404K Semi-Ai 2026-06-03 Tech Morning Brief - Optical Interconnect Re-rating, Memory Price Hikes, AI PC Spillover
AI infrastructure trading is spreading from single-card compute to connectivity, storage, power, and software security. Marvell Technology, HDDs, HBM, NAND, optical modules, server networking, and AI PCs are all providing incremental signals, while crowding, supply, and funding costs are becoming new constraints.
404K Semi-Ai | 2026-06-03
After-hours Summary
Semiconductors and AI infrastructure took over risk appetite. The Tech Evening Brief recorded QQQ up 45 bp, the semiconductor sector up 6%, Marvell Technology up 32% and another 6% after hours. Capital is no longer only chasing Nvidia. It is spreading along TPU, Vera CPU, HBM, HDDs, optical interconnect, copper cables, CPO, and power bottlenecks.
Software and platforms were more divergent. Palo Alto Networks’ earnings made the post-AI-deployment security demand tangible. MongoDB, Elastic, and UiPath still need to prove whether AI can convert into revenue, while Google was pressured by search data and financing headlines. The hardware chain has the greatest short-term elasticity; the software chain needs orders and retention to prove AI is more than a slogan.
The macro backdrop is not empty. SOX, Korean memory, and AI bottleneck trades have risen quickly, and both MacroCharts and BTIG warned about overheated concentration and options positioning. Data-center financing, copper prices, power architecture, and energy risk are lifting the cost floor for AI infrastructure.
Full AI/Semiconductor Value Chain
Compute GPU/ASIC/CPU
Nvidia Citi maintained Buy on June 2 with a $300 target price. The report placed Rubin GPUs, Vera CPUs, LPX inference trays, Spectrum-6, and BlueField-4 into the NVL72 system. The largest anchor is cumulative Blackwell and Rubin orders exceeding $1 trillion by 2027, while Vera CPU was also discussed separately with $20 billion of revenue visibility.
“Rubin GPUs, Vera CPUs, LPX inference trays, Spectrum-6 networking, and BlueField-4 storage are packaged as the NVL72 system.”
Marvell Technology Stifel maintained Buy and raised its target price to $321. The core point is not simply “connectivity devices”; after AI data centers are disaggregated, the connectivity layer becomes necessary infrastructure. Multiple materials cited the market reaction of a nearly one-third one-day gain, a close at $290.79, and market value of about $254.6 billion. Nvidia also invested $2 billion in the company in March.
“once computing is disaggregated and distributed across the data center, connectivity becomes the necessary layer”
MediaTek Morgan Stanley’s Greater China technology semiconductor report treated the N1X/N1 AI PC SoC as a key Computex validation point. Brand checks showed N1X AI PCs need to be priced at $2,899 and N1 models at $1,799. If 2026 shipments reach 5 million-8 million units, with a $40 license fee per unit, this could contribute 5%-10% of MediaTek’s 2026 EPS.
Intel Computex signals focused on 18A and rack-level CPU inference. Core Ultra Series 3 has more than 325 designs, including more than 130 edge designs based on 18A. Rackscale Blueprints can support up to 128 Xeon 6 CPUs in a single rack, 16,384 to 36,864 cores, 384TB DDR5, and about 100kW, emphasizing that agent inference will not be handled entirely by GPUs.
“Compute must be reinvented beyond the socket to rackscale systems”
Qualcomm Cowen relayed that the company has already received purchase orders for custom ASICs, CPUs, and customer-specific collaborations. The revenue ramp starts in fiscal 1Q27, with potential scale in the billions of dollars. This signal pulls Qualcomm from a handset SoC cycle recovery into custom AI infrastructure compute and the customer ASIC validation curve.
Arm Morgan Stanley interpreted RTX Spark, Grace-derived CPUs, Blackwell GPUs, and 128GB unified memory as catalysts for agentic edge computing. It believes successful N1/N1X products would shift the high-end Windows-on-Arm PC ecosystem from near-monopoly by Qualcomm to multi-party competition. The investment focus is IP licensing, royalties, and the edge AI developer ecosystem.
