Your briefing

Token demand and inference latency are reshaping the AI infrastructure market. The newsletter highlights explosive token growth — OpenAI shows a 320x YoY increase in reasoning tokens and Goldman reports China's enterprise token use rose 162% from June to December — while OpenRouter data indicates an acceleration since January 2026. UBS cites a 99.7% drop in token cost over three years but warns demand growth still makes compute capacity the limiting factor; inference latency has emerged as the next key bottleneck for agentic and coding workflows.

Nvidia's response is a disaggregated inference stack: Vera Rubin GPUs handle memory‑heavy prefill/KV cache work while Groq LPX (SRAM‑centric, low‑latency decode) accelerates token generation. Jensen Huang claims ~35x token‑generation performance improvement when fused via Dynamo; sampling is underway, Samsung will fabricate Groq LPX, and shipments are targeted around Q3 2026. UBS estimates LPX could add ~ $50B to C2027 data‑center revenue, supporting upside scenarios where Nvidia data‑center revenues approach ~$600B in C2027. Countervailing forces include hyperscaler ASIC programs (Meta’s MTIA roadmap and Broadcom as a key partner). JP Morgan projects Broadcom AI revenues of $65B+ in FY26 and $120B+ in FY27 with ~10 GW deployed, implying significant hyperscaler share gains. The authors model a 10x expansion in AI demand by 2031 as plausible and note humanoid robotics components as an additional long‑run growth vector, while warning that in‑house ASICs pose a medium‑term risk to Nvidia's share despite overall market expansion.

Developers are bypassing years-long grid interconnection delays by deploying mobile power plants—mobile gas turbines and jet engines—to run new data centers. xAI used semitrucked gas turbines in 2024 to energize Colossus in four months. Cleanview's 'Bypassing the Grid' documents dozens of permitted projects that could add dozens of gigawatts in 1–2 years, about 75% gas-fired, equal to five New York Cities of demand.

The webinar (presentation/webinar format) convened experts from DOE, Sandia, PNNL, Clean Energy States Alliance and IEEE to diagnose the energy‑storage interconnection bottleneck and surface policy, technical and standards solutions. Presenters documented a nationwide queue surge—driven by rapid declines in lithium‑ion cost (cited falling from ~$270/kWh mid‑2022 to ~$180/kWh by end‑2023), IRA‑driven project additions, and aggressive state decarbonization goals—which has produced years‑long wait times, high withdrawal rates (≈70%) and large, uneven interconnection upgrade costs (examples: PJM storage interconnection ≈$335/kW vs natural gas ≈$24/kW). Practical impacts at the state level were highlighted using Massachusetts (solar+storage = 93% of the queue) and national maps showing concentrated activity in Texas, California and the U.S. West.

Speakers outlined a layered solution set. Policy and market fixes include FERC’s 2023 reforms (cluster studies, readiness requirements, transparency and revised cost allocation), state experiments (California’s CPUC limited‑generation profiles; Oregon’s export‑capacity approach; New Jersey capacity‑based fees), and ERCOT’s connect‑and‑manage practice. Operational and technical remedies focus on better hosting‑capacity analysis, standardizing export‑control methods (non‑export, limited export, managing inadvertent export), and using group/cluster studies and flexible interconnection arrangements to speed outcomes. Standards and conformity work are essential: IEEE 1547 and its supplements (15479 for storage, 15474 for intentional islands) and the transmission IBR family (2800) are being updated to provide minimum performance requirements and commissioning/testing guidance, while UL 1741 certification and conformity assessment programs help utilities verify equipment. The DOE‑led i2x effort and lab partnerships (PNNL, Sandia, LBNL) have produced a transmission road map (volume 1) and are developing a distribution road map with stakeholder comment periods and targeted technical assistance to translate these reforms into implementable, equitable processes.

Citrini frames the April 20, 2026 Presidential Determination (2026‑10) as a formal Defense Production Act (Section 303) response to a deepening U.S. grid‑equipment bottleneck, invoking Executive Order 14156 (Jan 20, 2025) and classifying transformers, high‑voltage components, substations, control electronics, protective relays, capacitor banks and electrical core steel as essential to national defense. The note highlights accelerating backlog metrics — GE Vernova’s electrification segment posted a 1Q26 quarterly backlog addition nearly equal to the annual additions from 2022–2025 — and persistent transformer lead times that have exceeded 18 months. It details supplier dynamics: Hyosung’s Memphis plant is the only U.S. 765 kV capable facility; HD Hyundai Electric aims for 150 units/year from Montgomery by 2027; LS Electric opened in Bastrop, TX; Iljin secured a $333M U.S. order and its first European ultra‑HV contract. The Korean Big Four held a combined $23.9B backlog in Q3 2025 (about five to six years of work), justifying DPA action to shore up capacity.

Data centers are reshaping local fiscal landscapes and political choices: Greg Miller (Progress and Poverty, May 6, 2026) uses case studies—Loudoun County, Chandler AZ, Columbus OH and state incentive programs—to argue that data centers will be built and can be net fiscal positives if cities negotiate aggressively. He cites Loudoun’s ~$1.3B FY2027 data‑center revenue (≈45% of county receipts) and declining property tax rates as evidence of upside, while noting centers are high‑revenue but low‑employment land uses. Miller documents growing state giveaway costs (Virginia’s exemption rose from a $1.54M 2008 projection to $1.6B in FY2025; $2.7B forgiven 2015–2024) and examples of excessive local abatements (Google/Magellan: $300M facility, ~$54M 15‑year abatement for ~20 jobs). His prescription: permit and welcome centers for U.S. AI capacity, end broad state subsidies, and use local tools—closed‑loop water, guaranteed grid contributions, PILOTs or CBAs targeted at electricity relief and capital projects—to ensure community benefit.

The 100 MW Hyperscale AI Blueprint (Data Centre Dynamics, 27 Feb 2026) outlines a reference architecture for hyperscale AI data centers built to host 100 MW of high-density AI workloads. It emphasizes design and operational strategies that mitigate execution and operational risk and recommends prioritizing speed-to-market, capital efficiency, and infrastructure resilience for investable projects.

Utility Dive's April 23, 2026 Daily Dive newsletter aggregates industry moves: GE Vernova reported a roughly 100 GW gas-turbine backlog and said dollar-per-kW turbine pricing should accelerate in Q2, alongside large upticks in grid and wind equipment orders — a dynamic underscored by BNEF reporting a 66% surge in gas plant build costs. CenterPoint Energy told investors it will energize about 8 GW of hyperscaler data center load by 2029, citing Houston’s emergence as a preferred market. PJM’s market monitor opposed FERC waivers tied to Constellation’s Three Mile Island restart (needed to immediately deliver output), noting Constellation already has a 20-year contract to sell all attributes to Microsoft. Operational relief efforts include a planned 50-MW distributed battery for Guadalupe Valley Electric Cooperative; manufacturers warn transformer backlogs now run a year or more.

Data Centre Dynamics' 19 March 2026 livestream 'Emerging frontiers — Mapping the next data center hotspots' examines how power constraints, land scarcity and grid congestion in established hubs are redirecting investment to secondary cities and frontier economies. Speakers assess AI workload demand, sovereign cloud programs, connectivity upgrades and financing models for large-scale builds amid policy uncertainty and infrastructure risk.

Intel’s recent string of moves — market cap >$300B, joining Musk’s TeraFab (Apr 7, 2026) and partnerships with SambaNova and Google — signal a potential commercial turnaround centered on packaging and heterogeneous datacenter stacks. With TSMC CoWoS constrained, Intel’s EMIB packaging has seen accelerated demand through 2026; TeraFab’s 1 TW/year ambition offers Intel a whale customer if wafer starts materialize. The Intel–SambaNova stack pairs GPUs for prefill, Intel Xeon 6 as host/action CPUs and SambaNova SN50 RDUs for decode; the RDU’s SRAM+HBM+DRAM approach and “predetermined” data-paths aim to cut the number of decode chips versus SRAM-limited LPUs (a contrast drawn with Groq’s inelastic LPU design). Separately, Google+Intel deployments of Xeons plus IPUs (DPU-like offloads) reinforce x86’s role in heavy AI orchestration workloads.

Macro demand is amplifying these hardware bets: Samsung projects Q1 2026 profit of 57T won ($38B), a 755% YoY increase driven ~95% by memory, validating a multi‑year DRAM supercycle; Anthropic reported a $30B ARR (Apr 6, 2026) and booked 3.5 GW of Google TPU capacity starting 2027 while hedging across AWS/TPU/NVIDIA. Finally, Arm’s intent to sell finished AGI CPUs into China (rather than license IP) creates a fast route for Chinese datacenters to access high‑core, AI‑optimized server processors without transferring chip design IP — a notable strategic loophole with geopolitical implications.

Radical Candor: The Grid of the Future Runs on Transparency frames the next structural transformation of U.S. electric utilities around trust and data transparency. Michael Lee argues that utilities currently sell an “invisible” product — trust in reliable, affordable power — yet operate under cost-of-service regulation (COSR) incentives that reward capital deployment and information control. He cites concrete metrics: a J.D. Power 2025 customer satisfaction score of 499/1,000 and average residential bills up 34% since 2020 to $189/month. That divergence feeds political backlash even as utility stocks remain high. Lee contends that outcome-based Performance-Based Regulation (PBR) — which ties utility earnings to delivering reliability at lower customer cost — would flip incentives so utilities want third parties to site, dispatch, and optimize resources efficiently rather than protect rate-base-funded builds.

Lee lays out actionable, technical changes. He calls for a permissioned, tiered data platform exposing feeder-level loading, thermal constraints, live hosting-capacity maps, planned outages, and dynamic interconnection-pipeline status to authenticated developers, aggregators, and regulators under CEII/NDA frameworks. He documents the scale of the current mismatch: more than 10,000 queued projects (>2,000 GW) at end-2024, with ~70% of MISO projects historically not built, and cites PJM baseline transmission spend (e.g., $6.7B RTEP in 2024; 2025 projected $8–12B, five-year totals approaching $40B) as partly driven by suboptimal siting from opaque information. Operational reforms Lee recommends include standardizing and modularizing distribution hardware, treating long-lead procurements as financial options, normalizing utility ROE and tying upside to performance, and requiring line‑item engineering justification (hours of thermal exceedance, alternatives evaluated) for every capital dollar. He also highlights policy and consumer-friction risks: MassSave’s $4.5B 2025–27 plan funded via bill surcharges (policy charges in MA rose $15→$59/month, 2014–2025) undermines affordability and consent for mass-device enrollment. The post points to international and vendor precedents (AEMO’s Connections Simulation Tool, DOE i2X, Tapestry et al.) and argues that transparency, credentialed access, and platform-style information infrastructure are both technically feasible and essential to earn the mass-market permission needed for a distributed, flexible grid.

The newsletter examines wide uncertainty around Texas load growth driven largely by rapid data‑center expansion and a cluster of ERCOT reports, hearings, and policy moves in spring 2026. ERCOT’s preliminary 2026 Long‑Term Load Forecast shocked observers by projecting 368 GW by 2032 — over four times the current all‑time peak — a result Bloomberg‑quoted analyst Travis Kavulla called implausible. ERCOT CEO Pablo Vegas subsequently presented a far lower 111 GW by 2032 to the board, underscoring large model and assumption sensitivity. The piece frames two central challenges: how much incremental capacity and transmission will be required, and who should bear the cost, while noting ongoing industry conversations (including a podcast episode on batteries with Fluence VP Suzanne Leta) about how resources and markets must adapt.

WoodMac projects the US data center electrical-equipment market will expand from $20B in 2025 to $65B by 2030, driven by capacity growth from 24 GW to 110 GW (2026–2030), a 600 GW pipeline seeking power versus 183 GW with agreements, sustained 18–36 month lead times, and sharp hyperscale equipment and electricity demand growth.

