Equity research
Equity research
IMHO - memory is at a place now where I believe you will see more bad news than good news.
1. Margin sustainability2. Impact on demand of consumer electronics3. Chinese competition We know margins don't have much more room to grow, maybe peaking around 88-92%. Memory makers are selling near 90% gross margin on DRAM and it is becoming a question of how much longer can they keep up before the curve changes direction. Think of it like Nvidia. Nvidia faces the same questions - margins and competition.Bernstein: Meta AI Compute
> Bernstein estimates Meta currently has a massive ~20GW global footprint with another ~14GW coming online in the next few years.> Concurrently, reports indicate Google has throttled Meta's compute capacity due to its own constraints, raising questions about whether Meta truly has "excess" capacity right now.> Meta holds $35.2B in CoreWeave contracts (over one-third of CRWV’s backlog). Combined with Microsoft's ~$14B in contracts, nearly half of CoreWeave's backlog comes from companies that will likely be direct competitors by the time these contracts are up for renewal.CoreWeave's Major Backlog Highlights> Meta $Meta Platforms(META.US): $35.2B total (signed late 2025 and expanded April 2026 through Dec 2032, including early NVIDIA Vera Rubin deployments).> Microsoft $Microsoft(MSFT.US): Estimated ~$14B across a series of relationships (effectively ~9 contracts).> OpenAI: $22.4B total across multiple expansions signed in 2025.> Jane Street: $6B multi-year infrastructure deal signed in April 2026, which included a $1B equity investment.> NVIDIA $NVIDIA(NVDA.US): A backstop agreement through 2032 worth up to $6.3B to purchase unsold CRWV capacity.Renesas Electronics: AI Infra & Compute
The Rise of AI Servers and Inference Compute> Total general server shipments are expected to grow modestly (~1.3x) from 2024 to 2030. In contrast, AI Server volumes are projected to scale dramatically, growing ~2.4x, effectively doubling in volume over the period.> While GPUs have driven the initial training wave, the market is seeing a major pivot toward AI ASICs and CPUs as the industry shifts heavily toward inference. AI ASIC shipments are projected to grow ~3.0x by 2030, vastly outstripping GPU growth (~1.5x) and standard CPU growth (~1.7x).Revenue Drivers and Renesas’ TAM Expansion> The primary structural tailwinds include the grid-to-rack transformation, an increase in system complexity requiring advanced MCU control, and a crucial architectural shift to 800V and vertical power architectures.> Growth is driven heavily by Digital Power, followed by Memory Interfaces, and other supporting MCU components.The Power Bottleneck: Grid-to-Core Architecture> Grid Level: Managing power via Solid State Transformers (SST), Energy Storage Systems (ESS), and Uninterruptible Power Supplies (UPS) using Digital Controllers and high-voltage GaN (Gallium Nitride).> Rack Level: Stepping down power using 800V DC/DC converters, hot-swap controllers, and Power Supply Units (PSUs).> xPU Board Level: Delivering ultra-low voltage and high current directly to the processor via memory interfaces, high-speed optical module PMICs, and complex Vcore stages (utilizing integrated voltage regulators to optimize space and efficiency).Exponential Power Demand Metrics> Next-gen AI racks are projected to consume >1 megawatt (MW) of power, driving a >10x increase in power content per rack.> The AI Infra & Compute Power TAM is scaling almost exponentially toward 2030, with Voltage Regulator Modules (VRM) making up the vast majority of this value, followed by Intermediate Bus Converters (IBC) and AC/DC stages.> A single example leading next-gen AI board showcases the sheer complexity involved, requiring a GPU power solution with >10 Digital controllers and >100 Smart power stages, alongside a 48V IBC solution utilizing >30 MOSFETs. Dense packing drastically raises thermal constraints, making thermal performance and current density the ultimate competitive battlegrounds for silicon providers.This is bullish for many power semi names. $STMicroelectronics NV(STM.US) $Infineon Technologies(IFNNY.US) $Texas Instruments(TXN.US) $ON Semiconductor(ON.US) $Wolfspeed(WOLF.US)Rosenblatt note on Meta $Meta Platforms(META.US):
