Zero One Investment Research Daily Intelligence Brief, July 22, 2026

TSMC readies 2027 price hikes as NVIDIA moves to build its own AI network

Top Market Signals

TSMC's 2027 price hikes carry an AI premium

TSMC (2330 TT) has completed a fresh round of price negotiations with customers and will raise foundry prices from early 2027 by 5% to 10% on 7nm and more advanced nodes, according to Nikkei Asia reporting relayed by Taiwan press. The company declined to comment, saying only that its pricing is strategy-driven rather than opportunity-driven.

The structure matters more than the headline number. On top of the 5% to 10% base increase, customers placing additional high-performance computing orders will pay a further 10% to 15%, so the total increase on some leading-edge AI work can exceed 10%. Mature nodes, including 12nm, 16nm and 28nm, are being lifted by up to 10%, with some products below that cap. Mature process accounted for about 23% of last quarter's revenue, so the mature-node increase reaches a meaningful slice of the base.

TSMC has framed the timing as customer-friendly, starting negotiations in June, settling in July, and holding the new prices until early next year to give buyers time to adjust. Industry reports since last September have pointed to four straight years of increases running from 2026 through 2029, with the size varying by customer purchase tier.

TSMC is converting AI scarcity into durable pricing power without breaking its customers. Chairman C.C. Wei told last week's earnings call that the priority is customer success over maximum margin, noting he is content with a 68% gross margin even while memory makers run near 86%. The device layer absorbs the bill: Apple, whose iPhone 18 is built on TSMC's advanced nodes, faces another input-cost increase after already raising prices across much of its lineup.

Linked stocks: 2330 TT, TSM US, AAPL US

Sources: 台積電傳明年又要漲價10% 部分先進製程可能上調更多, 台積電傳2027年起漲價最高達10% HPC額外訂單加收溢價, Mobileworldlive

NVIDIA moves to buy America's dark fiber

NVIDIA (NVDA US) is planning to spend US$5bn to US$10bn over the next three years acquiring long-haul dark fiber across the United States, according to reporting cited by GuruFocus and Wolfe Research. The aim is to build its own high-speed optical backbone to link geographically dispersed data centers into a single large AI compute cluster, reducing its dependence on traditional carriers and hyperscale cloud providers.

Dark fiber is cable that was laid but never lit, with no transmission equipment attached. Much of the US inventory dates to the late-1990s telecom build-out that stranded roughly 97.5% of installed fiber underground after the 2000 dot-com bust. Lighting it now, and running 800 Gbps per channel while reusing 96 wavelengths, requires heavy deployment of DWDM (dense wavelength-division multiplexing) mux and demux units, fiber arrays, optical connectors and distribution frames at both ends. Wolfe Research says NVIDIA is buying routes configured with as many as 100 fiber pairs each.

The move extends NVIDIA from a chip supplier toward an AI-infrastructure and network integrator, a direction Jensen Huang has flagged in emphasizing that a GPU sits idle without fast data transport. It follows the company's investments in NVLink, silicon photonics and AI RAN, and its stake in Nokia.

The demand shifts from inside-the-data-center connectivity toward cross-region long-haul transport, and that lands on Taiwan's optical component makers. According to Economic Daily News, Browave is already an NVIDIA-certified supplier of the WDM mux, demux and optical-amplifier parts such a network needs; the paper also flagged a high-density fiber patch-panel and connector maker (6442 TT) for the point-to-point cabling and the suppliers of 800G and 1.6T long-reach transceivers for the data-center-to-data-center links.

Linked stocks: NVDA US, 6442 TT, NOK US

Sources: 輝達衝光纖網路擬收購美國暗光纖 引爆元件需求, 輝達衝光纖網路 瞄準傳輸痛點

China weighs barring its chip designers from TSMC

China's Ministry of Commerce is consulting domestic technology companies on a major expansion of export restrictions that could prohibit local chip designers from manufacturing at TSMC (2330 TT) and other foreign foundries, the Financial Times reported. Firms named in the discussions include Alibaba, ByteDance and Huawei.

The proposals reach beyond silicon. Regulators have discussed limiting transfers of AI training data out of China and restricting foreign users from downloading model weights, a change that would blunt the open-weight advantage held by DeepSeek and Moonshot over the closed flagship models from OpenAI and Anthropic. Beijing is also weighing tighter controls on foreign acquisitions of strategically important technology firms.

The foundry restriction is the most consequential and the most double-edged. Forcing Chinese designs onto SMIC (981 HK) would guarantee the domestic champion enough order volume to fund its research and capacity, but it would also deny Chinese companies the process-technology lead that TSMC holds over SMIC. The measures are aimed at keeping frontier AI work inside China, though they would also slow the global spread of Chinese AI standards.

This sits at the consultation stage, not the rulebook, and the read is directional rather than immediate. If enacted, it would remove a slice of TSMC's China-designer demand while accelerating domestic substitution, a negative for foreign-foundry access to that customer base and a support for SMIC utilization. For a broad read on AI supply chains, it signals that Beijing now wants leading-edge chip design kept and produced inside China rather than routed through the best available foundry.

Linked stocks: 2330 TT, 981 HK, BABA US

Sources: Tom's Hardware

SK Hynix eyes Intel's Ohio campus for US memory

SK Hynix (000660 KS) is in talks to buy Intel's (INTC US) chip campus in Ohio to house US memory production, the Korea JoongAng Daily reported. The move would give the world's largest HBM (high-bandwidth memory) maker a manufacturing footprint inside the United States as AI demand pulls memory supply toward its largest end market.