Texas is in the middle of simultaneous, rapid demand growth and an unprecedented pipeline of generation proposals, and the mismatch between generation ambitions and transmission capability shapes how projects are being built. ERCOT’s all‑time demand peak is 85.5 GW, yet the market entered 2026 with roughly 432 GW of generation in the interconnection queue and about 225 GW of new large‑load requests filed in 2025. Raina Hornaday — founder of Caprock Renewables and Fortress Microgrid and a developer responsible for more than 1 GW of Texas projects over two decades — describes transmission as a long lead, capital‑intensive constraint that “load doesn’t want to wait for.” Her utility‑scale portfolio includes 150–300 MW solar projects; one 300 MW site required a new substation and two years of construction, illustrating the multi‑year timelines for bulk transmission work.

Because transmission expansion lags, Hornaday and other developers are shifting toward distributed generation, co‑located solar+storage, microgrids and behind‑the‑meter supply for fast‑moving customers (notably AI/data centers and Permian loads). Site selection is increasingly nodal‑data driven — choosing nodes with attractive prices and substation access — and smaller projects (sub‑10 MW batteries or behind‑the‑meter builds) avoid the full interconnection study process, shortening lead times. Battery energy storage has become a critical stopgap: ERCOT’s Real‑Time Co‑Optimization Plus Batteries (launched December 2025) and a large buildout of storage (capacity roughly doubled in the prior year) have damped price volatility since the 2021 Uri event and provide fast dispatchable capacity while transmission is planned and built. Nevertheless, not all storage projects proceed; cancellations through early 2026 reflect permitting and siting opposition, local fire‑safety concerns, financing and supply chain issues.

Hornaday also highlights the local economic and workforce benefits of renewables (agrivoltaics, “solar sheep,” school‑district interactions), the loss of Chapter 313 incentives that changed project economics, and ongoing political risk — e.g., Senate Bill 819’s failed 2025 effort and likely future iterations. Her prescription: more education, standardized siting and permitting practices, and policy alignment so Texas can pair rapid load growth with flexibility resources and transmission investments required for long‑term reliability.

Sekar lays out why power delivery is the next “physics wall” for AI datacenters and why the industry is moving toward 800V DC rack architectures. The core physics is simple: for fixed power, lower voltage means much higher currents, which amplify resistive losses (I2R), force massive copper busbars, and erode voltage-headroom. Using his worked example, a 600 kW rack at 48V needs ~12,500 A, a 1% voltage-drop budget implies ~38 µΩ total resistance, ~6 kW dissipated, and roughly 1,000 cu. in (~147 kg) of copper busbar per rack. Moving to 800V reduces current to ~750 A, lets busbars be ~1"×1" (~84 cu. in, ~12.3 kg), and cuts dissipated power to ~0.56 kW.

Those physics drive architectural choices: keep voltages high as long as possible and perform point-of-load (PoL) conversion—vertical power delivery—close to GPUs. Conversion chains run from high-voltage grid lines (100 kV) down to medium/low voltages (10–30 kV substations, 400–480V AC halls) then to rack DC (48V historically, now 800V in new builds) and finally to VRM voltages (~0.8–1V). Each conversion stage adds loss (e.g., 97% per stage → ~91% over three stages), and isolation rules (SELV 60 V DC) force transformer-based topologies above 60 V, shaping which semiconductor and converter designs compete. Early 800V rollouts require greenfield or DC-sidecar approaches (OCP Mt. Diablo, NVIDIA Kyber), and analysts and vendors (Wood Mackenzie, SemiAnalysis, Citrini, ON Semi) expect sharply higher power‑semiconductor content and supply-chain shifts as deployments scale.

Mark Olsen, manager of reliability assessments at the North American Electric Reliability Corporation (NERC), presented a webinar (Clean Energy States Alliance, April 22, 2025) summarizing findings from NERC's 2024 Long‑Term Reliability Assessment and related reliability leadership work. The session was a technical briefing-style presentation that emphasized probabilistic methods (loss-of-load hours and expected unserved energy) and a 10-year planning horizon. Olsen showed NERC's risk map (high/elevated/normal categories) and explained that more than half of the continent faces growing energy risk driven primarily by a step-change in demand growth, continued retirements of thermal plants, and uncertainty/timing in bringing new resources online.

Olsen walked through specific metrics and implications: the interconnection queue grew by about 44 GW (12%) and is dominated by solar and battery projects, but actual adds lagged projections except for batteries. NERC accounted for roughly 78 GW of thermal retirements in its baseline and flagged another ~37 GW announced but not yet removed from planning, concentrating risk in market areas such as PJM and MISO. Winter reliability is emerging as a particular concern—winter peak growth is outpacing summer growth in many regions, winter shortfall events can last much longer (example winter events modeled up to ~12 hours), and short-duration batteries (2–4 hours) have limited ability to solve extended winter energy deficits. Transmission planning activity has increased (≈28,000 circuit miles in development vs a historical average near 18,000), but construction bottlenecks, permitting and siting issues (affecting ≈1,200 miles) slow realized delivery benefits. Olsen highlighted the interregional transfer capability study showing transmission can reduce some shortfalls but will not eliminate the need for additional resources. His recommendations for state and federal policymakers: pursue an all-of-the-above resource strategy, carefully manage generator deactivations, improve permitting and siting coordination, mature probabilistic and multi-weather-year modeling, account for grid-enhancing technologies, and plan for longer-duration storage and firm resources to cover growing winter and multi-day energy needs.

Virgo Network and TPU 8t/8i mark a major networking and optics-driven shift in Google's AI datacenter architecture: a single Virgo fabric now spans >130,000 TPUs, provides ~47 Pb/s bisection bandwidth, and cuts latency by roughly 40%. Google explicitly splits the stack into pod-level scale-up, Virgo scale-out across pods, and Jupiter for front end/storage, validating that the aggressive changes are to the DCN (scale-out) rather than the 3D torus. TPU 8t reports a multi‑fold DCN bandwidth increase and TPU 8i is called out as the complementary advancement for optics. The result is a clear move from compute-limited designs to bandwidth-limited systems, underscoring the need to co-design compute, storage, and networking for modern AI hypercomputers.

Distributed energy resources (DERs) — solar, residential batteries, EV chargers, smart thermostats — have demonstrable commercial value, but the distribution-layer infrastructure, data flows, and regulation prevent that value from being realized. The market evidence includes >5 million U.S. solar installs (SEIA, 2024) and firms like Base Power (25–50 kWh batteries per home; >$1.3B raised; Texas co-op deals by early 2026) and David Energy (aggregating thermostats, batteries, EV chargers across NY, NJ, MA, TX) that monetize wholesale signals where they exist. By contrast, distribution constraints (feeder/transformer thermal limits) are rarely instrumented, priced, or exposed, so DERs cannot respond to the most valuable local events.

Technical and structural gaps are clear: ERCOT’s ADER pilot revealed that residential hardware and telemetry were not grid-grade (requirements: 2 s telemetry, ±10% validation, 5-minute basepoints), producing only ~15 MW qualified by late 2024 until vendors built to spec and ERCOT raised the cap to 200 MW in 2025. OEM APIs and cloud backends are designed for consumer convenience, not low-latency per-device validation; OEM lock-in and warranty/fee restrictions add commercial friction. Static NEM and many community-solar crediting schemes ignore time and location value. A persistent “topology floor” — lack of real-time device-to-substation/feeder/phase mapping — plus utilities’ incentives under cost-of-service regulation to favor ratebase investments keeps valuable flexibility idle. The author argues PBR, better data sharing, grid-grade hardware/standards, dynamic operating envelopes (as piloted in Australia), and upgraded topology sensing are needed to unlock DER value and avoid unnecessary capital-intensive network builds.

US nuclear fuel supply chain faces deep import dependence and capacity shortfalls: the country runs ~97 GW (~20% of electricity) but may need 100–300 GW by 2050. Conversion (ConverDyn at ~7 kt/y), limited domestic uranium (<1%), and slow enrichment/fabrication buildout (DOE $2.7B; projects into 2034) create a gap McKinsey prices at $105–170B plus $20–45B if reprocessing proceeds.

Lumentum (LITE) is positioned as a central supplier for an industry transition toward optical scale-up, with management highlighting four revenue engines—Cloud Transceivers, OCS, Scale-Out CPO, and Scale-Up CPO—all relying on InP/EML, OCS engines, and UHP lasers. At OFC 2026 the company emphasized manufacturing scale: EML shipments grew >8x FY20–FY26E and Lumentum plans >50% more EML unit capacity by end-CY26. The firm is advancing 400G/lane Differential Drive EML for 3.2T (8×400G), arguing SiPho lacks bandwidth for 400G/lane, and projects ~85% CAGR in InP optical lane demand from CY26–CY30—implying an industry bottleneck at InP light sources. Commercially, 1.6T ramps to volume this summer, CW laser vertical integration is due in CQ3, and a multi-year OCS deal leaves >$400M backlog for 2H CY26 with a >$1B run rate target in 2027.

Stargate is a nationwide, $500 billion data‑center program led by OpenAI with Oracle and SoftBank that targets seven US sites totaling more than 9 GW of planned facility power — roughly the scale of New York City’s single‑hour peak demand — and an aggregate compute capacity the authors estimate at ~20 million H100‑equivalents. Abilene, TX is furthest along (~0.3 GW operational as of Apr 22, 2026; 250k H100‑eq; four of eight buildings live) and is expected to reach 1.2 GW by Q4 2026. The six other sites (Shackelford 2.0 GW, Doña Ana 2.2 GW, Milam 1.2 GW, Port Washington 1.3 GW, Saline Township 1.4 GW, Lordstown <0.3 GW) are mostly slated for Q4 2028 handovers. Developers are favoring on‑site natural‑gas microgrids and closed‑loop liquid cooling to shorten grid interconnection timelines and limit water evaporation, but those choices raise cost and regulatory scrutiny; procurement, financing, and local opposition remain key risks to the 2029 build‑out target.

DG Matrix's whitepaper (published 17 March 2026 by Data Centre Dynamics) analyzes how 'stranded power' within facilities limits AI data-center growth and shows that conventional distribution can't match dynamic AI workloads. It proposes software-defined power infrastructure to dynamically route energy across grid, storage and distributed resources, unlocking additional deployable compute within current utility limits and improving economics.

Qatar’s helium shutdown — a single industrial complex producing roughly a third of global helium went offline after Iranian missile strikes reported three weeks before March 29, 2026 — exposes a critical, unpriced dependency undergirding the global AI buildout. Helium is irreplaceable in EUV lithography, wafer cooling, and vacuum leak detection at fabs in South Korea and Taiwan; transport of liquid helium faces a 48‑day boil‑off window and regional shipping routes are closed. The timing is acute: the five largest U.S. cloud/AI providers plan $600–$700B capex in 2026 (75% for AI) and Goldman Sachs forecasts $1.15T hyperscaler capex across 2025–2027, all predicated on chip delivery. Nate outlines three propagation channels (helium, LNG/energy, and geopolitical compute-cost shifts), warns of memory shortages compounding delays, and argues the event could structurally advantage Chinese fabs if energy and supply realignments persist.

Earnings calls in the 2026 fourth-quarter season reveal a rapid, AI-fueled shift in U.S. electricity demand: tech companies locking long-term supply deals are driving both regulated utilities and merchant generators to plan for major load increases. CenterPoint Energy told investors that Greater Houston peak load should rise 50% between 2025 and 2029—an acceleration of its prior timetable—and the company now expects operating earnings growth at the high end of its 7–9% long-term target through 2028 and to beat a prior 11% CAGR base-rate revenue projection. Those company comments align with ERCOT’s August long-term forecast, which lifts summer peak from 87 GW in 2025 to 145 GW by 2031, explicitly counting about 24 GW of data-center load and 8.5 GW of crypto. The result is greater revenue visibility across the sector as AI demand reshapes planning and procurement.