"Bloomberg reported today that Meta was planning a cloud business to sell AI computing power. The story noted that the company is considering selling access to various AI models hosted on its existing AI infrastructure, as well as "raw" computing capacity, as part of its Meta Compute initiative. During their annual meeting (May 27), Meta CEO Mark Zuckerberg told shareholders leasing cloud computing business was possible if they overbuild their data center capacity, which at the time he indicated they had not. Additionally, questioning on Meta's 1Q results call was focused on the ROIC of their CAPEX, so we are not surprised by today's messaging from Bloomberg's "people familiar with the matter." Our checks show no change in hyperscale posture toward GPU compute procurement, with GPU shortages the norm right now across the industry. We also do not think that Meta has the right to resell any of the capacity that they have leased from CoreWeave through 2032 to third parties."No change in GPU compute procurement is bullish for $NVIDIA(NVDA.US) $AMD(AMD.US)Morgan Stanley: Memory Pricing
> 📈 3Q26 DRAM Outlook: TrendForce forecasts a 13–18% QoQ conventional DRAM price hike, perfectly matching GS estimates (+16% for Samsung).> 💻 PC & Server Strength: PC DRAM expected up 15–20% QoQ as OEMs aggressively build inventory. Server DRAM tracking up 13–18% QoQ, fueled by a widening DDR5 premium (now 22% over DDR4).> 📱 Mobile DRAM Cooling: Expected up 8–13% QoQ. A noticeable cooldown from massive 2Q spikes (73–85%) as suppliers ease hikes to support consumer electronics supply chains.> ⚠️ NAND Falling Short: TrendForce projects a 10–15% QoQ increase—missing Goldman's +21% estimate. While Enterprise SSDs hold strong (+18–23%), consumer NAND (eMMC/UFS) is soft at +5–10%.> 📊 Spot vs. Contract: Spot market continues to lead. DDR5 16Gb spot carries a 17% premium over contract, while older DDR4 8Gb spot sits at a steep 38% premium.Meta $Meta Platforms(META.US) is building a cloud business to sell excess AI capacity. What do I think of this?
I think this was coming. For months, I thought this was the only hyperscaler with an unclear AI strategy. You bought up all this compute for...what?But, I feel this move came under pressure from how their share price has been doing this year more than anything. I believe this is the right move for the business. They can turn around and become a neocloud.Wells Fargo has them generating $20B revenue from the resale opportunity. What does this mean for semis? This can get ugly fast. I believe Meta will still hit their CapEx target for FY26, but FY27 may paint a completely different picture. A drop in CapEx leads to an immediate selloff in the entire semi space. A market opportunity that becomes lost. You can check out my article on Meta from May in the replies 👇Intel $Intel(INTC.US) EMIB Supply Chain
Flip-Chip AssemblyBumpingPowertech Technology $6239.TW Amkor Technology $Amkor Tech(AMKR.US) Die BondASMPT $0522.HK Kulicke & Soffa $Kulicke and Soffa(KLIC.US) Laser MarkingE&R Engineering $8027.TWOPlasma CleaningE&R Engineering $8027.TWOEMIB SubstrateIC SubstrateIbiden $4062.T Unimicron $3037.TWO AT&S $ATS(ATS.US).VI Shinko ABF Film LaminationAjinomoto $2802.T Eternal Precision Mechanics $7795.TWOBridge Die BondToray $3402.TElectroplatingASMPT NEXXLaser via DrillingMitsubishi Electric $6503.TBaking Oven Group Up $6664.TWOOther ComponentsSilicon Capacitor AP Memory $6531.TW Samsung Electro-Mechanics $009150.KSSilicon Capacitor Foundry Powerchip $6770.TW United Microelectronics $United Microelectronics(UMC.US) Winbond $2344.TWInformation derived from Nomura Securities, but I included couple names in the supply chain I believe they missed - AT&S, AjinomotoCICC: SiC to the left, GaN to the right—third-generation semiconductors emerge as the inevitable solution for high-voltage architectures in data centers
