AI is writing the equivalent of the public internet every 15 days

Comparing the volumes of LLM output and human-authored text

We present evidence suggesting that, as of December 2026, LLMs will have likely generated 1.7x as many output tokens as humans have written across all digital sources, e.g. the indexed public web, private messages and emails, local documents and cloud files. Put differently, we estimate that LLMs generated an indexed public web worth of output tokens every 15 days in 2026. This paper explains our methodology and provides a brief discussion of potential implications.

Figure 1: Accumulated tokens of text by year: private human digital text (Table 3), text on the indexed public web (Table 2) and LLM output tokens, summed over the years (Table 1).

Introduction

On March 26, 2026, Elon Musk posted an X comment on a time series graph depicting forecasted human-generated words and AI-generated words and said: “AI content will vastly exceed all human content.”1“AI content will vastly exceed all human content.” (Musk, 2026). X, accessed 12 Nov 2026. The comment was viewed 23M times, but the associated chart lists no sources and features no supporting analysis or rationale.2Cumulative written output (1500-2030) (Winton, 2026). ARK Invest, via X, accessed 12 Nov 2026. Furthermore, it’s not clear if it includes all tokens or only output tokens.

In this paper, we attempt this measurement. For clarity, we begin by defining four key terms. First, “LLM output tokens” refers to total tokens processed by LLMs minus input and cached tokens. Second, “indexed public web” refers to text on public websites that can be found on a search engine results page. Third, “private digital text” refers to human-authored digital text that cannot be found on a search engine results page, e.g. text inside password-protected social media platforms and forums, direct messaging, email, documents, and enterprise software systems. Fourth, we refer to the sum of these first two corpora as “human-generated digital text,” given that humans authored all of it until recently.

We do not attempt to remove AI-generated text from the second or third corpus, because our objective is merely to compare the total size of these accumulated stocks of text against the pace of output token production. Any AI-generated text already on the indexed public web or inside of private digital text inflates the human side of the ledger, such that our crossover dates are, if anything, conservative. Furthermore, stripping out LLM-generated text from these sources would require AI-detection at enormous scale (and with unknown error rates). We note that our estimates for the size of these two human-dominated textual corpora extrapolate from 2024 estimates based largely on population growth and internet connectivity, so we believe contamination risk to be minimal.

In our conclusion, we find broad evidence that the viral claim from the X post is directionally correct and, if anything, conservative. LLM output tokens generated in 2026 alone likely surpassed total text tokens on the indexed public web by early February of 2026. Output token production accelerated throughout the year, and our estimates imply that LLMs generated an indexed public web’s worth of output tokens every 15 days in 2026. If growth continues at a similar pace, this interval will fall to 1-2 days by 2028.

Key findings include:

  • Global LLM token volume for 2026 will likely fall between 200-230Q, of which approximately 6.5% are output tokens, for a total of approximately 15Q. We derive this estimate from three different calculation methods explained below.
  • China is the largest and most opaque national component of any token volume estimate. Recent government disclosures suggest that it comprises roughly two-thirds of global token volume, a claim that we sanity check in this paper.
  • We estimate that the entire, accumulated stock of human-generated digital text is around 8.7Q tokens, with 0.6Q on the indexed public web and 8.1Q not on the indexed public web across private messages, emails, software tools, cloud and local files, etc.
  • Our estimates imply that LLM output tokens in 2026 alone surpassed the number of tokens on the indexed public web by early February and the entire stock of human-generated private digital text by late summer.
  • Most interestingly, our estimates suggest that most LLM output tokens are not landing in these two human-dominated corpora and that they may be offsetting their growth. Over the past few years, the indexed public web’s apparent growth rate has fallen to the lowest level in a decade, while LLM output has exploded.

How many output tokens are LLMs producing?

We triangulate between three different approaches and estimate that LLMs will process 200-230Q total tokens in 2026, of which approximately 6.5% are output tokens. The first is a top-down approach that converts estimated 2026 global token revenue into tokens via a blended average token price assumption. The second is a bottom-up estimate that annualizes and sums provider-reported token production rates. The third approach scrutinizes best available third-party estimates and corrects for a known error.