Greencoat Renewables’ Paul O'Donnell outlined how the yieldco model that dominated renewable finance for a decade has moved into a new phase of active operations, hybridization and growth into digital infrastructure. He traced the firm’s origins to 2012/13 and its listing in 2017, and reported total invested capital of about €13.5 billion, with just under 8 GW under management across roughly 445 assets. That long-term accumulation of government‑backed, contract-stable assets is giving way to a more complex market: higher interest rates, shifting policy after Europe’s energy crisis, and new demand drivers such as AI and data centre build‑out are forcing owners to become trading-savvy operators.

Paul described the practical response: long-term PPAs directly with tech companies, hybridizing wind/solar sites by adding batteries for arbitrage and grid services, and partnering to scale capabilities rather than trying to internalize every function. He cited a joint venture with a battery/storage provider (referred to in the interview as 'Coato') to access pricing and innovation — the partner, he said, had recently won roughly 80% of certain Italian opportunities. On storage he emphasized falling capex and attractive incremental costs when batteries are added to existing sites, and argued revenue stacks (day‑ahead arbitrage, balancing markets, capacity/flexibility services) now underpin investment cases. Turning to data centres, Paul said they already account for an estimated 22–23% of Ireland’s electricity and present both demand and policy tailwinds: governments want data centres to source renewables and provide grid backup/flexibility. He estimated a 100 MW data centre costs ~€1.5 billion to build plus €300–500 million of enabling energy infrastructure, and argued truly off‑grid, gas‑backed data centres are unlikely in Europe. Hosts agreed the strategic pivot — which they labelled 'yield to growth' — makes sense and welcomed energy investors entering the digital infra space, while also flagging structural risks for listed yieldcos (market marking, investor alignment and competition from fixed income). The market reaction to Greencoat’s March strategy — a roughly 20% share price rise — suggested investors rewarded the clearer growth and integration story.

The Data Center Cooling newsletter (published 24 March 2026) highlights accelerating moves toward high‑density, liquid‑cooled AI infrastructure: Nvidia’s new LPU‑based LPX rack is fully liquid cooled and specified to draw up to 160 kW per rack, matching the nearby Vera Rubin deployment and signalling rising per‑rack power densities. In M&A, Ecolab will buy CoolIT from KKR for $4.75 billion, a deal expected to close later in 2026 that expands Ecolab’s portfolio in direct‑to‑chip and in‑rack cooling. Funding activity continues: Frore Systems closed a $143 million round at a $1.6 billion valuation, promoting a proprietary coldplate it claims improves thermal efficiency for AI workloads. The issue also points readers to technical resources— a Zutacore whitepaper on waterless direct‑to‑chip cooling and an nVent showcase on secondary fluid network design—underscoring industry focus on liquid architectures and system‑level reliability.

Stanford CS153Systems Session 8 (video posted 2026-04-25) presents Scott Nolan of General Matter asserting that power constraints are the central bottleneck to AI scale. He maps AI factory demand, warns of stranded power, promotes nuclear as the baseload solution while highlighting enrichment and supply-chain gaps, and urges engineered sites, government support, and lessons from bitcoin mining to meet scaling timelines.

SemiAnalysis compares PJM (13‑state Eastern interconnection) and ERCOT (Texas) to assess whether AI datacenter growth is driving higher household electric bills. The analysis shows the immediate driver in PJM is not wholesale energy scarcity but the structure of PJM’s forward capacity market. PJM’s 2025/26 BRA cleared at about $270/MW‑day (versus $29/MW‑day in 2024/25), with regulators capping later auctions at $329/MW‑day. Because capacity is procured via a centrally modeled Variable Resource Requirement (VRR) curve determined by PJM’s own demand and supply forecasts, modest changes in projected datacenter load (PJM/IMM estimate ~7.9 GW incremental datacenter demand in 2025/26 and ~12 GW in 2026/27) re‑shaped the VRR near the clearing point and produced a simulated price shock. The IMM’s alternate runs show removing datacenters would reduce capacity payments by $9.33 billion (64%), demonstrating the auction’s sensitivity to forecast inputs. Translating the capped BRA into retail terms yields roughly $34/MWh of capacity cost (3.4¢/kWh), adding about $25–$30/month to an average PJM household (880 kWh/month, 40% load factor).

ERCOT’s experience contrasts sharply. ERCOT runs an energy‑only market with a real‑time Operating Reserve Demand Curve (ORDC) scarcity adder (price cap ~$5,000/MWh). Although ERCOT’s April 2025 Long‑Term Load Forecast showed an eye‑popping 77.9 GW of potential datacenter load by 2030, ERCOT applied aggressive haircuts (≈49–55% reductions) and delayed in‑service dates, and forward energy markets priced only an ~11–17% increase across 2026–2030 horizons—far below PJM’s capacity shock. Operational performance reinforced the divergence: Winter Storm Fern (Jan 24–27, 2026) saw PJM lose ~21 GW of generation despite high capacity payments, while ERCOT held with no major outages; DOE issued emergency orders and identified ~35 GW of backup generation at datacenters/industrial sites. SemiAnalysis attributes PJM’s outcome to market design and forecasting errors (including methodology changes that removed ~14 GW of gas capacity and ~35 GW of offered capacity declines over four years), construction and GPU supply‑chain delays that make PJM’s datacenter build‑out forecasts optimistic, and regulatory fragmentation (13 states + FERC) that slows reform. The piece concludes the principal fault lies in policy and market structure rather than in AI demand per se, and highlights regulatory asymmetry (ERCOT’s faster reforms like SB 6 versus PJM’s multi‑jurisdictional constraints) as a durable, investable difference for hyperscalers and power sector participants.

U.S. utility sector headlines (Utility Dive, May 6, 2026) center on grid strain from rapid large-load growth and the evolving role of nuclear and renewables. American Electric Power is weighing leaving PJM or SPP because slow generation interconnection is constraining its response to contracts for 63 GW of new large load by 2030. Xcel’s tariff agreement with Google is being positioned by CEO Bob Frenzel as a repeatable model to attract hyperscalers while accelerating renewables. TVA reports nuclear now supplies 41% of its power and interim CEO Mike Skaggs seeks clarity on new nuclear technologies and federal coordination. Duke added 2.7 GW of data-center contracts in Q1 (total 7.6 GW, ~66% under construction) and maintained a $103 billion capital plan. Meanwhile, NERC’s Level 3 alert requires seven prescribed actions by Aug. 3, 2026 to mitigate immediate risks from computational load losses.

AI Value Capture - The Shift To Model Labs focuses on how agentic AI, hardware advances, and software optimizations have recomposed profit pools so that frontier model labs (e.g., Anthropic) are now capturing the lion’s share of value. The brief documents a rapid structural change beginning December 2025: agentic workflows (code agents, multi-turn assistants) drive much higher token utility and demand, increasing model monetization while the cost of producing tokens has plunged. SemiAnalysis quantifies this with internal examples—annual Anthropic token runs up to $10.95M, per-employee consumption near 5B tokens/month, Claude Code input:output ratios ≈300:1 and cache hit rates >90%—and links those usage patterns to Anthropic ARR growth (reported from $9B to >$44B) and gross margin expansion (<40% → >70%).

The piece documents the technical levers behind the shift. New accelerators (Blackwell/GB300/VR NVL72) and ASICs (TPUv7, Trainium3) plus middleware improvements (wideEP, disaggregation, MTP) multiply tokens/sec/gpu: software alone can produce ~14x gains on B300; combined with hardware, SemiAnalysis reports GB300 NVL72 configurations delivering ~17x higher throughput vs H100 in FP8 and up to ~32x in FP4. Meanwhile memory scarcity and pricing have surged (DRAM up ~6x year-over-year, one-year H100 rental +40% since Oct 2025), making Nvidia’s socketed SOCAMM a crucial pricing lever—SemiAnalysis models SOCAMM at ~$8/GB in 1Q26 with potential to exceed $13/GB by end-2026 and assumes ~$10/GB as a working figure. System-level economics are surprising: capex per watt barely changes from GB300 ($37.4/W) to Rubin (VR NVL72 $38.1/W) despite large TDP and FLOP increases, leaving big room between cost-based rental floors (~$4.92/hr/GPU for a 15.6% IRR) and value-based ceilings (~$12.25/hr/GPU parity, conservative ~$9.63/hr). The authors present a ‘One Chart To Rule Them All’ that combines floor and ceiling constraints with Neocloud IRR curves to illustrate how incremental pricing by Nvidia or TSMC could shift captured value. Finally, SemiAnalysis argues that although TSMC and Nvidia could extract more given N3 and memory tightness (TSMC N3 >100% utilization expected H2 2026; DRAM fabs >90%), both firms have so far held pricing in part for strategic/regulatory reasons—meaning short-term value accrues disproportionately to model labs, inference providers, neoclouds and memory vendors unless system suppliers move to value-based pricing.

Power infrastructure — not GPUs — is emerging as the immediate choke point for the AI data‑center buildout. Sightline Climate data (cited on the All‑In podcast) shows 12 GW announced for US 2026 capacity across 140 projects while only ~5 GW is under construction; many projects lack concrete power plans. The largest single constraint is high‑voltage transformers and substation lead times, which have lengthened from ~24–30 months pre‑2020 to about five years today. That mismatch means hyperscalers that locked PPAs and electrical orders 3–4 years ago will get priority while others face multi‑year waits, idle shells, and depreciating GPU hardware.

Developers are increasingly using on‑site generation to avoid HV interconnection delays. BYOP and islanded microgrids are moving from stopgap to strategic choices: regulators approved Williams’ onsite natural‑gas plan for Meta in July 2025, and Oklahoma legalized self‑build power in 2025. Traditional gas turbines remain capacity‑constrained (GE reports ~3‑year waits; HSBC forecasts turbine capacity rising 70 GW→>90 GW by 2029), are noisy and polluting, and face long delivery schedules. Fuel cells — notably Bloom Energy’s SOFCs (native 800 V DC, 1.5 GW installed, Oracle order up to 2.85 GW, claimed 5 GW/yr manufacturing) — offer quieter, low‑emission, transformer‑avoiding alternatives. Goldman’s LCOE work shows fuel cells are sensitive to gas prices (≈40% more expensive at $4/MMBtu, but ~7% cheaper at $13.3/MMBtu). ABB reports switchgear capacity is expanding, making transformers and generation the primary bottlenecks for near‑term AI data‑center deployment.

As data centers pursue on‑site generation and self‑supply, utilities face new cost, planning and reliability risks, write Brandon Owens and Morgan Bazilian in Utility Dive’s March 21, 2026 Weekender. Industry disclosures predict a sizable share of new capacity could be partially or fully off‑grid by the end of the decade, complicating load forecasts and revenue recovery. The Energy Information Administration reported electric‑sector gas use fell 3% in 2025 due to more solar and batteries even as overall U.S. gas consumption reached a record; its modeling also shows a high‑demand data center buildout could raise 2025–2027 load growth by ~15% in Texas and 4.7% in PJM and, in an extreme case, contribute to a 79% ERCOT price spike in 2027. Utilities and regulators are responding with long‑term contracts and tariffs (e.g., PPL’s $275M settlement, 4.9% residential bill impact, 10‑year minimum agreements) and planning reforms such as FERC’s approval of SPP planning merger.

Belgium plans to nationalize all its nuclear reactors and has halted decommissioning immediately. An accord with ENGIE will define terms and start studies toward a full takeover of the Belgian nuclear park, a move the government calls necessary for secure, affordable, sustainable energy and reduced dependence on fossil imports.

