Briefing

Frontier AI training power demand is projected to reach 4–16 GW for the largest single training runs by 2030, per an EPRI-linked white paper (Aug 2025). The forecast combines 4–5x/year compute scaling with continued GPU efficiency gains (~26–40%/yr) and a slowdown in training-duration growth (10–20%/yr), yielding net power growth of ~2.2–2.9x/year and >100 GW total AI capacity globally.

The Stargate Project is a $500 billion initiative led by SoftBank and OpenAI to build U.S. AI infrastructure over four years, deploying $100 billion immediately to expand compute, create hundreds of thousands of jobs, and bolster national-security capabilities. Initial equity includes SoftBank, OpenAI, Oracle, and MGX; partners include NVIDIA, Microsoft, Arm, and Oracle; buildout begins in Texas and RFP/RFQ details were released Jan 31, 2025.

InferenceX v2 is SemiAnalysis’ expanded open-source benchmark suite for frontier LLM inference, aimed at measuring not just peak throughput but the full tradeoff curve between interactivity, throughput, cost, and energy efficiency across modern GPU systems. The new release covers all six recent NVIDIA western SKUs and all AMD western GPU SKUs from the past three years, using close to 1,000 GPUs per full sweep. Its headline result is that NVIDIA’s rack-scale Blackwell systems—especially GB200 and GB300 NVL72—massively outperform both Hopper and AMD once production-style inference techniques are enabled, particularly for mixture-of-experts models such as DeepSeek R1. SemiAnalysis argues that the real differentiator is not just chip FLOPS, but the ability to combine disaggregated prefill, wide expert parallelism, FP4 quantization, and mature multi-node serving software such as Dynamo and TensorRT-LLM. In that setup, GB200/GB300 NVL72 can deliver up to 98-100x the realized performance of strong H100 baselines and as much as 65x better tokens-per-dollar. The architectural reason is straightforward: NVL72 keeps 72 GPUs inside a single 900 GB/s-per-GPU NVLink domain, allowing expert-parallel all-to-all traffic and weight loading to stay on a far faster fabric than the 400-800 Gbit/s scale-out interconnect used between standard nodes.

AMD’s position is more nuanced. MI355X appears genuinely competitive in single-node FP8 and in FP8 disaggregated serving when compared specifically against NVIDIA running SGLang, and SemiAnalysis notes rapid improvement in AMD’s software stack over the prior two months, including ~2x gains in some DeepSeek R1 FP4 configurations and >20% throughput-per-GPU gains from MoRI in the 20-45 tok/s/user band. But the report’s core criticism is that AMD still lacks “composability”: isolated optimizations work, yet combining FP4, disaggregated serving, and wideEP causes performance to degrade sharply. That leaves MI355X far behind B200 once NVIDIA’s production stack, especially Dynamo + TRT-LLM, enters the picture. The report also critiques AMD’s upstream support model, noting MI355X still depends on an old forked vLLM 0.10.1 ROCm image while the official 0.15.1 image hard-fails, and citing insufficient CI hardware donations to projects like vLLM and SGLang.

The article is also valuable for its system-level economics. It shows why disaggregated prefill improves utilization by separating compute-heavy prefill from memory-bound decode, and why wideEP reduces redundant model weight loading across MoE deployments. It frames Anthropic’s “fast mode” as an inference scheduling decision rather than a hardware mystery: serving the same model at 2.5x higher tok/s/user naturally drives 6-12x higher cost per token because accelerator hourly cost is fixed while batching efficiency falls. Multi-token prediction is presented as the most powerful software lever in the current stack, often slashing cost per million tokens by multiples while preserving accuracy on benchmark checks like GSM8k. For anyone trying to understand who actually wins the AI infrastructure race, the main lesson is that rack-scale topology, interconnect bandwidth, software maturity, and inference orchestration now matter as much as the raw silicon.

ERCOT’s market design is becoming one of the most important pieces of Texas infrastructure policy because, in an energy-only market, reliability policy doubles as investment policy. In this interview, Potomac Economics deputy director Andrew Reimers explains that scarcity pricing and the operating reserve demand curve are supposed to send the signal for new generation and flexible capacity to get built. The problem is that Texas has spent the post-Uri period leaning toward more conservative operations—keeping more reserves online to reduce outage risk and even avoid politically toxic conservation alerts. That may improve short-run comfort, but it can also undermine the scarcity prices that tell investors the market needs more supply. Reimers argues this tension is especially acute in ERCOT because Texas is electrically isolated, has unusually high penetrations of wind, solar, and storage, and cannot rely on neighbors the way PJM, CAISO, or New York can.

