The Blade Weekly: AMD Wins the First 2GW AI Rack Order Outside NVIDIA
Week of July 19 to July 25, 2026 · Zero One Investment Research
Executive Summary
- AMD committed up to US$5bn to Anthropic against an order for up to 2 gigawatts of MI450 accelerators, the first training-scale commitment to a rack that is not NVIDIA's.
- NVIDIA prepaid Amkor US$1.5bn to expand Arizona packaging and planned US$5bn to US$10bn of dark fiber, paying cash upfront for what it does not manufacture.
- Korea's supply answer to a 2027 shortage is dated 2030: Samsung plans four new fabs, and SK Hynix reapproved a 7.1 trillion won expansion.
AMD's US$5bn into Anthropic wins the first 2GW rack order that did not go to NVIDIA
Every order of this size for a complete AI rack had gone to NVIDIA until last week. AMD says 8 of the 10 largest AI companies already run workloads on its Instinct GPUs, but none had committed years of purchases to a full AMD system at frontier-training scale.
- AMD funds its own customer: up to US$5bn into Anthropic in exchange for up to 2 gigawatts of Instinct MI450 accelerators, an order sized in the power the machines will draw, the unit data-center buildouts are planned in. The structure copies NVIDIA's deal with OpenAI: equity in the model developer, and years of committed chip purchases in return.
- Microsoft anchors the rack: on 20 July it committed to deploy Helios, AMD's 72-GPU rack system, in volume across its own data centers and on Azure.
- TSMC builds the challenger either way: the MI455X's compute dies are fabricated on TSMC's 2nm process, and AMD's Epyc Venice server CPU is the first high-performance part in mass production on that node.
NVIDIA held above 95% of data-center GPUs against AMD's roughly 4.5% going into last week, on CNBC's count, so one 2GW order does not move the share math. What it moves is the debate: with a training-scale buyer now signed and funded, the question shifts from whether a second lane opens to how quickly it fills. Helios ships in the second half, and Wiwynn and Wistron assemble racks for the same cloud customers now qualifying it.
NVIDIA prepaid Amkor US$1.5bn to move AI packaging onshore
NVIDIA is paying cash upfront for the parts of the AI system it does not manufacture.
- Amkor gets US$1.5bn before it builds: the 23 July agreement is a prepayment funding expanded advanced packaging and test capacity in Arizona, with the two co-developing packaging and test methods for next-generation AI platforms. Amkor rose about 16%.
- A US$5bn to US$10bn dark-fiber plan: three years of buying long-haul cable already laid but never lit, on Wolfe Research's reporting, to link dispersed data centers into one compute cluster. The routes are configured with as many as 100 fiber pairs each.
- NVIDIA now sells the server CPU too: it published full specifications for Vera, an 88-core processor it says runs roughly 50% faster than comparable x86 parts, and said it has already supplied the chips to OpenAI, Anthropic and SpaceX.
Taiwan's packaging and test houses sold off on the read-across, with ASE down more than 6% intraday the next session and ChipMOS and Tong Hsing off 6% to 7%. Nothing in their order books changed; what changed is that the customer funding new packaging capacity chose to fund it in Arizona. ASE and Powertech both hold results calls the week of 27 July, the first management commentary on what onshore US capacity means for their share.
Korea's answer to a 2027 memory shortage is dated 2030
Total memory demand could grow 50% to 60% in 2027 while the major suppliers have almost no capacity additions planned, SK Group chairman Chey Tae-won told a Korea Chamber of Commerce forum, calling current prices abnormally high and the scramble for supply chaos.
- The capacity approved last week arrives at the end of the decade: Samsung plans four new Korean fabs by 2030, plus a hybrid-bonding line at Pyeongtaek where equipment talks with Besi have reportedly stalled. SK Hynix reapproved a 7.1 trillion won expansion at Cheongju, about US$4.8bn, and denied a report that it would buy Intel's Ohio campus.
- The shortage is already in a module maker's profit line: Transcend reported 2Q26 net profit of NT$11.9bn, more than it earned in any full year of its history, with first-half profit up more than 24 times YoY.
- Nanya guided prices higher again: 3Q26 DRAM contract prices above 2Q26, with supply tight for several more quarters and fourth-quarter operations improving further.
Transcend shows what the shortage does to a company that buys DRAM and NAND and sells finished modules. It earns the spread between component cost and module price, so the same cycle that produced that quarter turns against it whenever memory prices roll over. A 3Q26 contract price at or below 2Q26 would be the first hard evidence the squeeze is easing, and nothing in Nanya's guidance last week pointed that way.
Alphabet raised 2026 capex to as much as US$205bn and said its compute supply still falls short
Three companies at different points of the chain reported the same demand within two days of each other.
- Alphabet lifted the capex guide again: to US$195bn to US$205bn from US$180bn to US$190bn, with 2027 spending to rise significantly again. It will lease third-party capacity from 3Q26 as a bridge, and cited contracted future revenue of US$514bn, roughly half of which converts within 24 months.
- AI demand finally pulled server CPUs: Intel's data center and AI unit grew 59% YoY, the fastest for that group in 15 years, and Intel raised planned capex above US$20bn to add capacity over two years.
- Supermicro booked more than US$60bn of orders in one quarter: and roughly doubled its 4QFY26 gross margin guide, to 15% to 17%. Orders are not revenue, and the company disclosed no dollar figure for the backlog itself.