The Great AI Silicon Shortage frames a 2026 supply shock driven by explosive AI compute demand and constrained advanced wafer and memory capacity. SemiAnalysis shows a rapid industry shift onto TSMC’s N3 family — NVIDIA’s Rubin (N3P), Google TPUv7 (N3), AWS Trainium3 (N3P), AMD MI350X/MI400 tiles and even Vera CPUs — leading AI workloads to claim ~60% of N3 wafer output in 2026 and an estimated ~86% in 2027. TSMC is running N3 above nameplate in H2 2026 and is accelerating capex after only surpassing its prior peak in 2025, but cleanroom space limits immediate scale‑up. Memory is the secondary bottleneck: HBM consumes roughly 3× wafer capacity per bit versus commodity DRAM (rising toward 4× with HBM4), customers are targeting ~11 Gb/s for HBM4 pins, and accelerators are increasing stack height (8‑Hi → 12‑Hi). SemiAnalysis quantifies reallocation scenarios — e.g., 5% of smartphone N3 wafer starts (from 437k) could yield ~0.1M Rubin GPUs — but notes packaging (CoWoS/OSAT/EMIB) and HBM availability must also be solved. The piece highlights market consequences: hyperscalers prioritizing N3 allocation, potential foundry diversification to Samsung/Intel, and elevated DRAM/HBM pricing and negotiations shaping 2027 supply dynamics.

Engineering the future of cooling — DCD streamed a three‑part broadcast on Tuesday, 10 March 2026 (9am, 10am, 11am ET) featuring Siemens, Nortek, Ark, Tillion, Fleet, Johnson Controls and Armada speakers. Sessions emphasized engineering methods (physics‑based simulation, digital twins, advanced controls), integrating liquid cooling from grey to white space, and modular/prefabricated solutions to scale high‑density AI deployments.

The Feb. 24, 2026 Utility Dive briefing highlights shifting resource and capital plans as load growth and data center demand reshape utility strategy. PJM’s proposed behind-the-meter reforms — part of a data center colocation effort — have drawn pushback from trade groups that say the changes would harm the economics of new CHP at industrial sites. Major utilities are responding with large capital programs: Dominion disclosed a $65 billion, five-year plan tied to data-center-driven load growth and the near-completion of Coastal Virginia Offshore Wind; Con Edison projects about $38 billion through 2030 for EVs and building electrification. In Texas, Entergy issued an RFP for gas-fired units deliverable by 2032 to meet ~1 GW of new demand. Separately, analysts report easing supply constraints for gas equipment after capacity expansions at GE Vernova, Siemens and Mitsubishi Heavy Industries.

Texas’s electricity system is entering a period of rapid, concentrated load growth—driven by data centers, population increases, and electrification—that is colliding with sharply higher capital costs and traditional utility incentives. Participants Micalah Spenrath, Joshua Rhodes and Matt Boms summarized the problem: transmission and distribution (T&D) spending in Texas has already averaged about $9 billion per year (roughly $5B distribution, $4B transmission) over the past decade, and utilities are filing major rate cases (the episode flags an Oncor $830M request and references resiliency programs of roughly $1B/year for some TDUs). At the same time, generation and hardware costs have escalated: combined‑cycle gas capex assumptions now often sit in the $2,500–$4,500/kW range (up from ~$1,000/kW previously) and pad‑mount transformer prices have reportedly risen ~200%. Those cost shifts increase the risk of short‑term “rate shock” if spending is front‑loaded while demand remains uncertain.

The panel argued the choice architecture matters: when infrastructure is built primarily for a single customer (“driveway” upgrades) those customers should bear the capital; when upgrades provide broad system value, costs can be socialized. They highlighted commercial examples—reported proposals where large tech loads (Meta, Microsoft) offer to fund dedicated plants (a cited ~366 MW El Paso gas plant)—and suggested borrowing mechanisms from Alberta where new loads pay upfront and are repaid over time if they persist. The Texas Advanced Energy Business Alliance (TAEBA) DER study shared on the call modeled ~2.5 GW of customer‑side DERs over 10 years and attributed about $1,850 in savings per ratepayer by combining T&D deferral benefits with wholesale market value. Panelists recommended policy fixes: expand price signals at the distribution edge, authorize larger utility pilots for virtual power plants and advanced DERs, and reorient incentives from pure cost‑of‑service toward performance‑based regulation. Those steps, they argued, can defer expensive poles‑and‑wires investments, align cost causation with the parties driving growth, and reduce the likelihood that broad ratepayers shoulder the risk of potentially transient load buildouts.

xAI handed its 220,000‑GPU Colossus 1 to Anthropic because the mixed-generation, heterogeneous cluster (≈150k H100, 50k H200, 20k GB200) was inefficient for distributed training: straggler effects, NCCL/ring‑topology latency at the 100k+ scale, and GB200 power‑smoothing issues that require a rewritten stack produced just ~11% MFU versus 40%+ at Meta/Google. xAI preserved training on Colossus 2, a homogeneous Blackwell fleet, and leased Colossus 1 as a single-tenant inference farm (300 MW) to Anthropic. Inference tolerates heterogeneity and removes multi‑tenant jitter, letting the asset earn an estimated ~$2.60/GPU‑hour or $5–6B/year—nearly hedging xAI’s annualized Q1‑2026 losses—and positioning SpaceXAI favorably ahead of a planned SpaceXAI IPO (~$1.75T) while materially boosting Anthropic’s deliverable capacity in the May 2026 compute race.

Professor David Rusic (Illinois Energy Prof) gives a presentation-style update on U.S. small modular reactor (SMR) progress, focusing on the two DOE-backed winners from the 2020 competition: X‑energy’s TRISO-fueled pebble-bed design and TerraPower’s Natrium sodium‑cooled reactor with molten‑salt thermal storage. He explains TRISO fuel as microscopic kernels coated with carbon/ceramic shells, assembled into golf‑ball‑size pebbles that permit continuous on-line refueling and strong containment of fission products; high neutron energies in these fourth‑generation concepts enable transmutation of long‑lived wastes. Natrium uses a sodium coolant and external molten‑salt storage so a plant can run continuously while dispatching electricity on demand from stored heat. Rusic walks through the multi‑year, capital‑intensive permitting sequence (site ownership, environmental impact statement, NRC preliminary safety analysis — ~1,700 pages for Natrium — non‑nuclear and nuclear construction permits, as‑built documentation, then operating license). He names sites and partners: Natrium’s Wyoming coal‑to‑nuclear conversion (Amazon committed as offtaker) and X‑energy’s Seadrift, Texas project adjacent to Dow Chemical, with X‑energy’s TRISO fuel assembly facility planned near Oak Ridge, TN. He judges Natrium’s schedule as potentially operational by 2029 (grid delivery ~2031) and X‑energy likely later (optimistic 2029, more likely ~2032), noting fuel‑fabrication and licensing remain key bottlenecks.

Nvidia CEO Jensen Huang told an audience at the Milken Institute Global Conference (interviewed Monday; article published May 5, 2026) that the recent rise of agentic AI — exemplified by Anthropic’s Claude Code — has made AI “useful” in the last several months and driven a dramatic surge in compute demand. He estimated agentic systems need roughly 1,000× the computation of generative models, causing even four- to five-year-old GPUs to appreciate in price. Huang detailed Vera Rubin server racks as ~3‑ton assemblies with ~1.5 million parts, silicon photonics, 3D packaging and liquid cooling, costing $4–5 million each. He said gross margins at AI-native firms have inflected positive over the past 3–6 months, spurring capacity races, and argued AI will create jobs, accelerate scientific discovery (what took months can take a day), and require massive reindustrialization — raising ambition by ~100×.

Cortex 2.0: Joe Tegtmeyer reports that his review of nearly 100 Giga Texas permits (shared with subscribers) includes one referencing DH5, implying at least five data halls. Tesla indicated partial operation in late April 2026 (likely DH1–3), with full build‑out expected this fall and a target capacity above 130,000 H100‑equivalent GPUs for FSD, Optimus, and other AI/robotics workloads; the DH5 permit covers structural steel for power, cooling, and rack infrastructure.

Power availability is fast becoming the defining constraint on data center growth as AI, cloud and high-performance workloads increase demand and strain grid capacity, deployment timelines and electrical architecture. DCD's Critical Power Supplement (23 March 2026) examines solutions — grid-interactive designs, on-site generation, distributed/hybrid electricals, solid-state and advanced UPS, modular builds, microgrids and intelligent energy management — to enable higher densities and resilience.

High-voltage transmission is surging in the U.S. after a decade-long lull: 888 miles of 345 kV+ lines were completed in 2024 (vs. ~4,000 in 2013). Planned spending jumped—PJM from $920M (2021) to ~$12B (2025)—and utilities plan $1.4T by 2030 (about half T&D). Some former boosters resist public cost-sharing even as private demand (data centers, Permian) is ready to pay.

Nvidia VP Ian Buck characterized the expanded Nvidia–AWS partnership as a multi‑chip, hyperscale effort that will deliver more than one million Nvidia GPUs through the end of 2027 and add Groq 3 LPUs plus ConnectX / Spectrum‑X networking for high‑performance east–west traffic. Buck said AWS and Nvidia are coordinating long‑range forecasting, with Vera Rubin ramps “in the second half” and roughly six months from silicon wafer to racks. The companies are jointly monitoring fleet health and swarming engineering to support large customers (Buck cited work with OpenAI), while Jensen Huang estimated LPUs could handle ~25% of certain ultra‑high‑throughput agentic inference workloads; the agreement also includes RTX PRO 4500 Blackwell Server acceleration for Amazon EMR and reflects AWS’s larger datacenter push amid a $200 billion 2026 capex outlook.

Katie Coleman, a leading Texas energy regulatory attorney and managing partner in O’Melveny & Myers’ Austin office, explains in a March 25, 2026 Energy Capital podcast how a wave of very large load interconnection requests—driven by AI data centers, hyperscalers, and third‑party site developers—has exposed transmission planning and cost‑allocation limits in ERCOT. Coleman emphasizes the market split: generation investment is borne by private capital under the energy‑only design, whereas transmission remains a regulated activity that the Public Utility Commission authorizes and ultimately pays for through customer rates. That asymmetry is creating a policy squeeze as planners try to decide which speculative or staged projects are “real” and how much transmission to build now versus later without unfairly loading existing ratepayers.

Coleman walks through how different industrial customers behave in scarcity events: more than half of industrial sites are high‑load, reliability‑constrained operations (not price responsive), roughly 25–30% can and do participate in demand response or ancillary services (Emergency Response Service, responsive reserve), and some operators can opportunistically curtail when prices spike. She disputes narratives that large users “profit” from emergency pricing, noting ancillary revenues are typically mitigation, not net profit, for firms that may spend hundreds of millions annually on power (up to 70% of production costs for some). Coleman points to Winter Storm URI (2021) as an example where outages resulted from generation performance (she cites about half the gas fleet offline and a healthy reserve margin beforehand) rather than the market construct. On policy, she highlights Senate Bill 6’s requirement for earlier financial commitments for interconnection studies, but warns a pending PUC proposal for $100,000/MW non‑refundable fees and other pay‑to‑play measures—plus the 75 MW threshold—could disadvantage traditional manufacturers. She urges stability and market predictability, and notes ERCOT work to allow interconnection credit when customers bring local generation as a crucial part of a pragmatic solution to serve rapid load growth while preserving market discipline.