> Our estimates suggest that by 2030, a single MW of data center deployment could require approximately 10,000 SiC devices and 21,000 GaN devices, corresponding to per-MW values of $220,000 and $49,000, respectively—indicating considerable market opportunities.> Between 2026 and 2030, rack-side (white zone) and server-room-side (gray zone) systems are expected to be progressively upgraded to 800V/±400V DC configurations, thereby driving demand for third-generation compound semiconductors. In the short term (1–2 years), transitional architectures led by 800V sidecar units may offer limited upside for SiC demand. However, as the following developments materialize: (1) blade-level voltage step-down or even high-density 800V-to-6V power conversion at the rack side, (2) centralized rectification at the server room side, and (3) implementation of solid-state transformer (SST) solutions, we believe the long-term optionality of SiC/GaN-related companies warrants proper pricing.> We believe Chinese companies have already established deep positions in the SiC/GaN space, with continuously strengthening competitiveness. Going forward, Chinese firms across the third-generation compound semiconductor value chain stand to benefit substantially from the adoption of high-voltage data center architectures.> Legacy data center racks average around 7 kW. In comparison, NVIDIA’s Hopper architecture consumes ~40 kW, Blackwell GB300 NVL72 reaches 134–140 kW, Rubin will exceed 200 kW, and future architectures (Kyber in 2027 and Feynman in 2028) are projected to scale to 600 kW and over 1 MW per rack.The industry is projected to move through four transitional stages to implement 800V DC setups:Phase 1 (2026/2027): White Space Retrofit with "Sidecar"Setup: Gray space (facility level) remains untouched. A custom 800V DC side-mounted power cabinet ("Sidecar") is added next to the IT rack to rectify AC to 800V DC locally.Semiconductor Impact: SiC demand is concentrated in the sidecar's front-end rectification/PFC modules.SiC Volume: ~1,594 units per MW.Phase 2 (2027/2028): Native 800V DC Computing (The GaN Inflection Point)Setup: Centralized low-voltage UPS systems are phased out in favor of distributed rack-level battery backup units (BBUs) and supercapacitors. The 800V DC bus connects directly to the compute blade.Semiconductor Impact: GaN replaces a massive portion of SiC for the main onboard power step-down stage (Intermediate Bus Converters) to meet strict space and thermal restrictions near the GPU. However, SiC finds rigid demand in high-voltage hot-swap protection and Solid-State Circuit Breakers (SSCBs).Volume: ~1,755 SiC units per MW; ~10,303 to 10,667 GaN units per MW.Phase 3 (2028/2029): Centralized Rectifiers in the Gray SpaceSetup: Power rectification moves entirely upstream into the gray space. Massive facility-level centralized rectifiers convert grid power to an 800V DC injection backbone distributed throughout the facility.Navitas GaN Transformation (June 2026): Driven by Navitas' GaNFast integration into NVIDIA’s MGX ecosystem, setups will begin converting 800V directly to 6V, entirely eliminating the traditional 48V intermediate step-down bus.Volume: SiC surges to ~6,948 units per MW (driven by massive industrial DC distribution, BBUs, and energy storage systems). GaN usage explodes to 20,800–21,600 units per MW.Phase 4 (Post-2029): The Final Solid-State Transformer (SST) ArchitectureSetup: Both the low-voltage transformer and low-voltage rectification stages are completely eliminated. Megawatt-scale Solid-State Transformers (SSTs) directly convert medium-voltage grid AC to 800V DC.Semiconductor Impact: High-frequency SST operation relies heavily on high-voltage SiC. The gray zone infrastructure captures 54% of the total SiC value chain.SiC Volume: Skyrockets to ~9,886 units per MW.Market Value> As setups transition from Phase 1 to Phase 4, the hardware value of SiC per MW jumps drastically from $10,000–$25,000 to approximately $270,000. This value growth is primarily driven by voltage rating upgrades (moving from 650V to 1200V and 3.3kV devices).