1. ASML: Expanding EUV Supply Capacity and D·E·F Tool Generation Transition — Structural Improvement in ASP and Margins
• Core Source
“ASML reconfirmed that it will expand EUV equipment supply capacity to a minimum of approximately 60 units in 2026 and a minimum of approximately 80 units in 2027, with internal capacity expandable to approximately 90 units, and potentially beyond.”
“The expansion from 60+ to 80+ EUV units represents approximately a doubling of 2025 shipment capacity and has been identified as the key growth driver for next year.”
“The last batch of D-tools (160wph / ~€170M ASP) is being shipped, while E-tools, which account for the majority of this year’s shipments, feature 220wph / ~€240M ASP, with a 230wph upgrade supporting further ASP increases. F-tools target approximately 260wph and higher ASP, and E-tools are expected to take the ‘lion’s share’ of next year’s shipments.”
“In the long term, Low-NA targets approximately 300wph by 2029, with 400–500wph also considered achievable through optical redesign and laser power improvements, with throughput gains expected to drive structural improvement in ASP and gross margins.”
“ASML indicated that DRAM could approach over 50% of total revenue over time, and large-scale cleanroom expansions point to meaningful DRAM tool shipments ahead.”
• Expected Impact
The message delivered by ASML management at the JPMorgan TMT Conference centers on two simultaneous dynamics: volume expansion on the supply side and ASP improvement driven by tool generation transition on the revenue side, together ushering in a phase of both quantitative and qualitative earnings improvement.
On volume, the 2027 EUV shipment target of 80+ units represents approximately a doubling of 2025 output. The current workhorse model is the E-tool (220wph, ASP ~€240M), with the successor F-tool targeting 260wph and higher ASP. The exit of the D-tool (160wph, ASP ~€170M) and the ongoing mix shift toward E- and F-tools alone creates a structural uplift in average revenue per unit. Longer term, Low-NA tools target 300wph by 2029, with 400–500wph potentially achievable through optical redesign and laser power enhancements.
On demand, the acceleration of EUV penetration in DRAM represents another growth axis. EUV is already used in an average of 5–6 layers at the 1c/1-gamma node, and lithography intensity is expected to increase further at subsequent nodes. ASML’s indication that DRAM could approach over 50% of total revenue signals a structural shift from a logic-dominated revenue mix to one where memory becomes a core pillar. Major DRAM customers including SK Hynix, Micron, and Samsung, all in the midst of large-scale cleanroom expansions, are likely to meaningfully scale up EUV demand from 2027 onward.
Regarding High-NA, the fact that TSMC’s A14 node will proceed with Low-NA may delay the High-NA adoption timeline slightly, but this paradoxically supports margins by driving higher Low-NA shipment volumes. ASML’s margin update is scheduled for July, and the market is closely watching for potential upside guidance revisions.
2. AI Infrastructure Investment: Global Combined CapEx Projected to Reach $2.75 Trillion by 2030
• Core Source
“U.S. Big Four (AMZN, META, GOOG, MSFT) combined capital expenditure is projected to surge from approximately $212.9 billion in 2024 to $362.6 billion in 2025 and $661.0 billion in 2026, before expanding further to $893.0 billion in 2027, $1.069 trillion in 2028, $1.223 trillion in 2029, and $1.379 trillion in 2030.”
“U.S. listed companies combined reach $415.1 billion in 2026 and $1.523 trillion in 2030, and adding RoW (excluding China) of $239.0 billion (2026E) to $939.0 billion (2030E), the global total is estimated at approximately $2.759 trillion in 2030.”
“AI accelerator revenue TAM is estimated to grow from $191.6 billion in 2025 to $396.4 billion in 2028 (CAGR 27.4%), with the 2028 combined data center semiconductor TAM of $563.2 billion comprised of GPU/Custom ASIC (including HBM) at 70.4%, networking at 21.1%, other (including storage) at 6.4%, and CPU at 2.1%.”