The major recent structural change is real-time co-optimization, which went live on Dec. 5, 2025 after years of advocacy from Potomac and others. Before RTC, operating reserves were procured in the day-ahead market and carried as real-time physical obligations, which could create severe price distortions because reserved capacity was effectively withheld from the real-time energy market. Reimers points to summer 2023, when ERCOT’s addition of ECRS pulled thousands of megawatts out of the energy stack and produced what Potomac estimated were billions of dollars in excess consumer costs. Potomac now supports NPRR 1309, which treats DRRS as a true operating reserve product, but strongly opposes NPRR 1310, which would layer a capacity-like resource adequacy mechanism into hourly reserve procurement. His core objection is methodological: real-time or day-ahead operating conditions are appropriate for pricing present scarcity, but not for setting durable forward signals to induce efficient future investment.

The conversation also highlights how batteries and forecasting are reshaping ERCOT’s operational logic. Solar has lowered prices during high-load daylight hours and shifted reliability stress to the evening net-load ramp, where roughly 15 GW of batteries now help bridge the loss of solar output. But Reimers says forecast error—not peak demand alone—is now the key variable, especially with wind, whose output is harder to predict. That is why Potomac favors a multi-interval real-time market that looks ahead one to two hours, allowing batteries to preserve state of charge when the system expects tighter conditions later. He argues ERCOT’s current four-hour duration requirement for non-spin is a crude substitute that can misprice storage and unintentionally push batteries into energy sales rather than reserve provision. More broadly, Reimers warns that out-of-market programs—from emergency reserves to residential demand response and firm-fuel constructs—can dilute the transparency and allocative efficiency of scarcity pricing by paying resources through side channels instead of through the core market.

Texas' power grid faces a January 23, 2026 winter storm expected milder than 2021's Uri; ERCOT showed ≥7 GW reserve and >9 GW thermal outages Friday, while battery capacity jumped to ~17,000 MW from 220 MW in five years.

Utility-sector headlines: AEP’s contracted large-load pipeline surged to 56 GW (Texas data-center pipeline 36 GW by December vs. 13 GW in Q3), while Exelon is planning $41.3B in capital spending with $12–17B more in transmission ‘line of sight.’ Treasury issued interim FEOC/MACR guidance and GAO warned DOE lacks an oversight plan for ~$21.5B in IIJA demonstration funds.

SpaceX's plan to build a 1,000,000‑satellite constellation as orbital data centers (reported 2026-02-05) promises massive distributed AI compute but intensifies collision and debris risks, requiring new traffic-management and debris‑mitigation engineering. Concurrently, India is offering zero taxes through 2047 to host foreign AI workloads, courting major cloud vendors despite power/water limits, while a conviction of ex‑Google engineer Linwei Ding underscores rising espionage threats to accelerator and data‑center IP.

Anthropic's 'Build AI in America' report sets concrete infrastructure targets and policy fixes to keep U.S. AI leadership: it projects 2 GW and 5 GW single-model training sites in 2027–2028, 20–25 GW of frontier training demand by 2028, and a need for ~50 GW total capacity; it recommends accelerated permitting, federal land use, transmission upgrades, and supply-chain and workforce programs.

Epoch AI's Frontier Data Centers database (updated Feb 19, 2026) compiles satellite imagery, permits, and public documents to estimate power, compute, timelines and costs for major U.S. AI campuses. It reports ~2.5M H100e-equivalent tracked (15–17% of shipped compute as of Nov 2025), an 80% confidence range of ±50%, GW-scale builds in 1–3.6 years, and site-level projections such as Fairwater at 3.3 GW.

AI data centers are becoming some of the largest infrastructure projects ever: concentrated campuses like OpenAI’s Stargate Abilene require roughly 1 GW of continuous power, enormous land (3.5 km²), and huge capital (~$32B). Aggregate US demand for frontier training could reach 20–30 GW by late 2027, driving siting toward jurisdictions with abundant, fast‑deployable power (e.g., Texas, China). What distinguishes AI centers is very high power density — racks such as NVIDIA’s NVL72 hold 72 GPUs and can draw >100 kW each — necessitating liquid cooling, chiller systems, and significant water management. Climate and security effects are currently localized (AI uses ~1% of US power now) but could grow to ~5% by 2027 if fossil fuels remain primary sources; satellite‑visible cooling infrastructure also complicates secrecy and physical security.

