U.S. AI Infrastructure Stocks This Week: Capital Rotates from the GPU Story to Supply Bottlenecks and Profit-Taking

I. What has happened

The core change in AI infrastructure trading in the US this week is not just “selling AI,” but the market beginning to reprice the entire AI infrastructure chain. Capital is moving out of the crowded, high-beta hardware story and rotating into storage, optical connectivity, semiconductor manufacturing equipment, power, and AI-native data centers — areas with tighter supply constraints and better order visibility.

1. Storage: from explosive demand to profit-taking

According to reports, ADRs related to SK hynix rose 27% in one day, but the move may have been strongly amplified by short-dated options trading and market maker hedging, and does not necessarily mean that fundamentals changed by the same degree in a single day.

Meanwhile, the storage group has already undergone a correction. JPMorgan research shows that investors are not uniformly bearish on storage demand; instead, they are focusing more on whether cloud companies’ capital expenditures will be revised upward again, whether HBM pricing can meet expectations, and whether storage manufacturers’ profit margins have already been fully reflected in prices.

JPMorgan believes that the scenario of average HBM selling prices rising 25%—30% in 2027 is more realistic than the scenario of prices doubling. Even so, DRAM supply is still viewed as extremely tight, while enterprise SSD demand is being supported by AI data centers and KV Cache Offload.

2. ASML: capacity guidance reinforces the supply-shortage trade

ASML reported revenue of 9.3 billion euro in Q2 2024, above the market forecast of 8.9 billion euro, and raised its full-year 2026 revenue guidance to 43 billion—45 billion euro.

Under the company’s capacity expansion plan, low-NA EUV capacity is expected to increase from about 65 machines in 2026 to about 85 machines in 2027, and to about 110 machines in 2028. Immersion DUV capacity is expected to increase from about 130 machines in 2026 to about 169 machines in 2027, and to about 220 machines in 2028.

The report said ASML expects advanced logic to rise about 25% in 2026, while memory is expected to rise about 75%, indicating that AI demand is simultaneously driving capacity investment in both logic chips and HBM-related products.

3. NVIDIA: the trade shifts from GPU to a full-stack platform

This week, the focus around NVIDIA was no longer just on GPU shipments, but on how the company is increasing its share of value inside each AI data center.

According to the presentation materials, AI labs account for about 20% of NVIDIA’s demand, while traditional hyperscalers account for about half of revenue. AI cloud, sovereign AI, industrial customers, and enterprise customers are becoming new sources of growth. For one major customer that previously relied mainly on ASICs, NVIDIA’s compute segment now accounts for nearly 50%.

NVIDIA emphasized that competition is not only about chip price, but also about total compute cost per token and deployment efficiency. The company’s product portfolio has also expanded from GPU to CPU, networking, interconnects, rack systems, and software.

4. Data centers and power: from rack counts to scarce capacity

The market is shifting its focus in data center infrastructure from traditional floor space to integrated high-density AI power delivery systems, liquid cooling, power, and networking.

An interpretation from a secondary report on NTT suggests that the company has about 20 billion USD backlog, equivalent to about 7.7 years of annual revenue. Global liquid cooling deployment is around 250MW, which can support GPU environments with up to 135kW per rack. The company’s advantage is described as vertical integration across data centers, submarine cables, and optical backbone networks.

However, the data above comes from a secondary article citing a Morgan Stanley report, and the available materials do not provide the company announcement or the full original report. Therefore, this should be viewed as an unverified trading signal, not a confirmed corporate fact.

5. IBM: the market punishes software as hardware absorbs the budget

According to reports, IBM fell 25.53% in one day, wiping out about 65 billion USD in market capitalization. Software growth slowed from 11% in Q1 to 5%, consulting was nearly flat, and infrastructure revenue declined 7% year over year.

Market analysis suggests that this stems from enterprises prioritizing purchases of GPU, server, and high-bandwidth memory, arguing that hardware spending may be crowding out budgets for software, middleware, and IT consulting. However, based only on the materials available, there is still not enough to conclude that IBM’s stock decline was entirely due to this factor.

II. Why this matters

1. The AI trade is spreading from a single main character to the entire supply chain

Changjiang Securities data show that after 2026, the trading share in the core AI basket increased to 11.23%, while the share of the basket excluding NVIDIA reached 7.40%. Storage, optical communications/optical interconnect, compute switching, and semiconductor manufacturing equipment became the main incremental contributors, while the trading share of data center networking and AI server/data center infrastructure declined.

The report also shows that concentration in the core AI basket is declining. In 2024, the average CR1 was 57.72%, but after 2026 this figure fell to 35.18%. This suggests that the trade is expanding from a single main character to multiple different infrastructure links.

2. The market is beginning to test whether “demand” can be converted into profits

The divergence within the storage group shows that investors are no longer satisfied with AI demand growth alone; they are starting to question whether that demand can translate into sustainable average selling prices, long-term contracts, and profits. As a result, storage stock volatility could be far greater than the actual change in fundamentals.

3. Supply constraints are becoming the new valuation anchor

ASML's production expansion plan shows that advanced manufacturing capacity, HBM and wafers all need to be booked years in advance. Shanghai Pudong International calls this style HALO trading and points out that capital is flowing into energy, power transmission networks, public utilities, critical manufacturing capacity, and semiconductor manufacturing equipment — these scarce areas.

That means the constraints on AI infrastructure are not just GPUs, but also power, advanced manufacturing, storage, liquid cooling, optical connectivity, and high-density data centers.

III. Evidence and trading implications

• Storage: DRAM supply is said to be tightening noticeably, while enterprise SSD demand is supported by AI data centers and KV Cache Offload. However, HBM prices and margins remain key to market rerating.
• Semiconductor manufacturing equipment: ASML has raised guidance and announced capacity expansion plans, reinforcing the view that investment in advanced logic and storage is still ongoing.
• NVIDIA: The company is evolving from a GPU vendor into a full-stack infrastructure platform, including GPUs, CPUs, networking, interconnects, rack systems, and software.
• Data centers: A high-density GPU environment requires integrated power, liquid cooling, and low-latency networking. However, the data related to NTT still needs to be verified.
• Relative value: The market may avoid traditional software companies, IT services, and consulting firms that are struggling to demonstrate short-term AI monetization, and instead favor chips, storage, networking, equipment, and AI data centers.

IV. What to watch next

1. Whether the latest earnings reports from major cloud providers confirm higher capex for 2026—2027.
2. Which storage, HBM and optical interconnect companies can turn orders and long-term contracts into sustainable profits.
3. Whether power, liquid cooling and AI-native data centers become a new investment theme for listed companies.
4. Whether compute rental prices continue to fall. Some market analysis shows B200 rental prices have dropped about 30% from the late-May peak.
5. Whether the slowdown in AI financing and private credit liquidity issues continue to increase debt pressure and equity fundraising pressure on AI companies.
6. Whether NVIDIA can withstand structural risks from ASICs, internal custom chips, competitors and export controls.

V. Conclusion

The most important thing this week is not the intraday move of any single stock, but the three-layer restructuring of the AI infrastructure trade: from the story of a single GPU to a full-stack platform, from the story of AI demand to supply constraints and monetization potential, and from software and asset-light valuations to scarce assets such as power, equipment, storage, optical interconnects and high-density data centers.

Source

Compilation of the provided research and media sources