“On the demand side, AI services revenue (token monetization) is projected to surge from $7.2 billion in 2024 to $43.8 billion in 2025E and $231.4 billion in 2026E, before reaching $431.7 billion in 2027E, $705.9 billion in 2028E, $998.6 billion in 2029E, and $1.287 trillion in 2030E.”
“By 2026E model revenue share, ChatGPT accounts for 14% ($32.5 billion), Claude 21% ($47.8 billion), Gemini 17% ($38.7 billion), MSFT API Foundry 11% ($25.9 billion), and Amazon Nova/Titan 13% ($29.0 billion), with Claude notably taking the top position in revenue share.”
• Expected Impact
The central thesis of Citi’s AI industry supply-demand model, published on May 22, is that 2026 marks a qualitative inflection point in the investment cycle — not merely another year of CapEx growth, but the first year in which AI services revenue rises to a level that begins to justify the scale of infrastructure investment.
On the supply side, the U.S. Big Four’s combined 2026 CapEx of $661 billion represents an approximately 82% year-on-year surge, the largest single-year increase on record, signaling that the AI infrastructure race has effectively shifted into all-out competition. By 2030, this figure is projected to reach $2.759 trillion on a global combined basis.
On the demand side, AI services revenue is forecast to surge more than fivefold in 2026 to $231.4 billion. Notably, Claude is projected to claim the top revenue share at 21% among AI models. This reflects Claude’s premium positioning with a higher monetization price per token, even as ChatGPT maintains an overwhelming lead in token production volume share of 45–63%.
Data center power demand is also set to surge dramatically. Power consumed by AI workloads is projected to expand more than sevenfold from 8.8GW in 2024 to 64.6GW in 2030, with AI inference workloads alone growing from 3.8GW to 45.2GW. These figures provide structural underpinning for a supercycle across the full data center infrastructure stack, spanning power, cooling, and real estate.
3. A Generational Shift in Data Center Power Architecture — From 48V to 800VDC
• Core Source
“As GPU cluster power density surges beyond 600kW per rack, the existing 48V/AC-based architecture has reached its physical limits.”
“Raising voltage from 48V to 800V reduces the current required for the same power delivery by approximately 15 times, which theoretically reduces resistive heat loss by more than 220 times. This significantly reduces copper cabling weight, improves overall system power efficiency by approximately 5%, and delivers annual power savings of over 50MW for a 1GW-scale data center.”
“Phase 1: White Space Retrofit (2026–2027) — Google and Meta are leading early adoption centered on the Open Compute Project’s ‘Diablo 400’ specification.”
“Phase 4: SST (Solid-State Transformer) Endgame (post-2029) — Integrating the transformer and rectifier into a single unit reduces volume by over 90% and pushes efficiency to approximately 99%.”
“The total electrical infrastructure cost per MW (approximately $3.6M–$4.8M) does not change significantly, but where that cost is deployed and which components capture it shifts entirely.”
• Expected Impact
The 800VDC transition analyzed by SemiAnalysis is not simply a voltage specification change — it represents a full restructuring of the data center power value chain. As AI server rack power density has exceeded 600kW, the existing 48V architecture has reached a physical limit that can no longer be resolved through incremental improvement.
The core logic of the transition is rooted in physics. When voltage rises approximately 17-fold from 48V to 800V, the current required to deliver the same power drops by approximately 15 times, and since resistive losses scale with the square of current, heat loss is theoretically reduced by more than 220 times. For a 1GW-scale data center, this translates to annual power savings of over 50MW.
The transition unfolds across four phases. Phase 1 (2026–2027) involves adding power sidecar racks within existing AC infrastructure, led by Google and Meta. Phase 2 (2027–2028) sees native 800VDC compute within server blades become mandatory. Phase 3 (2028–2029) converts the building-level power distribution to DC entirely, eliminating AC switchboards and PDUs. Phase 4 (post-2029) introduces solid-state transformers (SSTs) to replace iron-core transformers, cutting volume by over 90% and pushing efficiency to 99%.
