This Week’s U.S. AI Infrastructure Trade: Capital Shifts from the GPU Story to Supply Constraints and Profit Taking

1. What happened

The core change in this week’s U.S. AI infrastructure trade is not simply about “selling AI,” but about the market beginning to re-rate the AI infrastructure chain. Capital shifted from high-beta, crowded hardware stories to areas with stronger supply constraints and order visibility, such as storage, optical interconnects, semiconductor equipment, power, and AI-native data centers.

1. Storage sector: from demand surge to profit taking

SK Hynix-related ADRs reportedly rose 27% in one day, but that move may have been significantly amplified by short-dated options trading and dealer hedging, and does not necessarily mean fundamentals changed by that magnitude in a single day.

Meanwhile, the storage sector had already seen some correction. According to JPMorgan research, investors are not uniformly bearish on storage demand; rather, they are focusing more on whether cloud providers’ capex will be revised higher again, whether HBM pricing will reach expectations, and whether storage makers’ margins are already fully reflected in share prices.

JPMorgan sees a scenario in which 2027 HBM average selling prices rise 25%—30% as more realistic than a scenario in which prices double. That said, DRAM supply is still said to be clearly tight, and enterprise SSD demand is being supported by AI data centers and KV Cache Offload.

2. ASML: capacity expansion guidance reinforces the supply-bottleneck trade

ASML’s second-quarter 2024 revenue was 9.3 billion euros, above the market estimate of 8.9 billion euros, and it raised its full-year 2026 revenue guidance to 43 billion—45 billion euros.

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

Reports say ASML expects advanced logic business to grow about 25% and memory business about 75% in 2026, indicating that AI demand is boosting both logic chips and HBM-related investment.

3. NVIDIA: the trading logic is shifting from GPUs to a full-stack platform

This week’s focus on NVIDIA has moved beyond GPU shipment volumes to how much value share the company can capture within a single AI data center.

According to presentation materials, AI labs account for about 20% of NVIDIA demand, while traditional hyperscale cloud providers account for about half of revenue. AI cloud, sovereign AI, industrial customers, and enterprise customers are emerging as new sources of incremental demand. At one major customer that had previously relied primarily on ASICs, NVIDIA’s share of compute has already risen to nearly 50%.

NVIDIA emphasizes that competition is not only about chip pricing, but about total compute cost per token and deployment efficiency. Its product scope is also expanding from GPUs into CPUs, networking, interconnects, rack systems, and software.

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

Market attention on data center infrastructure is starting to shift from traditional floor space to the ability to provide high-density AI capacity, liquid cooling, power, and networking as an integrated package.

In an interpretation of a secondary report on NTT, the company’s order backlog is about $20 billion, equivalent to about 7.7 years of annual revenue. Global liquid cooling deployment is about 250MW, which can support GPU environments of up to 135kW per rack. Its advantage is described as vertical integration across data centers, submarine optical cables, and fiber backbone networks.

However, the figures above are a paraphrase from a secondary article of a Morgan Stanley report, and because the current materials do not include a company announcement or the full original report, they should be treated as a trade signal pending verification rather than confirmed corporate fact.

5. IBM: the market penalizes the risk that software budgets are being crowded out by hardware

IBM reportedly fell 25.53% in one day, erasing about $65 billion 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 links this to enterprises prioritizing purchases of GPUs, servers, and high-bandwidth memory, suggesting that hardware spending may be crowding out budgets for software, middleware, and IT consulting. However, based on the current materials alone, it is insufficient to conclude that IBM’s stock decline was caused solely by this factor.

2. Why it matters

1. AI trading is spreading from a single leader to the entire supply chain

According to an aggregate from Changjiang Securities, since 2026 the turnover share of the AI core basket has risen to 11.23%, while the share of non-NVIDIA baskets has increased to 7.40%. Storage, optical communications/photonics interconnects, compute transitions, and semiconductor equipment were the main incremental contributors, while the turnover share of data center networks and AI server/data center infrastructure actually declined.

The same report says the concentration of the AI core basket has also fallen: average CR1 was 57.72% in 2024, but has dropped to 35.18% since 2026. This indicates that the trade is spreading from a single leader to multiple infrastructure components.

2. The market has started testing whether “demand” can be converted into profit

The divergence in the storage sector shows that investors are no longer satisfied with AI demand growth alone; they are now asking whether that demand can translate into sustainable ASPs, long-term contracts, and profits. As a result, storage share volatility may far exceed the underlying change in fundamentals.

3. Supply bottlenecks are becoming the new support for valuation

ASML’s expansion plan shows that advanced manufacturing, HBM, and wafer capacity must be secured years ahead. Shanghai Pudong International summarizes this style as HALO trading, saying capital favors scarce-asset areas such as energy, transmission and distribution networks, utilities, critical manufacturing capacity, and semiconductor equipment.

This means the constraints in AI infrastructure are not only in GPUs, but also in power, advanced manufacturing, storage, liquid cooling, optical interconnects, and high-density data centers.

3. Evidence and trading implications

• Storage: DRAM supply is said to be clearly tight, and enterprise SSD demand is supported by AI data centers and KV Cache Offload. However, HBM pricing and margins remain key factors the market is re-evaluating.
• Semiconductor equipment: ASML raised guidance and announced expansion plans, reinforcing the view that investment in advanced logic and memory remains ongoing.
• NVIDIA: The company is moving from a GPU supplier to a full-stack infrastructure platform that includes GPUs, CPUs, networking, interconnects, rack systems, and software.
• Data centers: High-density GPU environments impose integrated requirements for power, liquid cooling, and low-latency networking. However, the NTT-related data still needs verification.
• Relative value: The market may prefer chips, storage, networking, equipment, and AI data centers while avoiding traditional software, IT services, and consulting companies that are harder to prove on short-term AI monetization.

4. What to watch next

1. Whether the latest earnings reports from major cloud providers confirm capex upgrades for 2026—2027.
2. Which of storage, HBM, and optical interconnect companies’ orders and long-term contracts will translate into confirmed profits.
3. Whether power, liquid cooling, and AI-native data centers become the next listed investment themes.
4. Whether compute rental prices continue to fall. One market analysis says B200 rental prices have dropped about 30% from the late-May peak.
5. Whether a slowdown in AI investment and private credit liquidity issues further increase financing pressure on AI companies’ debt and equity funding.
6. Whether NVIDIA can withstand structural risks from ASICs, in-house chips, competitors, and export restrictions.

5. Conclusion

The most important development this week is not the one-day rise or fall of any single stock, but the threefold reconfiguration of the AI infrastructure trade: from a standalone GPU story to a full-stack platform, from an AI demand story to supply bottlenecks and profit taking, and from software and asset-light valuations to scarce assets such as power, equipment, storage, optical interconnects, and high-density data centers.

Original Source

Integrated from the provided research and media sources