[E745] Fund explicitly avoiding enablers. By incorporating intangible assets, models rotate out of overvalued stocks while maintaining innovation exposure. Screen for quantitative AI ROI mentions on earnings calls rather than vague thematic usage.
[E444] Bill Gates' TerraPower gets approval to build new nuclear reactor. (Shared TechCrunch article)
[E338] Become absolute power user - prospect of having agents doing stuff while sleeping is dream come true. Always wished for superpower to multiply oneself.
[E337] Built 5-layer framework for AI research with OpenClaw bot Etta. Asked it to research Privacy Concerns as Structural Barrier to Corporate AI Adoption - report was outstanding. Very exciting developments coming to Inner Loop.
[E336] Using AI extensively for research and trade idea generation - allows covering much more ground. Found incredible winners from AI research with properly engineered pipeline of comprehensive specific prompts.
[E325] Uncertain on delivery timeline for grid energy. However, easy to pipe Permian Basin gas directly into turbines. As oil runs out, more NG available than can get to coast for liquification. Datacenters could be: gas in, fibre in, AI out - no grid needed.
[E312] $600B+ 2026 spending (Amazon $200B) sparks overinvestment fears. Markets focus on short-term cash-flow returns while bulls emphasize long-term civilizational build-out like electricity/railroads.
[E309] Shared Citi reports on AI Supply Bottleneck and US AI Risk highlighting datacenter stress.
[E4704] Security is consolidating upward into control planes rather than uniformly declining. As AI and automation increase system complexity, threat velocity, and identity sprawl, demand rises for platforms that can observe, correlate, enforce, and respond across multiple layers. Narrow point products that solve isolated problems without owning the broader policy or telemetry layer face structural decline.
[E4714] The investment edge is no longer simply 'SaaS is expensive and needs to come down.' The edge is recognizing that broad multiple compression and structural demotion are not the same thing. Some companies were dragged down because the category got repriced even though their strategic role in the stack remains strong — these may already be through the valuation damage.
[E4642] Observability and infrastructure intelligence occupy a structurally strong position. As systems become more distributed, automated, and AI-enabled, enterprises need continuous awareness across infrastructure, applications, workloads, anomalies, and performance. AI itself adds complexity, strengthening rather than weakening observability demand. These products survive harsh budget scrutiny because ROI is operationally legible.
[E4705] Systems of record represent the most resilient software category. These platforms define operational truth inside the enterprise — customer data, employee data, financial records, identity information, security states, infrastructure logs, contracts. AI typically strengthens this class because intelligence layered into a system of record makes it more central. 'The moat is not size. The moat is authority.'
[E4641] Identity, permissions, and policy layers may become the most strategically important software category in the agentic stack. As agents touch customer data, trigger workflows, approve tasks, and operate across applications, the layer that decides whether the agent is allowed to act becomes critical infrastructure. SightBringer positions this as 'sovereign infrastructure' for the coming software regime.
[E4395] Taiwan's semiconductor sector faces existential threat from the Hormuz closure. With 40% of Taiwan's grid powered by LNG and only 11 days of emergency stockpile, Taiwan may be forced to ration power to its industrial sector. Helium force majeure from Airgas compounds semiconductor production risks. Gromen asks: 'What is the value of the Mag-7 or NDX if Taiwan is forced to ration LNG or helium? (MUCH lower.)'
[E4601] Philadelphia Semiconductor Index (SOX) is the next 'Gradually, then suddenly' candidate with negative momentum divergence since Feb 2026. Risk to 7000 (350 for SMH), then 6200 (320 for SMH). Technical position suggests semiconductor selloff ahead.
[E4602] Micron (MU) was 160% above 200-Day MA on Jan 30, now only 54% above. Gained 300% in 120 days previously. Resistance at 450, first support at 350, second support at 250. Memory semiconductor showing classic blow-off top pattern.
[E4540] The Philadelphia Semiconductor Index (SOX) is the next 'Gradually, then suddenly' candidate with negative momentum divergence since Feb 2026. Roque recommends selling here with risk to 7000 (350 for SMH) then 6200 (320 for SMH). Micron was +160% above 200-Day MA on Jan 30, now 54% above with $450 resistance and $350 first support, $250 second support.
[E4541] Micron was 160% above its 200-Day MA on Jan 30, now only 54% above. Resistance at $450, first support at $350, second support at $250. The stock rose 300% in 120 days but momentum is fading — emblematic of the AI capex cycle froth unwinding.
[E4513] 22V published research on AI Goods (MS22AIG) and AI Services (MS22AIS) baskets with thesis that companies touching the physical world AND leveraging AI are in stronger position than pure service companies leveraging AI. This supports the 'atoms over bits' positioning within AI investment theme.
[E4446] MS On Site Power basket (MSXXONPW) is the biggest AI basket outperformer YTD at +38% despite suffering this week (-4.16%). Tremendous uptrend since summer 2025 breakout has resulted in overbought readings across the board. Pullback/retracement expected next. MS AI Power basket (MSXXAIPW) also down this week (-3.51%) but still +21% YTD with strong price action since summer 2025 breakout.
[E4447] 22V published report refreshing AI Goods (MS22AIG Index) and AI Services (MS22AIS Index) baskets with thesis that companies touching the physical world AND leveraging AI are in stronger position than pure service companies leveraging AI. This supports the view that physical infrastructure plays benefit more than pure software plays in the AI buildout.
[E4732] Capex cycle accelerating despite credit concerns. Hyperscalers committed to trillion-dollar buildout. Bottleneck constraints (power, chips, real estate) limiting deployment speed but not demand. 2026 marks transition from infrastructure planning to embodied AI rollout.
[E4259] The 15th FYP contains a wish list of mega infrastructure projects including national computational power, high voltage transmission, and state data centers. AI is explicitly listed as a priority sector alongside semiconductors, quantum computing, and embodied intelligence. The plan highlights the scale of PRC tech ambition with unprecedented state resource allocation to R&D.
[E4276] China has massively outpaced US power generation capacity additions. China's annual net additions reached ~550GW in 2024-2025 versus minimal US growth. Total Chinese electricity generation capacity stock now exceeds 3,500GW versus ~1,250GW for US utility-scale. Wood heard estimates in Beijing that electricity costs will fall to one-third of current levels in ten years as solar scales further.
[E4277] Jefferies China tech team projects China AI capex rising from Rmb519bn ($72bn) in 2025 to Rmb1.67tn ($242bn at current FX) by 2030. Wood has 'much greater confidence' that China AI capex will continue to increase versus US, where 2026 may mark cycle peak. Chinese enterprise tech confidence has improved since DeepSeek moment over one year ago.
[E4087] Data Infrastructure (Snowflake, MongoDB, Rubrik) will benefit from AI as a 'force multiplier.' As organizations deploy AI copilots and agents, they will create and query more data, demand faster response times, and tolerate less downtime. This raises the value of the foundational data layer that determines whether AI can run reliably and securely in production.
[E4082] Microsoft is positioned as best positioned in GS coverage to benefit from compounding AI product cycles, beginning with leadership in AI compute and extending to Copilot and agent orchestration at application and platform layers. Microsoft's domain experience graph showing how enterprise knowledge workers collaborate represents a powerful moat from years of observing workplace behavior.
