404K Tech Evening Brief 2026-06-03 - Memory Price Hikes, Interconnect Re-rating, Compute Financing
The marginal change in AI infrastructure is no longer about a single GPU. Memory, interconnect, power, racks, and enterprise deployment are all tightening at the same time. Memory price targets were raised sharply, while Marvell Technology, Credo, Ciena, and Cisco pushed the networking layer to the foreground. CoreWeave and hyperscale cloud providers continue to push data-center construction toward long-term leases and diversified financing.
404K | 2026-06-03
Pre-market Takeaways
AI infrastructure continues to broaden from “buy GPUs” to “who can secure memory, connect racks, supply power, and finance the buildout.” Morgan Stanley raised Micron’s price target from USD 520 to USD 1,050 and SanDisk’s from USD 1,100 to USD 1,750. ASML, Credo, Ciena, and Aspeed Technology also saw price-target increases, suggesting sell-side models are again incorporating 2027-2028 earnings duration.
The more important marginal change is the shift in bottlenecks. OpenAI, CoreWeave, Google/Blackstone, and Meta are all reinforcing the compute shortfall and the path toward assetization. A CoreWeave-related data center was financed with USD 900mn of high-yield debt, backed by a 15-year lease and approximately USD 2.2bn of revenue. Goldman Sachs also cautioned, however, that AI investment returns still depend on whether enterprises can genuinely translate AI into profit or cost savings.
AI/Semiconductor Full Value Chain
Compute GPU/ASIC/CPU
NVIDIA Industry developments show that agentic AI is pulling CPUs back into the compute discussion. Industry participants say the GPU-to-CPU ratio may move from 2:1 to 1:1 “within a few months.” This cross-validates Stratechery’s criticism of RTX Spark: adding local GPU capacity does not necessarily equal the optimal architecture for the agent era, where cloud inference and local CPU scheduling matter more.
“The ideal agent-era configuration is a strong local CPU plus cloud inference”
OpenAI Sarah Friar’s interview provided the hardest data on the compute gap: USD 122bn raised in March, ChatGPT weekly active users above 900mn, and Codex growing from nearly zero in January to more than 5mn weekend users. She believes compute will still be insufficient in 2026 and quite limited in 2027, while Oracle’s 1GW Michigan campus is not expected to contribute compute until late 2027 or early 2028.
“multiple CSPs shift capex into opex”
AMD Industry posts said Advancing AI will showcase a product portfolio more recognized by the market. Other industry commentary argued that after NVIDIA emphasized CPUs, AMD’s AI narrative will need to be validated more by GPU share gains. This clue is only pre-launch expectation management; the hard data still needs to come from server CPU and MI-series shipment cadence.
Aspeed Technology J.P. Morgan maintained Overweight and raised its price target from NT$21,600 to NT$22,000. The 3Q26 revenue guide of NT$4.1bn-4.3bn was below expectations, but backlog is strong, book-to-bill exceeds 2x, customers have already placed orders for 2027, and AI server BMC shipments are expected to rise from 4.1mn units in 2025 to 16.7mn units in 2028.
“book-to-bill ratio above 2x”
Intel Multiple materials point to CPU and foundry validation: 18A is now in full mass production; Core Ultra Series 3 has more than 325 designs, including more than 130 18A edge designs. Rackscale Blueprints supports up to 128 Xeon 6 CPUs, 384TB DDR5, and roughly 100kW power per rack. The claim that agent token consumption can increase by up to 1,000x reinforces the CPU/memory linkage.
HBM/DRAM/NAND/eSSD
Micron, SanDisk Morgan Stanley maintained an Attractive industry view, raised Micron’s price target from USD 520 to USD 1,050, and raised SanDisk’s from USD 1,100 to USD 1,750. Micron’s CY2027E adjusted EPS was raised from USD 76.87 to USD 113.85. DRAM prices are expected to rise 40% in MayQ and 15% in AugQ, and shortages may persist for another 2-3 years.
