Altman drew a sharp line between what he called exaggerated per-query water claims and the very real macro-scale energy buildout AI will require, arguing the infrastructure challenge is about power generation — not gallons per prompt.

At a moment when artificial intelligence is reshaping industries and straining infrastructure planning, Sam Altman is confronting one of the most persistent criticisms head-on: the environmental cost of AI.

Speaking on the sidelines of the India AI Impact Summit in an interview with The Indian Express, the OpenAI chief executive dismissed viral claims that ChatGPT consumes gallons of water per query as “completely untrue” and “totally insane,” arguing that such figures bear “no connection to reality.”

The remarks land amid intensifying scrutiny of data center expansion, resource use, and AI’s long-term sustainability.

The Water Narrative — and What It Misses

Concerns about AI’s water footprint stem largely from how data centers are cooled. Many traditional facilities rely on evaporative cooling systems that draw significant volumes of water to regulate temperatures for densely packed servers.

Yet the link between a single AI query and water consumption is not direct. Water use occurs at the infrastructure level — in cooling systems and, in some regions, in power generation itself — rather than at the level of an individual prompt.

Cooling technology is also evolving. Hyperscale operators are deploying closed-loop liquid systems, advanced air cooling, and even water-free designs in some new builds. Efficiency gains per compute unit have improved steadily, though rising overall demand may offset those gains.

A recent projection by water technology firm Xylem and Global Water Intelligence estimated that water drawn for cooling could more than triple over the next quarter-century as global computing expands. That forecast reflects aggregate growth, not per-interaction intensity.

Altman’s pushback suggests he views the viral framing — “gallons per query” — as a distortion that conflates systemic resource use with marginal consumption.

Energy: The Real Constraint

Where Altman acknowledged a legitimate concern is the electricity demand.

“Not per query, but in total — because the world is using so much AI … and we need to move towards nuclear or wind and solar very quickly,” he said.

The distinction he is drawing is fundamental to understanding AI’s environmental calculus.

AI systems consume energy at two primary stages:

1. Training: the compute-intensive process of building large models, often requiring massive parallel processing over weeks or months.
2. Inference: the ongoing use of trained models to generate outputs in response to user inputs.

Training can require substantial bursts of energy, but inference — especially once hardware and software are optimized — is far less energy-intensive per transaction. The challenge lies in scale. Billions of inferences across millions of users translate into persistent demand on grids.

According to a May report from the International Monetary Fund, global data center electricity consumption in 2023 had already reached levels comparable to those of Germany or France, shortly after the debut of ChatGPT.

That comparison underscores how quickly AI has shifted data centers from background infrastructure to frontline energy consumers.

The Human Brain Analogy

Altman also addressed comparisons drawn by Bill Gates, who has suggested that the human brain’s efficiency implies AI systems could become dramatically more energy-efficient over time.

Altman argued that many comparisons overlook the energy embedded in human development.

“It takes like 20 years of life, and all the food you eat before that time, before you get smart,” he said.

He suggested the more appropriate benchmark is energy consumed per response once a model is trained — and by that metric, he believes AI may already be competitive.

The analogy has sparked debate. Critics argue that equating human cognition with computational systems risks flattening ethical distinctions. Sridhar Vembu of Zoho Corporation publicly criticized the equivalence, saying he does not want to see technology equated with human beings.

Beyond philosophy, the exchange highlights a deeper issue: AI efficiency is often discussed without standardized metrics. Measuring energy per inference, per token generated, or per model lifecycle produces very different narratives.

Infrastructure, Investment, and Political Friction

The debate is unfolding as governments and corporations commit billions to new data center capacity. AI has become a strategic priority, intertwined with economic competitiveness and national security.

To accommodate growth, some governments are accelerating approval processes for new power generation — including nuclear, solar, and wind. Environmental advocates caution that rapid buildouts could complicate climate commitments if fossil fuels fill short-term supply gaps.

Local resistance is also mounting. In San Marcos, Texas, the city council recently rejected a proposed $1.5 billion data center after sustained public opposition over concerns about grid strain and rising electricity costs.

These disputes reveal a widening tension between national AI ambitions and local resource constraints. Data centers are capital-intensive, geographically concentrated, and highly visible infrastructure projects.

One of the central questions is whether technological efficiency can outpace demand growth.

Historically, improvements in chip design and software optimization have reduced energy use per computation. However, AI workloads are expanding so rapidly that total consumption continues to climb — a classic case of the rebound effect, where efficiency gains stimulate additional usage.

Altman’s call for accelerated nuclear and renewable deployment implicitly acknowledges that efficiency alone will not solve the energy equation. Expanded generation capacity appears inevitable if AI adoption continues at current rates.

Anthropic’s claim that 16 million Claude interactions were siphoned through 24,000 fake accounts reframes the AI race as a battle over inference access, safeguards, and chip supply — not just model training.

U.S. artificial intelligence firm Anthropic has accused three Chinese AI developers, DeepSeek, Moonshot AI, and MiniMax, of orchestrating what it describes as a coordinated, large-scale “distillation” campaign to extract capabilities from its Claude AI system.