HBM/DRAM/NAND/eSSD/HDD
Micron TrendForce signals showed that in 2027 HBM4 negotiations, HBM capacity per AI ASIC rises from 96GB/192GB to 216GB/288GB, while Rubin Ultra GPU could reach 384GB. For the three memory vendors, HBM’s share of DRAM wafer starts is expected to rise from 18% in 2025 to 30% in 2027, while its bit-supply share rises from 8% to 13%.
“HBM capacity per AI ASIC rises from 96GB/192GB to 216GB/288GB, while Rubin Ultra GPU could reach 384GB.”
SK hynix Multiple memory materials point to long-term tightness. At the group level, the company said memory wafer capacity could double over the next five years and shortages may persist until 2030. Separately, CLSA raised its target price on SK hynix from KRW 3.0 million to KRW 4.0 million and assigned “high-conviction outperform,” indicating that the sell side has already incorporated the supply mismatch between HBM and general DRAM into valuation.
Kioxia Holdings Morgan Stanley sharply raised its target price from JPY 70,000 to JPY 110,000 and maintained Overweight / Top Pick. The reason is that NAND is shifting from a cyclical product to an AI inference capacity layer. The company raised its CY25-28 flash EB demand CAGR from 20% to 22%, and inference-related demand CAGR from 69% to 86%. Average annual capex for FY3/27-FY3/29 is about JPY 470 billion, with stronger supply discipline.
“AI inference demand driving NAND growth at ~46% CAGR in datacenters”
Seagate Technology / Western Digital Citi raised HDD earnings estimates and target prices: Seagate’s target rose from $740 to $1,150, and Western Digital’s from $500 to $685. The report said industry EB demand is about a 25% CAGR, CY26 EB demand exceeds supply by several hundred EB, tightness should last through CY28, Western Digital’s ASP/TB rose 9% YoY last quarter, and customer long-term-agreement visibility extends as far as CY30.
Western Digital TMTB noted that HDD demand is decoupling from compute capex. About 90% of revenue comes from hyperscale/cloud customers, orders are placed 52 weeks ahead, LTAs run as far as 2032, gross margin has moved into the low 50% range, and incremental gross margin is 70%-100%. This translates “the continuous accumulation of AI-generated data” into HDD pricing power, not just cloud-server restocking.
“Compute power is reused... Storage keeps compounding”
Micron / Phison Electronics NewMaxxSSD signals provided hard data on storage density and controllers. Micron’s 6600 ION reaches 245TB per drive. Phison’s PCIe 6.0 X3 SSD controller offers 28GB/s bandwidth, 6.8 million IOPS, and supports up to 2PB per drive. E37T/PS5037-E37T targets PCIe 5.0 with 4.5W power consumption. AI storage is moving from capacity constraints to architecture constraints.
Optical Interconnect, PCB, Servers, Power Supply
Ciena Morgan Stanley maintained Equal-weight with a $405 target price. It acknowledged that demand is “hard to find negative data points” but believes supply and gross-margin constraints limit near-term upward revisions. The report expects FQ2 revenue could beat by 3%-5%, at about $1.55 billion-$1.58 billion, with FQ3 guidance potentially $50 million-$75 million above FQ2 and full-year revenue growth expectations moving up to 30%-32%.
“Demand environment clearly still bullish”
Credo Conference-call signals showed FY27 optical-module revenue is expected to exceed $600 million, mainly in the second half. This year is led by 800G; FY28 optical-module revenue could double or triple and start contributing CPO/NPO. Needham raised its target price from $220 to $275, while Stifel maintained $250. The core risk is that the largest customer may account for 34%.
Coherent-lite optical interconnect FundaAI placed this in the Google TPU / OCS optical-interconnect chain. The key is not replacing 2-20km DCI, but using coherent detection to recover OCS link-budget losses. 2.4T corresponds to 300G/lane x8 and is introduced in 2027. 3.2T corresponds to 400G/lane x8, with demand rising in 2028-2029. The incremental 2028 industry value is $15 billion-$20 billion.
“OCS introduces about 1.5-3 dB insertion loss per node.”