“DRAM has become the principal bottleneck for the AI buildout”
Kioxia Holdings Goldman Sachs maintained Buy with a 12-month price target of JPY 93,000. Management expects a 22% CAGR in NAND capacity from CY25 to CY28, a 46% CAGR in data-center NAND demand, eSSD share rising from roughly 10% to roughly 15%, and LTA coverage of about 50% over the next two years. Tight supply-demand conditions are expected to last at least until 2H27.
“inference AI... NAND (SSD) is positioned not just as simple storage”
Samsung Electronics, SK hynix Multiple industry developments elevate the memory upcycle to the Korean macro level, suggesting the two companies could together generate USD 2tn in pretax profit by 2030. TrendForce-related clues show HBM’s share of DRAM wafer starts at the top three memory vendors rising from 18% in 2025 to 30% in 2027, while HBM bit supply share rises from 8% to 13%.
Global memory chain Morgan Stanley’s “chip inflation” report makes clear how AI servers are crowding out memory demand. Servers’ share of DRAM demand is expected to rise from 37% in 2023 to 59% in 2028, while enterprise SSDs’ share of NAND demand rises from 18% to 65%. If supply is insufficient, PCs and smartphones may still face shortages of roughly 58mn units and 134mn units, respectively.
Phison, Micron SSD ecosystem NewMaxx-related clues show Phison’s PCIe 6.0 X3 SSD controller rated at 28GB/s and 6.8mn IOPS, with support for 2PB per drive. Micron’s 6600 ION 245TB SSD pushes enterprise eSSD capacity density to “a quarter petabyte in a single drive bay.” The storage bottleneck is spreading from HBM to NAND, controllers, and the protocol ecosystem.
Advanced Packaging/Foundry/Equipment
ASML Morgan Stanley maintained Overweight/Top Pick and raised its price target from EUR 1,400 to EUR 1,660. The report expects approximately 90 EUV deliveries next year and 104 systems in FY28, including 96 low-NA and 8 High-NA systems. FY28 EPS was raised from EUR 37.44 to EUR 51.92. The core debate is that the market is overly concerned about capacity bottlenecks.
“Capacity concerns overshadow earnings power”
Tokyo Electron Goldman Sachs maintained Buy with a JPY 62,000 price target. The company expects CY26 global WFE to grow more than 20% YoY and exceed USD 150bn. DRAM equipment demand was revised from +20% to more than +30%, NAND was revised from flat to modest growth to +40%, and the company targets gross margin above 50% over the next two years.
“CY26 WFE market... growth of over 20% yoy and... exceed $150bn”
ASE Technology, King Yuan Electronics, Foxconn Technology Goldman Sachs’ Taiwan Computex notes show AI testing and packaging momentum continuing to move higher. ASE’s LEAP revenue guide is USD 3.5bn for 2026, with at least another USD 1.9bn of growth in 2027. King Yuan Electronics’ AI revenue share is expected to rise from 27%-28% in 2025 to roughly 40% in 2026, with 2026 capex of NT$50bn.
Foxconn Technology Goldman Sachs said its AI/HPC FT handler market share exceeds 90%, and maintained Buy with a NT$12,000 price target. The company reiterated that 2026 capacity will expand by more than 40% YoY, with roughly 50% YoY growth in both 2027 and 2028. Demand is coming from CPUs, ASICs, GPUs, CPO, and EVs, while CPO electrical-test orders have entered phased shipments.
GlobalWafers Computex notes emphasized full utilization for 12-inch silicon wafers and SOI demand driven by silicon photonics. U.S. 300mm SOI expansion will triple existing SOI capacity. However, simultaneous expansion across six sites, depreciation, and ramp costs are pressuring near-term gross margin, indicating that the materials side benefits from AI but still needs advanced-node and silicon-photonics demand to materialize.
Optical Communications/High-speed Interconnect/Networking
Marvell Technology Multiple industry materials point to stock re-rating: the shares rose nearly one-third on Tuesday, closing at USD 290.79 with a market cap of approximately USD 254.6bn. NVIDIA invested USD 2bn in March and expanded cooperation. Jensen Huang appeared on stage and said Marvell could become the “next trillion-dollar company.” Photon Capital emphasized that the quotable original point is that the connectivity layer is “essential” to the evolution of AI data centers.
