Does Netflix use more water than ChatGPT?

The Comparison Everyone Asks, No Dataset Answers

It's the question that pops up in every debate about AI's environmental cost: does an evening of Netflix bingeing really use more water than an evening of ChatGPT? Here's the honest answer: nobody knows — and that itself is the story.

We dug through peer-reviewed studies, UN reports, and policy research looking for Netflix or video-streaming water data. It doesn't exist in any public source. Even a Bloomberg data-center water investigation that might have addressed the comparison was inaccessible. The streaming side of this matchup is a blank page.

The AI side, by contrast, is heavily documented. A peer-reviewed analysis estimates AI's 2025 water footprint at 312.5–764.6 billion liters — roughly the global annual consumption of bottled water. ChatGPT alone handles ~2.5 billion prompts per day, drawing about 383 GWh of electricity annually, according to a UN University report that puts its water footprint on par with the minimum annual domestic needs of ~500,000 people in Sub-Saharan Africa.

So why can't we just pull Netflix's numbers and settle it? Three structural problems:

• Operators don't separate AI workloads from non-AI workloads like streaming in their disclosures, making side-by-side comparisons impossible.
• Disclosure is voluntary and inconsistent — investigative research found Amazon disclosed no water data at all, and Microsoft only company-wide figures.
• Most water use is indirect — electricity generation consumed an estimated 12x more water than direct cooling in 2023 — so even good disclosures miss most of the footprint.

This is a transparency gap, not a verdict. Any confident "Netflix wins" or "ChatGPT wins" headline you've seen is built on numbers that don't exist publicly. As researcher Alex de Vries-Gao puts it, further disclosure from data center operators is "urgently required" to responsibly manage AI's environmental impact.

That gap is familiar territory for us at SkillzBench. We're building an independent lab for verifying and benchmarking AI agent skills precisely because unverifiable claims are endemic to this industry — if you can't measure it, you're trusting marketing. The comparison everyone asks about can only be answered when companies start publishing the data — and until then, the AI-side figures are the only documented half of the story.

Skills, proven. SkillzBench is pre-launch — join the waitlist at skillzbench.com to be first in line for independently verified, benchmarked agent skills.

What We Can Measure: ChatGPT's Water Footprint

Every time you ask ChatGPT a question, somewhere a data center quietly drinks roughly a bottle of water. That's the most relatable figure in the AI water debate, and it comes from UC Riverside scientists, who estimate that a single 100-word prompt consumes about 519 mL of water — most of it for cooling the servers doing the thinking.

Scale that up and the numbers get startling. According to a United Nations University report, ChatGPT handles roughly 2.5 billion prompts per day, drawing about 383 GWh of electricity annually. Its total water footprint, the report estimates, matches the minimum annual domestic water needs of about 500,000 people in Sub-Saharan Africa.

Zooming out, researchers publishing in a peer-reviewed analysis put AI's total 2025 water footprint at 312.5 to 764.6 billion liters — a range wide enough to swallow the global annual consumption of bottled water. That uncertainty isn't sloppiness; it reflects how little operators actually disclose.

The footprint breaks down roughly like this:

• Inference, not training, dominates — running deployed models accounts for 80–90% of AI's total energy consumption.
• Training still stings locally: GPT-3's training run directly evaporated an estimated 700,000 liters of freshwater in Microsoft's U.S. data centers.
• Indirect water use — consumed generating the electricity — is estimated at 12 times direct cooling use.
• Per-output costs vary by task: roughly 29 mL per AI-generated image, and 4.1 liters for a complex AI video.

Here's the honest caveat every reader deserves: these estimates swing by an order of magnitude depending on model and methodology. One widely cited GPT-3-era estimate put water use at ~500 mL per 10–50 responses, not per single response — and as trade reporting notes, current ChatGPT models likely use more water than GPT-3 did. The famous "one bottle per prompt" figure is a useful mental model, not a meter reading.

That variability is exactly the kind of problem SkillzBench exists to solve in its own domain — AI agent skills, where published research found roughly 26% carry at least one vulnerability. When vendors' claims can't be independently verified, whether about water footprints or skill quality, the honest answer is measurement, not marketing. Skills, proven — join the SkillzBench waitlist to see how we're building independent verification and benchmarking for agent skills.

The Three-Part Footprint: Why Any Fair Comparison Is Hard

Before you can fairly compare Netflix's water bill to ChatGPT's, you need to know what "water use" actually means for a data center. It turns out almost nobody counts all three layers — and that omission is exactly why the viral comparisons fall apart under scrutiny.

Part one: direct cooling water. This is the number most headlines quote — the water evaporated on-site to keep servers from overheating. U.S. data centers directly consumed 17 billion gallons (64 billion liters) in 2023, a figure projected to double or quadruple by 2028, according to investigative research on data center water use.

Part two: indirect electricity-generation water. Power plants consume water too, and data centers are voracious electricity users. Indirect water use through electricity generation was estimated at 12 times direct cooling use in 2023 — roughly 211 billion gallons, dwarfing the on-site figure. Policy research pegs indirect consumption at about 1.2 gallons per kWh, with 56% of data center electricity still coming from fossil fuels.