OpenAI’s Stargate program targets 10 GW of U.S. AI capacity by 2029 and reports it is already beyond halfway to that target one year after the Jan 2025 launch; the Abilene, Texas site is already training and serving frontier models. Stargate pledges to fund incremental generation, storage, and grid upgrades, minimize water use with closed-loop cooling, and create workforce pipelines via OpenAI Academies (first in Abilene, spring 2026).

OpenAI and SoftBank Group each committed $500 million to SB Energy and OpenAI leased 1.2 GW for a Milam County AI data center, part of SB Energy’s multi-gigawatt campus program. Initial facilities target service entry in 2026, emphasize reduced water usage, job creation, and pair OpenAI’s data center designs with SB Energy’s integrated energy delivery; Ares provided $800 million in preferred equity.

OpenAI and Broadcom announced a multi-year collaboration to co-develop racks featuring 10 GW of OpenAI‑designed AI accelerators paired with Broadcom Ethernet, PCIe and optical networking, aiming to start deployments in the second half of 2026 and finish by end of 2029. OpenAI intends to bake model-driven architecture choices into the custom silicon to improve performance and efficiency for large-scale training and inference.

Google's purchase of a roughly $1 billion, 300 MW / 100‑hour (30 GWh) iron‑air battery from Form Energy for its Pine Island data center implies an all-in capital cost near $33.33/kWh. The iron‑air system uses reversible rusting (iron + oxygen) to deliver long‑duration storage, aiming to beat lithium‑ion on multi‑day firming costs and target ~$20/kWh commercially.

Stargate is expanding via a new OpenAI–Oracle agreement to add 4.5 GW of U.S. AI data‑center capacity, bringing total Stargate development to over 5 GW and supporting more than 2 million chips. Oracle has started delivering Nvidia GB200 racks and OpenAI is running early training/inference workloads; OpenAI projects 100,000+ jobs and cites a prior White House $500B/10 GW commitment.

Standard Intelligence Team deployed a 30 PB, 100Gbps storage cluster (2,400 HDDs, 100 DS4246 chassis, 10 head nodes) to store ~90M hours of video for pretraining, costing ~$426.5k capex and $29.5k/month including depreciation and $17.5k/month internet+power. They used simple software (200-line Rust writer, nginx, SQLite, XFS), prioritized simplicity over redundancy, and report saturating their 100G link while avoiding complex systems like Ceph.

Nvidia’s shift to 800V DC power architecture is the centerpiece of Crucible Capital’s thesis about the next phase of AI datacenter design. The argument is that AI workloads have pushed facilities beyond the comfortable operating envelope of legacy AC distribution. Traditional datacenters were built around 415V/480V three-phase AC and modest rack densities of roughly 5-20 kW, but modern AI clusters are now driving rack power into the hundreds of kilowatts and, on the authors’ timeline, toward 1 MW-class systems with Rubin Ultra in 2027. In that regime, repeated AC-to-DC and DC-to-AC conversions become a serious efficiency penalty, with the article citing 10-20% aggregate losses. Nvidia’s proposed architecture converts grid AC to 800V DC once at the site perimeter, distributes DC through busways, and then steps down locally to 54V/12V at the rack. That reduces conversion stages, current, copper requirements, and thermal waste. The piece also emphasizes the strategic role of OCP and Nvidia reference architectures in standardizing not just GPUs, but the entire rack-power-cooling stack, effectively forcing suppliers such as Schneider Electric, SuperMicro, Vertiv, and others to align on Nvidia-led design choices.

The second major theme is that electrical redesign and cooling redesign are inseparable. As rack densities move from traditional 10-50 kW toward 100 kW-1 MW+, the rack itself becomes the constraint rather than the silicon. The article highlights several enabling technologies: solid-state transformers based on silicon carbide or gallium nitride, supercapacitors for sub-second transient smoothing, LFP battery systems for longer backup and load shifting, and 800V-compatible rack sidecars and power shelves. Claimed benefits include 30-50% smaller SST footprint, 150% more power flow through existing conductors, over 98% DC-DC conversion efficiency, up to 60% more compute space from removing legacy PSUs, and 45% less copper in redesigned racks. The article also surfaces a supply-chain angle, especially around gallium, noting China’s dominance in gallium separation as of 2024 and framing advanced power electronics as a critical-minerals problem as much as an electrical-engineering one.