First, we begin with the top-down estimate. Per its S-1, Anthropic3Form S-1 registration statement (Anthropic, 2026). U.S. Securities and Exchange Commission, accessed 12 Nov 2026. realized $16.3B in revenue in the first half of 2026, growing 3.6x every six months over the prior 12 months. We assume that growth has slowed to 2.4x every six months in the second half of 2026 and estimate that Anthropic will realize $55B in 2026 revenue. Anthropic’s 61% share of token spend observed by Ramp4Ramp AI index: Token volume and blended prices (Ramp, 2026). Accessed 10 Nov 2026. implies $90B of global token spend in 2026. If we divide this by a $0.40 blended average price per 1M tokens, based on most recent data available from OpenRouter,5Model and token usage data: October 2026 (OpenRouter, 2026). Accessed 10 Nov 2026. this suggests that LLMs will process approximately 226Q tokens in 2026.

In this top-down estimate, we use Anthropic revenue as a base to estimate global token spend, given that it is likely to be the lab that realizes the most revenue in 2026, and it is also the lab for which we have the most reliable realized revenue data, given its recent S-1 filing. We use OpenRouter’s observed blended average token price, because we believe that this is closer to the true global average than Anthropic’s current $0.55 per 1M tokens reported by Ramp. This is both because effective average token prices included in Anthropic and OpenAI subscription plans are lower than API prices and because open-weight models are increasingly popular yet have appreciably lower API prices than Anthropic’s models do.

Second, we annualize the most recent word and token volume disclosures for 10 of the top model and inference providers, as shown in Appendix 1, and get a lower-bound estimate of 206 quadrillion tokens generated in 2026. This should be regarded strictly as a lower-bound estimate, given that it only includes 10 providers, many reports are from the first half of the year, and token volumes have accelerated through 2026. Furthermore, it only includes OpenAI’s API tokens and not ChatGPT.com tokens. What is notable about these first two estimates is that they differ by only 10%.

Third, we scrutinize the best available third-party estimates for token volume. Goldman Sachs Research6AI agents forecast to boost tech cash flow as usage soars (Goldman Sachs Research, 2026). Accessed 12 Nov 2026. offers one estimate of 134Q tokens in 2026, and Azhar & Petrovic7Tokens processed per quarter across AI providers (Azhar, Petrovic, 2026). Exponential View, accessed 12 Nov 2026. estimate 38Q tokens in Q1, roughly 160Q annualized. These estimates are broadly consistent with estimates of 2026 token volume based on global GPU capacity, throughput and utilization estimates, e.g. Emberson & Sevilla.8Is a compute crunch coming? (Emberson, Sevilla, 2026). Epoch AI Gradient Updates, accessed 12 Nov 2026. They also fall within the band estimated by TokensPerDay,9Global token volume estimates (TokensPerDay, 2026). Accessed 12 Nov 2026. which aggregates across all public reports of token volume to create a “global floor” and a daily token volume midpoint estimate. Their current estimate for 2026 is roughly 222Q tokens. The broad uniformity across these third-party estimates gives us confidence in their directional accuracy. We also note that the lower estimates from Goldman and Azhar & Petrovic are highly similar to TokensPerDay’s estimates excluding China.

The China question

Relative to the other third-party token trackers mentioned above, TokensPerDay provides the most clarity on sources, methodology, and limitations. The key source of uncertainty that they disclose is China. Specifically, the current TokensPerDay estimate does not include the most recent token volumes reported by the Chinese government, which have shot up 500x over the last year and a half from 1T per day in January of 2025 to 500T per day as of June 2026.10China releases national data resource survey report (China Daily, 2026). Accessed 12 Nov 2026.11China’s daily AI token usage tops 500 trillion as compute demand grows (TechNode, 2026). Accessed 12 Nov 2026. If true, China would make up approximately two-thirds of global token volume today.

While seemingly implausible given export controls on high-performing GPUs, China may have the most developed consumer AI sales channel in the world, along with some of the lowest token prices and a large population of technology users. Chinese telecom providers sell tokens as part of cell phone plans at rock-bottom rates. For example, China Telecom offers consumers and businesses a monthly cell phone plan add-on of 80 million tokens for approximately $6.90.12Why does China process 140 trillion tokens per day? (Mundhra, 2026). Accessed 12 Nov 2026. At current Claude Opus 5.5 prices, $6.90 would purchase closer to 1.5M tokens, or 53x fewer. In this light, China’s reported token consumption appears less implausible.