[E4004] CapEx is backed by orders — hyperscalers surpassed $1 trillion in backlog orders, growing faster than CapEx in nominal terms. This is demand-led investment, not spending into a void. Energy is the binding constraint, not over-investment. IEA projects global data center electricity consumption to double by 2030.
[E4030] Energy is to the AI age what land was to agricultural age or capital to industrial age — the scarce resource around which everything orbits. The core metric is intelligence output per unit energy. Energy infrastructure is the foundational play: solar, fuel cells, cooling systems, grid build-out, and reshoring of critical supply chains.
[E4029] Energy is the binding constraint for AI, not over-investment. Elon Musk has warned AI buildout will run out of electricity. Microsoft has $80B in Azure orders unfulfillable due to power constraints. Meta building Prometheus, first gigawatt-scale data center with Manhattan-sized land footprint. US/Europe grid interconnection queues are years long.
[E4003] Hyperscalers deploying ~$600B in planned CapEx in 2026, up dramatically from 2025. Amazon guiding $200B likely posting negative FCF. Hyperscaler spending exceeds 1.3% of US GDP — higher than Manhattan Project (0.4%) or Apollo (0.7%). Microsoft has $80B in Azure orders it cannot fulfill due to power constraints.
[E5297] And this is the thing I want to make sure you guys realize because this gets into the risk with the hyperscalers that I've been debating people on all week.
[E5577] Data center buildout facing 125% surge in opposition/delays. Government moratorium and environmental challenges slowing capex. Oracle facing lawsuit for massive capex without equivalent revenue growth. Capex is still front-running revenue realization.
[E5182] SaaS disruption requires massive infrastructure capex for agent training and operation. Hyperscalers building out compute, power, and interconnect to support agent economies. SaaS pain = hyperscaler gain.
[E3717] Goldman forecasts AI-driven electricity demand will directly boost headline PCE inflation by 0.1pp in 2026, 0.07pp in 2027, and 0.05pp in 2028. Consumer electricity inflation expected to remain ~6% in 2026-2027 before decelerating to ~3.5% in 2028 on lower natural gas prices. Under higher passthrough scenario (50% vs baseline 33%), electricity inflation would hit 8% in 2026-2027.
[E3713] Goldman expects data centers to account for ~40% of total US power demand growth over the next five years, with data center share rising from 7% currently. Power demand growth projected at 2.6% annualized 2026-2030, with data centers contributing 1.2pp annually. This creates acute supply-demand imbalances in Midwest, California, Texas, and Mid-Atlantic regions where wholesale prices will spike.
[E3714] Power supply growth is structurally constrained by regulatory bottlenecks, equipment shortages, and labor scarcity. Median time for grid connection approval has risen to 4.5 years from under 2 years in 2000. Incremental natural gas capacity beyond scheduled additions cannot come online before 2030. Gas turbine and qualified labor shortages compound delays.
[E3730] Utility capex has become the most important driver of electricity inflation, with capex as a share of revenues increasing significantly since the 2010s due to grid upgrades, renewable energy transition, and retiring coal plant replacements. Capex costs flow through to consumer prices as utilities earn pre-set returns on invested assets.
[E3732] Computers, communication equipment, and medical instruments face the largest capex price pressures from higher electricity costs. These AI-adjacent sectors will see production cost increases that flow through the supply chain, contributing to equipment price inflation.
[E3715] Regional power market tightening will be severe due to extreme geographic concentration of data centers — about 1% of US counties account for ~70% of data center capacity. Midwest (MISO), California (CAISO), Texas (ERCOT), and Mid-Atlantic (PJM) face the largest wholesale price increases, with some regions seeing 15-25% price spikes based on supply-demand imbalance projections.
[E3716] Utility capex is expected to rise ~7% annually through 2029 as companies expand supply to meet surging power demand. PJM capacity auctions in late 2025 cleared at prices 2-3x higher than pre-pandemic levels, signaling significant investment requirements. Capex has already grown from ~25% to ~40% of utility revenues since early 2010s.
[E3718] Higher power prices will boost core PCE inflation by an additional 0.1pp in both 2026 and 2027, and 0.05pp in 2028, as businesses pass through higher electricity costs to consumers. Medical services, transportation services, and food services account for roughly half of the estimated core PCE boost. Total headline PCE impact is ~0.2pp in 2026 and ~0.15pp in 2027.
[E3719] Higher electricity prices will exert a ~0.2pp drag on consumer spending growth in 2026-2027 by lowering real disposable income. Lower-income households face disproportionate impact because electricity accounts for a greater share of their spending. Regions with high data center concentration (Virginia, Midwest, California) will see larger income drags than national average.
[E3723] Data centers employ very few workers relative to electricity consumption — Virginia data centers employ <5 workers per million kWh versus 50+ for US private sector average. Labor share of gross output for data center operations is ~5% versus ~32% for private sector. This means labor income from data centers will not offset higher electricity costs in concentrated regions.
[E3724] Policy responses are attempting to shift data center infrastructure costs away from residential consumers. Recent state policies (Oregon, Texas, Georgia, Kentucky, Virginia) require data centers >25-150MW to pay for transmission/distribution costs directly. However, Goldman expects non-AI customers to still bear ~33% of excess capex costs due to implementation gaps, loopholes, and market dynamics.
[E3726] Power outages represent a significant tail risk that could dramatically increase output losses. Power outages cost the US economy ~$8.4 per kWh lost — nearly 50x the average residential electricity price. If tight power markets increase outage duration or frequency, GDP losses would grow substantially beyond the baseline 0.1pp drag estimate.
[E3728] Capex pricing will also boost core inflation through equipment price increases. Computers, communication equipment, and medical instruments are the largest contributors to capex price passthrough. This adds to the direct and indirect consumer electricity cost passthrough, creating multiple channels for AI energy demand to flow into broader price levels.
[E3637] Nitto Boseki (3110 JP) T-Glass bottleneck confirmed by WSJ Feb 9 2026. Company has 75% YTD gain as capacity is essentially entirely booked by Apple and Nvidia vendors, leaving massive supply gap for rest of industry. Nikkei reports supply issues won't be alleviated until 2027 when Nan Ya Plastics partnership could produce ~20% of Nittobo's output.
[E3639] Asahi Kasei (3407 JP) has two emerging chokepoints: PIMEL photosensitive polyimide (PSPI) used in advanced semiconductor packaging (every HBM stack and AI accelerator needs it), and Q Glass quartz fabric for next-gen copper clad laminates. In May 2025, semiconductor industry panicked over rumored PIMEL supply cuts. Asahi invested ¥16B to double capacity by end of year.
[E3640] Photomask blanks are potential chokepoint: EUV blanks require ~40-50 alternating layers of molybdenum and silicon with sub-angstrom precision, zero defect tolerance. Market extremely concentrated between Hoya (7741 JP) and AGC (5201 JP). Extreme AI chip demand has stressed both leading and lagging edge processes. AGC trades at 15x forward earnings and <1x book value.
[E3641] Transition from tungsten to molybdenum in 3D NAND word lines (happening at 200+ layers) is one of most significant materials transitions in memory in a decade — tungsten doesn't work above 200 layers. Switching the metal requires entirely new CMP slurry chemistry. Entegris and Merck are the leaders providing both deposition precursors and CMP slurries.