> GaN addressable market value grows from $33k–$33.8k/MW in Phase 2 to $46.5k–$49.1k/MW in Phases 3/4.Silicon Carbide (SiC) Unit CostsThe unit cost for SiC scales aggressively as devices move from standard rack components to high-voltage, high-density infrastructure:Non-SST / In-Rack Devices: $2.5 – $3.5 per unit.UPS / PDU / BESS Peripherals: $3.5 – $5.0 per unit.800V Transitional / Early SST Stages: $10.0 base price per unit (utilizing mass-production 1200V-class SiC MOSFETs).Full SST Architectures:1200V SiC Modules: $15.0 – $20.0 base price per unit.3.3kV SiC Modules: $50.0 base price per unit (reflecting the severe technical premium for medium-voltage direct conversion).Gallium Nitride (GaN) Unit CostsGaN pricing is stratified by application scenario, specifically focusing on its proximity to the high-voltage bus and the GPU itself:Phase 2 (On-Blade Intermediate Bus Converters / IBC):High-Voltage IBC (800V --> 50V): $3.8 per unit(utilizing 650V integrated GaN ICs).Low-Voltage IBC (50V --> 12V): $2.2 per unit (utilizing100V-class GaN chiplets).Phases 3/4 (Centralized & GPU-Proximate): Navitas PDB Board System: $3.2 per unit.GPU-Proximate Embedded GaN: $2.2 per unit (scaled AI server order pricing).The SiC Step-Function ExplosionSiC exhibits massive value elasticity, meaning the financial value grows significantly faster than the physical unit count due to the transition to higher-priced, higher-voltage (1200kV -->3.3kV) devices.Volume Growth: Scalings jump from 1,594 units/MW (Phase 1) to 9,886 units/MW (Phase 4), representing a strong ~83.73% CAGR.Value Growth: Financial capture skyrockets from $2,000 – $15,000/MW (Phase 1) to $220,000/MW (Phase 4). This represents a staggering value CAGR of 144.78% to 379.14%.The GaN High-Volume, Steady-Value CurveConversely, GaN follows a high-volume, highly localized deployment strategy. Its volume scales rapidly, but because it operates at lower, more commoditized voltage classes closer to the compute blade, its dollar-value growth is more linear.Volume Growth: Unit counts double from 10,303 – 10,667 units/MW (Phase 2) to 20,800 – 21,600 units/MW (Phases 3/4), a CAGR of 95% – 110%.Value Growth: Financial capture shifts from $33,000 – $33,800/MW to $46,560 – $49,120/MW, representing a more modest value CAGR of 38% – 49%.iM Securities: Is TSMC Bottlenecking Nvidia's Short-Term Growth?
> CoWoS Wafer Revisions: Due to slower-than-expected capacity expansion by TSMC $Taiwan Semiconductor(TSM.US) and other CoWoS suppliers, iM Securities lowered its CY26 global CoWoS allocation forecast for AI accelerators from 1,380K to 1,096K wafers. Consequently, Nvidia's projected AI GPU production for the year has been cut from 11.14 million units (YoY +57%) to 9.24 million units (YoY +31%).> Rubin Delays: Most notably, production projections for the next-generation Rubin GPU have been slashed in half, from 3 million units down to 1.5 million units.> HBM Market Impact: Lower accelerator production drops total CY26 HBM demand from 4.89 billion GB to 4.23 billion GB. Memory manufacturers have adjusted their production down slightly to 4.33 billion GB, factoring in lower-than-expected demand and lower margins compared to conventional DRAM.> Nvidia’s upcoming next-generation AI accelerator, the Rubin Ultra (slated for CY27), is facing technical hurdles that may force a significant specification downgrade. While still under negotiation between Nvidia and memory vendors, a 384GB scaled-down version is being considered instead of the original 1TB (1,024GB) target.> CoWoS-L Size Limitations: The physical limit of the CoWoS-L interposer is 8,150 mm^2. The original 4-die plan requires 6,750mm^2. While mathematically possible, scaling the interposer up dramatically increases substrate warpage, concentrates stress on the corners, degrades bump fatigue life, and tanks packaging yields. Returning to a 2-die architecture automatically drops the maximum HBM cubes from 16 to 8.> TSMC's CoPoS Alternative is Too Late: TSMC’s next-gen solution to bypass this size limit is CoPoS (Chip on Panel on Substrate). However, because TSMC is only just beginning to select equipment and component vendors, mass production is not expected until the second half of 2028 (2H28). This leaves a packaging bottleneck that will stress Nvidia's growth through next year.> HBM4E Stacking Yields: Memory manufacturers are also struggling with the production yields of stacking 16-layer HBM4E, though this is flagged as a memory vendor issue rather than a TSMC-inflicted constraint.