Cooling is treated as the practical bottleneck. Air cooling is portrayed as maxing out around 15-20 kW per rack, making liquid systems mandatory for current and future AI clusters. Direct-to-chip liquid cooling is already standard for Nvidia H200/B200/B300 systems, and Nvidia’s warm-water approach at 45°C is positioned as especially important because it can reduce or eliminate chillers. The authors cite a specific startup example, Elkhorn, whose water-based system reportedly achieved a COP of about 11.8 at an AI datacenter in Newport, Washington, halving cooling power from roughly 170 kW/MW of IT load to 85 kW/MW and improving PUE from 1.18 to 1.09. Beyond facility mechanics, the article stresses software and maintenance: as rack density grows by a claimed 90x over a datacenter’s life, operators need sub-second telemetry, asset-level monitoring, and maintenance orchestration to protect hardware that can cost tens of millions of dollars per MW-scale cluster. Overall, the report is strongest as a synthesis of how AI demand is pushing simultaneous changes in power electronics, storage, cooling, supply chains, and operational software.

Datacenter CPUs are becoming strategically important again, not because GPUs stopped mattering, but because AI systems have become more CPU-hungry around the GPU cluster. SemiAnalysis argues that since late 2025, reinforcement learning, agentic inference, retrieval-heavy workloads, multimodal preprocessing, and the management of petabyte-scale data pipelines have all increased CPU intensity. In training, CPUs execute RL environments, compile and verify code, run math and physics simulations, and keep expensive accelerator clusters fed. In inference, agents generate far more external API calls, database lookups, and internet traffic than classic chatbot serving. The article’s most concrete infrastructure example is Microsoft’s Fairwater site for OpenAI, where a 48 MW CPU/storage building supports a 295 MW GPU cluster, implying CPU demand is no longer incidental to AI buildouts. That demand is also colliding with mainstream cloud consolidation: newer cloud-native CPUs can replace old Intel Cascade Lake fleets at 10:1 socket consolidation ratios or better while using less than one-fifth the power, freeing capacity for GPUs.

On product competitiveness, the piece is notably bullish on AMD and skeptical of Intel. It presents Venice as AMD’s strongest step yet: eight TSMC N2 Zen 6c CCDs, 256 cores, 16 memory channels, 1.64 TB/s with MRDIMM-12800, restored 4 MB L3 per Zen 6c core, and a claimed 1.7x perf/W gain over Turin. SemiAnalysis also highlights AMD’s willingness to keep serving the mainstream 8-channel enterprise segment with a Venice SP8 platform, just as Intel has reportedly cancelled Diamond Rapids-SP. Intel’s roadmap is framed as architecturally ambitious but commercially compromised: Diamond Rapids adopts a more AMD-like multi-die topology with four compute building blocks and two memory/I/O hub dies, but loses SMT on P-cores, which the authors think badly undermines datacenter throughput. Clearwater Forest, meanwhile, is treated as a costly learning vehicle for 18A and Foveros Direct rather than a volume winner, with only 17% better performance than Sierra Forest after a delay into H1 2026.

The broader market structure is also shifting. Hyperscalers are no longer just Arm licensees; they are some of the most capable datacenter CPU vendors. AWS Graviton5 doubles to 192 Neoverse V3 cores and feeds Trainium3 head nodes, Microsoft’s Cobalt 200 jumps to 132 Neoverse V3 cores and 50% better performance, and Google is splitting Axion between higher-performance C4A and scale-out N4A variants. NVIDIA is evolving from Grace to Vera with far higher coherent bandwidth and memory capacity to support GPU-centric systems, while Arm itself is crossing a line from IP licensor to CPU supplier with Phoenix for Meta and possibly OpenAI-linked systems. For anyone tracking AI infrastructure, the important takeaway is that the next datacenter bottleneck may not be just accelerators or power delivery, but the surrounding layers of CPUs, DRAM, packaging, and interconnect needed to make GPU clusters useful.

Dario Amodei’s core claim is that AI progress is still following a fairly simple scaling story, but that the world has not internalized how close that trajectory may be to transformative outcomes. He frames today’s systems as the continuation of a thesis he has held since 2017: intelligence emerges less from bespoke clever tricks than from pouring compute through sufficiently broad data distributions under objectives that scale cleanly. In his telling, the old pretraining scaling laws have not broken; instead, reinforcement learning now appears to be extending the same pattern. He points to public reports of log-linear performance improvements with more RL training on verifiable tasks such as AIME-style math and says Anthropic is seeing similar behavior across a wider range of domains. That leads him to treat the current pretraining-plus-RL stack not as a dead end but as evidence that the main recipe is still working.