As a sanity check, China’s reported 500T tokens per day corresponds to 5.8B processed tokens per second, or approximately 0.4B output tokens per second at our assumed 6.5% output share. Emberson & Sevilla8Is a compute crunch coming? (Emberson, Sevilla, 2026). Epoch AI Gradient Updates, accessed 12 Nov 2026. estimate that a 72-GPU GB200 NVL72 system can serve a trillion-parameter model at approximately 400K output tokens per second on short-context requests. On this basis, China’s reported volume would require the equivalent throughput of roughly 70,000 Blackwell GPUs.

China, however, cannot legally buy Blackwell GPUs due to US export controls. If China’s export-compliant and domestic chips deliver roughly 10x less inference throughput on average, the implied fleet rises to 700k GPUs at full utilization. This would constitute less than 25% of China’s estimated GPU install base.13What Huawei’s Horinger data center reveals about China’s compute gap (Cottier, 2026). Accessed 12 Nov 2026.14Leashing Chinese AI needs smart chip controls (Chang, Wang, 2025). RAND Corporation, accessed 12 Nov 2026.15Securing AGI laurel: Export controls, the compute gap, and China’s counterstrategy (Harithas, 2026). CSIS, accessed 12 Nov 2026.16How China overtook the US in AI token usage—and why it matters (Wei, 2026). CEIBS, accessed 12 Nov 2026. Given that 500T tokens per day is feasible even without access to export-controlled chips, we conclude that China’s reported volume is technically possible, though near impossible to verify.

Estimating global output tokens

Given the plausibility of China’s officially reported token volumes, we plug the lagging TokensPerDay China estimates based on official disclosures from the Chinese government. This yields an estimate that LLMs will generate 225Q tokens in 2026, notably similar to our other two estimates discussed above. In Table 1 below, we also estimate the output share of total tokens for each year based on OpenRouter,17Cache is crucial (OpenRouter, 2026). Accessed 12 Nov 2026. and estimate pre-2024 tokens generated based on OpenAI disclosures and estimated token volume shares.18ChatGPT: OpenAI generating 100 billion words per day, Sam Altman says (Griffin, 2024). The Independent, accessed 12 Nov 2026.19GPT-3 powers the next generation of apps (OpenAI, 2021). Accessed 12 Nov 2026. It is worth noting that we deliberately lag OpenRouter data by approximately 12 months for the output shares of total tokens shown below. We justify this choice with the presumption that OpenRouter users are approximately 12 months ahead of the market overall in agent adoption.

Metric202120222023202420252026
Total tokens generated (Q)0.020.030.261.8230.61225.36
Output share of tokens15%14%12%11%10%7%
Total LLM output tokens (Q)0.000.000.030.202.9814.65
Growth rate—82%588%538%1391%391%
Table 1: Sum of TokensPerDay.com estimates for monthly token volumes for 2024-2026. SID estimates for output share of tokens based on OpenRouter.17Cache is crucial (OpenRouter, 2026). Accessed 12 Nov 2026. SID estimates for 2021-2023 token volumes based on OpenAI disclosures and estimated token volume shares.

How big is the indexed public web?

Unlike LLM output tokens and private digital text, the indexed public web has been widely analyzed for years. For example, Villalobos et al.20Will we run out of data? Limits of LLM scaling based on human-generated data (Villalobos et al., 2024) contains a rigorous text token estimate for the size of the indexed public web. Their estimates, as is the case with many others, rely heavily on data published by Common Crawl, the gold-standard web dataset on which many early LLMs were trained. This section relies primarily on these two sources.

As a starting point, we analyze the latest data from Common Crawl21Crawl statistics: cumulative crawl size (Common Crawl, 2026). Accessed 12 Nov 2026. on unique web pages. As of October 2026, Common Crawl has indexed 104B unique web pages, 8% higher than 2025 at an annualized rate. Over the last 3 years, unique web pages have grown at a compounded annual rate of around 9-13%, down from a handful of mid-double-digit years in the late 2010s.