Texas energy regulation in spring 2026 is focused on managing rapid, AI‑driven load growth and who will pay for the necessary grid build‑out. A CSIS report by Arushi Sharma Frank frames a regulatory scramble: at an April Senate committee hearing PUCT Chair Thomas Gleeson and ERCOT CEO Pablo Vegas said they were processing ~410 GW of load applications (≈5× current ERCOT capacity), ~90% from data centers. Policymakers are responding with targeted measures — a Dispatchable Reliability Reserve Service (DRRS) (H.B.1500; PURA §39.159(d),(e)) to preserve revenue for gas and long‑duration storage, a ‘denominator effect’ approach to dilute fixed costs by expanding the load base, and a novel batching/flexible connection interconnection regime that enforces project maturity, bankable energization timelines, and as‑available service to accelerate energization. The newsletter also highlights local impacts: San Antonio Express‑News coverage shows renewables providing critical income for rural landowners, even amid political resistance, and flags other trends such as threats to solar and insurance cost pressure from climate risks.

The DCD Mission Critical Power whitepaper (DG Matrix, published 27 April 2026) argues that conventional peak-driven electrical architectures create stranded capacity as AI workloads raise power volatility. It presents software-defined power methods—coordinating grid supply, battery storage and distributed resources—as practical strategies to raise usable capacity and deploy more AI compute without new utility connections.

In the Mar 27, 2026 newsletter, several semiconductor and photonics shifts converged. TurboQuant — Google Research’s KV‑cache compression paper that had been on arXiv for about a year — resurfaced and spooked memory investors, contributing to a sharp pullback even as Micron reported FQ2 revenue of $8.05B, a record 75% gross margin, and FQ3 guidance of $8.8B. Arm unveiled its first in‑house Arm AGI CPU built on TSMC 3nm with 136 Neoverse V3 cores and a 300 W TDP; Arm described rack densities of 8,160 cores (air) and >45,000 cores (liquid) and projected $15B in AGI CPU revenue within five years, with Meta as an anchor customer. Supply stress is visible elsewhere: Intel and AMD raised CPU prices ~10–15% YTD and lead times widened to as much as six months. In photonics, Tower Semiconductor and Coherent demonstrated 400 Gbps/lane (420 Gb/s PAM4) on Tower’s silicon‑photonic process using Coherent InP CW lasers, potentially delaying the need for TFLN modulators; days later GlobalFoundries sued Tower over 11 patents, seeking import bans and prompting share price moves (Tower −7.5%, GF −4.6%). Applied Optoelectronics secured a $53M 800G order (Q2→mid‑Q3) after a $200M+ 1.6T win. Separately, helium spot prices surged over 50% after attacks on Qatar’s Ras Laffan (≈17% LNG capacity hit), spotlighting a material risk for lithography, ion implantation, and leak detection in fabs.

Sudden load losses at data centers prompted the North American Electric Reliability Corporation (NERC) to issue an alert and recommendations after a series of “widespread and unexpected” customer‑initiated reductions in 2024 and 2025. Utility Dive reports that multiple events saw 1,000 MW or more drop off the bulk power system in single incidents, raising concerns about coordination between large commercial customers and grid operators and the potential for cascading reliability impacts. NERC’s response flags the need for steps to mitigate similar future occurrences; the newsletter highlights the scale (≥1,000 MW per event) and timeframe (2024–2025) but does not detail the specific technical measures NERC recommended.

Europe's energy problem has resurfaced after the 2022 Russia–Ukraine shock and the 2026 Iran war, exposing uneven progress across member states. Spain has made measurable gains — a mix of nuclear, flexible gas plants and expanded solar and wind meant gas-set prices fell to 26% of hours in 2025 (from ~50% in 2019, Ember). By contrast Italy still relies heavily on gas (gas set the price in 85% of hours in 2025) and has averaged €130/MWh wholesale in 2026 to date versus Spain’s €47. The piece argues that expanding renewables helps but cannot be the whole solution: northern winters limit solar output, wind is variable and connecting offshore farms requires costly grids; batteries and backup generation create large fixed costs that policymakers must actively manage and decide who bears.

Utility Dive's March 27, 2026 briefing covers grid stress, policy debate and rapid data center growth. At FERC, PJM's data-center colocation proposal was sharply criticized by Vistra and the Data Center Coalition, which warned colocated onsite generation still faces curtailment risk. Winter Storm Fern revealed interregional transmission shortfalls and extreme price divergence — some hubs spiking into the “hundreds of dollars per MWh” while adjacent areas saw negative prices, underscoring calls for expanded interregional transfer capacity. Offshore, Coastal Virginia Offshore Wind has started producing electricity from a single 14.7-MW turbine as part of a 2.6-GW buildout, though PJM transmission upgrades are needed for full deliverability. Meanwhile FERC data show roughly 50 GW of data center capacity online by end-2025, with MISO growing fastest (≈43% annual growth since 2020) and a notable industry shift toward onsite power (≈>33% of developers target 100% onsite by 2030).

FundaAI’s weekly briefing synthesizes a catalyst‑heavy week in which Agentic AI demand materially repriced the hardware and space stacks. Key reported numbers: TSMC raised 2026 revenue growth guidance to >30% YoY with 2026 CapEx at the high end of $52–$56bn; ASML lifted 2026 revenue to €36–40bn and warned of a 2027 EUV shortage; Anthropic’s ARR climbed from >$9bn at end‑2025 to >$30bn in April 2026. Using bottom‑up SOTP and TPU rack BOM analysis, FundaAI models network port‑ratio shifts (OCS scale‑up ~1.5:1 → 2:1–10:1; scale‑out ~0.2:1 → 1:1) and finds TPU TCO >2× better than GB200, concluding that DCN/interconnect — not raw FLOPS — is the immediate bottleneck. Implications: expanded TAM for optics, transceivers, memory and semicap equipment, and major near‑term catalysts including earnings season and Google Cloud Next (Apr 22–24, 2026).

Futurum Group and NVIDIA's report frames AI as a five-layer stack—energy, chips, infrastructure, models, applications—and warns the five largest U.S. hyperscalers could spend up to $690B on infrastructure in 2026 (nearly double 2025). Energy and cooling now outpace silicon as the main bottleneck; Blackwell inference is ~35× cheaper than Hopper, even as token-heavy reasoning and agents drive rising compute needs and reshape workforces and national sovereignty.

Nabiax’s ADC campus in Madrid outlines a 100 MW, AI-ready data center designed for up to 100% liquid cooling and scheduled for Q4 2027 delivery. The DCD whitepaper (5 May 2026) emphasizes secured grid access, on-site substations and live construction to accelerate deployment and address power/permitting constraints for hyperscalers in Spain.

Bechtel and Kiewit were tapped as lead builders for a Japan-backed, $33 billion power-generation buildout in Pike County, Ohio, aimed at supplying roughly 10 gigawatts of new capacity for a forthcoming data center. First announced in October, the investment will channel billions into local infrastructure and construction work to create the generation capacity needed to serve hyperscale computing demand. Reported in Construction Dive's Weekender on March 28, 2026, the assignment positions two major U.S. contractors on a large-scale energy project tied directly to data‑center development and highlights the scale of private international capital flowing into U.S. power and data‑infrastructure projects.

Fred Stafford (@fredstaffordcs) warns that NERC's May 4, 2026 'essential action' Level 3 alert targets data-center loads—especially in Texas where data centers have destabilized the grid and need new standards. He reminds readers that one of three recent Level 3 alerts (last year) concerned inverter-based wind and solar, which have caused mass instabilities when built without key electronic controls; Latitude Media links the alert to AI-driven power demand.

The DCD EMEA weekly review (1 April 2026) highlights shifting demand and fresh supply moves across Europe and beyond. ByteDance reportedly cancelled plans to lease a second Irish data centre, removing a high‑profile hyperscaler demand signal for the market. On the supply side, Nebius has committed to a large‑scale 310 MW AI data centre in Lappeenranta, Finland, targeting initial capacity in 2027 — a project aimed squarely at high‑density AI workloads. In Sweden, Binero has acquired a leased E.On facility outside Stockholm (terms undisclosed), reflecting continued consolidation of regional operator assets. The newsletter also flags major financing and land‑acquisition activity — CoreWeave’s $8.5bn GPU loan and Hyperscale Data’s 48.5‑acre Michigan purchase — and includes industry content on rising rack densities and waterless liquid‑cooling approaches for AI factories.

Crucible Capital argues that the AI-era buildout has exposed decades of deferred industrial maintenance and just-in-time fragility, creating an investment opportunity they call the Industrial Resilience Stack. The fund maps five concrete layers where durable value accrues: (1) Dispatch & Optimization — software/market layers that operate batteries, grid and compute more efficiently (e.g., Shatterdome, Mercury Computing, which runs pilots with Duke and Dominion); (2) Verification & Ground Truth — real‑time telemetry and verified asset valuation for GPUs and cooling (e.g., Aravolta, Reliability Engine); (3) Financial Infrastructure — trade finance, hedging, and transaction ledgers for industrial materials (Matium, Pillar); (4) Materials & Supply — onshoring and new processing tech for critical minerals (Supra’s ion‑specific supramolecular receptors for gallium/scandium, TRISO fuel fabrication, laser‑infused copper/aluminum); and (5) Defense & Infrastructure Security — OT/IT convergence, persistent ISR (Hextronics/Hex) and AI-native forensics. Crucible backs companies that produce ground truth — verified pricing, provenance, and asset condition — because they believe execution commoditizes while verification gains value (citing Christian Catalini’s economics of AGI). The memo uses concrete metrics (1 MW+ rack densities, 18 GW battery installs in 2025, >90 GW pipeline, $3–4M/MW-year lost revenue, DoD 6% supplier visibility, 30% rise in cyberattacks in 2024) to argue the maintenance bill is imminent and that startups embedding into verification, finance, materials, and security will capture durable moats.

PJM’s proposal to create a colocation pathway for data centers has drawn sharp objections from generators and industry groups — notably Vistra and the Data Center Coalition — in comments filed with the Federal Energy Regulatory Commission. Critics say the proposal still exposes colocated customers to curtailment even when they bring on-site generation sized to meet their load, undermining the commercial value of the pathway and potentially deterring participation. That challenge matters as data center capacity has surged (FERC counted ~50 GW online by end of 2025, with the Midcontinent ISO seeing ~43% annual growth since 2020), creating urgent pressure on market rules, interconnection processes and reliability planning as operators decide how to allocate curtailment and reliability risk for large, flexible loads.

The DCD Critical Power Supplement (published 2 April 2026) argues that escalating AI, cloud and HPC demand is making power availability the defining limit on digital growth. It surveys technical responses — solid-state and advanced UPS designs, grid-interactive and hybrid electrical architectures, on-site generation, microgrids, modular infrastructure and intelligent energy management — to enable higher densities, faster 'speed to power' and greater resilience.

Utility Dive’s April 29, 2026 Daily Dive highlights major grid planning and market developments: PJM reopened its revamped interconnection queue (first cycle since 2022) with >800 projects totaling about 220 GW, led by roughly 106 GW of proposed gas‑fired generation. In the Midcontinent, MISO’s summer auction showed offered capacity climbing 3.4% YOY to 141 GW (from 136.3 GW), a supply surge—aided by solar additions—that has driven capacity prices down. On the West Coast, CAISO and PacifiCorp report confidence in EDAM’s imminent launch after 90 days of parallel operations validating the day‑ahead market design. Separately, TVA and Plus Power announced the Crawfish Creek 200 MW / 800 MWh battery project in Alabama, supporting TVA’s target of up to 1.5 GW of storage by 2029. An opinion piece warns an $80B PJM capacity bill risks grid strain without faster replacement generation, grid‑enhancing tech and permitting reform.

NERC's Level 3 alert (published May 5, 2026) highlights computational and data center loads as immediate reliability risks and requires certain transmission and distribution participants to complete seven defined actions by Aug. 3, 2026 to mitigate load-loss impacts. The newsletter also reports policy movement in Pennsylvania, where the state House unanimously approved an advanced transmission technology bill that could obligate utilities such as PPL Electric, PECO and FirstEnergy to deploy ATTs — part of a wave of similar laws across at least nine states. In California, the DOJ has subpoenaed Golden State Wind amid an investigation and anticipated litigation over offshore lease buyouts tied to the Trump lease deal. Corporate items include Schneider Electric attributing revenue growth partly to the AI boom, and Dominion Energy saying its 2.6‑GW Coastal Virginia Offshore Wind began producing in March, should be fully operational by 2027, and may save about $5 billion in fuel over a decade.

