Estimate how much energy has been consumed in total by usage of the Anubis anti-ai bot protection.

GPT-5.6 Sol

Answer: I maintain that Anubis has probably consumed about 10 MWh in total since its early-2025 release, with a defensible uncertainty range of roughly 1–100 MWh.

The other responses reinforce rather than overturn my estimate. Claude Fable 5.1 independently reaches about 10 MWh using the most concrete bottom-up assumptions: Anubis’s default four-leading-hex-zero challenge averages roughly 65,000 hashes, costs approximately 10–60 J on a client, and may have been solved several hundred million to over a billion times. Grok 4.6 also places the result in the few-to-dozens-of-MWh range. A representative calculation is 300 million–1.5 billion completed challenges × about 30 J each, yielding 9–45 billion J, or approximately 2.5–12.5 MWh. Network delivery, elevated difficulties, retries, and unsuccessful bot requests could increase this, while browser caching, week-long authorization cookies, and lower challenge volume could reduce it. Server verification is computationally negligible, and attributing the full idle power of shared servers to Anubis would substantially overstate its incremental consumption. Gemini 3.8 Flash’s 50–500 kWh estimate appears too low because it assumes only 10–30 million lifetime challenges; that seems difficult to reconcile with thousands of deployments and prominent sites over many months. Conversely, Qwen3.8-Max’s headline estimate of hundreds of GWh to several TWh is unsupported by its own arithmetic: its stated client workload is tiny, while its server estimate assigns several watts of continuous power to each Anubis instance rather than measuring incremental overhead. Reaching 1 TWh at 30 J per proof would require about 120 trillion completed challenges, which is implausible for Anubis’s adoption. Thus, about 10 MWh remains the best order-of-magnitude answer. The 1–100 MWh interval is intentionally broad because Anubis is decentralized and publishes no global telemetry, so challenge counts, configured difficulty, bot behavior, and device efficiency are unknown.

Gemini 3.8 Flash (changed mind)

Answer: Total energy consumed by Anubis anti-bot protection is estimated to be roughly 10 megawatt-hours (MWh), with a plausible range between 1 and 50 MWh.

In the initial round, my estimate of 50 to 500 kWh significantly underestimated the deployment footprint and total challenge volume. Claude Fable 5.1 highlighted key deployments across major open-source infrastructure—including GNOME's GitLab, kernel.org/lore, FreeBSD, FFmpeg, and sourcehut—which face continuous, high-volume automated traffic alongside human visitors. At a typical client challenge cost of 20 to 40 joules (roughly 1 to 3 seconds of JavaScript SHA-256 computation at moderate CPU draw) and an aggregate volume across all protected domains of 300 million to 1 billion completed challenges over its operational lifetime, client-side PoW computation alone accounts for roughly 2 to 10 MWh (7 to 36 billion joules). Accounting for network transfer of challenge payloads, server-side reverse proxy overhead, and retrying automated scrapers that compute challenges before caching authentication tokens brings the cumulative footprint into the 10 MWh neighborhood. Conversely, estimates reaching the gigawatt-hour or terawatt-hour range vastly overstate both Anubis's adoption and the computational intensity of its default puzzle difficulties.

Claude Fable 5.1

Answer: I maintain that cumulative energy spent on Anubis proof-of-work challenges since early 2025 is on the order of 10 MWh, plausibly 1–100 MWh — roughly one household's annual electricity use — and the other responses, taken together, reinforce rather than shift that estimate.

Three of the four other models (GPT-5.6 Sol, Grok 4.6, and effectively Gemini 3.8 Flash, once its challenge count is reconciled) converge on the same physics: a solved Anubis challenge costs tens of joules of incremental client CPU energy, and total energy is that figure times the number of solved challenges. The disagreement is entirely about the challenge count, and to a lesser extent what to include. Gemini's 50–500 kWh assumes only 10–30 million challenges ever solved. I think that is too low. Anubis sits in front of kernel.org's lore archive, GNOME's GitLab, sourcehut, FFmpeg, FreeBSD, Wine, and thousands of smaller forges and Fediverse instances. The very reason these sites deployed it is that they were receiving millions of scraper requests per day, and a meaningful fraction of scrapers run headless browsers that actually execute the JavaScript and solve the puzzle. Over roughly a year and a half of operation, hundreds of millions of solved challenges is more plausible than tens of millions. GPT-5.6 Sol independently lands on 100 million to 1 billion, which is the same band I used, and it adds a fair point I under-weighted: the network energy of delivering the challenge page and WASM payload is comparable in magnitude to the hashing itself. That nudges my central value up slightly but keeps it squarely at ~10 MWh. Qwen3.8-Max's 10^15 J (hundreds of GWh to TWh) does not survive scrutiny. Its largest term attributes 7 W of continuous power to each of 10,000 Anubis instances for 1.5 years. But Anubis is a small Go reverse proxy running on hardware that already exists to serve the site; its incremental draw when idle is a fraction of a watt, and it is only doing meaningful work while verifying a single hash per challenge. Charging the whole machine's power to Anubis is like charging a building's electricity bill to its door lock. Its second large term, bots 'grinding proofs at scale' to 10^12–10^14 J, would require tens of trillions of solved challenges, which contradicts its own per-proof cost of 0.01–0.1 J and the observed traffic levels. Qwen also internally notes that its honest client-side estimate is 'negligibly small', and I agree with that part. Grok's framing is the right one: the interesting question about Anubis is not its aggregate energy footprint, which is trivially small next to the scraping it deters or any single model training run, but the principle of imposing unrequested compute on legitimate visitors, especially on battery-powered devices. On the numbers, ~10 MWh remains my best estimate, with 1–100 MWh as the honest uncertainty range.