Where Amodei departs from many skeptics is in his willingness to map those scaling curves directly onto very aggressive capability forecasts. He says he is around 90% confident that within 10 years we will have what he calls “a country of geniuses in a data center,” and his personal hunch is much shorter—roughly 1-3 years. He is most confident on tasks with verifiable feedback, especially software engineering, where he says end-to-end coding is effectively guaranteed inside a decade and likely 1-2 years away. He is somewhat less certain on domains where verification is weaker—scientific discovery, Mars mission planning, novel writing—but he argues that generalization from verifiable to less verifiable domains is already visible. On the contentious issue of continual learning, he takes a surprisingly minimalist stance: these systems may not need human-like lifetime learning to become economically dominant. Broad pretraining, RL generalization, and long-context in-context learning may already cover most of the gap, with continual learning potentially arriving as an extra capability rather than a prerequisite.

A major theme of the interview is the distinction between capability growth and economic diffusion. Amodei repeatedly argues that the technology curve and the adoption curve are both exponential, but not identical. He dismisses the idea that diffusion nullifies AI progress, yet insists deployment will still be bottlenecked by enterprise procurement, compliance, security review, organizational change management, and the physical pace of closing loops in real workflows. As evidence that adoption is already very fast, he offers Anthropic’s revenue numbers: roughly $100 million in 2023, $1 billion in 2024, and $9-10 billion in 2025, with another few billion allegedly added in January 2026 alone. Even so, he says that is not the same as instant absorption of AGI-scale capability into GDP. The same logic underlies his defense of Anthropic’s compute posture. If model capability reaches “country of geniuses” status in 2026 or 2027 but monetization lags by 1-5 years, overcommitting to trillion-dollar annual compute purchases could bankrupt a lab that is directionally right but off by a single year.

That leads into his economic model of frontier labs, which he portrays less as speculative money furnaces than as businesses whose losses are largely a timing artifact of compute scale-up. His stylized picture is that individual model generations can already have strong positive gross margins on inference, but firms remain unprofitable because every profitable model finances a much larger next model. In a more mature equilibrium, he expects a small-number-of-firms market analogous to cloud infrastructure: very high barriers to entry, differentiated products, positive but not monopolistic margins, and a meaningful though not dominant share of compute continuously devoted to R&D. He pushes back on the view that profitable AI labs must be underinvesting, arguing that scaling returns are approximately log-linear, so there is a rational interior optimum rather than a need to push 95-100% of resources into training. Notably, he also quantifies the physical side of the buildout: industry AI power demand at 10-15 GW in 2026, rising 3x yearly toward ~300 GW by 2029, implying multi-trillion-dollar annual compute expenditures if that trajectory persists.

On policy and geopolitics, Amodei is simultaneously pro-build and hawkish. He rejects a blanket 10-year moratorium on state AI laws absent federal replacement, arguing that the world may face meaningful AI-enabled bioterrorism and autonomy risks well before then. His preferred sequence is transparency first, then targeted regulation such as mandatory biological-risk classifiers if threat evidence hardens. He also worries that AI may create offense-dominant equilibria in which one actor can do catastrophic harm, making traditional balance-of-power assumptions unreliable. That concern carries into his China stance: he does not think both the US and China should simply race to build symmetric “countries of geniuses,” because advanced AI could stabilize authoritarian control internally and destabilize deterrence externally. At the same time, he distinguishes restricting frontier compute and chips from restricting downstream benefits, suggesting AI-enabled drugs and development should spread widely, especially to the developing world. The broader implication is that Amodei sees the bottleneck after AGI less in invention than in governance, distribution, and institutional adaptation—and believes those questions are arriving on a timeline measured in years, not decades.

Crusoe is scaling energy storage for data center infrastructure with two distinct battery strategies: long-duration storage from Form Energy and second-life EV battery systems from Redwood Materials. The 12 GWh Form purchase is a notable commercial validation of 100-hour iron-air storage, while Crusoe’s existing 12 MW / 63 MWh microgrid and planned 8 MW expansion show how reused EV packs may complement large-scale backup and load-shifting needs for AI-oriented power demand.

China’s March 2026 climate-and-energy policy package reinforces a familiar but increasingly formalized strategy: rapid clean-energy expansion, continued attention to energy security, and gradual legal codification of long-term decarbonization. The final 15th five-year plan changed little from the earlier draft, but it sharpened the institutional backdrop by referencing the newly passed ecological and environmental code and by adding geothermal energy to the list of promoted technologies. National Energy Administration head Wang Hongzhi framed 2026-2030 as both the decisive period for peaking carbon emissions and a critical phase for building a “new energy system,” emphasizing market-based pricing reform to enable fossil-fuel replacement without compromising reliability.