Corning AI data-center fiber is today’s most direct bottleneck evidence in optical materials. Industry posts said AI data centers require 36 times more fiber than standard server designs, and glass shortages have pushed cable lead times to one year. The Springboard plan points to annualized sales of $20 billion in 2026, $30 billion in 2028, and $40 billion in 2030.
“AI data centers require 36 times more fiber than standard server designs”
Hewlett Packard Enterprise Citi said F2Q26 revenue was $10.7 billion, up 40% YoY, and non-GAAP EPS was $0.79, up 108% YoY. Quarterly AI systems orders were $1.8 billion, cumulative AI bookings were $16.4 billion, and backlog was $5.9 billion. Morgan Stanley separately raised its target price sharply from $33 to $71, indicating that server pricing and the networking business have taken over from GPUs as validation points.
Microchip Technology / STMicroelectronics Data-center control, analog, storage control, and high-speed interconnect are starting to be re-rated. Microchip disclosed CY25 data-center revenue of $302.7 million, with CY26 guidance of about $500 million, up about 65% YoY. STMicroelectronics raised its FY26 data-center revenue target from more than $500 million to $1 billion, and some materials said 2027 could double again.
Jiaze Terminal / FIT Hon Teng Connector and rack value content comes from platform transitions. Jiaze’s server-end 2026 shipments are expected to rise 15%-20% YoY, and SOCAMM production starts in 3Q26. FIT Hon Teng’s AI server revenue mainly comes from data interconnect and power. Management guides total CY27-28 revenue to mid-to-high 20% YoY growth, with cloud/data-center share rising to the low 30% range.
AI data-center power chain SemiVision noted that GB200/GB300 rack power has already exceeded 200kW, and future Kyber racks are moving toward 1MW. 800V HVDC can reduce copper weight by about 45% and improve end-to-end efficiency by about 5%, while SST system efficiency can exceed 98%. This shows AI capex has moved from chip procurement into power architecture and copper-consumption constraints.
“The bottleneck is no longer only compute, memory, or networking, but power delivery.”
Internet/Platforms
Google TMTB recorded that the stock fell 4% on $80 billion equity financing and weaker search data. Yipit and M-sci believe third-party search data within the quarter may be 2-3 percentage points below expectations, while Oxford DP forecasts it 2-3 percentage points above expectations. The same material said GCP may approach 80% growth. The platform divergence centers on the tug-of-war between search cash flow and cloud AI velocity.
Microsoft Build signals expanded the platform from Copilot to agent-native devices, Cobalt 200 Arm CPU, Azure HorizonDB, and Rayfin. Other materials said Microsoft released seven MAI models, with Transcribe-1.5 five times faster, and MAI running on the Maia 200 base.
“Microsoft released seven MAI AI models.”
OpenAI CFO Sarah Friar’s interview illustrated operating intensity: $122 billion of March financing for compute, data centers, and business-model optionality; ChatGPT weekly active users above 900 million; Codex rising from nearly zero in January to more than 5 million users by the weekend; and a revenue structure close to 50/50 between consumers and enterprises. The AI platform narrative is still using user density to support capex.
“ChatGPT weekly active users exceed 900 million, and Codex rose from nearly zero in January to more than 5 million users by the weekend.”
Anthropic Publicly supported execution of U.S. AI policy. The material had no financial figures, but the implication is that leading model companies are becoming more deeply tied to policy frameworks. Combined with market observations that Amazon’s $8 billion investment in Anthropic is worth about $74 billion on paper, platform value has extended from pure model capability to cooperation across cloud, regulation, energy, and national infrastructure.
CoreWeave Is the hard anchor for AI cloud-platform financing pressure. A related data-center issuer sold $900 million of five-year high-yield bonds at a 7.5% yield, and a 15-year lease corresponds to about $2.2 billion of revenue. Ben Bajarin also placed the company’s disclosed backlog of about $99.4 billion and run-rate revenue target of $18.5 billion-$19.0 billion into a “monetizable megawatts” framework.
“The same 1MW can reach $7 million-$13 million in the hands of GPU cloud / full-stack neocloud providers.”
Oracle Related materials linked the $20 billion revenue visibility for Nvidia’s Vera CPU to Oracle order support. The point is not a single cloud earnings report, but that hyperscale customers are starting to procure CPUs, GPUs, networking, and database platforms as a package. This reinforces the main line that AI agents require more general-purpose compute orchestration.
