A single vendor’s model portfolio can therefore span nearly a 20x energy range depending on which variant users choose or organizations deploy

Decision Lens

The energy intensity gap between competing AI models is no longer a sustainability footnote — it is a capacity planning variable. A market-driven shift in AI model usage can swing per-query power demand by more than an order of magnitude without a single new data center being added to the portfolio.

90-Second Brief

Today, independent research reveals a stark efficiency gap between the two dominant AI models. OpenAI’s GPT-4o consumes approximately 0.30 Wh per query; Anthropic’s Claude 3 Opus consumes roughly 4.05 Wh per query, a difference exceeding 13x. At the reported volume of over one billion ChatGPT queries per day, that gap translates to approximately 300 MWh of daily electricity consumption for one model alone. A politically driven user migration event is already underway, and if it holds, infrastructure teams planning capacity around current load profiles face material underestimation risk.

What’s Actually Happening

Two distinct developments are converging in ways that carry direct implications for power infrastructure planning. First, independent benchmarking details that AI model efficiency varies dramatically — not just between companies, but within model families. Claude 3 Haiku, Anthropic’s lighter variant, consumes around 0.22 Wh per query, while Claude 3 Opus runs at 4.05 Wh. A single vendor’s model portfolio can therefore span nearly a 20x energy range depending on which variant users choose or organizations deploy.

Second, a politically driven user migration is underway. Following OpenAI’s deal to run its AI tools inside U.S. Department of Defense classified systems — finalized in late February 2026, with safety limits on weapons use and mass surveillance — and the White House instructing agencies to stop using Anthropic’s technology over national security concerns, Claude surged in public usage. Third-party app analytics indicated that Claude experienced a significant rise in U.S. App Store rankings from late January to early March 2026, eventually reaching the number one free app position. Company statements and reporting indicated significant growth in Claude’s user base and paid subscriptions during early 2026. Uninstalls of the ChatGPT app spiked following the Pentagon announcement.

This is not a marginal shift. ChatGPT handles hundreds of millions to over one billion queries daily. If a meaningful fraction of that volume migrates to a more energy-intensive model variant, aggregate power draw across AI-serving infrastructure rises without any change in total active users. The model, not the headcount, becomes the energy variable.

At the data center level, this matters because AI inference — the continuous process of responding to user queries — represents the dominant and growing share of daily energy footprint for AI workloads. Unlike model training, which is a discrete and schedulable event, inference is demand-driven and continuous. Its energy profile is determined by which model variant is running, not simply how many users are connected.

Why It Matters for Global Heads of Data Center Energy?

  • From a budgetary standpoint, load forecasts anchored to the energy intensity of today’s dominant model mix carry embedded error risk. A 13x per-query variance across available models means a shift in market share — driven by policy, geopolitics, or user preference — translates directly into unplanned energy cost exposure at the portfolio level.

  • From an operational standpoint, power density planning assumptions for AI inference workloads need to account for model-mix sensitivity. Tenants running Claude 3 Opus-class workloads versus GPT-4o-class workloads draw fundamentally different power profiles from the same rack count — a variable most current infrastructure designs do not explicitly model.

  • From a regulatory standpoint, the projected growth of data centers to 3–4% of global power consumption by 2030 is premised on current model efficiency profiles. If energy-intensive model variants gain market share, those projections understate demand, with direct implications for interconnection queue positions, utility load growth filings, and state-level regulatory planning.

  • From a competitive standpoint, operators who can accurately forecast and provision for model-mix-driven load variability will hold a differentiation advantage in securing hyperscaler and AI-native tenant contracts, where power delivery reliability and capacity headroom are primary selection criteria.

  • From a workforce standpoint, energy procurement and infrastructure planning teams need to incorporate AI model efficiency tracking into load forecasting workflows — a capability that does not currently sit in most energy operations functions.

The Forward View

Over the next 30–90 days, watch whether the Claude usage surge stabilizes or continues to compound. If paid subscription growth and daily active usage metrics confirm a structural shift rather than a news-cycle spike, the energy intensity of AI inference workloads running across major cloud and colocation facilities will need to be re-baselined. The DoD-OpenAI classified systems integration will also drive procurement of additional secure compute capacity, adding a new demand signal in markets adjacent to government and defense infrastructure clusters.

Peer Moves

OpenAI secured classified system access within the DoD in late February 2026, expanding its operational footprint into a demand segment with distinct uptime and security requirements. Anthropic, simultaneously removed from federal agency use, pivoted to commercial user growth — with measurable market share impact within weeks, accelerating a compute load shift that infrastructure operators now need to track.

What We’re Uncertain About?

  • Model mix transparency: AI vendors do not publicly disclose which model variants are serving the majority of queries at any given time, making real-time energy intensity estimation for inference workloads effectively impossible from outside the system. Resolution depends on voluntary disclosure or regulatory reporting mandates.

  • Durability of the migration shift: Whether the Claude usage surge reflects a permanent reallocation of AI workloads or a temporary spike driven by news events remains unresolved. App download and subscription data lag actual compute utilization by weeks; sustained load shift will only become visible in utility consumption data over the next quarter.

  • Training versus inference energy split: The benchmarked research focuses on inference energy use per query. Full lifecycle emissions — including training runs estimated to exceed 500 metric tons of CO₂ equivalent per model version — are not publicly disclosed by either company, leaving total infrastructure energy planning incomplete.

One Question to Bring to Your Team

If Claude 3 Opus-class workloads replaced 20% of our current GPT-4o-equivalent AI inference load, what would that do to our peak power draw assumptions — and do our current tenant contracts give us the headroom to manage that shift without breaching capacity commitments?

Sources

  • Carboncredits — ChatGPT vs Claude AI: Carbon Footprints, Pentagon Deal, and Energy Impact (Link)

Tags