“connectivity becomes the necessary layer”
Credo J.P. Morgan raised its price target from USD 230 to USD 250, while BofA raised its target from USD 210 to USD 252. F4Q26 revenue was USD 437mn, up 157% YoY. FY27 management guidance calls for revenue growth of 80%+ YoY to more than USD 2.4bn. Optical revenue is expected to exceed USD 600mn, while AEC remains the core growth engine.
“AECs as a ‘core growth engine’”
Coherent simplified optical interconnect chain FundaAI believes the solution is not merely a 2-20km DCI/LR replacement, but a link-budget solution paired with OCS in Google TPU/AI clusters. 2.4T corresponds to 300G/lane x8 and is expected to be introduced in 2027; 3.2T corresponds to 400G/lane x8 and is expected to ramp in 2028-2029. The incremental industry value in 2028 is estimated at USD 15bn-20bn.
Ciena J.P. Morgan maintained Overweight and raised its December 2026 price target from USD 550 to USD 635. F2Q26 is expected to deliver revenue of USD 1.53bn and EPS of USD 1.51. FY26/FY27/FY28 revenue is estimated at USD 6.3bn, USD 8.259bn, and USD 10.802bn, respectively. The key issue is not revenue, but whether gross margin can remain near 44%.
“Margins Than Revenues”
Cisco Morgan Stanley maintained Overweight with a USD 120 price target, while Goldman Sachs raised its target from USD 116 to USD 125 but maintained Neutral. Cisco Live 2026 emphasized that agent task traffic is 450% higher than human traffic and that AI-related network traffic will grow 3x over the next three years. Silicon One, Acacia, and pluggable optical modules continue to carry the AI networking upgrade narrative.
“agents are expected to generate 450% more traffic than humans”
Wiwynn Goldman Sachs maintained Buy with a NT$7,147 price target; the share price on the report date was NT$5,475. The company showcased NVIDIA Vera Rubin NVL72, AMD Helios rack-scale solutions, CPO optical interconnect, high-voltage DC, and 6kW-class dual-sided liquid cooling. 2027E EPS is NT$398.71, implying only 13.7x P/E.
“CPO in AI server racks scale up”
Power/Servers/Financing
Kehua Data BofA initiated coverage with Buy and a RMB 46 price target. China’s AIDC power-system market is expected to grow from roughly RMB 5bn in 2025 to roughly RMB 16bn in 2030, a 25% CAGR. The company is China’s No. 2 UPS supplier and the world’s No. 2 PCS supplier, with 2025-2028 net-profit CAGR expected at 36%.
SemiVision 800V architecture The article describes the power bottleneck directly: GB200/GB300 rack power has already exceeded 200kW, and future Kyber racks are moving toward 1MW. Under the traditional 54V architecture, a 1MW rack would require about 200kg of copper busbars; 800V HVDC can reduce copper weight by roughly 45% and improve end-to-end efficiency by about 5%.
“The bottleneck is no longer only compute, memory, or networking, but power delivery”
CoreWeave A related data center was financed with USD 900mn of high-yield debt, with a 5-year tenor and a 7.5% yield. The project near Chicago is leased entirely to CoreWeave for 15 years, corresponding to roughly USD 2.2bn of revenue. AI data-center developers have raised more than USD 27bn through high-yield debt this year, reflecting strong demand but also amplified credit structures.
Google/Blackstone TPU Cloud Goldman Sachs’ AI project tracker shows Blackstone and Google setting up a TPU Cloud joint venture with initial capacity of 500MW. Blackstone committed USD 5bn of equity, and the first capacity is expected to come online in 2027. AI cloud competition is expanding from GPU leasing to TPU, project financing, and cloud software stacks.
Global infrastructure capital Goldman Sachs’ alternative-investment report expects the four major hyperscalers, Meta, Microsoft, Amazon, and Google, to spend a combined USD 5.3tn in capex from FY2025 to FY2030. It also gives a baseline estimate of USD 7.6tn in combined compute, data-center, and power capex from 2026 to 2031. Average data-center construction cost is about USD 11.3mn per MW.