TeraWulf / Google Rittenhouse Research added that TeraWulf expects to deliver about 500MW of TPU-based data-center capacity to Google by the end of 2026, and speculated that another 500MW could be added annually in 2027-2029. This signal is useful as a validation point for platform companies’ TPU capacity outsourcing and power-project management capability.
OpenRouter Tomasz Tunguz recorded that open-weight models account for 69.1% of token volume, while closed-source models account for 30.9%. He also noted that AWS can still generate more than 30% operating income even in commoditized businesses such as cloud computing and storage. The platform implication is that model open-sourcing does not necessarily eliminate profit; the key is control over routing, scale, tools, and cloud resources.
“Open-weight models account for 69.1% of token volume, while closed-source models account for 30.9%.”
SpaceX IPO Q&A framed commercial space as vertical AI infrastructure. The S-1 addressable market was $28.5 trillion, of which $26.5 trillion is AI-related. Falcon 9 launch cost has fallen about 95% since 2008. If Starship falls below $100/kg, orbital compute could be about 25% cheaper than terrestrial compute, but heat dissipation and mass budgets remain real constraints.
Meta The Information-related updates said that after employee opposition, Meta scaled back some functions of an employee tracking tool and added privacy protections, partial exemptions, and a 30-minute pause option. The platform takeaway is that using employee desktop activity to train AI agents will hit privacy and governance boundaries; AI efficiency cannot be explained only by “more internal data is better.”
Uber Rittenhouse Research defined it as an “AI loser” trade sample, saying the stock is down 13% year to date and questioning the mismatch between engineering investment and new features, UX/UI improvement. Internet-platform AI valuation no longer depends only on whether management talks about AI; it depends on whether product iteration, cost efficiency, and the core scenario experience are genuinely improving.
Dan Loeb / Third Point The interview gave a platform-capital perspective. Third Point manages about $25 billion of assets, and the AI infrastructure chain covers power, chips, semiconductor equipment, storage, hyperscale data centers, foundation models, software, and applications. Platform investing is no longer just buying model companies; it is identifying who can connect credit, chips, and customer demand.
Goldman Sachs AI ROI framework Discussed potential total AI spending of $7 trillion-$8 trillion, with some hyperscale cloud providers spending about $200 billion in annual capex. The conclusion is not that capex goes to zero, but that the next two years need to prove whether enterprises can convert AI into profit, cost reduction, or sustainable competitive advantage. Platform stocks need to rotate from “imagination” back to ROI.
“AI investment boom: when will returns appear?”
AI customer-service platforms The Giga ML case showed that application platforms still have real wedges. Traditional customer-service bots hand off to humans about 10%-15% of the time, while AI experiences can reach 60%-70%, and leading customers target 90%-95%. An eight-person team won a DoorDash pilot and had no downtime for three months, showing that enterprise agent-platform opportunities lie in quantifiable KPIs, not concept demos.
Location platforms Counterpoint Research re-rated maps from navigation tools to real-time contextual-intelligence engines. HERE remained the leader for the ninth consecutive year, while TomTom, Google, and Mapbox were also listed as leaders and pioneers. This signal belongs in the platform section because AI location intelligence will connect mobility, robotics, enterprise dispatching, and agent decision-making.
Software/SaaS
Palo Alto Networks Earnings were the strongest anchor for the software sector. Fiscal 3Q revenue was $3.0 billion, up 31% YoY, and adjusted EPS was $0.85. NGS ARR was $8.1 billion, up 60% YoY, RPO was $18.4 billion, up 36% YoY, and full-year revenue guidance was raised to $11.415 billion-$11.425 billion. AI deployment security demand is starting to enter orders and visible revenue.
“NGS ARR was $8.1 billion, up 60% YoY; RPO was $18.4 billion, up 36% YoY.”
Palo Alto Networks The reverse constraint is M&A and valuation. Gross margin fell from 72.9% to 67.6%, and GAAP operating income/loss swung from a $219 million profit to a $183 million loss. Forward P/E is about 80.69. The market is willing to pay an AI security premium, but the follow-up depends on margin repair.