[E3616] Nuclear fuel cycle faces structural constraints. SOLS controls ~20% of global conversion capacity with Cameco, Orano, and Rosatom as only competitors. Russian supply increasingly sanctioned. Spot $64/kgU doesn't incorporate 60+ reactors under construction, SMR pipeline, or AI-driven nuclear PPAs from Microsoft, Amazon, and Google. US target of 400 GW by 2050 represents 4x current installed base.
[E3626] Semiconductor materials bottlenecks spreading from obvious (GPUs, HBM packaging) to invisible (chemicals, filters). Entegris Q4 2025 confirmed liquid filtration for advanced nodes hit another record quarter. Molybdenum transition from tungsten in 3D NAND word lines (200+ layers) is most significant materials transition in memory in a decade. Over 60% of Entegris revenue now from advanced nodes.
[E3612] AI hyperscaler capex creates colliding demand across data centers, defense, and aerospace, constraining shared supply chains. Jet engines and natural gas turbines use identical input materials. Trump's $1.5T defense budget, NATO buildout, and aerospace backlogs compete with AI infrastructure for the same physical inputs — beryllium, titanium, tungsten, superalloys, and electrical steel.
[E3617] Nuclear demand is accelerating beyond current 440-reactor global fleet: 60+ reactors under construction, SMR pipeline, AI-driven nuclear PPAs from Microsoft/Amazon/Google, and US target of 400 GW by 2050 (4x current installed base). If conversion demand grows even 10-15% from new builds while Western supply stays capped at 35-37k MTU, spot prices could reach $80-100/kgU.
[E3615] Data centers compete directly with aluminum smelters for power contracts. AI data centers pay up to $115/MWh while aluminum smelters require $30-40/MWh contracts to operate profitably. This creates an odd situation where AI demand driver simultaneously constrains domestic aluminum production, benefiting existing producers with locked-in power contracts.
[E3622] T-Glass bottleneck confirmed by WSJ (Feb 9). Nitto Boseki controls 70-90% of high-end electronic glass fiber, up 75% YTD. Operating profit +26.4% YoY, profits surged 272.7%. Capacity booked by Apple and Nvidia vendors leaving 'supply gap' for rest of industry. Supply issues won't alleviate until 2027 Nan Ya partnership. Upstream materials catching up to downstream valuations as bottlenecks migrate up supply chain.
[E3613] Gas turbine demand from data centers is the fastest growing demand category for superalloys. Siemens Energy nearly doubled gas turbines sold in 2025. GE Vernova gas power equipment orders more than doubled in Q3 2025. Each heavy gas turbine contains hundreds of superalloy components requiring replacement every 1000-2000 days. ATI Inc energy segment growing as 'very profitable part of overall portfolio.'
[E3642] Entegris (ENTG) confirms commoditized semiconductor upstream seeing price increases at chokepoints. Over 60% of revenue now from advanced nodes. CEO confirmed liquid filtration for advanced nodes had another record Q4 2025 quarter. Advanced packaging wave driving outperformance in names like ASE Technology and Amkor, with higher margins resulting in upstream spending.
[E3638] Upstream AI materials suppliers significantly underperformed midstream/downstream over long-term lookback, creating catch-up opportunity as bottlenecks work up supply chain. Since December publication, upstream basket has outperformed midstream — exactly what expected as market realizes molecules and powders can become hard constraint. Resonac (4004 JP) up 40% YTD.
[E3631] Electronic chemicals at intersection of qualification moats and structural tailwinds. Arkema (AKE FP) Kynar PVDF franchise is dominant lithium-ion battery binder, targeting €1B battery sales and €100M+ data center sales by 2030. Olin (OLN US) only back-integrated US epoxy resin producer for PCB laminates, transformer coils. Orbia (ORBIA* MM) Dura-Line is leading HDPE conduit manufacturer for data center fiber/power infrastructure.
[E3552] ASMPT (522 HK) is key bonding-equipment supplier with deep SK Hynix relationship. SK Hynix placed order for 30+ TC bonders, targeting ~100 total for HBM4 with meaningful ASMPT portion. Samsung reportedly in discussions with ASMPT for TC bonders. ASMPT supplies both TC bonders and hybrid bonding tools to TSMC. CXMT could use ASMPT TC bonders for HBM production given Hong Kong/Singapore ties.
[E3363] Emerging markets supply approximately 70% of global semiconductor and AI hardware exports, positioning them as structural beneficiaries of the AI build-out. Leadership is expected to expand beyond EM Asia into Latin America and frontier tech economies as AI adoption spreads.
[E3505] HBM4E will introduce hybrid bonding, moving from thermal compression to direct copper-to-copper pad bonding. Samsung confirmed hybrid bonding adoption in HBM4E with yields reportedly highly mature. BESI is positioned as key beneficiary — potentially securing 50-60% of SK Hynix demand and 80%+ of Micron's hybrid bonding equipment needs.
[E3521] DDR4 is facing severe global shortages, trading at premium to DDR5. Samsung, SK Hynix, Micron, and CXMT are phasing out DDR4 as they pivot to DDR5/HBM. Memory buyers engaged in panic buying in Q1. Citrini expects DDR4 pricing to rise through Q2 and possibly Q3, with double-digit QoQ increases. DDR3 shortages are now emerging as spillover.
[E3441] Tesla's Optimus V3 unveiling Q1 2026, targeting 50,000 units by end of 2026, then 1 million, then 10 million, with 50-100 million units annually by 2028 at $20,000 each. Musk says Optimus could account for 80% of Tesla's value. Figure AI, 1X Technologies, Boston Dynamics, Agility Robotics, and Unitree ($16,000 humanoid) are all racing to deploy AGI-level intelligence in physical bodies.
[E3453] World models transform robotics training by enabling simulation at scale. A robot can experience a million scenarios simultaneously, each in its own simulation. Figure AI, 1X Technologies, Agility Robotics, and Boston Dynamics all building on Nvidia Cosmos platform. Nvidia Cosmos Transfer generates photorealistic factory environments; Cosmos Reason enables physics intuition for robots.
[E3459] The intelligence-led cycle has inverted traditional relationships: semiconductor sales now lead the cycle rather than following it because computing power drives everything downstream. Taiwan Semi sales lead Japan orders, which lead PMI. The next leg is second-order CapEx: power grids, data centers, cooling systems, power distribution. This is the 2026 story PMIs will respond to.
[E3520] UMC partnered with imec for silicon photonics process supporting 800G and 1.6T pluggable optics, with risk production planned for 2026-2027. The imec iSiPP300 process supports LiNbO₃ modulators enabling 400G+ per lane performance. UMC is trading as a commodity player but building specialty portfolio in power semiconductors, mixed-signal, and photonics.
[E3559] Silicon photonics addresses AI cluster bandwidth constraints as east-west data center traffic explodes. Scaling copper beyond 224Gb/s is practically impossible — higher speeds shrink electrical reach and require more retimers adding power and cost. PICs enable 400+ Gbps with light pulses through silicon waveguides. Co-packed optics may eventually integrate into GPU/ASIC architecture but thermal management challenges remain.