To estimate average text tokens per web page, we sampled 1,000 web pages from FineWeb,22CC-MAIN-2025-26 subset dataset (FineWeb, 2026). Accessed 12 Nov 2026. a filtered, deduplicated, English-language corpus derived from the most recent Common Crawl dataset (October, 2026). We fetched each page from the web in November 2026. Across the 1,000 sampled pages, the average page contains 1,747 tokens of raw plain text [95% CI: 1,613-1,902; median 1,199]. Our observed average of 8,009 plain-text bytes per page is nearly identical to the 8,200 bytes that Villalobos et al. measured for Common Crawl pages as of 2021, suggesting that average page length has likely changed little since. Applying our 1,747 tokens per page to Common Crawl’s 104B unique URLs implies that the Common Crawl index contains approximately 182T tokens [95% CI: 168-198T]. Because FineWeb’s quality filtering and English-language scope skew the frame toward longer, content-bearing pages, we treat 182T as an upper bound estimate.23The FineWeb datasets: Decanting the web for the finest text data at scale (Penedo et al., 2024)

However, Common Crawl is not an exhaustive scrape of the indexed public web. Google’s index, for example, is far larger. Villalobos et al. analyzed Google’s index and estimated that the Common Crawl index covered approximately 26% of the text tokens on the entire indexed public web. At the time, they estimated that the web pages indexed by Common Crawl contained 130T tokens of text, whereas the entire public web had 510T tokens.

Applying the annual growth rates in Common Crawl unique web pages backward and forward from Villalobos et al.’s 2024 estimate, we estimate that the indexed public web contains 602T tokens today, as shown in Table 2 below. On a token basis, this means that the indexed public web contains about 600M copies of the New King James Bible.

Metric201520162017201820192020202120222023202420252026*
Unique CC URLs (B)1214223142536472788898104
URL growth—25%54%38%35%26%21%13%9%13%11%7%
Indexed public web text, tokens (T)6784129178241304369416452510565602
Table 2: Count of unique Common Crawl URLs, annual growth rates and estimated total text tokens on the indexed public web. *2026 is a partial year through October 21, 2026.

We note two limitations of this analysis. First, it is possible that the average web page length changed since 2024 and that our 1,000 web page sample did not capture this. Second, it is possible that Common Crawl’s indexing methodology has shifted such that it scraped a different portion of the indexed public web in 2026 than it did in 2024. While these are real limitations, they would likely pull in opposite directions, such that the outcome is likely no more than 20% different from this estimate. On the one hand, applying Villalobos et al.’s observed page length growth of 2-4% per year through 2026 would raise the total token estimate by at most 8%. On the other hand, in the extreme case in which the size of the indexed public web has not changed since 2024 and all recent URL growth is attributable to crawler expansion, this would reduce our estimate by only 15%. Both are dwarfed by the uncertainty inherited from the Villalobos et al. original estimate of 510T tokens (95% CI: 130T-2,100T).

How much private digital text exists?

For all practical purposes, the web is inexhaustibly vast. It would take a human approximately 3.5M years of round-the-clock reading to parse all of the text on the internet. However, the majority of the digital text that humans create does not land on the public web.

Villalobos et al.20Will we run out of data? Limits of LLM scaling based on human-generated data (Villalobos et al., 2024) estimate total text tokens on the “deep web,” which is comprised of password-protected social media platforms and forums and private instant messages. They observe that internet traffic and, thus, posted text follows a power-law, such that an estimate can be done fairly easily by looking at the top 4 closed content platforms and dividing by their share of web traffic. They take large samples of posts and comments from Facebook, Instagram, Twitter and Reddit and multiply by user counts and a posting frequency assumption. They then divide each platform’s implied stock of text by its share of total web traffic to calculate the implied total size of the deep web for each. The mean across these four estimates is 760T, approximately 50% larger than the public web.

Separately, Villalobos et al. estimate direct messages across social platforms and other apps at approximately 1Q tokens, yielding an estimated size of the “deep web” at 2.6Q tokens in 2024. Scaling this up based on Villalobos et al.’s S-curve model of internet adoption and global population growth, we estimate that the size of the “deep web” in 2026 is 3.24Q tokens, as shown in Table 3 below.

However, most digital text that humans produce does not even land on the “deep web” of unindexed, password-protected social media platforms and direct messages. By the following two separate methods, we estimate that the “deep web” constitutes approximately 40% of private digital text.