The briefing’s most important new development is legal, not just technological. China’s ecological and environmental code gives statutory footing to the “dual-carbon” goals of peaking by 2030 and neutrality by 2060, while embedding total-emissions caps, carbon-intensity controls, carbon trading, footprint management, and climate-related enforcement into a more rules-based governance structure. At the same time, China signed the global pledge to triple nuclear capacity by 2050, even though its own nuclear rollout has undershot recent targets, reaching only 62GW by end-2025 versus a 70GW target. The Middle East oil shock underscores why Beijing continues to pursue a diversified system: crude stockpiles and renewables cushion import risk, but oil and coal still retain strategic roles in petrochemicals, industrial feedstocks, and power-system flexibility.

Nvidia’s 2026 strategy, as described by Jensen Huang in this Stratechery interview after GTC, is to expand the meaning of accelerated computing from chips and training clusters into full-stack AI factories. Huang’s core thesis is that agents will not live solely inside chat interfaces or model APIs; they will increasingly operate the software humans already use, from SQL databases to desktop applications and EDA tools. That pushes Nvidia beyond selling GPUs into accelerating legacy software, designing CPUs, integrating networking and storage, and helping customers construct entire data-center-scale systems. He quantified the scale of the new buildout in unusually concrete terms: a 1 GW AI factory could cost $50-60 billion, of which roughly $15-17 billion is just land, power, and shell. His point is that customers will not risk tens of billions of dollars unless Nvidia can de-risk throughput, utilization, and system integration across cooling, power, networking, and compute. That framing is particularly relevant to the data-center and infrastructure buildout around AI, because Huang treats power as the fundamental economic constraint and argues system-level co-design is the only way to maximize intelligence produced per watt and per dollar.

On the technical side, Huang drew a line between the first wave of generative AI and the current wave of useful AI. He said reasoning, reflection, retrieval, and search improved enough over the last year to ground models, reduce hallucinations, and enable paid applications—especially coding agents. His description of coding as a separate modality was notable: code must be generated in coherent chunks and validated by execution, not merely by token probability. He also argued that transformers alone are not enough for future workloads, citing quadratic attention costs, KV-cache bloat, and the need for new architectures for continuous motion, geometric symmetry, and physical constraints. He named Nemotron 3’s transformer-plus-SSM hybrid and cuEquivariance as examples. At the infrastructure layer, he defended Nvidia’s CPU effort as complementary to GPUs rather than a reversal: the goal is to prevent expensive accelerators from sitting idle. Vera, he said, emphasizes very high single-thread performance and 3x higher bandwidth-per-CPU than prior designs to support agent tool use over NVLink. The Groq acquisition fits the same logic: Nvidia wants finer-grained heterogeneous inference, including disaggregating pieces of decode attention, to push latency-sensitive coding and enterprise agent workloads beyond what a general GPU-only setup can economically deliver.

The interview also highlighted constraints and geopolitics. Huang said the bottleneck in 2026 is not one thing but everything at once—power, fabs, supply chain, and site readiness—though he sounded confident in Nvidia’s planning across hundreds of suppliers. More striking was his China argument: keeping an American AI stack present in China is, in his view, strategically essential because Chinese labs and open-source communities are too important to ignore. He explicitly praised DeepSeek, Kimi, and Qwen as technically meaningful contributors and warned that exclusion could let rival ecosystems harden across chips, platforms, models, and applications. That makes the conversation highly relevant not just as an Nvidia profile, but as a window into how the dominant AI infrastructure vendor thinks about power markets, data-center economics, hardware/software co-design, and U.S.-China competition.

Epoch AI’s Frontier Data Centers Hub is a useful new primary-source-style dataset on the physical buildout of frontier AI infrastructure, focusing on power, compute, construction timing, and capital intensity. The most interesting contribution is methodological: rather than relying mainly on company announcements, the project estimates capacity from high-resolution satellite imagery of cooling systems and corroborates those estimates with permits and hardware assumptions, giving outside observers a way to independently track opaque hyperscale projects. The numbers underscore how quickly the industry is moving from large campuses to nation-scale power loads: multiple facilities are expected to exceed 1 GW in 2026, and the leading site, xAI’s Colossus 2, is projected at 1.4 million H100-equivalents. Epoch also highlights phased commissioning, projecting Microsoft Fairwater’s first phase in March 2026 and full operation by September 2027, while estimating top-end capital costs above $100 billion for the largest campuses.

Texas power grid performance during the early-2026 winter storm benefited from rapid additions of solar and battery storage and winterization measures, with batteries injecting power at the morning peak and reducing prices. ERCOT now forecasts peak demand climbing from ~87 GW (2025) to ~145 GW (2031), driven largely by data centers (5,302 MW added since 2022; >24,000 MW projected).