Alphabet's Q1 2026 results, reported April 29, were unambiguously strong: revenue was $109.9 billion (22% YoY, 19% in constant currency) and net income rose 81% to $62.6 billion, with EPS of $5.11 versus a $2.62 consensus. Google Cloud posted $20.2 billion in revenue, growing 63% YoY and beating the $18.05 billion Street target; management characterized Cloud as compute‑constrained and said enterprise AI demand has become the primary growth driver. The quarter marked Alphabet's 11th consecutive double‑digit revenue growth quarter, backlog nearly doubled sequentially to $462 billion, and operating margin expanded to 36.1%, prompting a >7% after‑hours share pop.

Data center growth and grid policy headlines dominated Utility Dive’s March 26, 2026 Daily Dive. FERC found 50 GW of data center capacity online at the end of 2025, with MISO exhibiting roughly 43% annual growth since 2020 and rapid increases also in ERCOT, SPP and the Southeast—pressuring transmission and local planning. More than one‑third of developers now expect 100% onsite power by 2030. At the state level, Michigan and New York lawmakers are advancing virtual power plant bills that would bar utility ownership of participating DERs and require reasonable access for third‑party aggregators. The DOE opened a $50 million tribal energy funding opportunity prioritizing projects that supply firm, reliable power. Separately, the EIA reported utility‑scale solar and wind hit a record 17% of U.S. generation in 2025 (19% including small‑scale solar).

Cadence's whitepaper (published 1 May 2026 on Data Centre Dynamics) warns that AI workloads are forcing higher rack densities, rapid power swings, and cooling and capacity shortfalls that risk uptime. It advocates digital twins for planning/commissioning/operations, liquid cooling for dense racks, and smarter infrastructure planning to cut stranded capacity and improve efficiency.

Wendell Hortop (GB Market Lead, Modo Energy) and host Quenton presented a Transmission Today interview-style market update focused on what happened to battery storage revenues in Great Britain over the recent winter and what it implies for the year ahead. Modo’s index (annualized revenue per MW) averaged about £27,000/MW/yr from November to February excluding Capacity Market payments; adding typical Capacity Market value (≈£13,000/MW/yr on average) pushed some assets toward ~£40,000/MW/yr. Those figures contrast with 2023’s ~£57,000/MW/yr and 2022’s ~£153,000/MW/yr—showing a steep drop driven by lower wholesale volatility, reduced peak prices, and growing battery supply (≈3.9 GW online today with 2–3 GW expected this year and >10 GW plausible by 2027).

Hortop attributed the revenue squeeze to three interacting forces: saturation of frequency response (its share fell from >90% historically to ~5–10% this winter), a materially smaller intraday spread (≈£60–70/MWh average this winter, about a third of prior winters), and lower system peak demand (average peak ~37 GW). Interconnectors (≈9.2 GW total, including the 1.4 GW Viking Link) and returning French nuclear and Norwegian hydro suppressed UK peak prices. Operational practice in the control room amplified the problem—batteries were often ‘skipped’ in favour of single, simpler-to-dispatch units (e.g., Dinorwig or pumped hydro). Positive developments include the Open Balancing Platform (launched December), introduction of bulk dispatch, and the March change increasing the dispatch visibility from 15 to 30 minutes. The long-term aim is no fixed‑duration rule: dispatch decisions driven by real‑time state‑of‑charge and algorithmic least‑cost selection. Hortop’s outlook: winter hurt revenues but the Balancing Mechanism should become the main growth channel for battery value as more capacity joins the BM, and summer negative‑price events could restore arbitrage spreads—while asset returns will continue to depend heavily on location and contract terms.

Hyperscalers are dramatically accelerating AI‑driven infrastructure spend: Microsoft, Meta, Amazon and Alphabet collectively guided roughly $700–710B of 2026 capex (about double 2025’s ~$370B), with Microsoft (~$190B), Meta (125–145B), Alphabet (180–190B) and Amazon (~$200B) leading. Much of the increase reflects component price inflation (Microsoft cites ~$25B) and capacity additions; Q1/Q3 year‑over‑year quarter moves show capex roughly doubling (e.g., Alphabet Q1 2025→Q1 2026: $17B→$36B). Two‑thirds of some recent spend is on short‑lived GPUs/CPUs, accelerating depreciation and squeezing free cash flow (Amazon TTM FCF collapsed to ~$1B; Microsoft and Alphabet FCF down 22% and 38%). The picture is clouded by non‑cash equity gains (Alphabet $36.8B, Amazon $16.8B, Microsoft $5.9B) and large off‑balance‑sheet commitments (Moody’s ~$662B leases, Meta +$107B commitments in one quarter), revealing significant prepayments, SPV funding and financial engineering alongside real AI engineering investment.

Cerebras builds wafer-scale AI accelerators (WSE3) that keep large SRAM on‑chip to attack the memory‑bandwidth bottleneck in autoregressive decoding. The WSE3 — a TSMC 5nm wafer with ~4 trillion transistors and ~900k AI cores — exposes 44 GB of SRAM and ~21 PB/s bandwidth; Cerebras cites up to 15× latency improvements versus an Nvidia B200 for inference and extreme speedups (authors claim >1,000× in niche workloads). To address SRAM capacity limits, Cerebras clusters wafers (≈45 CS‑3s to host a 1T‑parameter model; theoretical max 2,048 wafers) and implements manufacturing redundancy to route around defects, at high capital and power cost (node ~$2–3M, ~27 kW).

Commercially the company has two anchor partnerships: a binding AWS term sheet to deliver a disaggregated inference stack pairing CS‑3 with Trainium3 (targeting ~5× token throughput in the same footprint and AWS Bedrock availability), and an OpenAI MRA signed 24 Dec 2025 (deliveries from 23 Jan 2026; Codex‑Spark public 12 Feb 2026). OpenAI committed 750 MW (~$20B over 3 years, ≈$6.7B/yr) with expansion optionality to 2 GW; financing includes a $1B loan and a 33.5M‑share warrant. The article flags two structural weaknesses — SRAM capacity and very high cost per deployment — and positions Cerebras as a low‑latency, premium offering for customers who pay for speed, while noting current modest scale (Abu Dhabi revenues), ~40% gross margin, and ~‑40% EBIT driven by R&D and equity compensation.

U.S. uranium supply chain is heavily import-dependent: in 2023 foreign suppliers met ~95% of reactor purchases, while domestic mines produced 677,000 lb U3O8 in 2024 (~1.4% of demand). Critical processing and enrichment capacity is singular and partly foreign-owned, there is no commercial HALEU in the U.S., and a projected rise to 4M+ lb by 2030 still falls far short of the 433M lb needed from 2025–2035; Congress has pledged $2.7B but rebuilding will take decades (USGS Apr 2026).

Google’s 1Q26 update shows AI-driven cloud acceleration: GCP revenue rose +63% YoY and backlog jumped to $460B (nearly doubled QoQ). Management lifted 2026 CapEx guidance to $180–190B (from $175–185B) and signaled 2027 CapEx will be meaningfully higher, saying the business is currently compute-constrained and that some Cloud revenue was lost for lack of capacity. At-scale TPU monetization has begun in 1Q26 with deliveries into customer data centers; Google is offering its custom TPUs alongside the Axion CPU and NVIDIA GPUs (including the Vera Rubin NVL72). Gemini Enterprise paid MAUs increased +40% QoQ, and management believes AI will expand ad monetization for longer, more complex queries. (AWS/MSFT detail in the post was behind a paywall.)

Google’s new 850 MW Texas data center will be powered via a co‑located AES Clean Energy buildout that, according to an AES filing in December, comprises 600 MW of solar and 945 MW of wind. Thanks to a recent Texas rule change allowing a co‑located load to use a generation project’s interconnection, Google can reportedly piggyback on AES’ grid connection and avoid ERCOT’s huge large‑load queue (225 GW at end‑of‑2025), cutting deployment time to about 18 months. AES’s wind phase is expected online August 2027. The author contrasts this path with Amazon and Meta’s move to build on‑site gas plants to meet fast timelines, arguing Google demonstrates you can achieve similar speed and near round‑the‑clock clean power by combining co‑located renewables with grid reliability; the claim is based on Cleanview‑sourced documents and the AES filing.

Tae Kim argues (Key Context, published 2026-03-30) that short‑term macro fears — from the Iran war to “peak AI capex” talk — obscure a structural pivot: inference is now the dominant workload. Citing Bill Dally and on‑the‑ground GTC conversations with Meta/Google/Nvidia engineers, Kim reports ~90% of data‑center power is going to inference and hyperscalers’ planned “hundreds of billions” in capex still can’t meet demand. Agentic AI (notably coding assistants) multiplies token demand roughly 15× per Arm CEO René Haas and has driven CPU orchestration needs from ~30M to ~120M cores per GW; Intel is locking multi‑year supply deals. Nvidia is central — networking sales +263% YoY, Groq acquisition for ultra‑low‑latency inference, and Vera Rubin GPUs slated for 2026 — while model improvements (Anthropic “Mythos,” larger context windows, multimodal data) and Amazon’s GPU rearchitected ads point to sustained, large‑scale demand (U.S. knowledge work ≈ $6.2T).

Eversource Energy is taking an oppositional stance toward new data center load growth: CEO Joe Nolan said in May 2026 the assets provide “no value” to residential or other customers and will push up rates. That stance sits alongside system-level alarm — NERC issued a Level 3 alert saying computational loads pose “immediate risks” and ordering seven actions by Aug. 3, 2026 to mitigate data center-related load losses. The sector is also wrestling with interconnection and market strain: AEP, with ~63 GW of contracted large load through 2030, is weighing exits from PJM and SPP over slow generation interconnection, while PJM is floating capacity-market overhaul options. Meanwhile, Dominion’s 2.6 GW Coastal Virginia Offshore Wind has started producing (March 2026), aims for full operation by 2027, and expects about $5 billion in fuel savings over a decade even as Q1 energy costs jumped 67%.

Inference compute has moved from supporting experimentation to becoming the dominant production constraint across AI stacks. AINews positions recent comments from industry figures (Noam Brown, Sam Altman, Jensen Huang) and Intel’s Q1 signals as evidence of an "inference inflection": inference token and compute needs have exploded (Jensen cited ~10,000× per‑task compute growth and suggested an aggregate feeling of up to 1,000,000× over two years, with usage up ~100×). SemiAnalysis and pod transcripts underline a CPU refresh cycle issue — enterprises bought ~ $100B of CPUs in 2020–21, creating underinvestment and rising utilization now that simulation, RL gyms, and production agents increasingly run on CPUs.

The market response is visible across models, runtimes, and kernels. The newsletter highlights new model families (Mistral Medium 3.5 dense ~128B with 128K context; IBM Granite 4.1 open weights 30B/8B/3B with cost‑efficient Granite 8B), vendor moves toward prefills/decode disaggregation and specialized inference hardware, and software/hardware co‑design wins: vLLM+Blackwell throughput (DeepSeek V3.2 at 230 tok/s, 0.96s TTFT), Qwen FlashQLA’s 2–3× forward speedups, and serving fixes (e.g., GLM‑5 LayerSplit improving prefill throughput by up to 132%). The piece argues the combination of rising inference demand, falling open‑model pricing, and harness engineering (agent harnesses improving pass@1 and token efficiency) is reshaping procurement, architecture, and economics for AI in 2026.