Internet/Platforms
Meta Morgan Stanley maintained Top Pick/Overweight with a USD 775 price target. The report acknowledges market concerns over more than USD 380bn of combined capex in 2027-2028, but argues that Meta AI search, subscriptions, ad estimate revisions, and Neocloud leasing could each contribute roughly USD 1-3 of EPS in 2028. Leasing out 1GW of compute could contribute USD 2.97 of EPS.
“compute is a scarce resource”
Meta AI search Model assumptions include 1bn MAUs, one query per day, a 10% commercial-query rate, and USD 0.30 per commercial query. This could generate roughly USD 11bn of annual revenue and USD 2.89 of incremental EPS. If usage frequency and MAUs expand, the search revenue path could exceed USD 30bn.
“definitely on the table”
Amazon Leo BofA maintained Buy with a USD 310 price target. Leo already has more than 300 satellites in orbit and is expected to conduct 17 launches in 2026. The initial constellation is expected to cost roughly USD 25bn by 2028, with cumulative 2024-2029 spending of about USD 35.477bn. The 2032 consumer broadband revenue opportunity is estimated at about USD 14.04bn.
Amazon/AWS and LEO networks The report estimates Leo needs roughly 13mn users to break even. On the enterprise side, partnerships already include Delta, the U.S. Department of Defense, Airbus, Apple, AT&T, Verizon, Vodafone, and JetBlue. The implication for AWS is not just broadband, but a long-term entry point into enterprise, government, and private connectivity.
OpenAI Interview materials show consumer and enterprise revenue are now close to a 50/50 split. Free users average about 7 conversation rounds per day; USD 20 Plus users average about 3x free users, and Pro users about 11x. OpenAI is diversifying compute procurement across Oracle, CoreWeave, Microsoft, GCP, and AWS, while continuing to view NVIDIA as a preferred partner.
Google/Blackstone The 500MW TPU Cloud plan extends Google’s AI platform from internal capex into a cloud asset financed by external capital. This will add competition for CoreWeave, but also proves that the AI cloud market has not stopped at a single GPU-cloud model and is migrating toward combinations of TPU, software services, and project capital.
SpaceX IPO Q&A materials frame the company as a vertically integrated AI infrastructure platform. The total addressable market discussed in the S-1 is USD 28.5tn, of which USD 26.5tn is AI-related. If Starship launch costs fall below USD 100/kg, orbital compute could be roughly 25% cheaper than ground compute under certain assumptions, but Starship reuse and heat dissipation remain key execution risks.
Microsoft Project Solara Stratechery believes Microsoft is not putting agents into traditional PCs, but building an enterprise agent-device platform with “cloud as the hub and multiple devices as spokes.” Qualcomm and MediaTek are already chip partners, and the platform is based on Android rather than Windows. The investment implication is that enterprise context and device entry points matter more than the stand-alone AI PC narrative.
“Project Solara’s next computer is not a single device, but a coordinated system of devices”
Uber Enterprise AI cost governance is becoming a new variable for platform companies. Wall St Engine recorded that after Uber exhausted its 2026 AI budget in four months, it capped AI coding-tool spending at USD 1,500 per employee per month. This suggests enterprise AI adoption is moving faster than budgeting systems, and the next phase will shift from trials to ROI control.
Chinese internet platforms Goldman Sachs’ China strategy report downgraded H shares/MSCI China from Overweight to Market Weight. Nine large internet companies account for roughly 35% of MSCI China’s market-cap and earnings weight, and are down a combined 22% year to date. Food delivery and instant retail subsidy losses reached RMB 34bn in 1Q26 and have totaled RMB 180bn since 2Q25, slowing earnings recovery.
MiniMax Although it sits in the model application layer, MiniMax is the clearest AI monetization signal within China’s platform ecosystem. J.P. Morgan maintained Overweight with a HKD 1,100 price target. M3 supports up to a 1mn-token context window; at 1mn tokens, prefill is accelerated 9.7x and decoding 15.6x. Whether pricing premia can hold depends on user validation.
Hong Kong and Asian capital markets Morgan Stanley’s interview with Gokul Laroia said Hong Kong raised nearly USD 40bn last year, with biotech, industrial automation, AI, metals, and mining as the main themes. China and India are the two largest IPO pipelines in Asia, while Taiwan and Korea benefit from underlying AI themes.