General Matter, led by Scott Nolan, is rebuilding U.S. uranium enrichment capacity by reviving Paducah and winning a $900M DOE contract to avoid a 2028 supply cliff after Russian restrictions; the move addresses a supply chain where the U.S. fell from ~86% of global enrichment to near zero commercially, threatens a quarter of reactor fuel, and focuses on LEU/HALEU needs (20% enrichment threshold).

Texas’ grid five years after Winter Storm Uri is materially stronger: regulators introduced stringent winterization standards (ambient-temperature specs up to −17°C in the Panhandle), ERCOT hired hundreds of inspectors, and the PUC gained enforcement power including fines up to $1M/day per incident. Uri’s cascading failure — a rapid loss of ~4 GW in under 30 minutes that triggered frequency collapse and four days of outages — motivated these changes.

The system has also added dispatchable battery capacity (from ~1 GW in 2021 to ~11–15 GW now), which McAdams says would have arrested frequency freefall and stabilised the system. The defining issue ahead is unprecedented load growth: ERCOT’s interconnection queue includes roughly 2,000 projects (~435 GW) and surged ~300% last year. Policy and market innovations — ADER virtual power plants, nodal settlement for DERs, residential demand-response telemetry — are positioned to mobilize distributed resources, reduce peak stress, defer wires, and lower customer bills if properly implemented.

The January 8, 2026 Energy Capital roundtable (hosts Matt Boms, Joshua Rhodes, Micalah Spenrath) framed 2025 as a year in which uncertain, rapid load growth — especially from data centers and AI-related facilities — upended conventional planning. Panelists contrasted ERCOT's ~85.5 GW peak today with roughly 225 GW of large-load requests in the queue by 2030, warning that timelines for transmission, interconnection, and generation can't keep pace without new approaches. Speakers argued speed is itself a grid constraint and recommended prioritizing fast, local flexibility: distributed energy resources, demand response, backup power, and distributed batteries (ADER pilots with vendors like Tesla were cited). They also flagged policy levers (Senate Bill 6, PUC dockets), private PPAs and co‑located solar+storage as near-term solutions, and stressed that high near-term capital costs (e.g., gas plants ~2.5x, transformers ~2x) require fair cost-sharing and community-focused siting to preserve reliability and equity.

Cleanview's December 2025 report (Michael Thomas) finds 1,891 U.S. power projects canceled in 2025, eliminating 266 GW of planned capacity — about 150% of ERCOT — and an estimated $400 billion in investment. Causes include local opposition, transmission shortfalls, battery saturation, tariffs and federal policy; ERCOT approved a $9.4B, 765-kV transmission "superhighway".

The Texas large-load queue hit 225 GW (up from 99 GW in Feb), adding ~30 GW in two months; only 6.6 GW are energized with 0.6 GW due in 2026. PUC filed an interconnection draft (comments due Dec 19) and ERCOT will vote on the eastern 765-kV plan.

Stanwell’s Tarong Battery entered commercial operations in February 2026: a 300 MW, two‑hour (600 MWh) lithium‑ion system of 164 Tesla Megapacks costing $514 million, sited beside Tarong coal station to deliver sub‑second firming and raise the site to 2.1 GW. The build responds to high coal outages in Apr–Sep 2025 and is one piece of Stanwell’s plan for 5 GW of batteries by 2035.

Power required to train frontier AI models has increased roughly 2.1x/year (90% CI 1.9–2.2x) across 45 frontier models since 2010, per Epoch AI (Sep 19, 2024); analysis uses 211 models and a log-linear fit.

OpenAI launched a Request for Proposals on January 15, 2026 to accelerate U.S. manufacturing for AI infrastructure, building on its Stargate initiative (nearly one year old) that has pledged capacity already well over halfway toward a 10‑gigawatt commitment. The RFP targets domestic production of data‑center inputs, compute/power/cooling hardware, consumer‑electronics modules, and robotics components; proposals are due June 2026.

OpenAI argues the US must rapidly expand electricity generation and skilled labor to preserve AI leadership, noting China added 429 GW in 2024 versus the US's 51 GW and urging a 100 GW/year national build. It cites user growth from >400M to >800M weekly, a $500B/10 GW Stargate pledge (nearly 7 GW and $400B in three years), and a 20% skilled-trades demand over five years.

AMD and OpenAI announced a multi-year, multi-generation partnership to deploy 6 gigawatts of AMD Instinct GPUs, beginning with a 1 GW roll-out of MI450 series racks in 2H 2026. AMD granted OpenAI warrants for up to 160 million shares tied to deployment, share-price, and technical milestones; AMD forecasts 'tens of billions' in revenue and accretive non-GAAP EPS.