Anthropic said the three labs created more than 24,000 fraudulent accounts that generated over 16 million interactions with Claude. The queries, it alleged, were designed to systematically replicate some of Claude’s most advanced features, including agentic reasoning, tool use, and coding — areas considered differentiators among frontier AI systems.

The accusations land amid heightened geopolitical tension over artificial intelligence, particularly as Washington reassesses export controls on advanced semiconductors to China and as Chinese AI labs close the performance gap with U.S. counterparts.

Distillation as a competitive shortcut

Distillation is a common technique in machine learning in which a large, high-performing model acts as a “teacher” to train a smaller “student” model. Within a single company, it is used to compress models for lower-cost deployment while retaining much of their capability.

Across companies, however, the method becomes controversial. By querying a rival’s model at scale and using the responses as training data, a competitor can approximate performance without replicating the comprehensive research, compute expenditure, or alignment work that went into the original system.

Anthropic said the alleged campaigns varied in focus and scale. It tracked more than 150,000 exchanges linked to DeepSeek that appeared aimed at strengthening foundational reasoning and alignment, including workarounds for policy-sensitive prompts. Moonshot AI allegedly generated more than 3.4 million exchanges focused on agentic reasoning, tool integration, coding, data analysis, and computer vision. MiniMax accounted for roughly 13 million exchanges targeting agentic coding and orchestration, with Anthropic claiming it observed traffic being redirected toward the latest Claude release shortly after launch.

The scale matters since sixteen million interactions represent not casual usage but what Anthropic characterizes as industrialized extraction.

DeepSeek, in particular, has drawn scrutiny since releasing its open-source R1 reasoning model last year, which analysts said approached the performance of leading U.S. frontier labs at a fraction of the cost. The company is reportedly preparing DeepSeek V4, a new model said to outperform both Claude and OpenAI’s ChatGPT in certain coding benchmarks. Earlier this month, OpenAI also accused DeepSeek in a memo to U.S. lawmakers of using distillation techniques to mimic its systems.

Export controls and compute leverage

The dispute is unfolding alongside debate over access to high-end chips. Last month, the Trump administration allowed U.S. firms, including Nvidia, to export advanced AI processors such as the H200 to China, loosening earlier restrictions.

Anthropic linked the alleged distillation campaigns to computing power. “The scale of extraction… requires access to advanced chips,” the company said in a blog post.

It argued that export controls serve a dual function: limiting direct model training and constraining the compute needed for high-volume distillation.

This framing shifts the chip debate. Historically, export controls focused on preventing Chinese firms from training large frontier models from scratch. Anthropic’s claim suggests that even if direct training is restricted, access to sufficient inference compute could enable large-scale replication via API querying.

Policy analysts say this complicates enforcement. Distillation occurs through legitimate product interfaces — paid or public APIs — rather than through overt hacking. That creates a grey zone between normal usage and systematic extraction.

National security dimension

Anthropic also framed the issue as one of security. The company said U.S. developers build safeguards into frontier systems to prevent misuse in areas such as bioweapons design or malicious cyber operations.

“Models built through illicit distillation are unlikely to retain those safeguards,” Anthropic wrote, warning that dangerous capabilities could proliferate if protections are stripped out during replication.

It pointed to the possibility of authoritarian governments deploying advanced AI for offensive cyber operations, disinformation campaigns, and mass surveillance — risks that increase if such systems are open-sourced without embedded safety layers.

Dmitri Alperovitch, chairman of the Silverado Policy Accelerator and co-founder of CrowdStrike, told TechCrunch the allegations were unsurprising.

“It’s been clear for a while now that part of the reason for the rapid progress of Chinese AI models has been theft via distillation of U.S. frontier models. Now we know this for a fact,” he said. “This should give us even more compelling reasons to refuse to sell any AI chips to any of these [companies], which would only advantage them further.”

Anthropic said it will continue investing in defensive measures to make distillation harder to execute and easier to detect, while calling for “a coordinated response across the AI industry, cloud providers, and policymakers.”

Such coordination could include tighter API rate limits, improved anomaly detection, contractual enforcement mechanisms, and shared threat intelligence among AI labs. Cloud providers — which host the infrastructure underpinning both U.S. and Chinese AI workloads — may also face pressure to monitor and flag high-volume extraction patterns.

The broader stakes extend beyond one company. If frontier AI capabilities can be replicated rapidly through sustained querying, the competitive moat built on research expenditure and chip access narrows. In that scenario, advantage may hinge less on breakthrough architecture and more on distribution control, access management, and compute governance.

At the same time, aggressive restrictions carry trade-offs. Limiting chip exports could affect U.S. semiconductor revenues and accelerate domestic Chinese chip development. Restricting API access could constrain legitimate global customers and developers.

Anthropic’s allegations therefore crystallize a central tension in the global AI race: openness versus control. The tools that make AI widely usable, APIs, cloud access, and scalable inference, also create vectors for replication. As Chinese labs close the performance gap with U.S. peers, the contest increasingly revolves not just around building the most advanced model, but protecting it.