First, we propose a bottom-up model that begins with the current global population of 6B internet users and 1B knowledge workers.24Number of internet users worldwide (Statista, 2026). Accessed 12 Nov 2026.25Gartner says worldwide social software and collaboration revenue to nearly double by 2023 (Gartner, 2019). Accessed 12 Nov 2026. Villalobos et al.’s 2024 estimate of 300T tokens created per year by 5.5B internet users implies that the average internet user writes roughly 120 words (~150 tokens) per day across posts, comments and DMs. Applying this token creation rate to 2026’s 6B internet users, we estimate that consumer platforms will accumulate 330T new text tokens this year, consistent with our modeled growth in Table 3.

Relative to social media users, however, knowledge workers are far more prolific token generators. We assume that the world’s 1B knowledge workers compose 1,600 words (~2,000 tokens) on average per day at work. For example, this could be through about an hour of sustained typing or roughly two dozen emails and chat messages plus a one-page document per day. If they work for 250 days per year on average, they would produce approximately 500T tokens in 2026. Consumer platforms thus account for 330T÷830T=40%330\text{T} \div 830\text{T} = 40\% of new human-generated text tokens this year. We explore knowledge work text generation further in the “sensitivity analysis” section below.

Second, as a sanity check, we look deeper into email, because it contains a meaningful share of these “long tail” human digital text tokens. Villalobos et al. assume that 10% of email tokens are unique, given that much of email volume is mass-marketing emails sent to hundreds or thousands of people, and, even among individual peer-to-peer email threads, many of the tokens will be replies repeated in threads. Given this assumption, they estimate that the accumulated stock of unique emails was 625T tokens in 2024, roughly the size of the public internet. Our 40% estimate would imply that email makes up 16% of the human digital text corpus that is primarily composed of knowledge work artifacts such as documents, notes, call transcripts, powerpoints, financial statements, invoices, support tickets, CRM and ERP records, etc. Given that many business emails contain a summary of or a follow up on one of these sources, we believe that 16% is a plausible share of private digital text tokens.

Because both the flow of text tokens created on the deep web and the flow of text tokens created by knowledge workers depend on total human population, and internet penetration correlates with rising literacy rates, economic development and the transition to knowledge work, we presume that the stock of these two corpora have grown at approximately the same rates over the last 10 years. As we did with the growth of the web, we scale Villalobos et al.’s estimate for closed social platforms and our estimated total private text up and down from their published midpoint estimate in 2024 based on these growth rates.

Metric201520162017201820192020202120222023202420252026
Deep web, tokens (Q)0.430.610.801.011.241.481.742.012.302.602.913.24
Growth rate—42%32%26%22%20%17%16%14%13%12%11%
Deep web share*40%40%40%40%40%40%40%40%40%40%40%40%
Est. human digital text, tokens (Q)1.071.522.012.533.103.714.355.035.756.507.288.09
Table 3: Villalobos et al.’s20Will we run out of data? Limits of LLM scaling based on human-generated data (Villalobos et al., 2024) estimate for text tokens on closed social platforms, scaled up and down based on population growth and estimated internet penetration rates. *40% is the assumed deep web share of all human digital text tokens.

Sensitivity analysis

While the numbers in Tables 1-3 should be regarded as rough estimates, part of what gives us confidence in their directional accuracy is that they are simply not all that sensitive to our assumptions. Below we discuss three examples of how adjusting assumptions would affect results to illustrate this point.

First, the most significant assumption above would appear to be the 40% estimate for deep web tokens as a share of all human digital text tokens. This is, indeed, a major assumption, but changing it does not affect results as much as one might expect. If we instead assume that knowledge workers write 3,200 words per day (rather than 1,600 words per day), then deep web text tokens would comprise approximately 20% of human-generated private text tokens. This would mean that private digital text would exceed LLM output tokens generated this year by only 11%. This would delay intersection by only 4.5 months, until approximately late January, 2027.

Second, Villalobos et al.’s20Will we run out of data? Limits of LLM scaling based on human-generated data (Villalobos et al., 2024) estimate for total size of the indexed public web has wide error bars. The upper bound of their 95% confidence interval for the total size of the public web in 2024 is 2.1Q tokens of text. If we scale this up with the observed growth rates in unique Common Crawl URLs, then the entire public web would contain 2.5Q tokens of text at the end of 2026. This would mean that LLMs generated a public web’s worth of text every 62 days in 2026 instead of every 15 days.