OpenAI's Stargate initiative, joined on Oct 1, 2025 by Samsung and SK, commits to scaling advanced DRAM production (targeting 900,000 wafer starts/month) and expanding AI data-center capacity in Korea through an MoU with MSIT and partnerships with SK Telecom and Samsung affiliates; deployments include ChatGPT Enterprise and APIs to support enterprise workflows and regional AI growth.

OpenAI submitted a May 7, 2025 response to the DOE RFI arguing the U.S. must act quickly to build AI supercomputing hubs, citing “hundreds of billions” in private capital and potential for “tens of thousands” of jobs. The proposal highlights Stargate (first campus in Abilene, Texas), a Jan 2025 LANL partnership, and recommends streamlined permitting, lease predictability, and financial incentives for co‑located federal‑land deployments.

OpenAI’s Nov 4, 2024 comments to the NTIA argue that data center policy is critical for U.S. competitiveness: their outside forecast estimates a single 5 GW data center would support ~40,000 jobs and contribute $17–$20 billion in state GDP. OpenAI highlights $175 billion in global infrastructure capital and urges US-focused investments in energy, nuclear, and semiconductor capacity to retain AI leadership.

An OpenAI backend engineer (blog post published 2022-12-08) describes managing the company’s large-scale supercomputing clusters—debugging NUMA, GPUDirect, CPU pinning, and NIC issues (e.g., >30 Gbps kernel panic) and shipping fixes from hotfixes to upstream kernel changes; one tweak reportedly saved ~6 days of compute across the fleet per week, speeding research.

The Stargate Michigan campus in Saline Township, announced 2025-10-30, will be developed by Related Digital with construction starting in early 2026 and >2,500 union construction jobs. It's part of OpenAI's 4.5 GW Oracle partnership and—together with six U.S. Stargate sites (Oracle, SoftBank)—pushes planned capacity past 8 GW and >$450B investment, using closed-loop cooling and DTE's excess transmission capacity.

OpenAI and NVIDIA announced a strategic partnership (Sept 22, 2025) to deploy a minimum of 10 GW of NVIDIA systems—characterized as millions of GPUs—to power training and inference for next-generation models. NVIDIA may invest up to $100 billion incrementally tied to gigawatt deployments; the initial 1 GW phase is slated for H2 2026 on the Vera Rubin platform, with roadmap co-optimization and partner coordination (Microsoft, Oracle, SoftBank).

Stargate Norway is OpenAI’s announced Narvik data-center project (2025-07-31) built with Nscale and Aker as a 50/50 JV, targeting 230 MW of compute (expandable by 290 MW) and 100,000 NVIDIA GPUs by end-2026. The site will be fully renewable (hydropower), employ closed-loop direct-to-chip liquid cooling, and pipe waste heat to local low-carbon industry while prioritizing regional researchers and startups.

The Waratah Super Battery (850 MW, 1,680 MWh) suffered a transformer failure in mid-October 2025 when an internal fault caused winding damage and a tank rupture; the site is operating at ~350 MW/740 MWh while Wilson Transformer Company manufactures a replacement due Q3 2026. Consolidated Power Projects and Wilson completed a design review and remediation work is planned to return full SIPS capability by end-2026.

Decentralized training: Epoch AI models a 10 GW cluster assembled from 23 datacenters tied to spare capacity at U.S. gas plants and linked by a 4,800 km fiber ring. For a 72T-parameter (≈1.1 Pbit per sync) model they target ~250 ms all-reduce (6 Pbps per link + ~24 ms propagation) to keep network overhead <5%, with network cost ≈$410M versus $90B in datacenter build costs.

Data movement bottlenecks constrain LLM training scale: Epoch AI models show efficient, 3-month training runs break down past ~2e28 FLOP and hit a latency-imposed upper bound near 2e31 FLOP. The analysis models intra-GPU HBM bandwidth and inter-GPU communication across tensor/data/pipeline/expert parallelism, uses Chinchilla D=20N, and numeric examples (tlat=9 μs, nlayer=120, B=4M) to quantify limits and mitigations.

Epoch AI models four bottlenecks—power, chip manufacturing (packaging/HBM), data availability (text+multimodal), and the latency wall—using public forecasts and uncertainty ranges. They find on‑trend 4x/year compute growth could reach ≈2e29 FLOP by 2030 (median), with ~100M H100‑equivalents and ~6 GW power plausible; upper/lower bounds span ~1e28–1e32 FLOP depending on grid, fab, and data assumptions.