The April 29, 2026 Load Management briefing highlights major operational and market moves across U.S. power systems: CAISO’s new EDAM day‑ahead market is expected to launch smoothly after 90 days of parallel operations, per PacifiCorp’s Mike Wilding. Retail innovation in Texas continues as Octopus Energy and Lunar Energy introduce plans that pair deeply discounted home batteries with grid‑support services and backup capability. Utilities are also managing shifting demand and costs — Consumers Energy’s sales were lifted by industrial loads even as emergency coal plant orders raised costs and prompted analyst questions about regulatory strategy following DTE’s comments on pausing future rate requests. On infrastructure, CenterPoint targets 8 GW of data center load by 2029, and TVA plus Plus Power announced a 200 MW / 800 MWh Crawfish Creek BESS as part of a path to 1.5 GW of storage. Meanwhile EIA data show retail electricity revenue per kWh jumped sharply in several states (Virginia +26.3%, Ohio +21.9%, Pennsylvania +19.5%) and averaged a 9% U.S. price increase in February.

AI chips sit at the center of modern AI progress: specialized accelerators (Nvidia Blackwell/Hopper, Google TPU, Amazon Trainium, Huawei Ascend) are designed by a few firms but mostly fabricated by TSMC, with critical upstream suppliers such as ASML (EUV) and HBM vendors Samsung, SK Hynix, and Micron. By Q4 2025 roughly 20 million AI chips had been shipped by the five largest designers, with Nvidia supplying ~50% of units and >66% of deployed compute capacity. Metrics matter more than sticker price: although flagship chips rose from $5,700 (2016) to $34,000 (2022), computation per dollar improved dramatically (H100 ≈17× P100), and compute-per-dollar has doubled ~every 2.5 years. Chips consume ~1,000 W each and total datacenter power reached tens of gigawatts by late 2025; energy efficiency has improved ~40% annually, but aggregate electricity use still grows because deployment outpaces per-chip gains.

Large-load tariffs have proliferated as states step up oversight of data center growth: regulators approved 29 such tariffs in 2025 and many more proposals are underway (Utility Dive, Apr 4, 2026). The tariffs aim to allocate grid and transmission costs associated with big, concentrated electricity demand, but experts say it is too early to assess effectiveness. Pushback is mounting from ratepayer advocates and municipal utilities — for example AMP and the Ohio Consumers’ Counsel contested the return-on-equity and proposed rates for a $1.1 billion AEP–FirstEnergy transmission project at FERC, citing exposure of customers to data center-related costs. The shift comes amid broader grid pressures: analysts expect near‑term retail price increases (LBNL/Brattle), FERC reported a 22% fall in solar builds in 2025, and industry forecasts that 1 in 3 data center campuses could exceed 1 GW by 2035, all factors that heighten regulatory scrutiny and affordability concerns.

Scale AI cooling efficiently promotes an on‑demand episode from Data Centre Dynamics' Management & Operations Channel (published 19 Mar 2026) that guides operators on deploying liquid cooling at scale for AI‑driven high‑density racks. It emphasizes improving energy efficiency, cutting power and water impact, early design validation to prevent inefficiencies, and meeting sustainability targets while maintaining performance.

CERAweek conversations captured a Texas grid under rapid and uncertain expansion, where traditional 10–20 year planning cycles have compressed to 12–18 months. NRG’s Gin Kinney described a strategic shift toward speed‑to‑power and flexibility: the company has sourced 5.4 GW of gas turbines, secured Kiewit construction capacity, and is building three gas plants supported in part by the Texas Energy Fund (TEF). At the same time NRG is scaling a one‑gigawatt virtual power plant (VPP) in Texas by aggregating smart thermostats and other distributed assets, using AI and predictive analytics in market operations and home automation to shave peaks, lower customer bills, and avoid some transmission investments.

The wider market picture driving those moves is highly unsettled. ERCOT’s queue and forecasts are volatile — panelists cited the large‑load queue growing roughly 300% last year and headline figures as high as ~410 GW of prospective large loads compared with current summer peaks of ~85.5 GW. Regulators and operators are responding with batch‑zero queue processes, SB6 implementation rulemaking (PUCT Project No. 58317), and pilots such as ERCOT’s ADER program to target nodal congestion. The consequence: many hyperscalers and data centers are adopting BYOP (bring‑your‑own‑power) or bridge‑power strategies for speed to commercial operation; a striking example is a ~350 MW facility near El Paso temporarily powered by ~800 small ~450 kW generators. That fragmentation raises logistics, maintenance, and cost‑allocation questions.

Taken together, the episode argues that distributed flexibility (VPPs, residential smart homes, EVs as potential storage) can substitute for some new centralized generation and defer T&D build—if it can be scaled, contracted, and remunerated correctly. Speakers emphasized the need for cross‑sector partnerships (generators, retailers, T&D, aggregators), clearer policy/PUC direction, and robust interconnection processes to reconcile ambitious load growth claims with the physical limits of supply chains, permitting, and transmission build timelines.

Epoch AI's AI Chip Owners explorer (launched 2026-04-06) maps who owns the world's leading AI chips by distributing previously estimated sales volumes (Nvidia, Google TPU, Amazon Trainium, AMD, Huawei) across major owners using analyst estimates, company disclosures, capex data, and frontier data-center analysis. The study finds US hyperscalers control over 70% of AI compute, led by Google (≈5 million H100-equivalents, ~25% of global compute) with Google TPUs contributing nearly 4 million H100-equivalents (CI 3.1M–4.5M). Mainland China holds just over 5% of cumulative compute as of end-2025; illicit imports may have added tens of thousands to >100,000 A100/H100s in 2024 per prior reports but are unlikely to close the gap. The explorer includes interactive visualizations, a methodology and dataset for researchers and policymakers.

Subject: Optical and AI infrastructure scale‑up — Micron HBM4, OCI MSA, Credo, Lumentum, NVIDIA GB300, Nebius‑Meta deal. The newsletter summarizes OFC and GTC headlines (Mar 2026) that signal simultaneous scale‑up in memory, optics, and enterprise AI infrastructure. Micron is now publicly supplying HBM4 to NVIDIA’s Vera Rubin (NVL576) and was praised by Jensen Huang as delivering the “industry’s fastest HBM4,” reversing earlier rumors that Micron would not ship HBM4; Micron’s stock was ~62% higher YTD.

On optics, industry players formed the OCI MSA to standardize Optical Compute Interconnects and avoid vendor lock‑in, with Lightmatter pursuing complementary OCP CPO specs. Vendors pushed new hardware: Credo released Cardinal 1.6T and Robin 400/800G DSPs plus ZeroFlap 800G transceivers; Lumentum demoed 1060 nm VCSELs for a “slow‑but‑wide” parallel‑lane scaling approach (GaAs VCSEL arrays as an alternative to scarce InP EMLs). In infrastructure, NVIDIA’s DGX Station GB300 targets on‑prem enterprise develop‑to‑deploy inference, and Nebius struck a five‑year agreement with Meta for up to $27B of AI capacity ( $12B committed, up to $15B additional), with Vera Rubin cluster deliveries starting early 2027.

Anthropic’s growth narrative has gone hyperbolic: the AI lab reported a >$30 billion annualized revenue run rate in early April 2026 after rising from ~$1B at the start of 2025 and $9B at year‑end 2025, with intermediate milestones of $14B and $19B in February. The company recently capped peak‑usage and limited third‑party harnessing of subscriber compute while signing a 3.5‑GW deal with Google Cloud (up to ~1M TPUs) to be manufactured by Broadcom, capacity that will mostly arrive in 2027 and that Wilhelm values at roughly $35B by his GW cost heuristic. Wilhelm cautions that Anthropic’s headline revenue may overstate GAAP‑style economics because of how cloud‑partner revenue is counted.

At the same time, geopolitics and economics are converging: U.S. legislators’ MATCH Act would curtail exports of critical SME (DUV/EUV and cryogenic etch tools), pressuring ASML and likely accelerating Chinese tooling programs. OpenAI has proposed a policy package—higher capital taxes, targeted automated‑labor levies, wage‑linked incentives, a public wealth fund, portable benefits, and AI access as public infrastructure—framed as a new social contract to manage predicted labor disruption (Dario: 50% of entry‑level white‑collar roles disrupted in 1–5 years) and political risks voiced by investors like Vinod Khosla. The piece ties rapid commercial AI scale, scarce compute, infrastructure externalities, and tax/political choices together as central determinants of whether AI’s economic gains translate into broad prosperity or concentrated capital power.

Across techniques they find typical trade spans of order 1–2 OOM in each axis: (a) Chinchilla-style scaling can trade ~1.2 OOM training for ~0.7 OOM inference; (b) MCTS exhibits an S-curve where at low performance ~1 OOM training ↔ ~1.6 OOM inference (and the trade shrinks as models near perfect play); (c) IMP pruning to ≈10% density saved ≈1 OOM inference at ~0.7 OOM extra training because of multi-round re-training; and (d) resampling with cheap verification (e.g., code generation pass@k) extrapolates to much larger spans (authors estimate up to ~3 OOM training savings vs ~6 OOM inference increase) but caution low confidence in extrapolation. The authors categorize three tradeoff archetypes—saturating/constant span (e.g., Chinchilla), saturating/decreasing span (e.g., MCTS, limited n@k), and non-saturating/increasing span (resampling with effectively unlimited trials)—and show that some techniques combine roughly additively but often with diminishing returns, so realistic combined spans are ~2–3 OOM. Methodologically they fit empirical curves (smoothly broken power laws, log relations) and prioritize Pareto frontiers for optimal TC/IC choices. The paper emphasizes governance implications: because aggregate inference cost often exceeds training cost in deployed systems, commercial models will be biased toward low-IC configurations, but researchers or well-resourced actors can invest IC to simulate larger-model capabilities for evaluation, internal use, or small-customer deployments—affecting safety assessments and policy proposals aimed at controlling frontier capabilities.

Matthew Yglesias (Mar 24, 2026) argues that the apparent muted move in headline oil benchmarks masks acute real-world disruption: more than 20% of global oil transits the Strait of Hormuz, and a supply loss that large would, by his reckoning, require prices nearer $150/barrel to force a 20% demand cut versus the roughly $100/barrel price cited. Short‑run gasoline elasticity is low, so consumption falls mainly via product shortages, not immediate consumer behavior. The disruption disproportionately affects heavier crudes feeding Asian refineries, causing rationing (India’s commercial LPG curbs), jet‑fuel‑related flight cancellations, naphtha and marine‑fuel shortages, and refinery windfall margins. Futures are restrained by a 'TACO' expectation that President Trump would politically intervene if U.S. pump prices spike, while post‑sanctions Urals crude doubled in price and niche blends trade ≈$10/bbl over Brent.

The Nuclear Company and Brookfield announced a joint venture on 2026-05-04 to deliver GW-scale nuclear power, including Westinghouse reactors and management of AP1000 completions in South Carolina. The author, serving as Chief of Staff to CEO Jonathan Webb and working with CNO @TheNuclearJoe, emphasizes repeatable, proven designs, disciplined project management, on-time/on-budget delivery, and use of emerging technologies.

AI CPUs are driving a new surge in DRAM demand as the industry pivots from training-centered GPU farms to inference architectures that require large "context memory." Vendors aim to pack 300–400 GB of commodity DDR5 into next‑generation CPUs — up to four times current CPU capacities — while GPUs and accelerators are also ballooning HBM capacity (NVIDIA Vera Rubin 288 GB, AMD MI400 432 GB, Google TPU 8i 288 GB). Server ratios are shifting from 8:1 (GPU:CPU) toward 1:4 and even 1:1 for inference coordination. Spot-price evidence shows DDR5 holding a premium (DDR5 +2.8% vs DDR4 −16% in April on a 16 GB basis), and an industry source estimates roughly a 10 percentage‑point supply shortfall. The post concludes that memory suppliers (Samsung, SK hynix) are unlikely to catch up in 2026, pushing the shortage into 2027.