[E3554] PDF Solutions (PDFS) provides manufacturing intelligence for fabs via Exensio platform — normalizing data across FDC, test, assembly, packaging into single semantic model. Has upside to Intel 18A node and advanced packaging (EMIB/Foveros). Uses 'forward deployed engineers' model making contracts stickier. Not generic enterprise software so relatively isolated from SaaS selloff.
[E3551] Powertech (6239 TT) has DRAM testing capacity nearly fully utilized with NAND also high. Guided 2026 capex over NT$40B. Q1 2026 expected much better than Q1 2025 despite normal seasonality weakness. PiFO (CoWoS-L alternative) and FOPLP give Powertech a path to incremental packaging orders, with revenue opportunity visible around 2027. Meta reportedly showed interest in PTI PiFO.
[E3560] Data centers transitioning to 800 high-voltage DC architecture, boosting demand for power semiconductors just as TSMC and Samsung cut lagging edge production. Global Foundries licensed GaN technology (650V and 80V) from TSMC specifically for data center power including 800V HVDC. This creates dual tailwind of photonics ramp and 800V DC power conversion wave.
[E3501] Memory, storage, semicap, and advanced packaging sectors have performed exceptionally well, reinforcing the supply/demand dynamics described in Citrini's 'Post-Traumatic Supply Disorder' thesis. HBM content per GPU is rising fast — NVIDIA's Blackwell Ultra GB300 contains $4k of HBM3E. The ability to scale MEMS manufacturing and solve hybrid bonding physics will ultimately determine the pace of AI advancement.
[E3522] NOR flash supply is tightening with prices rising. TSMC and Samsung cutting 8-inch capacity contributes to projected 2026 200mm output decline. Macronix (2337 TT) reduced NOR production to increase MLC NAND output, with ~30% NOR price hike reported in Q1 2026. Samsung announced MLC end-of-life in March 2025 with final shipments June 2026, causing global MLC NAND capacity to drop ~41.7% YoY in 2026.
[E3524] Cohu (COHU) saw ~75% utilization in Q3 with accelerating AI data center demand. Management raised 2025 Neon HBM inspection tool revenue to $10-11M with first HBM4-configured shipment (likely to Micron). Eclipse handler selected for next-gen AI processor testing at leading US manufacturer (likely Intel) handling ~3,000W power dissipation.
[E3502] TSMC expects to spend $190-200B in capex over the next 3 years. Intel could ramp to 100k wafers per month of 18A-P + 14A capacity by 2029-30. Intel Foundry investments are front-loaded in 1H with greater equipment purchases, suggesting Intel is becoming more confident in securing foundry customers despite market disappointment with Q4 earnings.
[E3503] AI workloads require memory across the full stack. NVIDIA GPUs keep 'hottest' data in HBM while pushing 'colder' Key-Value cache to DRAM and SSDs. This structurally lifts demand for HBM, DRAM, and SSDs as video and 'world model' workloads grow. AI labs pushing toward 100M+ token contexts require far more working memory during inference.
[E3504] At 3nm and below, semiconductor wafers require 1,500+ individual process steps, each representing yield loss risk and subsystem supplier content opportunity. Each new process node increases fab cost by ~30%. Advanced packaging equipment growth extends beyond test/assembly to include gas, vacuum, and thermal systems enabling capacity expansion.
[E3509] Silicon photonics is emerging as the only viable path forward for AI data center bandwidth. Copper cabling hits a physics wall at 224Gb/s per lane. Silicon photonics enables 400+ Gbps transmission. The market for photonic components is expected to inflect within two years, with risk production planned for 2026-2027.
[E3506] Memory testing requires MEMS-based probe technology to contact thousands of microbumps (40µm pitch) without damage. FormFactor (FORM) and Micronics Japan (6871 JP) are near-exclusive suppliers of this critical probe card technology. The market is rewarding companies that validate chips and memory at extreme speeds and densities.
[E3507] Citrini created a 16-name Memory Testing basket covering capital equipment (automated test equipment, wafer probes, thermal controls) and consumables (probe cards, test slots). Key names include Advantest (6857 JP), Teradyne (TER), FormFactor (FORM), Cohu (COHU), Aehr (AEHR), and socket suppliers WinWay (6515 TT), LEENO (058470 KS), ISC (095340 KQ), Enplas (6961 JP).
[E3508] Subsystems are 'picks and shovels for the picks and shovels.' AMAT, KLAC, TEL, LRCX tools are assembled from hundreds of subcomponents — vacuums, fluid delivery, wafer handling, gas boxes. Citrini called out ICHR and UCTT as subsystem beneficiaries back in July 2025. Subsystem vendors are positioned to outperform the WFE market they serve.
[E3510] ASM International (ASM NA) has highest exposure to Atomic Layer Deposition (ALD), which becomes exponentially more important as the industry transitions from FinFET to GAAFET. More 3D surfaces and tighter tolerances make film uniformity critical. AI chips as most compute-dense and power-hungry will be first/largest volume using GAAFET processes.
[E3511] Global Foundries (GFS) is assembling a vertically integrated silicon photonics platform through its Singapore Advanced Micro Foundry acquisition, Cisco partnership, and R&D hub. GFS licensed GaN technology from TSMC for 800V HVDC data center power with Navitas and ON Semi as customers expected to begin manufacturing in 2026. This dual tailwind of photonics and power creates re-rating potential.
[E3315] Brazil is authorizing multi-gigawatt grid connections to single AI campuses at a scale that is 'increasingly unfeasible in OECD power markets.' These campuses function as programmable load centers absorbing excess renewable generation that would otherwise be curtailed. Sub-80ms latency to North America enables Brazilian compute to serve global AI markets without performance degradation.
[E3344] Visser introduces the concept of 'Green Compute arbitrage' — converting surplus renewable energy and mineral endowment into globally tradable AI workloads. Brazil's data center campuses function as programmable load centers absorbing excess solar and wind generation, effectively 'turning renewable electrons into exportable intelligence.' The country becomes a host economy for AI compute, not merely a supplier.
[E3330] Visser argues 'the future has finally run out of alternative venues' — in a world where AI is constrained by power, materials, and capital rather than algorithms, Brazil offers all three at scale and at the right point in the cycle. Data centers, transmission, rare-earth processing, and grid-scale storage are long-duration assets whose valuations are acutely sensitive to real discount rates now declining in Brazil.
[E3316] AI training is 'energy-inelastic' — large models go where power is abundant, reliable, and affordable rather than waiting for grid upgrades. Brazil operates an overwhelmingly renewable energy matrix (hydro, wind, expanding solar) with substantial surplus capacity on a curtailment-adjusted basis. This surplus would be stranded in other jurisdictions but can be monetized through AI compute infrastructure.
[E3314] Visser argues the US and Europe face binding physical constraints on AI buildout — multi-year interconnection queues, rising marginal electricity costs, and political friction around new infrastructure. This creates a structural opportunity for jurisdictions like Brazil with surplus renewable capacity. AI training is 'energy-inelastic' and will relocate to where power is abundant, reliable, and affordable rather than wait for grid upgrades.
[E3206] The three big software prints from Microsoft, ServiceNow, and SAP last week signaled growth is stable at best but likely still moderating, triggering massive software sector selloff. Growth rates must accelerate or estimates revise upward to bend the negative narrative; more partnerships between SaaS firms and OpenAI/Anthropic would help signal coexistence.
