satellite product revenue up 147% and gross margin expanding from 47.3% to 56.9%. The company will spend more than NT$1bn on capacity this year after about NT$800m last year, as first-tier satellite customers raise volume requirements and new design projects open. (27 Jul 2026) Source: 昇達科自結上半年每股賺8.15元 超越去年全年7.83元