From an investment standpoint, the critical insight is that total infrastructure cost is maintained, but the beneficiaries change fundamentally. The share going to legacy central UPS and large transformers shrinks, while HVDC power racks, SSTs, and battery racks in close proximity to servers capture rising value. In particular, SiC MOSFET-based power semiconductors capable of high-voltage, high-speed switching, and SSCB (Solid-State Circuit Breaker) makers capable of interrupting current within microseconds without arcing, emerge as the new class of beneficiaries. With Google and Meta leading early OCP-standard deployments from 2026–2027, the order pipeline for relevant supply chains is already beginning to materialize.
4. AMD Commences Mass Production of EPYC Venice on TSMC 2nm
• Core Source
“AMD announced that it has begun mass production of the 6th generation EPYC CPU codenamed ‘Venice.’ This product represents an important milestone in AMD’s 2nm process collaboration with TSMC. AMD also plans to conduct future production of this product at TSMC’s wafer fab in Arizona, USA.”
“AMD’s Lisa Su stated that AI infrastructure demand has entered a new phase of expansion. She explained that surging inference and Agentic AI demand is driving supply shortages not only in GPUs but across CPUs, advanced packaging, substrates, memory, and rack-scale systems.”
“She particularly emphasized that CPU demand is far stronger than expected and that the importance of CPUs within AI infrastructure has risen sharply alongside the expansion of inference. AMD plans to gradually scale up CPU supply this year and significantly expand production capacity from 2027 onward.”
“AMD (+3.99%) rose after CEO Lisa Su announced that inference demand has grown substantially, highlighting the surging importance of CPUs, and projected annual growth of 35% or more over the next five years.”
• Expected Impact
AMD’s announcement that it has commenced mass production of the 6th generation EPYC CPU ‘Venice’ on TSMC’s 2nm process carries two simultaneous implications: a technology milestone in the AMD-TSMC partnership, and a fundamental shift in CPU demand structure in the AI era.
On the technology side, TSMC’s 2nm (N2) process is the most advanced node currently in mass production. Venice’s entry into N2 production confirms that AMD maintains process leadership in its core server CPU lineup. AMD’s plan to also produce the same product at TSMC’s Arizona fab signals a clear commitment to U.S.-based AI semiconductor supply chain localization.
On the demand side, a more structural shift is underway. The proliferation of AI inference and Agentic AI is transforming the AI compute architecture from GPU-centric to a CPU-GPU collaborative structure. While CPU utilization relative to GPU was minimal in traditional AI training environments, in inference and agent workloads, the role of CPUs has exploded — handling scheduling, network management, KV cache processing, and more. Lisa Su’s statement that “CPU demand is far stronger than expected,” paired with a projection of 35%+ annual growth over the next five years, is a direct confirmation of this structural shift. The fact that NVIDIA simultaneously declared entry into the CPU market with a $200 billion TAM target for its Vera CPU signals that CPUs have been elevated to strategic assets within AI infrastructure — a shift now acknowledged across the industry.
5. UBS Raises Micron Target Price from $535 to $1,625 — LTA Redefines Memory from Cyclical to Structural Growth
• Core Source
“As long-term supply agreements (LTAs) become firmly established across the industry, we are raising our 2027–2029 estimates again. During this period, EPS is expected to comfortably exceed $100, and Micron (MU) is expected to generate over $400 billion in free cash flow (FCF) over the same period.”
“Our supply chain survey of long-term supply agreements (LTAs) across the memory industry suggests that up to 30% of total industry DDR volume is likely to soon be locked in at prices slightly below current levels.”
“We are raising our 2027/2028/2029 EPS estimates to $155/$167/$117, respectively, from $133/$122/$77 previously.”