> Total AI Accelerator chip volume for CY26 is projected to hit 17,865K (17.87M) units, reflecting a 51% YoY growth.> Nvidia $NVIDIA(NVDA.US) commands 56% of the total CoWoS capacity allocation (640K wafers), yielding 9,242K chips (+31% YoY).> Broadcom $Broadcom(AVGO.US) captures 272K CoWoS wafers (a massive 206% increase), resulting in 5,354K chips (+70% YoY). This is heavily anchored by Google's TPU, which takes 75% of Broadcom's share (4,027K chips). Meta's MTIA accounts for 1,017K chips.> AMD $AMD(AMD.US) AI GPUs take up 70K wafers, translating to 875K chips (+22% YoY), split between MI350X (560K) and MI400X (315K).> Global HBM demand for CY26 is expected to reach 4,234 million GB (4.23B GB), marking an explosive 95% YoY growth.> Nvidia alone consumes 2,427M GB (roughly 57% of total market demand).> Broadcom represents 1,162M GB of HBM demand (+115% YoY), dominated heavily by the TPU at 851M GB.> AMD accounts for 282M GB of demand, with the upcoming MI400X utilizing ultra-dense HBM4 configurations (384GB per accelerator using 48GB cubes).Tesla accelerates mass production, Chinese manufacturers race ahead collectively—is the 'EV moment' for humanoid robots arriving?
According to Nomura, Tesla is transitioning its Optimus Gen 3 from initial assembly into full-scale production, establishing the baseline for global manufacturing velocity. The company has adjusted its annualized capacity target for the Fremont facility (repurposed from the Model S/X production lines) upward to roughly 70,000 units. Furthermore, Tesla intends to build an additional 70,000 units of capacity in Austin by 2028, aiming for a long-term total capacity of 1.5 million units. Supply chain data indicates that 2026 Optimus shipments will reach approximately 25,000 units (+/- 10,000), with weekly production goals in September potentially rising to 1,000 units. UBS Group identifies the official rollout window for the Optimus Gen 3 between July and August, marking it as a pivotal market driver for the latter half of the year.Outside of China, manufacturers continue to trail domestic production speeds. Figure AI and Boston Dynamics are both tracking annual shipments of roughly 500 to 1,000 units, while other international builders are averaging 100 to 200 units. Notably, Figure’s BotQ line has publicly achieved a manufacturing takt time of one unit per hour.Nomura has upwardly revised its 2026 shipment forecast for Chinese humanoid robots to 40,000–50,000 units. This growth is propelled by intensified government procurement for embodied AI bases and a late-year consumer demand inflection point triggered by the introduction of cheaper models.Market share is distributed across distinct tiers:Tier 1 (Top 2 firms): Each shipped roughly 10,000 to 15,000 units (representing a 2x to 3x year-over-year increase).Tier 2: Several mid-tier companies shipped approximately 3,000 units each.Tier 3: Smaller players shipped between 500 and 1,000 units each.Downstream application allocations for 2026 shipments are structured as follows:Consumer Applications: 30%Performance & Entertainment: 30%Government Procurement (Data Collection): 20%Education: 15%Commercial & Industrial Applications: 3%–5%UBS Group notes that state policy is actively shifting the Chinese robotics industry from experimental showcases to practical operational deployment across industrial, logistics, healthcare, and residential environments. The Ministry of Industry and Information Technology (MIIT) intends to deploy more than 10,000 units across over 100 scenarios by late 2027, while the Shanghai municipal government targets 100,000 factory-deployed units by 2030.Investment Trends and Market DragStrategic capital activity in the global robotics sector intensified during the first half of 2026. Despite this momentum, overall market sentiment faces clear headwinds. UBS data reveals that China’s Humanoid Robotics Index has lagged behind the broader machinery index by roughly 9 percentage points year-to-date. This underperformance stems from delays in Optimus Gen 3 mass production, weak trading momentum, and capital shifting toward data centers and commercial aerospace.Upcoming CatalystsUBS highlights several milestones in the second half of the year capable of shifting market momentum:> The official mid-summer launch of Optimus Gen 3 (July–August).