SemiAnalysis proposes a practical, bottom‑up Total Cost of Ownership (TCO) framework for GPU clusters that goes beyond headline $/GPU‑hr. Their monthly TCO formula sums GPU rental, storage (hot/warm/cold), networking, control plane, support uplift, plus two categories of implicit costs: Goodput Expense (lost useful work from failures and inefficiencies) and amortized engineering expenses for setup and debugging. The firm published two interactive tools — a GPU Cluster TCO Calculator and a Goodput Calculator — and populated defaults from an August 2025 pricing snapshot, hands‑on tests of 80+ neoclouds, and interviews with >150 users.

They evaluate three representative provider tiers (Gold‑tier, Hyperscaler, Silver‑tier) across three scenarios: Large LLM Pretrain (5,184 NVL72 GPUs, ~80% allocated; 500 TiB hot + 10 PiB cold storage; $4/GPU‑hr input), Multimodal RL Research (2,048 B200 cluster with ~12 TB/GPU storage; neocloud $2.40 vs hyperscaler $3.10/GPU‑hr), and Inference Endpoints (512 GPUs, 1 TB/GPU). Key results: holding GPU price equal, gold‑tier TCO was roughly 5–15% lower than silver in large training; in the pretrain scenario multi‑year cost ratios were Gold 1.00x, Hyperscaler 1.10x, Silver 1.15x. Goodput modeling—with formulas for checkpoint‑hot, checkpoint‑cold, and fault‑tolerant cases—shows large jobs suffer disproportionately from MTBF, detection latency, repair time, checkpoint frequency and blast radius: in the Large Pretrain example SemiAnalysis reports goodput losses of ~6.14% (gold), 10.53% (hyperscaler), and 20.91% (silver).

On fault tolerance, they compare TorchFT (Meta open source; easier recovery but >10% comms overhead due to GLOO vs NCCL), AWS SageMaker HyperPod checkpointless (announced Dec 2025; AWS claims ≈1m45s recovery and uses model redundancy over EFA), and TorchPass (commercial; no perf hit but requires idle spare nodes — their test used 32 spare GPUs, ~0.62% of cluster). The analysis concludes that hidden line items (support tiers, orchestration premiums, setup/debug engineering, and goodput losses) often make ostensibly cheaper GPU‑hour offers more expensive in practice. ClusterMAX 2.1 (Apr 2026) adds several providers (Core42, BitDeer, FPT, Radiant/Ori, etc.), and SemiAnalysis plans broader ClusterMAX 3.0 testing and MTBF data collection this summer.

The presenter flags concrete safety and program‑management concerns: Valor’s higher operating temperature (~900°C) raises unanswered questions about fuel integrity, oxygen ingress, and neutron‑accelerated material aging that would normally be explored in NRC failure‑mode analyses. He also notes an operational incident tied to Valor’s head of operations in 2021 as a data point on quality culture. Economically, the talk argues SMRs struggle without mass production — citing China’s heavy investment in ~1 GW pressurized water reactors — and predicts many startups will fail for market reasons rather than regulatory blockage. The recommended middle path is to streamline but not eliminate expert review (for example, require an initial NRC phase review) so innovators can iterate while retaining independent safety assurance.

Flashing Orange, by Doomberg (Apr 16, 2026), warns that Europe is facing a high-risk natural gas crunch: the EU finished winter 2025–2026 with very low inventories while a war in Iran has effectively closed the Strait of Hormuz and taken ~20% of global LNG offline. Doomberg aggregates EU+2 production and consumption data to show the region produces under half its gas needs, leaving a deficit near 20 bcf/d (roughly the Permian Basin's output). The newsletter contrasts this vulnerability with Germany's lost nuclear capacity — 19 reactors that once delivered ~170 TWh/year — calling the phase-out a geopolitical own goal. The author says the refilling season leaves little room for error, expects higher energy prices and continued deindustrialization, and reports having mapped major inflows and risks (the detailed findings are paywalled).

Microsoft's Q3 FY2026 results beat expectations across revenue, cloud, and profit metrics: revenue totaled $82.9B (+18% YoY), Microsoft Cloud brought in $54.5B (+29% YoY), and Azure accelerated to 40% YoY growth (39% in constant currency), outperforming prior guidance of 37–38% CC. Operating income reached $38.4B (+20% YoY) with a 46.3% margin, and GAAP diluted EPS was $4.27 (+23% YoY) after a Q2 spike tied to an OpenAI investment gain (non‑GAAP trends show underlying growth). Q3 CapEx including finance leases was $31.9B (below the $34.9B consensus), but CFO Amy Hood's FY2026 CapEx guidance of $190B — roughly $35B above prior market expectations — drove a negative market reaction. Commercial remaining performance obligations surged to $627B (+99% YoY), the largest backlog Microsoft has reported.

Canada announced a national nuclear framework at the Canadian Nuclear Association conference on April 29, 2026, with a final strategy due by end‑2026, committing to full‑spectrum nuclear policy (new builds, exports, fuel‑cycle expansion, fission and fusion R&D). Measures include DND’s $40M microreactor assessment and $2.2B for Chalk River over 10 years; Canada already supplies ~24% of global uranium.

DCD's April 24, 2026 Cloud & Hybrid monthly newsletter spotlights major shifts in AI infrastructure and enterprise cloud strategy. Google unveiled eighth‑generation TPUs — two distinct, purpose‑built chips for training and inference — aimed at boosting performance for large model workloads. In enterprise moves, Stellantis signed a five‑year deal with Microsoft Azure to migrate workloads and target a 60% reduction in its data‑center footprint. Anthropic is scouting data‑centre leases in Europe and Australia as it scales capacity and OpenAI retreats from Stargate Europe. The issue also flags industry developments including Vast Data’s $1bn Series F at a $30bn valuation, Google breaking ground on a Kronstorf, Austria data centre, and an NTT guide mapping India’s rapid hyperscale expansion, underscoring ongoing geographic and vendor shifts in compute and real‑estate strategy.

Om Malik's May 4, 2026 essay (5,000-word), "Say Hello to the Internet of AI," argues AI is 'supersizing' network scale and the capital needed for competitive advantage. Drawing on interviews with networking and infrastructure insiders, he finds the major impacts are in data centers, cloud and backbone (physical pipes), shifting demand profiles and ownership models—more infrastructure-centric than consumer-bandwidth-driven.

Scale for triple‑digit rack densities (DCD, 8 May 2026) promotes a whitepaper arguing AI workloads are pushing data centers toward triple‑digit rack densities and gigawatt‑scale expansion. It advises treating infrastructure as a single 'unit of compute' by aligning power, thermal and IT layers to manage extreme densification and enable next‑gen high‑performance deployments.

Cerebras is pursuing an IPO after re‑filing with materially different dynamics: strong top‑line growth and strategic hyperscaler/lab deals have reduced its historical concentration risk. The company posted $510.0M in 2025 revenue (hardware $358.4M, cloud/services $151.6M), with sequential quarterly growth of +31% and +26% in the latest two quarters, though gross margin slipped to 39.0% as costs rose. Historically dominated by Abu Dhabi customers (G42/MBZUAI), Cerebras closed a January 2026 multi‑year agreement with OpenAI to deploy 750 MW (option to scale to 2.0 GW) that includes a $1B term loan and warrants, and a March 2026 AWS partnership to deploy CS‑3 systems integrated with Trainium3 for disaggregated inference. With $24.6B RPOs and near‑breakeven results excluding stock‑based comp, the IPO case rests on continued expansion of AI inference demand and adoption of purpose‑built wafer‑scale inference chips.

Load Management weekly (Utility Dive, May 6, 2026) spotlights immediate operational and planning pressures on U.S. grids: NERC raised a Level 3 alert centered on rapidly growing computational/data center loads and ordered specific participants to complete seven actions by Aug. 3, 2026 to mitigate risk. Retail programs and local DERs are presented as practical load‑management responses — TXU Energy’s Texas EV plan provides Ford drivers 15 hours/day of “free” home charging and Ford reports shifting 515 MWh to off‑peak in 2025 — while Ann Arbor’s new Sustainable Energy Utility is deploying locally sited solar and batteries to improve reliability and lower subscriber costs. Meanwhile, American Electric Power is reconsidering membership in PJM and SPP amid slow interconnection timelines as its systems absorb contracted large loads totaling 63 GW by 2030. The newsletter also highlights calls (eg., Jigar Shah) to scale front‑of‑meter storage and strengthen cyber resilience as complementary strategies.

ARM’s Mar 24, 2026 Analyst Day reframed the AI data center around CPUs, arguing the shift to "agentic" workloads will multiply CPU deployment (from ~30M to ~120M cores per gigawatt) and double the Data Center CPU TAM to >$100B by FY2031. Management unveiled a strategic move from IP licensing to Compute Subsystems and announced a commercial ARM AGI CPU (not yet shipping), two rack designs (air: 60 CPUs / 8,160 cores; liquid: 336 CPUs / 45,696 cores), and customer endorsements from Meta and OpenAI — which together catalyzed market moves in ARM, Intel and AMD shares after the event.

Beyond The Hype criticizes both product and financial narratives. The AGI CPU’s disclosed specs are described as pedestrian and not competitive with AMD Venice except possibly on TDP; ARM’s 2x perf-per-watt claim vs x86 is labeled implausible without third‑party benchmarks (expected H2 2026). The newsletter also disputes ARM’s revenue and royalty math: royalty CAGR is projected to accelerate from 14% to 20%, AGI CPU revenue is small until FY28 (~$1B), and management forecasts $15B of $25B coming from AGI CPUs by FY2031 — a mix the author says would compress margins, contradicts claimed >65% operating margins, and ignores hyperscaler preference for in‑house ASICs, Qualcomm legal setbacks, Apple’s low rates, a 2026 memory-driven device decline, and potential Ampere acquisition costs.

Alphabet and Microsoft led Thursday’s earnings beat driven largely by enterprise AI demand: Alphabet reported Q1 2026 revenue of $109.9B (vs $107.1B expected) and EPS $5.11, with Google Cloud at $20.0B (63% YoY), operating profit rising to $6.6B, a cloud backlog >$460B, API traffic at 16 billion tokens/min, Gemma 4 at 50M downloads, and raised 2026 capex guidance to $180–195B; Waymo now does 500k paid rides/week. Microsoft posted $82.9B revenue and $4.27 EPS, Azure growth ~40%, Microsoft Cloud $54B (+29% YoY), an AI business >$37B ARR (+123%), >20M Microsoft 365 Copilot paid seats, GitHub Copilot in ~140k orgs, $31.9B quarter capex and ~ $190B calendar-year capex target. The newsletter also flags a congressional probe into U.S. firms using Chinese models (Anysphere/Cursor used Moonshot Kimi K2.5; Airbnb used Alibaba Qwen), a move that could favor domestic labs but raise costs and slow inference-price declines.

TPU v8, Intel's comeback, and Marvell's Polariton acquisition are the focus of Vikram Sekar's Apr 24, 2026 TWiC. Google introduced two TPU v8 SKUs — v8t for training and v8i for inference — noting the 8i has substantially larger on‑chip SRAM (3× the training chip) and higher HBM capacity, and uses Arm Axion CPUs as head nodes. Google is also shifting networking: Virgo (OCS, high radix) replaces Jupiter for scale‑out, while a Boardfly approach connects 4 chips/board × 8 boards/rack × 36 groups/pod for 1,152 TPUs per pod; training retains a 3D torus. Intel’s AI business now represents ~60% of revenue with 40% YoY growth, plus EMIB packaging demand and improving 18A yields; Eric Demers was hired as Chief GPU architect in 2026. Marvell bought Polariton to deploy SPP‑based plasmonic phase shifters that promise lower capacitance, faster modulation, and reduced power for photonics.