[E3263] US DOE launched Nuclear Lifecycle Innovation Campuses to expand domestic enrichment, conversion, and recycling capacity, directly addressing nuclear infrastructure constraints for AI power demand. This policy momentum supports the thesis that nuclear development is critical for meeting AI data center energy requirements.
[E3205] Semis/hardware companies are disclosing AI revenue impacts materially larger than software firms and growing at accelerated clips. Microsoft's Azure acceleration in Q3/March 2025 boosted the 'long-infra' trade. Infrastructure, data, and security software basket up 63% since July 2024 versus SaaS basket down 18%.
[E3244] Cantor Fitzgerald's uranium supply-demand model explicitly includes AI/Data Centre demand as a distinct demand category alongside reactor requirements. Their updated model projects a cumulative uranium supply-demand deficit of 1.19 billion lbs U3O8 over 2026-2040, with AI and data center electricity demand contributing to the structural deficit that supports higher uranium prices.
[E3245] Cantor Fitzgerald's uranium supply-demand model explicitly includes AI/Data Centre demand as a separate demand category alongside reactor requirements, reflecting the growing role of nuclear power in meeting AI infrastructure energy needs. Their 1.19 billion lb U3O8 cumulative deficit projection (2026-2040) incorporates this AI-driven power demand as a structural driver.
[E3108] AI represents the first example of Reed's Law (Metcalfe's Law squared) — an exponential network built on Internet infrastructure at the fastest adoption pace ever. AI will scale knowledge and human imagination infinitely, similar to how nuclear energy can scale energy infinitely. This productivity miracle is the only way out of the debt trap created by the Magic Formula.
[E3163] Hundreds of billions of dollars are flowing into AI infrastructure (chips, data centers, power, model training). This spending pulls resources, energy, and capital into one sector which can look inflationary in the short run. But once adopted at scale, AI productivity gains will be profoundly deflationary — Q3 2025 US nonfarm business productivity rose at 4.9% annualized rate while unit labor costs fell 1.9%, explicitly linked by economists to AI investment.
[E4765] Oracle and Microsoft reporting massive RPO (remaining performance obligations) backlogs that exceed capex capacity constraints. Bottlenecked data center construction forces hyperscalers to choose between rapid revenue recognition and sustainable capex deployment, creating earnings uncertainty.
[E3019] China's 15th Five-Year Plan (2026-2030) expected to prioritise technological leadership and self-sufficiency in AI and chip manufacturing. Taiwan holds dominant position in cutting-edge chip capacity with approximately 60%, while China leads in legacy chip capacity. This creates strategic vulnerability and investment implications for semiconductor supply chains.
[E3003] Demand for AI continues rising, enabling hyperscale cloud companies to invest an expected half-trillion dollars in data centres this year. Monthly token usage increased more than 100-fold in 18 months to October 2025 per Google. However, AI has most complex supply chain in history with bottlenecks in HBM components, energy supply pressure, and talent shortages for both programming and facility construction.
[E3008] AI agents powered by latest models now perform better than professionals with 14 years experience at various tasks. GPT-5.2 and Claude Opus 4.5 exceed 60% parity with industry experts. However, widening misalignment between demand and capacity, with more bottlenecks in High Bandwidth Memory, energy supplies, and talent shortages.
[E2676] Every notes the Genesis Mission aims to 'unleash a new age of AI-accelerated innovation' but AI firms have been told they 'are not allowed to pass on their soaring electricity demand to consumers via utility bills, and must pay their own way.' This positions AI capex as subject to state direction under the national security umbrella, with energy constraints being managed through policy intervention rather than market mechanisms.
[E2801] Humanoid and non-humanoid robots will drive massive demand for power, lithium, copper, and aluminium. Elon Musk's projection of 10bn humanoid plus 10bn non-humanoid robots by 2040 would require a doubling of global power output to recharge and 16x lithium supply, 3.5x aluminium, and 3x copper supply. China service robot output is up ~80% y/y.
[E2773] Industrial software is making rapid progress enabling testing of manufacturing processes and product designs through computer simulations dubbed 'digital twins.' UBS envisions a world of 'physical AI' with potential to transform traditional manufacturing, where robotics technology continues to improve alongside AI to move assembly lines from rigid production systems into adaptive, interconnected, responsive systems.
[E2853] UBS highlights EM equities feature leaders in key AI value-chain segments including chip and memory producers in Taiwan and South Korea. AI adoption driving earnings growth expectations of 20% for MSCI EM, highest among major indices. Index composition now more tech-heavy and less commodity-focused, justifying above-average valuations.
[E2767] Collaborative robots (cobots) now account for over 10% of total industrial robot installations and the share is expected to grow. International Federation of Robotics estimates show industrial robot installations projected to exceed 700,000 units annually by 2027, up from around 500,000 in 2021, indicating strong capex deployment in physical AI infrastructure.
[E2761] UBS argues AI and automation are becoming the backbone of modern manufacturing assembly lines, with industrial robot installations projected to continue rising significantly. The global stock of industrial robots has climbed to over 4 million units, with the US ranking third globally behind China and Japan. The automotive industry remains the top US industry for robot installations, though the US still has only 30 robots per 1,000 manufacturing jobs, indicating substantial room for further automation investment.
[E2800] Citi projects data center buildouts will add ~150ktpa of copper demand growth in both 2026 and 2027, with the sector growing at ~45% in 2026 and ~20% in 2027. Data center copper demand will level out at ~850ktpa toward 2030. Currently only 1.5% of global copper consumption, expected to reach 2.4% by 2027.
[E2775] UBS highlights that electric power usage will rise further as the auto industry continues to transform toward EVs and autonomous vehicles. The report references UBS's 'Power and resources' TRIO as capturing investment opportunities related to the tech-enabled modern assembly line, implying energy infrastructure is a key constraint and investment theme for manufacturing transformation.
[E2810] Aluminium smelter capacity growth is limited by China's capacity cap and power availability constraints, with data center competition for power creating structural supply constraints. Copper and aluminium are 'THE Energy Transition and AI trades within commodities' due to their unique combination of structural demand drivers and supply constraints.
[E2734] Gromen reframes 'AI bubble' risk: the danger is material constraints, not weak use cases. Current data-center construction may represent the last tranche completable under existing critical mineral supply. China dominates AI infrastructure inputs (rare earths, gallium, germanium) while US tries to simultaneously build AI infrastructure, electrical grids, and military capacity.
[E2774] UBS recommends their Artificial Intelligence and Power and resources TRIOs as capturing ample opportunities throughout the tech-enabled modern assembly line and manufacturing process. This represents an explicit portfolio positioning view linking AI infrastructure buildout to investment opportunities in both digital and physical layers.
[E2790] UBS argues that owning the application layer is every bit as important as the infrastructure layer in transformational technology investing. As demonstrated with electricity and the assembly line, one transformational technology can beget another — this lesson is instructive for seeking AI and automation investment opportunities beyond pure infrastructure plays.