Part three: chip manufacturing. Every AI accelerator begins life in a fab that drinks enormous volumes of Ultra-Pure Water. Industry reporting estimates roughly 2,200 gallons of Ultra-Pure Water per AI chip — a cost incurred before a single prompt is ever processed.

A fair comparison, then, requires counting all three layers for both sides:

• Direct cooling water evaporated on-site at data centers and streaming infrastructure
• Indirect water consumed generating the electricity that powers both
• Manufacturing water embedded in the chips and hardware each depends on

Here is where the comparison collapses: nobody has published those numbers for Netflix. The peer-reviewed research on AI's footprint is candid that operators don't even separate AI workloads from non-AI workloads, and disclosure remains voluntary and inconsistent. Streaming's water footprint is simply absent from the public record.

This is a measurement problem, not a marketing problem. At SkillzBench, we take the same view of AI agent skills: until something is independently measured against a consistent framework, any claim about it — positive or negative — is just a headline. Without a shared methodology, the Netflix-vs-ChatGPT debate is unanswerable — and anyone offering a confident verdict is comparing numbers that were never collected.

Alarming Figures vs Honest Counterpoints

The scariest AI water statistics are real — but so are the counterpoints, and honest analysis demands both. More than 160 new AI data centers have opened in water-scarce U.S. regions in just the past three years, and some facilities draw over 25% of local community water supplies.

Google's Council Bluffs, Iowa operation consumed roughly 1 billion gallons in 2024 alone — enough to supply all of Iowa's residential water for five days, according to investigative reporting on data center water use. Those are the numbers that make headlines, and they should worry anyone living near a hyperscale facility in a dry climate.

Yet the picture changes dramatically depending on geography. A UK survey of 73 English data centers found that 51% use waterless cooling, and 64% consume under 10,000 m³ per year — less than a typical leisure centre. In cooler climates with closed-loop designs, data centers can be genuinely light water users.

The deeper trap is assuming green choices are green across the board. The UNU-INWEH assessment found that switching from coal to bioenergy cuts carbon emissions by roughly 70% but increases water consumption more than thirty-fold. Lead author Dr. Miriam Aczel put it bluntly: choices that look greenest from a carbon perspective often end up worse for water or land.

This is why local water stress, not aggregate consumption, is the real story. As Mara Pusic of the Environmental Law Institute notes, the challenge for many communities isn't how much water is used total — it's when, where, and who bears the cost when supplies run low.

Three takeaways for reading these numbers responsibly:

• U.S. figures and UK figures describe different facilities, climates, and cooling technologies — neither invalidates the other.
• Indirect water use (from electricity generation) was estimated at 12x direct cooling use in 2023, so "waterless" cooling doesn't mean zero water.
• A data center sited next to a leisure centre in England and one in Arizona's drought belt are entirely different environmental stories.

The uncomfortable truth is that we can't verify most of these claims independently — Amazon disclosed no water data at all, and Microsoft reports only company-wide figures. That verification gap is exactly why SkillzBench exists in the AI tooling space: when vendors won't show their numbers, independent testing has to fill the void. Claims deserve evidence, whether the subject is water footprints or agent skills.

The Real Lesson: Unverifiable Claims Are the Problem

We can't answer "Netflix vs. ChatGPT" for one uncomfortable reason: nobody publishing the numbers has to prove them. The water debate isn't really a data problem — it's a verification problem, and it's one that runs through the entire AI industry.

Data center operators don't separate AI workloads from anything else running in the same buildings, which means researchers can't isolate AI's footprint even when operators do disclose. And disclosure is voluntary: water policy specialists McCauley and Scanlan found that Amazon published no water data at all, while Microsoft reported only company-wide figures. That's why estimates of AI's 2025 water footprint span an enormous range — 312.5 to 764.6 billion liters, a spread wide enough to hide a small country's consumption inside it.

The experts aren't shy about what's needed. Alex de Vries-Gao, whose work underpins those footprint estimates, has called for urgent disclosures from data center operators so the industry can "responsibly manage the growing environmental impact of AI systems." Until then, every headline number — including the ones in this article — is an estimate built on trust, not verification.

The pattern looks familiar to us at SkillzBench, because we see the same dynamic in AI agent skills:

• Vendors and authors make performance claims nobody independently tests.
• Published research found roughly 26% of skills carry at least one vulnerability — a risk most users only discover after installing.
• The market rewards confident claims, not verified ones, so unverifiable claims keep winning.

That's exactly the gap SkillzBench exists to close. We're building independent verification for agent skills — pinning every version to a commit SHA, security-scanning and license-checking each one, and benchmarking every skill in a sandbox with the same model, with and without the skill, so the Lift it actually delivers is measured rather than asserted. Ratings never mix with performance, and money never touches scores.

Verification beats trust. The water question stays unanswerable until operators disclose; skill claims stay unverifiable until someone tests them. We're building the testing side.

Skills, proven. Join the SkillzBench waitlist at skillzbench.com to get early access to independently verified, benchmarked agent skills.