MongoDB Morgan Stanley maintained Overweight and raised its target price from $335 to $380. Fiscal 1Q revenue was $688 million, up 25% YoY. Atlas grew 29%, and the midpoint of FY27 revenue growth rose from 17% to 19.5%. The core divergence is that AI-native customers are emerging, but AI benefits in OLTP databases are being released more slowly than at Datadog and Snowflake.
“Core business is solid, AI-native customers are beginning to form key use cases, and the growth inflection is closer.”
GitLab TMTB recorded fiscal 1Q EPS of $0.23 above $0.21 and revenue of $264.2 million above $254.2 million, but billings, deferred revenue, and RPO were all below expectations, and the company cut 14% of staff. The investment implication is that demand for the developer toolchain is not absent, but there is still a gap between income-statement improvement and order visibility.
Elastic Morgan Stanley’s New Stack review showed Q4 total revenue of $451 million, up 14% in constant currency, and Cloud revenue of $217 million, up 19%. cRPO rose to $1.20 billion and total RPO to $1.98 billion. Customers with ACV above $100,000 using AI features increased from 470+ to 600+, but in-year acceleration requires Cloud commitments to convert into revenue.
“case for acceleration remains a debate”
UiPath Revenue was $418 million, up 17% YoY and about 6% above expectations. Non-GAAP operating profit was $92 million, with a 22% margin, and ARR was $1.901 billion, up 12% YoY. Sixteen of the top 20 deals included AI. Maestro was positioned as an orchestration layer across agents, robots, APIs, systems, and humans, but Morgan Stanley cut its target price from $17 to $15.
PagerDuty Remains a show-me story. Q1 revenue was $121 million, up 1.0% YoY, ARR was flat at $496 million, NNARR was negative $2.7 million, and DBNRR fell to 97%. The positives are 24.5% operating margin, FY27 EPS guidance raised to $1.30, and usage-based products already contributing 10% of ARR.
“ARR acceleration still not fully proven”
Snowflake FundaAI’s weekly report said fiscal 1Q product revenue grew 34% YoY, full-year product revenue guidance was raised from 27% to 31%, and Cortex Code is expected to contribute about 2% incremental revenue and is described as the fastest-growing new product in company history. Data-platform software elasticity comes from AI developer tools, not only single-point improvement in traditional warehouse usage.
Marvell Technology / Semtech-related software read-through FundaAI also placed Marvell Technology’s interconnect business growing 70% YoY, data center growing 50%, and Semtech optical-chain upside into an AI SaaS weekly report. This shows software investors are also using the hardware chain to validate AI application demand, which will raise the near-term correlation between SaaS valuation and semiconductor cycle strength.
MiniMax Morgan Stanley maintained Overweight with a HK$1,100 target price. M3 is described as an open-weight large model with coding/agentic capability, native multimodality, and up to a 1 million-token context. Sparse Attention is more than 9x faster in prefill and more than 15x faster in decoding at a 1 million context, while API pricing doubled versus M2.7.
“Up to a 1 million-token context window”
Codex The product line expanded from code assistance to hosted sharing sites, plugins/skills, and visual annotation feedback. The materials had no revenue figures, but the direction is clear: agents will not only write code inside IDEs, but also connect documents, slides, spreadsheets, web prototypes, and feedback loops into enterprise workflows. The boundaries of collaboration software will be redefined.
Block / Shopify / Stripe Morgan Stanley’s SMB survey showed the top three payment-processing usage rates were Square 52%, PayPal 45%, and Stripe 29%. Block’s target price was raised from $96 to $98, and FY27 Square gross profit growth expectations were raised to 16.5%. The divergence in software-enabled payments lies in the coexistence of fee-rate sensitivity and AI workflow attachment.
Intuit Was cut to Sell in market materials on Goldman Sachs concerns over competition in tax, with the stock down more than 10%. This line belongs in software as a reminder that AI is not incremental for every vertical software category. The more standardized tax, forms, customer service, and similar workflows are, the more easily new agents can attack the profit pool.