For Taiwan the capex line matters more than the order line, because Alphabet's 2027 commitment extends order visibility for the rack assemblers by another year. Foxconn supplies both NVIDIA GPU racks and Google TPU trays, putting its combined share of the two platforms above 40%, and Quanta carries AI order visibility into 2027 and 2028 while adding three US sites. Microsoft and Meta report on 29 July.
TSMC settled 2027 prices in July: 5% to 10% on advanced nodes, plus 10% to 15% on extra HPC orders
TSMC has completed a round of price negotiations and will raise foundry prices from early 2027 by 5% to 10% on 7nm and more advanced nodes, per Nikkei Asia reporting relayed by Taiwan press. Customers placing additional high-performance computing orders pay a further 10% to 15% on that work. Mature nodes rise by up to 10%, and mature process was about 23% of last quarter's revenue, so the smaller increase still reaches a meaningful part of the base.
This continues a program the industry has described since last September, of four straight years of increases running from 2026 through 2029, sized by customer purchase tier. TSMC opened these talks in June, settled them in July, and holds the new prices until early next year. Chairman C.C. Wei told the 2Q26 earnings call he is content with a 68% gross margin while memory makers run near 86%. Apple, whose iPhone 18 uses TSMC's advanced nodes, absorbs another input cost increase.
In brief: China's chip fight moved from models to export rules
Washington accused Moonshot AI of large-scale distillation of Anthropic's Fable model, and Treasury Secretary Scott Bessent said sanctions and Entity List designations remain on the table. Beijing, separately, is consulting domestic companies on rules that could bar Alibaba, ByteDance and Huawei from manufacturing at TSMC, a proposal still at the consultation stage rather than in the rulebook. DeepSeek founder Liang Wenfeng told investors Huawei's Atlas 950 SuperPoD can substitute for NVIDIA's GB300 racks at roughly four Huawei chips per NVIDIA chip, a ratio that caps how fast Chinese buyers can switch.
Last week's rotation: AI server assemblers up to 28.8%, Japan's tool makers down double digits
The buying went to the companies assembling systems now, and the selling to the layer that equips them. Wistron rose 28.8% after building the first US-made NVIDIA GB300 board, with Wiwynn, BizLink and Inventec behind it, while Japan's semiconductor equipment makers fell as a group alongside the wafer maker Sumco at 21.5% and Korea's Hanmi Semiconductor, which builds HBM packaging equipment, at 17.5%. The steepest of those declines came on Friday, and nothing in last week's coverage identifies a single cause for them. STMicroelectronics fell 17.0% after profit and guidance came in below expectations, and UMC and Walsin extended the previous week's unwind.
| Company | Ticker | Group | 1-week |
|---|---|---|---|
| Soitec | SOI FP | Engineered substrates | +32.0% |
| Wistron | 3231 TT | AI server ODM | +28.8% |
| BizLink | 3665 TT | Interconnect | +24.9% |
| Wiwynn | 6669 TT | AI server ODM | +24.0% |
| Transcend Information | 2451 TT | Memory modules | +14.8% |
| Advantech | 2395 TT | Edge AI / industrial PC | +13.3% |
| MediaTek | 2454 TT | ASIC / SoC | +11.3% |
| Inventec | 2356 TT | Server ODM | +10.3% |
| ASUSTeK | 2357 TT | PC / servers | +10.1% |
| Novatek | 3034 TT | Display driver IC | +10.1% |
| Lite-On Technology | 2301 TT | Server power | +9.2% |
| Gigabyte Technology | 2376 TT | PC / servers | +9.1% |
| Sumco | 3436 JT | Silicon wafer | -21.5% |
| Hanmi Semiconductor | 042700 KS | HBM equipment | -17.5% |
| STMicroelectronics | STM US | Analog / MCU | -17.0% |
| Screen Holdings | 7735 JT | Semicap equipment | -15.6% |
| Disco | 6146 JT | Semicap equipment | -13.7% |
| Tokyo Electron | 8035 JT | Semicap equipment | -11.6% |
| Vicor | VICR US | Power modules | -11.2% |
| UMC | 2303 TT | Mature-node foundry | -11.1% |
| Walsin Technology | 2492 TT | Passives | -10.9% |
Week-over-week changes to the Friday July 24 close. Taiwan, Japan, Korea, France and the United States all traded that day, so the table shares one as-of date and includes the reaction to Intel's and Alphabet's Thursday-night results. Coverage-grade names ranked by one-week change; off-theme holdings excluded. Source: Zero One Investment Research; prices via public market data.
Week Ahead / What to Watch
- SK Hynix 2Q26 results (29 July). The first major memory print since the selloff, and the read on HBM4 shipments and third-quarter contract pricing.
- Microsoft and Meta 2Q26 earnings (29 July). Capex guidance from two of the largest cloud buyers, against the raise Alphabet just delivered.
- ASE and Powertech results (week of 27 July). Whether onshore US packaging capacity means eventual share loss, or simply meets demand Taiwan cannot absorb alone.
- Yageo 2Q26 earnings call (29 July). First read on whether AI-server MLCC demand is offsetting a roughly 43% drawdown since early July.
- Apple and Amazon 2Q26 earnings (31 July). AWS capex and custom-silicon commentary against the TPU and Trainium order visibility Foxconn and Quanta carry into 2027.