AI Giants Pivot to Smaller, Faster Data Centers
TREE NEWS reports: OpenAI and Anthropic are shifting part of their computing demand toward 20 to 30 megawatt (MW) data centers, a significant departure from the hundreds of megawatts and gigawatt-scale infrastructure deals that have defined the AI buildout over the past year. Anthropic has already begun talks in the UK and the Nordic region for projects of this size, while OpenAI is exploring similar opportunities in the Nordics, with negotiations also underway in the United States.
The shift comes as AI computing demand migrates from model training toward inference services. Smaller facilities can be brought online faster and are better suited to distributed processing of inference tasks. Structure Research projects that inference workloads will overtake training loads in global data center capacity by 2027, rising to 37% by 2030.
From Mega-Projects to Multi-Site Strategies
Over the past year, both companies signed massive infrastructure agreements. Anthropic struck a roughly $45 billion cloud computing deal with Nscale to lease approximately 460 MW of capacity at a West Virginia data center. OpenAI’s Stargate project has exceeded its initial 10 GW commitment, with additional development plans of 3 GW in Georgia and 8 GW in Ohio.
Now, both firms are supplementing those mega-projects with more flexible capacity sources. OpenAI said it is “building a diversified compute portfolio to meet growing global AI demand,” noting that different workloads require different infrastructure and that the company weighs demand, performance, reliability, timelines, and cost.
Speed Becomes the New Currency in Compute Procurement
Jabez Tan, research director at Structure Research, said the core advantage of smaller projects is faster access to available capacity. Securing a few megawatts at an already-powered site is often more practical than waiting for a single location to release large-scale capacity. For workloads that can run across sites, multiple smaller facilities can aggregate into substantial total capacity.
Large data center construction faces long lead times. Major projects in the US and other markets encounter local community resistance, while parts of Europe face land and power constraints. For AI companies, securing distributed resources first can bridge the gap before large projects come online.
Crusoe, which built a large data center campus for OpenAI in Texas, is increasing investment in smaller data centers. Such facilities are faster and cheaper to build, helping mitigate delays from large projects. Crusoe announced $3.9 billion in funding at a $30.9 billion valuation.
Inference Demand Drives Distributed Compute
The changing structure of AI workloads is another key driver. Large model training typically requires massive numbers of chips working together in a single cluster, while inference tasks can be split across smaller clusters handling independent requests, making them more suitable for distributed deployment.
JLL data shows inference workloads accounted for 9% of global data center workloads in 2025, with training at 14%. By 2027, inference is expected to surpass training, rising to 37% by 2030, while training’s share falls to 13%.
This trend is already influencing data center construction. In February, Nvidia announced it would collaborate with data center industry participants to study small data center solutions for distributed inference. As inference demand grows, AI infrastructure competition is shifting from pure hyperscale pursuit toward speed of acquisition, deployment flexibility, and distributed capabilities.
Market Implications
- Data center REITs and infrastructure providers: Smaller, distributed facilities could benefit regional operators and edge computing providers, while traditional hyperscale-focused REITs may face slower growth in mega-deal pipelines.
- Semiconductor demand: Distributed inference favors a broader mix of chips, potentially benefiting Nvidia’s inference-optimized offerings and competitors focused on edge AI.
- Energy and utilities: Smaller sites may ease pressure on single-grid locations but increase overall distributed power demand, benefiting renewable energy providers and grid infrastructure companies.
- Crypto and decentralized compute: The pivot toward distributed inference validates the thesis behind decentralized GPU networks, which could see increased interest as AI firms seek flexible, location-agnostic compute.
- Real estate and construction: Faster, cheaper builds could favor modular data center construction firms and repurposed industrial sites.
Key Takeaways for Investors
- The AI infrastructure trade is broadening beyond hyperscale mega-projects to include smaller, distributed facilities — investors should watch regional data center operators and edge computing plays.
- Inference is becoming the dominant workload, which changes chip demand profiles and favors companies positioned for distributed, lower-latency compute.
- Speed to capacity is now a competitive advantage, meaning firms that can deliver powered sites quickly may capture premium pricing.
- Decentralized compute networks could benefit as AI companies embrace multi-site, flexible infrastructure strategies.