OpenRouter / model platforms Rising open-weight model traffic share coexists with per-token pricing increases over the past one to two quarters, showing that AI software business models do not depend only on whether models are closed-source. Developer routing, tool calling, context length, latency, cost ceilings, and enterprise permissions jointly determine who can convert tokens into profit.
Consumer Electronics/Smart Vehicles
XPeng Morgan Stanley said that after the GX launch, the company shifted from a “show-me story” to a “believe-me story.” It maintained Overweight, with a HK$96 Hong Kong target price and a $25 ADR target price. 2026-2028 delivery forecasts are 442,000, 560,000, and 626,000 vehicles. Demand for the high-end GX Ultra is viewed as key evidence of renewed high-end brand recognition.
“After the GX launch, the company moved from the stage of ‘needing to prove itself’ to one where ‘the market can believe again.’”
Apple Counterpoint Research dissected the M5 Pro and believes the Pro series has entered a dual-die / chiplet architecture. GPU AI throughput reaches 4x that of M4 Pro, maximum memory bandwidth is 307GB/s, and unified memory goes up to 64GB. The consumer-electronics investment point is shifting from a pure replacement cycle to whether edge AI models can truly use local memory and chiplet architecture.
Microsoft / Nvidia RTX Spark Morgan Stanley said the platform combines Nvidia Grace CPU architecture, Blackwell RTX graphics compute, and up to 128GB unified memory, allowing local operation of large models with up to 120 billion parameters. The first batch includes Surface Laptop Ultra and other OEM models, launching within the year. Windows is moving from an application entry point to a personal AI agent terminal.
“Run large models with up to 120 billion parameters locally”
Nvidia RTX Spark / N1X Ming-Chi Kuo believes the market should not only look at code names and specifications. The real determinants of the edge AI replacement cycle are the operating system, cloud/local LLM switching, agent harness, cross-application workflows, and sandbox. He expects RTX Spark devices to remain a niche notebook market over the next two years, and commercialization needs verification through real software experiences.
AI PC price bands Signals from Max Weinbach and LeoSkie both place N1X AI PCs at $2,899 and N1 models at $1,799. If 32GB can run some Qwen 3.6 35B-A3B tasks, a 128GB version may replace part of AI subscription services. The key for consumer electronics is not “can it run,” but whether local inference is economical versus cloud subscriptions.
“AI PCs with N1X need to be priced at $2,899, while N1 models are priced at $1,799.”
Qualcomm Benefits from both AI PCs and customer ASICs. The Windows-on-Arm ecosystem is no longer defined only by Qualcomm, but if Qualcomm can convert CPU, NPU, and custom ASIC orders into fiscal 2027 revenue, it may also move from the handset cycle into edge AI and customized compute for data-center customers. The June 24 event needs attention to order value and customer names.
Arm After RTX Spark enters the high-end Windows-on-Arm PC ecosystem, Arm’s edge AI narrative expands from handsets to notebooks, robotics, and local inference. Morgan Stanley emphasized local AI execution, 128GB unified memory, and 1 petaflop AI performance. The investment question is whether developers are willing to localize agent workflows on Arm PCs.
Lenovo Group Existing materials mentioned AI-driven growth, server OPM rising to 3.6%, and PC OPM staying at 7.0% despite rising storage costs. As AI PC price bands are being re-discussed today, the focus for global OEMs such as Lenovo is whether they can defend margins amid memory price increases, N1/N1X platforms, and Windows Agent updates.
Phison / Intel AI PC NewMaxxSSD recorded that Phison is working with Intel to bring larger-scale local AI workloads to Intel AI PC platforms. It is not a financial disclosure, but it explains why SSD controllers, low-power PCIe 5.0/6.0, and the local data layer will become key selling points for edge AI PCs beyond CPU/GPU.
OpenAI / Opal Wired reported that OpenAI led Opal’s $40 million Series B. Opal is shifting from high-end webcams to AI audio products, expected to launch in the next three to four months, with Samsung also on the shareholder list. The consumer-electronics implication is that AI hardware is not only in PCs; it will also seek entry points through cameras, audio, and standalone agent-control devices.
“Opal is renaming itself Opal Electronics and is developing an AI-powered audio product.”