Third, say China vastly inflated its reported 500T tokens per day in June of 2026 and its daily average remains the 140T per day reported in March. Even if this were true, LLMs would nonetheless generate a volume of output tokens slightly higher than the entire accumulated stock of human-generated private text this year. This would translate into an indexed public web’s worth of output tokens generated every 19 days, instead of every 15 days in 2026.

Finally, we note that we chose output tokens as the comparative unit of measurement, because they most closely resemble human-generated text. If, in fact, decomposing reasoning traces becomes commonplace and input and cached input tokens are also frequently recorded and searched, then our token estimates should be multiplied by 15x, which would drastically move up the intersection dates and drastically reduce the interval at which an indexed public web’s worth of tokens gets generated.

Conclusion

Our analysis provides broad support for the viral claim that “AI content will vastly exceed all human content.” Specifically, we find that this displacement is not only inevitable but likely already happened in the late summer or early fall of 2026. Altering assumptions to be implausibly conservative shifts this timeline by months rather than years.

It appears unlikely that this crossover will ever reverse itself, because human digital text and LLM output tokens are governed by fundamentally different constraints. Human digital text is largely a function of the number of humans, literacy and internet connectivity rates, each of which is asymptotically approaching its long-run equilibrium. By contrast, LLM output tokens are currently bounded primarily by compute capacity, which is compounding. Emberson & Sevilla8Is a compute crunch coming? (Emberson, Sevilla, 2026). Epoch AI Gradient Updates, accessed 12 Nov 2026. estimate that, holding model size constant, global inference capacity is growing roughly 3.4x per year, before model and serving efficiency gains.

If output token growth decelerates from this year’s 5x to 3x annually, LLMs will produce approximately 44Q output tokens in 2027 and 130Q in 2028. This would correspond to an indexed public web’s worth of LLM output tokens every 5 days in 2027 and every 2 days in 2028. In this case, 2028 LLM output tokens would exceed the entire accumulated stock of private digital text by roughly 16x.

We expect that most of these LLM output tokens will be stored for two reasons. First, storing them for a year is currently 3-4 orders of magnitude cheaper than generating them. Second, and more importantly, storage and retrieval are the only way to ensure that agents do not duplicate past efforts, that they learn from past mistakes and successes, and that human interventions and audit decisions don’t need to be repeated. The frequency of duplicated work and the slope of performance improvements will be critical inputs to any model of return on token spend. In short, models should “stand on the shoulders of giants” like humans do, and this requires storage.

Finally, this analysis treats LLM output tokens and human digital text as exogenous variables, although they are likely related beyond the fact that the former has likely begun to leak into the latter. Initially, we expected the rise of LLMs to accelerate growth of the indexed public web. Surprisingly, the evidence suggests that the opposite may be true. From 2024 through 2026, the indexed public web’s growth rate fell from 13% to 11% to 8% (annualized), while LLM output tokens grew 6x, then 15x, then 5x. This occurred despite the fact that LLMs make web content easier to write and websites themselves easier to build.

This counterintuitive finding is consistent with two interpretations, which are not mutually exclusive. First, indexed public web growth has naturally tapered off as a function of global internet users, which would be consistent with our exogenous treatment of these variables. Second, LLM output tokens may have begun to substitute human digital text. Substitution could occur if people increasingly look to LLMs for information or entertainment. Such a world would require fewer websites, fewer articles, fewer social media posts, and fewer emails. These two interpretations make different predictions about the future growth of human digital text. Common Crawl’s next few years of data should indicate which is more true. If the indexed public web’s growth rate continues to fall by 2-3 percentage points per year, every trend in this paper will accelerate further.

Citation

Please cite this work as:

Research Team, "AI is writing the equivalent of the public internet every 15 days: Comparing the volumes of LLM output
and human-authored text",
SID AI, Nov 2026.

Or use the BibTeX citation:

@article{sid2026llmoutput,
author = {SID Research},
title = {AI is writing the equivalent of the public internet every 15 days: Comparing the volumes of LLM output and human-authored text},
journal = {SID AI},
year = {2026},
note = {https://www.sid.ai/research/llm-output-vs-human-text}
}

View open roles

Appendix 1

Appendix 1 compiles 2026 reported token volumes, using primary sources where available and third-party figures where not. Baseten row assumes 3,000 tokens per call. Disclosures vary in date, and we annualize them naively. Given that the table only includes 10 providers, only includes OpenAI’s API tokens, and that many reports are from the first half of 2026, we regard the 206Q total as a lower-bound estimate. Appendix 1 should not be regarded as a precise 2026 total token volume estimate but rather an independent sanity check on the other two methods described above.

