inthesingularity.ai
Wire · AI demand · AT&T

W.133AT&T’s 45 billion tokens a day is the figure it gave in July

AT&T processes about 45 billion AI tokens a day, SiliconANGLE reported on November 14, citing an executive who spoke at an Amazon Web Services event. AT&T published the same figure on July 23. About 27 billion was reported in late February and early March, and 8 billion for 2025. At current list prices, 45 billion tokens a day costs from about $2.4M a year on a small open model to about $296M on the most expensive model we priced, on an assumed mix of input and output. AT&T has not said what it spends.

Tokens a day, AT&T, November
≈45B
the same as its July 23 figure, 114 days before the article
Growth from its 2025 figure
5.6×
from 8B; 1.67× the 27B reported in February and March
A year of it at list prices
$2.4M–296M
Gemma 4 31B to GPT-6 Astra, a 122-fold range; 80% input assumed
Tokens a year at that rate
16.4T
16 times Microsoft’s trillion-token customer mark

What was reported

The figure comes from an article built on a conference talk. SiliconANGLE reported on November 14, citing Luke Corbin, an assistant vice president at AT&T Business who spoke at an Amazon Web Services event in Boston earlier this month, that AT&T processes about 45 billion AI tokens daily across the company. Corbin said at the event that its production AI investments are returning fivefold within a year. The article’s own narration adds that AT&T has nearly 500 agentic workflows in production and plans to triple that number next year, and that its agentic platform handles more than 1 million transactions a month.

AT&T published the same number in July. Andy Markus, its chief data and AI officer, wrote on the company’s blog on July 23 that it processes an average of 45 billion tokens per day. The November figure is “about 45 billion”. Either usage has been level for about four months or the July statistic was repeated; the article does not say which.

The article names no models, vendors or costs. Amazon Web Services appears only as the host of the event. The article gives the token figure as across the company. It does not say whether the workflow count and the platform’s transaction count cover all of AT&T, so the figures should not be divided into each other. The article says agents run autonomously only once accuracy is in the mid- to high-90% range.

AT&T’s stated volume rose from 8 billion to 45 billion tokens a day by July, and has been repeated at about that level since

Tokens AT&T executives have said the company processes per day, billions, at the four dates we found. 2025: “last year”, by Mark Austin’s account to Fierce Network. February–March 2026: 27 billion, from VentureBeat on February 26 and SiliconANGLE on March 5, where it was given for the Ask AT&T platform. July 23: AT&T’s blog, an average. November (pink): the figure SiliconANGLE attributed to Luke Corbin, “about 45 billion”. The bars are statements, not a measured series, and are not evenly spaced in time.

010B20B30B40B50B2025: about 8 billion tokens a day. Mark Austin, vice president of data science and AI, describing last year (Fierce Network, 2026)8B2025“last year”Feb–Mar 2026: about 27 billion tokens a day. VentureBeat’s article of February 26, in its own narration; Mark Austin on theCUBE, March 5, for the Ask AT&T platform (VentureBeat; SiliconANGLE)27BFeb–Mar 2026Feb 26, Mar 5Jul 23, 2026: about 45 billion tokens a day. Andy Markus, chief data and AI officer, on AT&T’s blog: an average of 45 billion tokens per day (AT&T)45BJul 23, 2026AT&T blogNov 2026: about 45 billion tokens a day. Luke Corbin, assistant vice president, as cited by SiliconANGLE: about 45 billion, across the company (SiliconANGLE, Nov 14)45BNov 2026SiliconANGLE
010B20B30B40B50B2025: about 8 billion tokens a day. Mark Austin, vice president of data science and AI, describing last year (Fierce Network, 2026)8B2025Feb–Mar 2026: about 27 billion tokens a day. VentureBeat’s article of February 26, in its own narration; Mark Austin on theCUBE, March 5, for the Ask AT&T platform (VentureBeat; SiliconANGLE)27BFeb–MarJul 23, 2026: about 45 billion tokens a day. Andy Markus, chief data and AI officer, on AT&T’s blog: an average of 45 billion tokens per day (AT&T)45BJul 23Nov 2026: about 45 billion tokens a day. Luke Corbin, assistant vice president, as cited by SiliconANGLE: about 45 billion, across the company (SiliconANGLE, Nov 14)45BNov
The four statements
DateTokens a dayWho said itSource
20258 billionMark Austin, vice president of data science and AI, describing last yearFierce Network, 2026
Feb–Mar 202627 billionVentureBeat’s article of February 26, in its own narration; Mark Austin on theCUBE, March 5, for the Ask AT&T platformVentureBeat; SiliconANGLE
Jul 23, 202645 billionAndy Markus, chief data and AI officer, on AT&T’s blog: an average of 45 billion tokens per dayAT&T
Nov 202645 billionLuke Corbin, assistant vice president, as cited by SiliconANGLE: about 45 billion, across the companySiliconANGLE, Nov 14

How fast it grew, and what has been said since July

The growth came before July. Forty-five billion is 5.6 times the 8 billion AT&T gives for 2025 and 1.67 times the 27 billion Mark Austin gave for its Ask AT&T platform on March 5, 140 days before the July post. From July 23 to the November article, 114 days, the same round figure was repeated. In an undated Fierce Network article written after July, Austin put it at upwards of 45 billion.