[E2766] NVIDIA CEO Jensen Huang proclaimed in January 2026 that 'The ChatGPT moment for robotics is here,' signaling that physical AI and robotics are reaching an inflection point. UBS sees industrial software enabling 'digital twins' for cheaper and faster manufacturing process testing, with cobots now exceeding 10% of total industrial robot installations and expected to grow further as labor shortages intensify.
[E2645] Apollo's productivity decomposition shows Total Factor Productivity growth historically follows capital deepening phases. The chart spanning 1975-2024 demonstrates TFP gains lag capital intensity contributions, implying current AI capex cycle will eventually translate into broader productivity improvements — supporting the thesis that AI infrastructure investment is a necessary precursor to productivity gains.
[E2659] Census Bureau data shows Information sector leads AI adoption at approximately 18%, followed by Finance & Insurance and Professional Services at 12-14%. Construction and Utilities lag at under 5% adoption. This sector dispersion pattern suggests AI capex benefits flow disproportionately to tech-adjacent sectors in the near term.
[E2644] Professional, Scientific & Technical Services leads AI adoption at ~35%, followed by Information at ~28% and Finance & Insurance at ~22% per Census Bureau data. These knowledge-intensive sectors are the early adopters, while physical sectors like Construction, Agriculture, and Manufacturing lag at under 10% adoption.
[E2637] Hyperscaler capex as a share of GDP is significantly higher today than telecom capex during the dot-com bubble. Apollo compares current hyperscaler capital expenditure (Oracle, Microsoft, Meta, Amazon, Google) to telecom companies (Level 3, WorldCom, Global Crossing, Nortel, Verizon, AT&T, Nokia, Cisco, Williams, XO Communications) during 1996-2003, showing current AI infrastructure investment exceeds the prior cycle's peak intensity.
[E2636] Apollo's Slok argues we are in the early stages of a productivity boom analogous to the PC and internet adoption cycles. AI adoption is explicitly compared to these prior general-purpose technologies, suggesting sustained capex and infrastructure buildout ahead. The framing positions AI infrastructure investment as a multi-year secular trend, not a one-time capex spike.
[E2642] CFO survey data shows strong productivity effects from AI adoption: 67% of firms report increased labor productivity, 62% improved decision-making speed/accuracy, 81% increased customer satisfaction/retention, and 65% more time spent on high value-add tasks. This supports the thesis that AI adopters are seeing measurable business benefits.
[E2615] The surge in AI capex is contributing to real economy liquidity demands that may crowd out financial markets in 2026. Combined with pending tax cuts and 5.4% Q4 2025 GDP growth, the expanding real economy appetite means Fed liquidity injections may be insufficient to support both real and financial economies — creating headwinds for financial assets like Bitcoin.
[E2288] Hyperscalers are transitioning from asset-light to asset-heavy businesses, now investing directly in energy. Alphabet acquired a renewable energy developer for US$4.75bn in December — the first time a tech company has bought an energy developer in-house, making energy capex rather than opex. Alphabet can afford this with US$141bn cash net of debt.
[E2473] Morgan Stanley projects AI infrastructure spending will reach $2.8 trillion globally for new data center construction through 2028, with US data center developers facing a 10-20% power shortfall. Total data center power demand will grow ~130 GW by 2028, with US accounting for 55% (74 GW). AI Infrastructure stocks outperformed with 59% returns in 2025.
[E2340] Estimated $1.5 trillion of private sector capital expenditures on generative AI represents massive investment enabled by US economic wealth. Technology investment contributed ~0.5 percentage points to 2025 US GDP growth of ~2.1%. AI-related narrowly defined capex contributed ~0.1% to overall GDP. Technology accounts for ~25% of US GDP growth in 2025.
[E2282] Wood argues AI-related capex continues to dominate while non-AI investment contracts. Real AI-related capex (information processing equipment, software, data center construction) rose 15.3% YoY in 3Q25 and 25% over two years, while real non-residential private fixed investment excluding AI declined 2.8% YoY and was down 2.4% over two years. The capex divergence is stark and accelerating.
[E2492] MS views AI capex financing risks as manageable. KKR argues data centers different from 1990s fiber overbuild — construction requires customer contracts, idle capacity is expensive, and power constraints prevent unconstrained builds. 'Temporary overbuilds behave like rolling upgrades rather than stranded assets.'
[E2474] Compute demand is growing exponentially faster than supply. EpochAI projects 2.6x annual increase in compute for LLM training through 2030, with largest training runs drawing 4-16 GW by 2030. Token demand increased >2,200% from Nov 2024 to Nov 2025 per OpenRouter data. NVIDIA's projected CAGR of ~210% still falls short of demand growth.
[E2518] CoreWeave re-contracted older NVIDIA Hopper GPUs at 95% of original pricing, indicating scarcity. Google guaranteed a data center lease for Anthropic via Hut8 at ~18.5% unlevered yield, implying a ~300% power price premium. These transactions demonstrate AI players willing to pay significant premia for compute and power access.
[E2500] Google exec stated the company likely needs to double compute every 6 months, resulting in >1,000x in 4-5 years. METR data shows doubling in average 'work' duration per LLM query every 7 months. This demand growth significantly exceeds projected NVIDIA chip supply growth of ~210% CAGR.
[E2531] Coding is now the dominant token-weighted use case, exceeding 50% of token volume by late 2025 vs 11% in early 2025. Proprietary models dominate coding assistance. LLM pricing has been only a modest indicator of usage, indicating greater model performance capabilities for computationally-intensive models command premium regardless of cost.
[E2285] Memory makers hold unprecedented pricing power in the AI capex cycle. Hynix's next fab will cost an estimated US$80bn to construct, while Micron has broken ground on a US$100bn 'mega fab' in Upstate New York. Hynix is reportedly asking major customers like Nvidia to contribute to production costs to lock in supply, with negotiations underway.
[E2519] MS projects 2025-28 data center capex: $710B AI workload + $337B non-AI = $1.05T total through 2028. Cumulative AI spending 2026-28 alone is $1.49T. GPT-5.2 scored 71% on GDPVal tasks (December 2025) vs Grok 4's 24% (July 2025), showing rapid frontier LLM improvement trajectory.
[E2283] TSMC announced capex guidance of US$52-56bn for 2026, up from US$41bn in 2025 — a 27-37% increase. This confirms the AI infrastructure buildout is accelerating rather than peaking, despite Wood's thesis that investors will eventually question returns on AI capex.
[E2284] The four major hyperscalers' capex is projected to increase from US$360bn in 2025 to US$480bn in 2026, a 33% increase according to consensus estimates. Wood notes this divergence between hyperscaler underperformance and chipmaker outperformance 'cannot go on forever' but is being ignored while capex projections rise.
[E2571] Archer cites data center construction for AI and cloud computing as a key driver of copper demand. Data centers use enormous amounts of copper in their electrical systems, and construction is growing exponentially. This positions copper as a critical input to the AI infrastructure buildout, linking the commodity thesis to the AI capex cycle.
[E2511] GPT-5.2 scored 71% on OpenAI's GDPval (measuring model performance on economically valuable tasks vs human experts), up from Grok 4's 24% in July 2025. Anthropic's model scored 48%. Frontier models complete GDPval tasks ~100x faster and ~100x cheaper than industry experts.