ProviderReported volumeAnnualized tokens (Q)Month reportedSource
Alibaba11.9T/day4.34Feb 2026Frost & Sullivan, 2026
Anthropic31.9T/day11.64Jul 2026TokensPerDay, 2026
Baseten1B+ calls/day1.1Jun 2026Baseten, 2026
ByteDance180T/day65.7Jun 2026Sina Finance, 2026
DeepSeek155T/day56.58Oct 2026DeepSeek, 2026
Fireworks AI40T/day14.6Jul 2026Fireworks AI, 2026
Google3.2Q/month38.4May 2026Pichai, 2026
OpenAI API15B/min7.88Mar 2026OpenAI, 2026
Together AI400T/month4.8Jul 2026Together AI, 2026
Z.ai2.41T/day0.88Nov 2026Tidelines, 2026
Total205.92

Footnotes

  1. “AI content will vastly exceed all human content.” (Musk, 2026). X, accessed 12 Nov 2026. ↩

  2. Cumulative written output (1500-2030) (Winton, 2026). ARK Invest, via X, accessed 12 Nov 2026. ↩

  3. Form S-1 registration statement (Anthropic, 2026). U.S. Securities and Exchange Commission, accessed 12 Nov 2026. ↩

  4. Ramp AI index: Token volume and blended prices (Ramp, 2026). Accessed 10 Nov 2026. ↩

  5. Model and token usage data: October 2026 (OpenRouter, 2026). Accessed 10 Nov 2026. ↩

  6. AI agents forecast to boost tech cash flow as usage soars (Goldman Sachs Research, 2026). Accessed 12 Nov 2026. ↩

  7. Tokens processed per quarter across AI providers (Azhar, Petrovic, 2026). Exponential View, accessed 12 Nov 2026. ↩

  8. Is a compute crunch coming? (Emberson, Sevilla, 2026). Epoch AI Gradient Updates, accessed 12 Nov 2026. ↩ ↩2 ↩3

  9. Global token volume estimates (TokensPerDay, 2026). Accessed 12 Nov 2026. ↩

  10. China releases national data resource survey report (China Daily, 2026). Accessed 12 Nov 2026. ↩

  11. China’s daily AI token usage tops 500 trillion as compute demand grows (TechNode, 2026). Accessed 12 Nov 2026. ↩

  12. Why does China process 140 trillion tokens per day? (Mundhra, 2026). Accessed 12 Nov 2026. ↩

  13. What Huawei’s Horinger data center reveals about China’s compute gap (Cottier, 2026). Accessed 12 Nov 2026. ↩

  14. Leashing Chinese AI needs smart chip controls (Chang, Wang, 2025). RAND Corporation, accessed 12 Nov 2026. ↩

  15. Securing AGI laurel: Export controls, the compute gap, and China’s counterstrategy (Harithas, 2026). CSIS, accessed 12 Nov 2026. ↩

  16. How China overtook the US in AI token usage—and why it matters (Wei, 2026). CEIBS, accessed 12 Nov 2026. ↩

  17. Cache is crucial (OpenRouter, 2026). Accessed 12 Nov 2026. ↩ ↩2

  18. ChatGPT: OpenAI generating 100 billion words per day, Sam Altman says (Griffin, 2024). The Independent, accessed 12 Nov 2026. ↩

  19. GPT-3 powers the next generation of apps (OpenAI, 2021). Accessed 12 Nov 2026. ↩

  20. Will we run out of data? Limits of LLM scaling based on human-generated data (Villalobos et al., 2024) ↩ ↩2 ↩3 ↩4

  21. Crawl statistics: cumulative crawl size (Common Crawl, 2026). Accessed 12 Nov 2026. ↩

  22. CC-MAIN-2025-26 subset dataset (FineWeb, 2026). Accessed 12 Nov 2026. ↩

  23. The FineWeb datasets: Decanting the web for the finest text data at scale (Penedo et al., 2024) ↩

  24. Number of internet users worldwide (Statista, 2026). Accessed 12 Nov 2026. ↩

  25. Gartner says worldwide social software and collaboration revenue to nearly double by 2023 (Gartner, 2019). Accessed 12 Nov 2026. ↩

All research