VentureBeat’s February article carries both early figures. Its headline and text give 8 billion a day as AT&T’s average, and the text then says 27 billion, a more than threefold increase in months. Fierce Network places the 8 billion in “last year”. We read 8 billion as a 2025 figure and 27 billion as the early-2026 one.

The stated return rose with the volume. Markus told CIO.com in January that AT&T was getting a two-times return on its AI investment. By August the company was saying fivefold within the year, the figure Corbin repeated in November. Tripling nearly 500 workflows would be nearly 1,500 next year; AT&T has given no token forecast to go with it.

At list prices a year of 45 billion tokens a day costs $2.4M on Gemma 4 31B and $296M on GPT-6 Astra

What 45 billion tokens a day for 365 days would cost at each model’s list price, $ millions, if 80% of the tokens were input and 20% output. The split is our assumption: AT&T does not disclose it. Strong grey: small models. Faint grey: mid-priced. Pink: the two most expensive we priced. Prices are per million tokens as read on November 16, 2026, and ignore caching, batch discounts, negotiated rates and self-hosting, all of which lower the bill. The table below gives the all-input and all-output bounds.

SmallMid-pricedMost expensive
0$100M$200M$300M$400MGemma 4 31Bopen weights, via OpenRouter; $0.10 in, $0.34 outGemma 4 31B: $0.10 input and $0.34 output per million tokens; blended $0.148; 45 billion tokens a day for 365 days: about $2.4M (all input $1.6M, all output $5.6M)$2.4M a year · $0.148 blendedGPT-6 LunaOpenAI; $0.10 in, $0.50 outGPT-6 Luna: $0.10 input and $0.50 output per million tokens; blended $0.180; 45 billion tokens a day for 365 days: about $3.0M (all input $1.6M, all output $8.2M)$3.0M a year · $0.180 blendedGemini 3.8 FlashGoogle, via OpenRouter; $0.75 in, $3.75 outGemini 3.8 Flash: $0.75 input and $3.75 output per million tokens; blended $1.350; 45 billion tokens a day for 365 days: about $22.2M (all input $12.3M, all output $61.6M)$22.2M a year · $1.35 blendedClaude Sonnet 5.5Anthropic; $2 in, $10 outClaude Sonnet 5.5: $2 input and $10 output per million tokens; blended $3.600; 45 billion tokens a day for 365 days: about $59.1M (all input $32.9M, all output $164.3M)$59.1M a year · $3.60 blendedClaude Fable 5.2Anthropic; $8 in, $40 outClaude Fable 5.2: $8 input and $40 output per million tokens; blended $14.400; 45 billion tokens a day for 365 days: about $236.5M (all input $131.4M, all output $657.0M)$237M a year · $14.40 blendedGPT-6 AstraOpenAI; $10 in, $50 outGPT-6 Astra: $10 input and $50 output per million tokens; blended $18.000; 45 billion tokens a day for 365 days: about $295.7M (all input $164.3M, all output $821.3M)$296M a year · $18.00 blended
0$200M$400MGemma 4 31B$0.10 / $0.34Gemma 4 31B: $0.10 input and $0.34 output per million tokens; blended $0.148; 45 billion tokens a day for 365 days: about $2.4M (all input $1.6M, all output $5.6M)$2.4MGPT-6 Luna$0.10 / $0.50GPT-6 Luna: $0.10 input and $0.50 output per million tokens; blended $0.180; 45 billion tokens a day for 365 days: about $3.0M (all input $1.6M, all output $8.2M)$3.0MGemini 3.8 Flash$0.75 / $3.75Gemini 3.8 Flash: $0.75 input and $3.75 output per million tokens; blended $1.350; 45 billion tokens a day for 365 days: about $22.2M (all input $12.3M, all output $61.6M)$22.2MClaude Sonnet 5.5$2 / $10Claude Sonnet 5.5: $2 input and $10 output per million tokens; blended $3.600; 45 billion tokens a day for 365 days: about $59.1M (all input $32.9M, all output $164.3M)$59.1MClaude Fable 5.2$8 / $40Claude Fable 5.2: $8 input and $40 output per million tokens; blended $14.400; 45 billion tokens a day for 365 days: about $236.5M (all input $131.4M, all output $657.0M)$237MGPT-6 Astra$10 / $50GPT-6 Astra: $10 input and $50 output per million tokens; blended $18.000; 45 billion tokens a day for 365 days: about $295.7M (all input $164.3M, all output $821.3M)$296M
Prices and bounds
ModelInput, $/MOutput, $/MBlended 80/20A year, blendedAll inputAll output
Gemma 4 31B$0.10$0.34$0.148$2.4M$1.6M$5.6M
GPT-6 Luna$0.10$0.50$0.180$3.0M$1.6M$8.2M
Gemini 3.8 Flash$0.75$3.75$1.350$22.2M$12.3M$61.6M
Claude Sonnet 5.5$2$10$3.600$59.1M$32.9M$164M
Claude Fable 5.2$8$40$14.400$237M$131M$657M
GPT-6 Astra$10$50$18.000$296M$164M$821M

What it would cost

Each dollar per million tokens is about $16.4M a year at this volume. On an 80/20 mix of input and output, a year costs about $2.4M on Gemma 4 31B at OpenRouter’s listed price; that is the open model AT&T’s telecom model, OTel 2.0, is post-trained from, according to the GSMA. It costs about $59.1M on Claude Sonnet 5.5, $237M on Claude Fable 5.2 and $296M on GPT-6 Astra. The most expensive is 122 times the cheapest.