[E2371] UBS notes that AI-related global capex expansion will persist, further supporting commodity demand—particularly for industrial metals and critical minerals. Corporate profit upswings historically fuel increased capital expenditures, with G7 economies forecast to strengthen. Copper should specifically benefit from increasing global demand in areas such as electric vehicles, renewables, and data centers.
[E2481] Morgan Stanley estimates nuclear renaissance worth $2.2 trillion through 2050 (up from $1.5T projected last year), adding 586.5 GW globally (~47% above current 398 GW capacity). China leads with ~270 GW and US with ~150 GW. Hyperscalers increasingly willing to pay premium for nuclear's price stability and reliability.
[E2527] Michael Dell's chart shows orders of magnitude greater computational intensity for emerging AI use cases: video generation, robotics, and deep research require dramatically more compute than baseline chat/code completion. METR now highlights a doubling in average duration of 'work' performed by LLM per customer query every 7 months.
[E2521] Nuclear renaissance projected to be worth $2.2 trillion through 2050 (up from $1.5T prior), with 586.5GW of new capacity adding ~47% to current 398GW global nuclear capacity. China (~270GW) and US (~150GW) to lead growth. Korea could win up to 39% of global nuclear market after adjusting for accessibility.
[E5148] AI buildout creates materials and energy bottlenecks absent in prior tech cycles. Data center capex requires mining, manufacturing, and permitting that takes years; supply-constrained commodity bull.
[E2269] AI data centers are copper-dense across power, cooling, and redundancy systems. Big Tech is now contracting forward tonnage rather than spot buying. Author positions copper demand from AI infrastructure as structural and compounding, not linear — AI needs grid, grid needs copper, and AI also needs copper directly. This makes copper an operating constraint on AI expansion.
[E4740] 2026 capex cycle accelerates as embodied AI adoption begins. Data center power requirements doubling again. Inference dominates compute allocation. GPU shortage persisting through 2026 as supply chains struggle to meet demand.
[E4916] 2026 real AI inflection: transition from app-based LLMs to agent-based inference across physical systems, robots, autonomous devices. High bandwidth memory (HBM) sold out through 2025-2026. Micron tripled since June 5 note predicting this shift. Cisco and Micron critical beneficiaries as infrastructure layer most investable in AI value chain.
[E5042] Hyperscaler capex acceleration confirmed via Oracle/Microsoft orders but ROIC sustainability questioned; inference capacity shortage permanent structural demand; capex-to-revenue conversion efficiency deteriorating despite volume; misallocation of resources likely.
[E6615] FFTT recommends buying US industrial equities, especially electrical infrastructure companies, as part of their portfolio positioning. This aligns with structural demand for power infrastructure driven by AI capex and broader electrification trends, though Gromen's primary rationale centers on weaker USD benefiting US industrials heading into the 2024 election cycle.
[E6641] PJM grid auction hit record prices with 8-10 year power plant construction timelines. Gromen notes 'there is simply no new capacity to meet new loads' and the solution requires data center builders to bring their own generation. Electrical grid bottlenecks prevent effective AI data center buildout and reshoring.
[E6718] AI infrastructure demands measured in trillions identified as a key driver of rising real capital costs. Grid constraints becoming political liability by 2026 midterms. Amazon and Berkshire Hathaway cited as unable to get datacenters powered despite 4-year head start, suggesting electrical grid bottleneck is insurmountable for the broader AI capex cycle.
[E6879] FFTT highlights an electricity crisis driven by AI/data center demand: AI servers require 5-9x more power than traditional servers, transformer lead times are extending to 2 years, and power capacity needs to grow 160-260% by 2050. Big Tech is investing billions in nuclear power to address shortages, requiring massive infrastructure investment over 2+ years.
[E6971] Larry Ellison (Oracle) stated they are designing a data center exceeding one gigawatt with building permits for 3 small modular nuclear reactors to power it. Microsoft is also investing billions in nuclear power for data centers. This illustrates the extreme scale of AI infrastructure energy demand creating a new structural electricity demand wave.
[E7079] AI semiconductor demand is constrained by 3-7 year electricity hookup delays, limiting actual deployment. Data centers in NVDA's hometown sit empty awaiting power, with some projects delayed beyond 2030. Satya Nadella quoted: 'The new AI moat is physical infrastructure and energy access. You can train a better model in 6 months. You cannot build a powered data center in 6 months.' Water availability cited as the next constraint for water-cooled systems.
[E7135] Microsoft reports 35% of new code is AI-written and claims $500M in AI-driven savings. Meta increased market cap by $1.5T over 3 years while simultaneously cutting 5% of workforce. These data points illustrate AI capex translating into measurable productivity gains even as employment is reduced, supporting the AI infrastructure investment thesis.
[E8264] US electricity generation capacity remained flat from 1999-2020 despite $31T in debt increase, but must now expand dramatically to support AI infrastructure requirements. Gromen is bullish on the GRID ETF, arguing infrastructure investment is a necessity in the global technology race, not optional spending.
[E8674] US manufacturing construction spending doubled to $190B annually, with factories requiring massive electrical grid infrastructure. Some projects face 8-15 year grid connection delays, creating structural multi-year demand for electrical equipment and highlighting severe infrastructure bottlenecks constraining industrial expansion.
[E9030] AI capex contributed more to US GDP growth in the past two quarters than ALL consumer spending combined, representing an extraordinary infrastructure spending boom. However, Gromen flags META's 22-year depreciation schedule on hardware with 2-3 year actual lifespans as evidence of potential earnings overstatement and AI bubble risk in the capex cycle.
[E6453] Microsoft CEO admitted compute (chips) isn't the AI bottleneck — energy and data center space are. Companies have GPU gluts they can't deploy due to power infrastructure constraints. Gromen argues the 'old moat was model quality and algorithm improvements; the new moat is physical infrastructure and energy access. You can train a better model in 6 months. You cannot build a powered data center in 6 months.' PAVE electrical infrastructure favored over NDX.
[E7908] Exponential AI electricity demand is creating structural copper and nuclear power shortages, identified as a key forward-looking catalyst. NVIDIA named as primary entity beneficiary. The energy constraint on AI infrastructure buildout is framed as both an investment opportunity and a bottleneck forcing capital allocation toward power generation assets.
[E8127] US grid operators warn of potential summer 2025 electricity shortages in regions like Ohio. AI electricity demand is projected to reach 50%+ of US grid capacity by 2028, potentially requiring a 'Manhattan Project' scale response. This forces difficult resource allocation choices between domestic infrastructure investment and military commitments like AUKUS submarine production.
[E8140] Munger identifies a clear need for a 'really big national grid' and believes Berkshire Hathaway will be a major participant when the infrastructure buildout happens, though government delays are frustrating. This supports the thesis that energy infrastructure spending is a structural bottleneck requiring massive capital investment, with utilities and grid operators as key beneficiaries.
[E5005] Genesis Mission executive order establishes public-private AI infrastructure via DOE labs; VLM/VLA compute needs exploding 5-20x beyond LLM training; TPU architectural advantage emerging; bottleneck shift from training to inference and multimodal vision-based models.
[E4822] Global AI infrastructure buildout estimated at $5 trillion over 5 years with Nvidia capturing 35-60% market share due to unassailable moat. Q3 2025 revenues at $57B (+62% YoY) represent continuation of vertical growth trajectory despite multiple compression in high-multiple names.
