SemiAnalysis' Jeremie Eliahou Ontiveros on the supply/demand dynamics of AI and data centers

Episode Summary

Executive Summary: The episode argues that AI demand for data centers and power remains strong despite DeepSeek-related fears, because model efficiency gains are being offset by larger training runs, synthetic data, and rising inference usage. Jeremy of SemiAnalysis says the real bottlenecks are power, data center space, networking, and execution speed. He sees NVIDIA, power/electrical suppliers, and select Bitcoin miners with usable sites as beneficiaries, while warning that adoption and monetization will ultimately determine how long the cycle lasts.

Main Topics: AI demand is still outrunning data center supply (Priority: 5/5): Jeremy explains that AI, especially NVIDIA-driven accelerator demand, has radically changed the data center market and that supply is still lagging demand across both power and physical space. DeepSeek and efficiency gains do not break the trend (Priority: 5/5): The conversation centers on whether DeepSeek implies lower compute demand. Jeremy argues efficiency gains are normal and are being absorbed by more usage, better models, and more inference demand rather than reducing total spend. Training vs inference and the need for more compute (Priority: 5/5): Jeremy distinguishes between training and inference, saying both are expanding. Training still requires much larger clusters, while inference will increasingly depend on scale-up networking and memory capacity. NVIDIA’s moat may be stronger than bears expect (Priority: 4/5): The discussion highlights NVIDIA’s hardware, software, and networking advantages, especially for inference and reasoning workloads where scale-up networking and NVLink matter more. Power, equipment, and the data center supply chain as beneficiaries (Priority: 4/5): Beyond utilities, the hosts discuss which parts of the value chain benefit most: electrical gear, cooling, infrastructure, nat gas, and time-to-power solutions, with companies like Vertiv, Schneider, and Eaton mentioned as examples. Bitcoin miners as AI infrastructure converts (Priority: 5/5): The episode spends substantial time on whether Bitcoin miners can repurpose stranded power and existing facilities into AI data centers, with Core Scientific, Applied Digital, IREN, and TeraWulf discussed as case studies. Monetization, adoption, and eventual cycle risk (Priority: 4/5): Andrew and Jeremy discuss what could eventually end the upcycle: slower AI adoption, weak consumer willingness to pay, failed monetization, or a sharp demand/expectations reset once the market stops revising capex upward.

Key Arguments: AI demand is not a one-time spike; it is being driven by continual improvement in model capability, which creates more training and inference demand rather than less. DeepSeek is not evidence of demand destruction; efficiency improvements are already a longstanding trend, and other models such as Gemini 2.0 Flash are reportedly cheaper and better. Even efficient AI labs still need more compute because they require synthetic data, reinforcement learning, and better evaluation to maintain model quality. Inference demand matters as much as training, and firms with strong networking and scale-up architectures may have durable moats. NVIDIA’s moat is not only GPUs but also networking and system-level engineering, especially around GB200 NVL72 and NVLink. The current AI boom is an upcycle in which earnings and capex estimates keep getting revised higher, making valuation less important in the near term. Bitcoin miners are attractive only if they can credibly convert stranded power and sites into AI data center capacity; otherwise, their economics remain weak. The best AI infrastructure plays may be those that already have real sites, power, and experienced teams, not just speculative claims. The U.S. remains the best place to build large AI data centers because it has the labor, supply chain, property rights, and hyperscaler ecosystem. Long-term risk comes from adoption and monetization: if users stop paying or AI revenue fails to scale, the spending boom could eventually reverse.

Data Points: Global data center IT capacity growth (2019-2023): ~4 gigawatts per year - Jeremy says the pre-2024 global industry added about 4 GW of IT capacity annually. NVIDIA-related demand added in 2024: 5 gigawatts - He claims NVIDIA alone adds about 5 GW of demand in 2024. Total AI demand in 2024: 7-8 gigawatts - He estimates GenAI demand, including custom ASICs, at roughly 7 to 8 GW. Largest current GPU clusters: ~100,000 Hopper GPUs - Used to illustrate the current scale of frontier training clusters. Largest current cluster power: ~130 megawatts IT power - Power draw associated with a 100,000-GPU cluster. Planned next-gen cluster scale: Gigawatt-scale data centers by ~2027 - Big AI labs are reportedly planning single sites at gigawatt scale. GPT-3 inference cost decline: About $1,200 lower over two years - Jeremy uses this to show how quickly AI inference costs have fallen. ChatGPT user growth: 200 million users in about two months - Cited as evidence of mass adoption and platform potential. Bitcoin miner infrastructure cost: ~$0.5 million per megawatt - Typical cost for Bitcoin mining data center physical infrastructure. AI data center infrastructure cost: $10 million+ per megawatt - Used to show how much more capital intensive AI conversion is. Core Scientific deal scale: 16-megawatt deal with CoreWeave - Mentioned as an earlier proof point before the larger AI conversion. Applied Digital North Dakota project: 600 megawatts - Described as a major existing site with about 100 MW already built. TeraWulf AI contract: 70 megawatts - Referenced as a meaningful but smaller AI-related contract. Backup power duration in data centers: 5 to 10 minutes - Typical UPS battery duration before generators take over.

Pivotal Quotes: "“DeepSeek is not changing the trend. If anything, DeepSeek is actually below trend.”" — Jeremy: Used to argue that efficiency gains do not reduce overall compute demand. "“The issue can be overcome somewhat easily.”" — Jeremy: Said in response to concerns that intermittent or curtailed power makes some Bitcoin miner sites unusable for AI. "“Why do you need 100% state-of-the-art performance all the time?”" — Andrew: A skeptical investor question about whether hyperscalers can simply use cheaper, slightly less capable in-house AI chips instead of NVIDIA.

Implications: Listeners should watch adoption, monetization, and capex revisions more than headlines. The likely winners are firms with real power, sites, and technical execution; the key risk is not efficiency alone, but demand failing to justify today’s huge buildout.