Ambarella / Vishay Precision / Ouster MoMoMacro listed machine-side perception as a scarce layer of physical AI. AMBA has about $3.4 billion of market value and trades at 7.5x sales. On May 28, the Hanwha deal locked in more than $800 million of edge AI chip supply over the next 10 years. VPG’s Q1 orders exceeded $100 million for the first time since 2022 and included its first $1 million humanoid-robot orders.
Hesai Technology Market discussion asked whether it could become a supply-chain mapping related to Nvidia’s Isaac GR00T humanoid-robotics platform. The logic is that as a Sharpa contract manufacturer, it could receive attached upside from Wave hands-related shipments. This signal has no order or revenue figures and is suitable only as an observation point for robotics-platform sensing/execution chains; it should not be extrapolated into certain earnings.
Yaskawa Electric Morgan Stanley IR Day feedback showed the company targets F3/30 revenue of JPY 650 billion and OPM above 15%, with about JPY 250 billion of investment over four years for AI, automation, and production-base reinforcement. But Morgan Stanley’s JPY 4,700 target price is below the June 1 share price of JPY 7,140, indicating that the physical AI narrative is strong while profit delivery remains insufficient.
Mitsubishi Heavy Industries / Preferred Networks The cooperation places AI into critical infrastructure and defense industrial systems, targeting a capital and business alliance agreement within F3/27. It is not consumer electronics, but it belongs to edge-side diffusion of smart devices and industrial AI. The investment divergence is that AI can improve equipment intelligence, but may also reduce power demand indirectly through server-efficiency improvements.
Location platforms and smart-vehicle ecosystem HERE remained the leader in location platforms for the ninth consecutive year. TomTom, Google, and Mapbox were listed as leaders and pioneers, while Baidu and Amap were listed as strong challengers in China’s mobility ecosystem. AI maps are moving from navigation into a real-time contextual layer for robotics, autonomous driving, enterprise dispatching, and agent decision-making.
Other Developments
Market concentration Kobeissi Letter recorded that the U.S. technology sector rose 42% over the past two months, the largest two-month gain in 24 years. SOX rose 66% over the same period, significantly outperforming the S&P 500’s 16% and the Dow’s 10%. The top 10 stocks contributed about 65% of the S&P 500’s gain since the March 30 low, and half of them were semiconductor stocks.
“The U.S. technology sector rose 42% over the past two months, the largest two-month gain in 24 years.”
Semiconductor crowding MacroCharts said SOX rose 80% in two months. If the S&P sees a standard 9%-12% pullback, the SOX base case could fall 30%, pointing to about 9k. After past episodes of extreme options speculation, SOX fell 19%, 28%, and 17%, respectively. Current intraday downside beta at times reached 8, warning that memory and optical-module trades need to guard against crowding.
Market breadth BTIG’s view warned that the S&P 500 has kept rising while breadth deteriorated. Over the past six sessions, the index rose each day while decliners outnumbered advancers every day, a pattern not seen since 1996. Over the past five sessions, only technology rose 5.9% and materials rose 0.40%, while communication services, REITs, consumer staples, and utilities all fell more than 3%.
“The index rose for six consecutive days, with more decliners than advancers every day.”
Korean memory leverage MacroCharts said assets in the Korean 2x SK hynix ETF have grown 10x year to date, accounting for nearly 9% of Hong Kong ETF assets, and due to swap-limit constraints, 27% has been allocated to call options. This explains why, beyond SK hynix and Samsung, Korean shadow equities and options positioning can amplify memory-cycle volatility.
Labor data ZeroHedge recorded JOLTS job openings at 7.618 million, significantly above expectations and the prior 6.866 million. This will pressure easing expectations and expose high-valuation AI trades to rising-rate risk. For the tech morning brief, the meaning is that AI capex is still strong, but the discount rate may not cooperate.
Eurozone inflation May CPI rose to 3.2% YoY, core inflation was 2.5%, and services inflation was 3.5%. The market has almost locked in a 25 bp ECB hike on June 11. This is not a technology-company fundamental, but it affects global long-duration growth stocks and data-center financing costs.