Software/SaaS
Goldman Sachs AI investment-return framework The report directly raises the debate over USD 7tn-8tn of potential total AI spending. The key question is not whether consumers will use AI, but whether enterprises can “make or save money implementing AI.” The report says economic value has so far mainly accrued to semiconductor companies, while the model layer, cloud service providers, and enterprise applications are still searching for sustainable profit pools.
“all of the economic value has continued to accrue to the semiconductor companies”
Palo Alto Networks Multiple materials recorded Q3 FY2026 revenue of USD 3.0bn, up 31% YoY; NGS ARR of USD 8.1bn, up 60% YoY; and RPO of USD 18.4bn, up 36% YoY. However, GAAP operating loss was USD 183mn, and gross margin declined from 72.9% to 67.6%. AI security demand is strong, but earnings quality still depends on acquisition integration.
“customers turn to us to secure their AI deployments at scale”
GitLab Q1 FY2027 revenue was USD 264.2mn, up 23% YoY; cRPO was USD 724.1mn, up 24% YoY; and net retention was 117%. At the same time, the company cut roughly 14% of staff and exited 22 countries, showing that the growth narrative for agent development platforms has entered the efficiency-delivery phase.
“agentic era is creating structural tailwinds”
Snowflake BTIG reiterated Buy and raised its price target to USD 325, citing AI Agent workloads, Cortex Code, and CoWork as potential drivers of above-consensus FY2027 revenue guidance. This price-target change comes only from a clear investment-bank action clue; the body text does not extend it into a financial forecast.
Giga ML Founder interviews show AI customer-service automation KPIs have risen to 60%-70% deflection rates, with leading customers targeting 90%-95%. An eight-person team won DoorDash, and a three-month pilot “never went down.”
“deflection rates are closer to 10 to 15%... 60 to 70%... 90 to 95%”
Giga ML delivery bottleneck The interview locates the real bottleneck in enterprise AI applications at the forward deployed engineer. Customer strategy, process documentation, dashboards, and meeting requirements all need to be productized. AI coding tools allow the current engineering team to be roughly 6-7x smaller than it would be without tools, but delivery automation remains the key to software-company scaling.
“the biggest bottleneck... is forward deployed engineer”
MiniMax M3 J.P. Morgan states the tension clearly: M3 Standard charges RMB 4.2 per million input tokens and RMB 16.8 per million output tokens in the 0-512k input range, rising to RMB 8.4/RMB 33.6 in the 512k-1mn range. DeepSeek V4-Pro output pricing is lower. If actual quality cannot create enough differentiation, user-migration risk will increase.
“M3’s API price is significantly higher than M2.7/M2.5”
Microsoft MAI models Stratechery recorded Microsoft releasing seven MAI models. MAI-Thinking-1 was tied with Claude Sonnet 4.6 in blind tests, while the Excel agent use case was said to match GPT 5.4 at 10x better cost efficiency. This shows software platforms are competing for model control, not just renting external models.
AI programming and Codex OpenAI CFO interview materials and supplemental information both point to accelerating Codex adoption. Weekly active users grew from nearly zero in January to more than 5mn, while another material said Codex had 4mn weekly active users and had grown 5x since February. Developer tools have expanded from code writing to knowledge-worker automation.
Katy Huberty AI adopter framework Morgan Stanley believes AI adopters are expanding margins at twice the pace of the market index, MSCI World, and the S&P 500. The areas with the largest rate of change are shifting from pure technology to consumer, apparel, durables, and autos. Software companies need to prove not only functionality, but margin improvement.
AI control plane Anthropic/OpenAI-related materials repeatedly mention that there is “no single best model” and that enterprises want to sit on the control plane, because model prices, service terms, regulation, and performance change quickly. This creates incremental opportunities in data orchestration, model routing, token-cost governance, and enterprise permission systems.
“the data isn’t ready to be agented yet”
Microsoft Agent OS NewMaxx recorded Microsoft releasing a device operating system that “does not run applications, only AI Agents.” This is not financial data, but the SaaS implication is direct: the application entry point may migrate from the AppLovin store to the agent orchestration layer, making identity, permissions, context, and tool calling the new software boundaries.