The logic is downstream of where SK Hynix's revenue now sits. Roughly 65% of its sales are generated in the US, and management has publicly warned of "chipflation" from AI memory scarcity while weighing US expansion. Acquiring a partly-built advanced campus rather than starting greenfield would shorten the path to onshore output.

For Intel, a sale would convert an idle asset into cash and further narrow its manufacturing ambitions under Lip-Bu Tan, who has been paring the company back toward the projects it can fund. For SK Hynix, an Ohio base would sit alongside its HBM4 push and reduce the geographic concentration risk that comes with producing the AI cycle's scarcest component almost entirely in Korea.

The read-across is that the memory build-out is following the logic buyers set earlier this year, when customers began funding supplier capacity directly. Localizing production near US hyperscalers is the next rung, moving memory manufacturing closer to the buyers driving the AI cycle.

Linked stocks: 000660 KS, SKHY US, INTC US

Sources: Google News

NVIDIA's Vera CPU takes aim at Intel and AMD

NVIDIA (NVDA US) released full specifications for Vera, its in-house server CPU, and said it has already supplied the chips to OpenAI, Anthropic and SpaceX, according to reports from qz.com and others. The 88-core, 176-thread processor is designed to pair with Rubin GPUs, and NVIDIA positions it as delivering roughly 50% higher performance than comparable x86 parts.

Vera opens a new front in the server-processor market that Intel (INTC US) and AMD (AMD US) have split between them. By offering a CPU tuned to feed its own accelerators, NVIDIA can sell a more complete rack and capture the host-processor content that previously went to x86 suppliers. The disclosure of full specs and named early customers signals the part is shipping, not roadmapped.

The competitive stakes are largest for the x86 incumbents. AMD has been taking server-CPU share, reported near 46% of the market, precisely as NVIDIA introduces an Arm-based alternative bundled with the accelerators data centers are buying anyway. Intel, still rebuilding its data-center franchise, now faces a third serious CPU competitor inside the AI rack.

For the broader chain, Vera confirms that the AI compute platform is consolidating under whoever controls the accelerator. NVIDIA is extending from GPUs into CPUs and, as its dark-fiber move shows, into the network between data centers, steadily absorbing more of the system's value into a single vendor.

Linked stocks: NVDA US, INTC US, AMD US

Sources: Google News, Nvidia details Vera CPU for AI servers

Intel's foundry lands its first named customer

Intel (INTC US) has signed Fortinet (FTNT US) as the first publicly named external customer for its foundry business under chief executive Lip-Bu Tan, and will manufacture Fortinet's FortiSP6 security processor, according to reports across trade and financial press. Intel shares reclaimed the US$100 level on the news.

A named external customer is the milestone Intel Foundry has lacked. The unit has struggled to convert its 18A process into committed third-party volume, and only days ago was reported to have pulled the bulk of its own Nova Lake desktop compute tiles back from TSMC to 18A. Winning outside silicon, even a single security chip, gives the foundry a reference customer it can point to when courting the next.

The win should be kept in proportion. A security processor is far smaller volume than the mobile or AI-accelerator sockets that fund a leading-edge foundry, and one customer does not validate a business that needs many. It does, however, mark the first external proof point for the strategy Lip-Bu Tan has staked the company on.

For the foundry landscape, Intel can now begin to compete for the security and networking silicon that sits a tier below TSMC's flagship AI work. Whether that widens into the higher-volume sockets is the test the next several quarters will set.

Linked stocks: INTC US, FTNT US, 2330 TT

Sources: Google News

先聲 First Word: Exclusives from Chinese-Language Sources

Nanya Tech builds a custom DRAM answer to HBM for edge AI

Nanya Technology (2408 TT) is developing custom high-I/O DRAM, which it calls UWIO DRAM, combined with wafer-on-wafer 3D stacking that places memory directly on a CPU or ASIC, senior vice president Wu Chih-hsiang said. He said the architecture can lift memory bandwidth five to ten times while cutting energy per bit to a tenth of HBM, aimed at power-sensitive AI inference in PCs, phones and robots rather than data-center training. It positions Nanya to sell into the on-device inference wave instead of competing head-on in HBM. (21 Jul 2026) Source: 南亞科攻客製化DRAM

Chenbro rides AWS ASIC chassis orders to upgraded estimates

Chenbro (8210 TT) posted June revenue up 29% MoM, ahead of estimates, on rising AWS ASIC server-chassis orders and a richer HGX chassis mix, prompting brokers to lift its 2Q26 gross margin estimate to 31.8%. Third-quarter revenue is now modeled up 20% QoQ to NT$9.39bn as the AWS T3 rack ramps, with full-year EPS revised up 75% YoY. It shows how ASIC server build-outs feed Taiwan's chassis makers. (22 Jul 2026) Source: 勤誠瞄準逾150天

Largan buys land twice in a month for its CPO push

Largan Precision (3008 TT) spent NT$561m on land and buildings in Taichung, its second land purchase within a month after a NT$628m buy in late June, taking the total to about NT$1.19bn. Chairman Lin En-ping has said the company would acquire more land if its co-packaged optics business succeeds, and it is already sampling a four-layer fiber-array component to a customer this month. The repeat land-buying signals Largan is provisioning capacity ahead of CPO orders. (22 Jul 2026) Source: 大立光擴產再度獵地 斥資5.6億取得台中南屯土地建物