“In conclusion, we are raising our target price from $535 to $1,625. This is based on applying approximately 15x NTM P/E to our 2029E EPS of $117, discounted back one year, replacing our prior SoTP methodology.”
“This means that even if DDR prices in the variable-price segment fall by approximately 50%, Micron can still achieve EPS of over $100.”
• Expected Impact
The significance of this UBS report lies less in the magnitude of the target price increase and more in the logic behind the change in valuation methodology. UBS has shifted from its prior Sum-of-the-Parts (SoTP) approach to directly applying a 15x P/E multiple to 2029 EPS estimates. This is a declaration that memory will no longer be treated as an extreme cyclical sector, but will instead be assigned the same multiples as general semiconductor companies.
The foundation of this argument is the structural proliferation of LTAs. UBS’s supply chain survey indicates that up to 30% of total industry DDR volume is expected to be locked into ‘enhanced LTAs’ spanning 3–5 years, and hyperscalers are estimated to have already secured 60–70% of their server DDR5 volume through LTAs. This creates a structure in which LTAs elevate memory’s earnings visibility to a level comparable to a SaaS model.
Crucially, UBS argues that even if DDR prices fall 50%, Micron can sustain EPS above $100. This means the volume locked in through LTAs acts as a structural floor, directly refuting the logic behind the traditionally low multiples applied to memory stocks — namely, cycle risk. UBS openly acknowledges that its $1,625 target is a significant outlier relative to the street consensus, while projecting that the multiple re-rating of the memory industry will spread across the sector. The forecast period for DRAM supply shortages has also been extended from Q4 2027 to beyond 2028.
6. Pony.ai 1Q26 Robotaxi Revenue Surges +395% YoY — Annual Target Raised to 3.5x, Commercialization Is Now Backed by Numbers
• Core Source
“According to Pony.ai’s 2026 Q1 results, quarterly revenue was 236 million yuan, up +145% YoY. Gross profit was 38.36 million yuan, up +140.1% YoY.”
“Robotaxi revenue reached 59.12 million yuan, a record quarterly high, surging +395.4% YoY and +28.7% QoQ. Passenger fare revenue also rose +456.5% YoY.”
“As of May 2026, the Robotaxi fleet size has exceeded 1,700 vehicles, and the company has raised its annual targets to reflect the strong growth momentum.”
“2026 annual Robotaxi revenue is targeted to grow more than 3.5x versus 2025, with fleet size to expand to over 3,500 vehicles, covering more than 20 cities domestically and internationally.”
“Driven by strong results and upward revision of growth targets, the company’s stock surged 9.3% in pre-market trading.”
• Expected Impact
Pony.ai’s 1Q26 results are significant in that they mark the quarter in which the growth curve of the robotaxi industry has, for the first time, begun to steepen meaningfully, as confirmed by concrete numbers. The figures of +395% YoY in robotaxi revenue and +456% YoY in passenger fares are not simply a reflection of a low base effect. The QoQ growth of +28.7% demonstrates that sequential growth momentum is being sustained.
Management’s decision to raise the annual guidance from 3x to 3.5x+ is also noteworthy. This was not a routine conservative upward revision, but a judgment made after directly observing the fleet’s actual demand absorption rate — already exceeding 1,700 vehicles in operation as of May 2026. The plan to more than double the fleet to over 3,500 vehicles by year-end was also announced.
From an investment perspective, the broader implication of these results is that the robotaxi narrative, centered on Waymo in the U.S., is now materializing in the Chinese market as well. At the same moment Waymo announced a temporary suspension of highway rides and a pause in Atlanta operations, Pony.ai signaled expansion to over 20 cities domestically and internationally. The competitive landscape in autonomous driving commercialization is simultaneously accelerating across both the U.S. and Chinese markets, and the fact that fleet size, fare revenue, and profitability improvement are all moving in a positive direction simultaneously is precisely what has drawn renewed market attention.
Comment [0]