> Progress regarding Unitree’s IPO.> The World Artificial Intelligence Conference in Shanghai (July).> The World Robot Conference in Beijing (August).Nomura highlights a major paradigm shift in how systems are trained: data acquisition is migrating toward "robot-free" methodologies, such as teleoperation, Universal Manipulation Interface (UMI), and first-person perspective (Ego) collection. These alternative techniques cost only 20% of physical robot-based methods and operate at a much faster pace.As a hybrid training model consisting of 90% non-robotic data and 10% physical robot data becomes the industry norm, the demand for physical hardware dedicated to data collection will contract through 2026. Government entities remain the primary backers of these large-scale data collection hubs; establishing a single facility with a 1,000-unit capacity demands an investment of RMB 50–100 million and carries a 3-to-5-year payback period. While this reduces immediate hardware procurement pressure for OEMs, it elevates the demand for cross-platform standardization, data consistency, and high quality. Actual industrial integration still faces verification friction regarding cycle times, precision, and training costs.Contract Manufacturing vs. In-House Supply ChainsField research indicates a clear trend toward full-system contract manufacturing. Because hardware supply chains require heavy capital and feature long payback horizons, the majority of second- and third-tier OEMs are opting to outsource production. Over the long term, contract manufacturing's market share is expected to expand. Conversely, OEMs maintaining vertically integrated supply chains (such as Unitree Robotics) gain a direct pricing advantage by lowering their bill-of-materials costs.Entrants coming from automotive supply chain backgrounds hold a long-term competitive edge due to their large-scale assembly experience and end-to-end management capabilities. Meanwhile, 3C electronics manufacturers currently hold a minor advantage in brand equity.Nomura projects that mainstream, full-sized humanoid models will retail between RMB 150,000 and 300,000, while compact models will cost between RMB 10,000 and 100,000. Though prices have fallen by over 50% year-over-year in 2026, the rate of price deflation is projected to slow down in 2027.Global Ecosystem ExpansionsStrategic moves from major technology and automotive players are accelerating across the industry:NVIDIA & OpenAI: NVIDIA $NVIDIA(NVDA.US) is actively funding an ecosystem that includes Chinese startup Unitree and South Korean enterprises like Hyundai. Concurrently, OpenAI renewed its focus on the sector in June by forming an "OpenAI Robotics" unit, initiating aggressive recruitment for AI, systems, and hardware engineering talent.Corporate Financing & SPACs: Germany’s Neura Robotics secured a Series C funding round of up to USD 1.4 billion, aiming to manufacture 6,000 units this year and top 10,000 units annually by 2027. Agility Robotics announced its intention to go public via a merger with Churchill Capital Corp XI, establishing a pre-transaction valuation of USD 2.5 billion, with closure targeted for Q4 2026.Automotive Integration: In June 2026, XPeng’s CEO assumed direct leadership of the company’s robotics branch, aiming to mass-produce its "IRON" humanoid robot by the end of the year. BYD also confirmed its own humanoid development program, planning an internal factory deployment of up to 20,000 units by year-end—marking one of the largest operational rollouts globally.Physical AI is interesting, and will be exciting to watch but don't expect the same bottlenecks in physical AI as you see in today's AI.
When you spend $700B in a single year, there will be bottlenecks because this rate of spending was not expected 3 years ago. It's the largest investment in mankind history, that could exceed the US interstate highway and railroads combined by next year. Hard to imagine humanoids spend will be even 10% of total CapEx spend for this year by 2030 domestically. Don't try bottleneck hunting in this area. It won't be as successful.