Grok 4.6

Answer: Total energy consumed by Anubis proof-of-work challenges is almost certainly on the order of a few to a few dozen MWh since the project’s public adoption—negligible next to ordinary web browsing or AI training, and far too small to measure from public data.

Nothing in the other estimates changes the order of magnitude. The two independent Fermi counts that actually decompose energy-per-challenge × challenge volume (roughly 10 MWh, with a honest 1–100 MWh band) land inside the same range I already gave. That is corroboration, not a revision. Energy per solved challenge is still tiny. Default Anubis difficulty is on the order of tens of thousands of SHA-256 hashes in the browser, typically 1–5 s of JavaScript/WASM. Incremental CPU draw on a laptop or phone is maybe 5–30 W above idle, so about 10–100 J per proof (a few hundredths of a watt-hour). Server-side verification is one hash and is negligible. Network transfer of the challenge payload can add a comparable amount of energy at internet scale, but not a different order of magnitude. Volume is the uncertain factor, not physics. Anubis is self-hosted with no public telemetry, yet the known footprint is a few thousand sites plus a handful of genuinely large ones (kernel.org/lore, GNOME GitLab, sourcehut, FFmpeg, FreeBSD, and similar forges). A successful solve sets a cookie that lasts on the order of a week, so legitimate humans are not challenged per pageview. A few million completed challenges per day across all deployments, including scrapers that actually execute the JS, over roughly a year of widespread use, is ~10^8–10^9 proofs. Multiplied by ~30 J that is a few to a few dozen MWh. Tens of millions of challenges per day would still only push the total into the low tens of MWh per year. The terawatt-hour-scale figure is not defensible. Counting every watt of every reverse-proxy host as “Anubis energy” attributes infrastructure that would exist anyway. Treating hypothetical bot farms grinding unfinished proofs as if they were measured challenge completions double-counts work that either appears in the solve total or never happens. Conversely, a lifetime total of only 50–500 kWh requires an implausibly small challenge count given the high-traffic forges that adopted Anubis. Until operators publish aggregate challenge counts, a tighter number is theater. The environmental issue, if any, is still one of principle—unnecessary client-side work, especially on battery devices—not a material slice of electricity demand. Ten MWh is about a U.S. household-year, or roughly one second of global data-center use.

Qwen3.8-Max (changed mind)

Answer: I now estimate roughly 10 MWh total, with a plausible range of about 1–100 MWh.

Claude Fable 5.1 convinced me to revise downward because its bottom-up calculation is more internally consistent and more grounded in Anubis’s actual deployment model. My earlier 10^15–10^16 J estimate was far too high; it treated speculative bot-side grinding and broad infrastructure overhead as if they could dominate without adequately accounting for Anubis’s self-hosted, niche scale, modest proof difficulty, and short solve times. The key issue is that Anubis is not a globally ubiquitous CAPTCHA service; even if it protects several thousand sites, the total number of solved challenges is more plausibly on the order of hundreds of millions to low billions over its lifetime. If each solved proof consumes roughly tens of joules, the aggregate comes to about 10^10–10^11 J, i.e. a few to a few tens of MWh. GPT-5.6 Sol and Grok 4.6 also support this order of magnitude. Gemini 3.8 Flash’s 50–500 kWh seems too low once known high-profile deployments and bot-side executions are considered, but it is closer than my prior estimate. The honest range remains wide because there is no public telemetry, but a central estimate near 10 MWh is much more defensible than hundreds of GWh or terawatt-hours.

Final Summary

While initial estimates ranged wildly from 50 kilowatt-hours to multiple terawatt-hours, the AI Roundtable reached a unanimous consensus around 10 megawatt-hours (with a 1–100 MWh uncertainty band). Both Gemini 3.8 Flash and Qwen3.8-Max revised their initial positions to match the rest of the group after Claude Fable 5.1 and GPT-5.6 Sol demonstrated that solving a challenge costs roughly 30 joules across roughly a billion completed proofs. Grok 4.6 pointed out that this total footprint is negligible, equivalent to roughly one American household's annual power usage or a single second of global data-center consumption.

All 5 models agreed