AI CapEx capital pool Ben Bajarin cited Morgan Stanley as saying the AI build-out cycle may require about $10 trillion, while global asset-owner capital is about $256 trillion, implying construction scale of about 4%. The conclusion is that the total amount of capital is not the bottleneck; the real bottlenecks are strategic allocation, project quality, power, and supply-chain execution.
“Capital is not the bottleneck. Strategic allocation is key.”
Data-center construction cost A Cowen panel said demand scale is “unprecedented,” with customers shifting from near-term capacity to five-year-plus planning. Construction cost has risen to $13 million-$19 million per MW, above about $12 million/MW before 2H25. Power, grid interconnection, natural-gas pipelines, electricians, and general-contractor capacity have all become AI infrastructure variables.
CoreWeave credit financing The $900 million high-yield bond, 7.5% yield, and 15-year lease corresponding to about $2.2 billion of revenue show that AI data-center developers are starting to use the credit market to support expansion. Related developer junk-bond financing has exceeded $27 billion this year. The follow-up is whether lease tenor, GPU depreciation, and customer renewals match.
Nuclear power and data centers Multiple observation-pool materials framed AI’s need for 24/7 clean baseload power as the core of the nuclear narrative. But another group of materials warned that new nuclear power often takes five to seven years, with high execution, regulatory, and cost risks. For AI data centers, the more realistic short- and medium-term power mix may still be gas, backup power, SST, HVDC, and demand response.
Copper prices and power materials ZeroHedge cited HSBC and Goldman Sachs as saying LME copper is about $13,832/ton, close to historical highs. Goldman cut its 2026 global mine-supply forecast by 350,000 tons and raised its end-2026 / 2027 average LME copper price forecasts to $13,735 / $13,800. AI grids, HVDC, and data-center copper consumption will amplify materials sensitivity.
“The 2026 global mine-supply forecast was cut by 350,000 tons, about 1.5% of global mine supply.”
WSTS semiconductor forecast Background materials said the 2026 global semiconductor market would grow 89.9% YoY to $1.5112 trillion, with memory growing about 3.5x to $803.9 billion. The figure comes from social-media relay and cannot serve alone as a valuation anchor, but it explains why capital is willing to continue revising memory and semiconductor equipment upward.
AI ROI bear case Where’s Your Ed At emphasized that token costs may still consume enterprise ROI. Uber exhausted its full-year token budget in four months; some companies spent $500 million in a single month because they had not set limits; GitHub Copilot users consumed quota quickly after the shift to token-based billing. It is a reverse stress test for Goldman’s “ROI validation over the next two years.”
Gold and reserve assets Whale Factor recorded that by end-2025, gold accounted for 27% of central-bank reserves, while U.S. Treasuries fell to 22%. This signal does not directly determine technology stocks, but it explains the macro backdrop facing high-valuation growth stocks: capital is chasing AI while also using gold to hedge sovereign-debt uncertainty.
Digital-asset risk appetite BTC fell below $69,000, down 3.5% intraday, wiping out more than $49 billion of market value within 24 hours. At the same time, Schwab plans to launch spot crypto trading for financial advisors in 2027. Digital assets are not this report’s main line, but they can reflect that high-beta risk appetite has not kept pace with AI semiconductors.
Investment-bank Target Price Changes Over the Past 12 Hours
DirectionInstitutionDateTargetTarget-price changeRating actionRaisedStifel2026-06-03Marvell TechnologyRaised to $321 (prior value undisclosed)Maintained BuyRaisedNeedham2026-06-03Credo$220 -> $275Maintained BuyRaisedMorgan Stanley2026-06-03Hewlett Packard Enterprise$33 -> $71Maintained Equal-weightRaisedMorgan Stanley2026-06-03Kioxia HoldingsJPY 70,000 -> JPY 110,000Maintained Overweight / Top PickRaisedCiti2026-06-03Seagate Technology$740 -> $1,150Maintained BuyRaisedCiti2026-06-03Western Digital$500 -> $685Maintained BuyRaisedMorgan Stanley2026-06-03MongoDB$335 -> $380Maintained OverweightRaisedBofA2026-06-03STMicroelectronics$73 -> $83Maintained Neutral