[E5358] Now they're complaining companies are spending too much on com on capex.
[E5356] Short-term RPO of 400 billion shows that our need to build infrastructure is very high.
[E5345] Looking ahead, we expect our fullear capex to be approximately 125 billion in 2025.
[E5357] I remember when they complained companies were spending too much on buybacks and not enough on capex.
[E5019] Sam Altman factory producing 1 gigawatt AI infrastructure weekly; supply constraints not demand (unlike dot-com bubble); compounding exponential growth every 6 months; capital goods orders and industrial metals demand surge ongoing.
[E4749] Now, in terms of the demand side, we got a ton of things this week on this ch...
[E4893] Critical semiconductor shortage in edge devices requiring NPUs (Neural Processing Units). Nvidia-Intel partnership and Tesla-Samsung deals signal shift from GPU training to edge device deployment. ASML monopoly on EUV lithography critical for advanced chips; stock unchanged since 2021 despite PMI stagnation but will surge as NPU demand grows.
[E4906] Retail traders driving discovery of battery/power/semiconductor names before institutional adoption. Small-cap and mid-cap generator and battery companies (EOS, Generrack, Bloom Energy) moving on fundamentals before analyst coverage. Social media coordination enabling efficient discovery of supply-chain beneficiaries.
[E4904] Semiconductor/hardware supply chains stretched 5+ years. DRAM pricing frozen by Micron due to AI demand. Hard drive lead times extended to 12 months. ASML EUV monopoly critical for advanced NPU manufacturing. Supply bottleneck will persist and widen, supporting sustained capex spending and hardware company valuations.
[E5048] Oracle Q1 order increase $317B (+359% YoY) signals AI capex acceleration not deceleration; RPO at $455B; inference capacity shortage critical; cloud contracts with OpenAI, XAI, Meta, Nvidia, AMD booked; power requirements equal 90% Japan electricity or 155% Germany total consumption.
[E5170] Data center buildout accelerating with massive infrastructure needs for power, cooling, transformers, and materials. Supply chain constraints on gas turbines, transformers, semiconductors creating multi-year bull for commodities and industrials.
[E4778] Data center power demand doubled per CEO commentary. Global capex revised upward significantly. PMIs surging globally (US 51.0, Eurozone 51.1, Japan +6mo). Supply constraints, not demand weakness. Rate cuts expected despite PMI rises.
[E4858] Energy constraints becoming binding constraint on AI capex. Goldman Sachs projects capacity crunch 2026-2027. Microsoft committing to Three Mile Island nuclear power. Deregulation critical bottleneck—DOGE targeting 50% regulation elimination by 2026 to accelerate infrastructure buildout.
[E5084] Data center capex remains structural despite Deep Seek concerns. Caterpillar, Parker Hannifin, Cummins all reporting record backlogs and data center demand visibility through 2027.
[E5189] AI infrastructure buildout creating unprecedented demand for data centers, energy systems, and semiconductors. Energy bottleneck is primary constraint; companies reporting back orders through 2030+ with record backlogs.
[E5157] AI capex buildout requires power and industrial infrastructure chronically undersupplied. Data center construction highest since 2022; capital goods breaking out above CRB; industrial cycle without prior excess capacity.
[E5521] AI electricity demand bottleneck is structural limiting factor not chips. 92 gigawatts additional power needed but only 2 nuclear plants in last 30 years built. Natural gas and fossil fuels must supply baseload before fusion/fission solutions arrive. Power constraint becomes key scarcity lever.
[E4816] Comparing AI capex to telecom 5G is misleading. Hyperscalers high-margin (30-40%), low-leverage, massive cash generation. More analogous to 1990s semiconductor boom and early AWS/Azure buildouts (both generated exceptional returns). Capex cycle likely 5-7 years, not 2-3 year bust scenario.
[E4802] Veo 3 video generation requires 150x normal compute (100x for video, 50% more for audio). Demand off the charts; millions of videos generated daily. Inference moment emerging as bottleneck for all AI applications. Power grid and compute capacity the limiting constraint for AI expansion.
[E4882] 2026 will see collision between infinite AI demand and finite data center capacity. Goldman Sachs projects capacity crunch 2026-2027. Power grid constraints binding earlier than compute constraints. Capex cycle likely reverses mid-2026 as ROI clarity forces discipline. Current hyperscaler spending unsustainable without efficiency breakthroughs.
[E5064] Power is the limiting factor for AI infrastructure expansion. Jensen Huang stated every future data center will be power-limited, requiring massive capex buildout beyond existing capacity estimates.
[E5495] DRAM and NAND memory prices are rising for first time since semiconductors peaked; Micron, Samsung, SK Hynix leading beneficiaries. Memory chip shortage critical for AI infrastructure buildout and robotics infrastructure stage.
[E4785] McKinsey estimates $7T global data center spend by 2030 to meet AI compute demand. Inference explosion (Veo 3 video generation = 100x normal compute) creating bandwidth bottleneck. H200 HBM prices spiking; all NAND flash 2026 production already sold out.
[E4938] Meta raising AI capex guidance to $65B annually, doubling prior year spend, for data center buildout with 1.3M+ GPUs. Compared to Stargate 2M+ GPU scale. Both represent unprecedented hardware buildout phase, not just software/training. Implies sustained semiconductor demand, power infrastructure, construction/PMI tailwinds for years ahead.
[E4959] Semiconductor index (SMH) broke out of 7-month consolidation following June peak. MACD buy signal generated. Earnings growth expected to surprise as PMI inflection drives capex on hardware (not just software). Semis correlated with hardware phase, not LLM training phase. SK Hynix and other memory companies positioned for cycle as AI memory demands surge.
[E4952] Stargate announced as massive data center buildout requiring approximately 1/3 of all GPUs created in 2024 (for one facility). Combined with Meta's $65B capex plan and other hyperscaler projects, represents unprecedented hardware infrastructure phase. Buildout driven by inference scaling and edge device (robotics/humanoid) requirements. Supply chain constraints (semiconductors, power, cooling) will persist for years.
[E4757] AI agent adoption accelerating faster than analyst consensus. Bank earnings surprised to upside on AI productivity gains. BCG reports 1/3 of companies budgeting $25M+ annual AI spend. Visser forecasts sustained profit margin expansion as quarterly agent integration scales through 2025.
[E4981] AI capex cycle transitioning from LLM training to inference and embodied agents. Robotics buildout requires hardware not just code. 10 billion robots planned by Musk. Data centers being built across multiple companies (Meta 65B, Stargate 500B+). Hardware buildout in physical economy breaking 4-year software dominance narrative.
[E5127] AI capex secular theme unaffected by correction. Google maintaining $75B AI spending, Nvidia bringing production back onshore. Tariff disruption temporary vs AI productivity permanent.
[E5469] back for a uh pre-election video uh this one uh we'll go through what happened last week kind of what the markets are talking about how the economic data was uh I'll start talking about the structural rise and rates fall this has been going on for for four years but uh I I only started doing videos
[E5479] the the company showing massive Revenue uh I'm sorry massive capex spend but not seeing the Revenue come through their earnings are decelerating in terms of the growth side so Nvidia becomes more important and this is the 100 day rate of change which really just says to me that their earnings announ