Consumer Electronics/Smart Vehicles
Horizon Robotics Goldman Sachs maintained Buy but lowered its 12-month price target from HKD 15.30 to HKD 13.14. The company plans to release HSD V2.0 in 3Q26. The 560 TOPS J6P is key to high-end intelligent-driving penetration, but 2026/2027E net profit forecasts were revised down to losses of RMB 4.512bn/RMB 80mn.
“J6P 560 TOPS”
Horizon J6 series Goldman Sachs provided vehicle launch evidence: J6E has been adopted by SAIC MG 4X highway NOA models, J6B by GAC Toyota Bozhi 3X, and J6M-based CARIZON solutions have shipped to SAIC ID.UNYX 06/07. The product-mix upgrade is valid, but the 2026E R&D expense ratio remains as high as 89.1%.
Hesai Technology Goldman Sachs maintained Buy, with an ADR price target of USD 35 and an H-share price target of HKD 273. The company was included in the supply chain for NVIDIA’s Isaac GR00T humanoid-robot reference platform. The Sharpa Wave five-finger hand has 22-25 degrees of freedom, and the related supply framework agreement has a 2026 annual transaction cap of RMB 100mn.
“Isaac GR00T Reference Humanoid Robot”
Hesai robotics modules Management guided 2026E strategic growth plan revenue of RMB 100mn, including KOSMO spatial-intelligence devices and upcoming robotic actuator modules. Goldman Sachs expects 2026E revenue growth of 50% and LiDAR shipments up 101% YoY, but blended ASP down 27% YoY.
NVIDIA AI PC Stratechery believes the full RTX Spark/N1X configuration has at most 20 Arm CPU cores, 6,144 CUDA-core Blackwell GPU, 128GB LPDDR5X, and 300GB/s memory bandwidth, and can run 120bn-parameter models and a 1mn-token context. But the true bottleneck in the agent era is CPU scheduling and coordination with cloud inference.
Microsoft Project Solara devices The platform already has two hardware designs. Chip partners include Qualcomm and MediaTek, and it is based on Android rather than Windows. The investment implication is not “AI PCs replace PCs,” but that enterprise agent devices may form a new terminal entry point around context, identity permissions, and cloud inference.
Transmission of memory price hikes to end devices The global technology “chip inflation” report estimates that if incremental memory costs are passed through while preserving gross margin, smartphones, PCs, servers, and storage arrays would need price increases of roughly 34%, 67%, 83%, and 114%, respectively. This would pressure non-AI end demand and may also force lower-memory configurations and product segmentation.
AI PC memory content Supplemental memory information says NVIDIA’s upcoming AI PC architecture is expected to require substantially higher memory content. Combined with clues that DDR5 RDIMM profitability was at one point higher than HBM, edge AI does not only drive NPU/GPU demand; it may also continue to push up LPDDR, DDR5, and local SSD configurations.
Auto and robotics adopters Katy Huberty’s interview mentioned that the sectors with the largest change in AI adoption speed include consumer, durables, and autos. This echoes the two sell-side reports on Horizon and Hesai. Smart vehicles/robots are no longer just an “edge story,” but a joint validation of models, sensors, actuators, and automaker commercialization.
Other Developments
Market crowding MacroCharts believes the first wave of AI/semiconductor bottleneck trades has already seen crowding and leverage risk, with the SOX rebounding 80% over two months. If the S&P 500 sees a standard 9%-12% pullback, the author estimates the SOX could fall 30% in the base case, and warns that demand for 2x leveraged SK hynix ETFs in Korea and call options could amplify volatility.
“Capital may shift rapidly: first voluntarily, then forcibly”
Goldman Sachs alternative investments The report says infrastructure fundraising hit a record USD 221bn in 2025. As of May 2026, 695 infrastructure funds globally were fundraising, with aggregate targets of USD 555bn. AI data centers are pulling real assets, private infrastructure, real estate, and utility financing into the technology investment framework.
“data centers represent two-thirds of deal activity in digital infrastructure”
High-yield technology credit J.P. Morgan’s technology and telecom weekly shows Google/Blackstone’s 500MW TPU Cloud, CoreWeave’s roughly USD 100bn backlog, and APLD’s 300MW/15-year/USD 7.5bn lease all validating AI compute demand. But a single APLD project may require roughly USD 3bn of capital, making credit-financing capacity the bottleneck.
CoreWeave competitive landscape The weekly report believes CoreWeave currently has roughly 1GW of capacity and is expected to reach 3.5GW in 2027, with contracts signed with three of the four major hyperscalers. Blackstone/Google entering neocloud will increase competition, but it also shows large cloud providers are willing to separate capital organization from the compute software stack.
Geopolitical energy risk Kuwait’s military said air-defense systems were intercepting hostile missile and drone attacks. Macro supplemental information also recorded that the probability of a U.S.-Iran agreement before the end of June fell from above 75% to 27%, and the probability of the Strait of Hormuz reopening was only 22%. This is not about technology-company fundamentals, but it affects expectations for oil prices, inflation, and data-center power costs.
“AI-driven stock rally is becoming increasingly vulnerable”
U.S. credit and housing Goldman Sachs Global Markets Daily noted that the 6.5% mortgage rate remains in the 80th percentile since 2000, while HY Building Products has underperformed the HY index by 100bp year to date. Data-center construction is a source of long-term demand within nonresidential construction, but housing and consumer pressure are appearing through corporate-credit differentiation.
China strategy Goldman Sachs maintained Overweight on A shares and downgraded H shares/MSCI China to Market Weight. Its 2026 EPS growth forecast for A shares was raised from 16% to 20%, and the CSI300 target was raised from 5,300 to 5,500. The report prefers Power, Infrastructure, Semi, and Physical AI over offshore internet index beta.
China real estate Expert calls showed that secondhand-home price declines across 80 cities narrowed in May 2026, while destocking cycles in leading tier-1 and tier-2 cities were 9-32 months. Small units in Shanghai and upgrade-oriented homes in Shenzhen stabilized first. This item is only macro and consumer-confidence context and does not dominate the main technology-brief thread.
Pet consumption Morgan Stanley audio transcripts show U.S. pet-industry annual growth is expected to slow from nearly 9% in 2019-2025 to about 4%, with services accounting for slightly more than 40% of industry spending in 2025. This shows consumer budget pressure is real, contrasting with technology-stock valuation expansion.
Bitcoin/quantum risk Quantum-computing materials argue that large-scale threats are more likely from the mid-2040s to mid-2060s, but roughly 25% of network wallets are exposed and USD 400bn-500bn of market-cap incentives will pull migration discussions forward. This is not the day’s main line and is included only as a long-term security-infrastructure observation.
Investment-bank Price-target Changes Over the Past 12 Hours
DirectionInstitutionDateCompanyPrice-target changeRating actionUpMorgan Stanley2026-06-03MicronUSD 520 -> USD 1,050Maintained OverweightUpMorgan Stanley2026-06-03SanDiskUSD 1,100 -> USD 1,750Maintained OverweightUpMorgan Stanley; J.P. Morgan2026-06-03ASMLMorgan Stanley: EUR 1,400 -> EUR 1,660; J.P. Morgan: raised to EUR 1,900 (prior value not disclosed)Morgan Stanley maintained Overweight/Top Pick; J.P. Morgan raised estimatesUpJ.P. Morgan; BofA2026-06-03CredoJ.P. Morgan: USD 230 -> USD 250; BofA: USD 210 -> USD 252J.P. Morgan maintained Overweight; BofA maintained BuyUpJ.P. Morgan2026-06-03Aspeed TechnologyNT$21,600 -> NT$22,000Maintained OverweightDownGoldman Sachs2026-06-03Horizon RoboticsHKD 15.30 -> HKD 13.14Maintained BuyUpJ.P. Morgan2026-06-03CienaUSD 550 -> USD 635Maintained OverweightUpGoldman Sachs2026-06-03CiscoUSD 116 -> USD 125Maintained NeutralUpBTIG2026-06-03SnowflakeRaised to USD 325 (prior value not disclosed)Maintained Buy