Even the top of that range is small beside AT&T’s costs. AT&T reported $28.2B of operating expenses in the third quarter, including a $2.6B impairment, and expects an additional $4B of annual cost savings by 2028. $296M is 7.4% of that savings target and 0.26% of the quarter’s operating expenses times four. AT&T has not published its AI spending.

The range is why AT&T talks about routing. Its July post says its AI gateway cuts costs by as much as 90%, and VentureBeat reported savings of up to 90% in February after AT&T rebuilt its orchestration layer. Executives have put open models’ share of its usage at between about a quarter and 40% this year in different interviews, with targets of 65% to 80%. At list prices a tenfold saving needs only a shift from a mid-priced model to a small one.

At 16.4 trillion tokens a year AT&T would clear the highest of three customer thresholds Microsoft and Alphabet have cited

Tokens a year, trillions. Faint grey: thresholds two vendors have used to count large customers. Microsoft said on April 29 that over 300 Foundry customers were on track to process over one trillion tokens this year; Alphabet said on its third-quarter call that more than 2,600 customers had used at least 100 billion tokens in 12 months and nearly 75 at least 10 trillion on Gemini Enterprise. Pink: AT&T’s 45 billion a day multiplied by 365. AT&T says it uses Microsoft Foundry and runs open models, but it has not said how its tokens divide between vendors, and nothing says it is among the customers counted.

05T10T15T20TThresholds vendors use to count large customersAlphabet: 100 billionmore than 2,600 customers, in 12 monthsAlphabet: 100 billion tokens: more than 2,600 customers, in 12 months0.1TMicrosoft Foundry: 1 trillionover 300 customers on track, this yearMicrosoft Foundry: 1 trillion tokens: over 300 customers on track, this year1TAlphabet: 10 trillionnearly 75, on Gemini Enterprise, in 12 monthsAlphabet: 10 trillion tokens: nearly 75, on Gemini Enterprise, in 12 months10TAT&TAT&T, 45 billion a day × 365its own figure; vendor split not disclosed45 billion × 365 = 16.425 trillion tokens a year16.4T
05T10T15T20TVendor thresholdsAlphabet, 100B2,600+ customersAlphabet: 100 billion tokens: more than 2,600 customers, in 12 months0.1TMicrosoft, 1T300+ customersMicrosoft Foundry: 1 trillion tokens: over 300 customers on track, this year1TAlphabet, 10Tnearly 75Alphabet: 10 trillion tokens: nearly 75, on Gemini Enterprise, in 12 months10TAT&T45B a day × 365split undisclosed45 billion × 365 = 16.425 trillion tokens a year16.4T

How large it is

Large for one customer, small for a platform. At 45 billion a day AT&T would process 16.4 trillion tokens in a year, or about 1.37 trillion a month. That is 16 times the one-trillion mark Microsoft used for its largest Foundry customers and above the 10 trillion Alphabet used for its top tier of nearly 75, which it counted on Gemini Enterprise.

Google’s own APIs process AT&T’s day in about 1.4 minutes. Alphabet said in October that its first-party model APIs handle nearly 32 billion tokens a minute, about 46.1 trillion a day. ByteDance said in June that its Doubao models had passed 180 trillion tokens of daily usage, by one account of its conference; AT&T’s figure is 0.025% of that.

What this changes on inthesingularity.ai

  • How fast is AI demand growing? Lead: one large enterprise’s stated volume rose 5.6 times from 2025 to July 2026 and was stated at the same level in November. A second enterprise series would show whether a pause, if it is one, is common.
  • How fast are token prices falling? Lead: the same volume spans $2.4M to $296M a year across list prices, so an enterprise’s bill depends more on its model mix than on its token count.
  • Who makes money serving open models? Lead: AT&T’s executives describe open models at a quarter to 40% of usage this year with targets of 65% to 80%; its telecom model is post-trained from Gemma 4 31B, the GSMA says.
  • Lead for a tracker: AT&T’s next stated tokens-per-day figure. A number above 45 billion would show growth has resumed or never stopped; none higher has been published since July 23, though Mark Austin told Fierce Network the figure was upwards of 45 billion.

Sources and method

ItemFigureSource
The reportAbout 45 billion AI tokens daily across the company, according to Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business; Corbin, at an Amazon Web Services event in Boston: production AI investments returning fivefold within a year. In the article’s narration: nearly 500 agentic workflows in production, to triple next year; more than 1 million transactions a month on the agentic platform; autonomy only after accuracy reaches the mid- to high-90% rangeSiliconANGLE, Nov 14
The reportSummary of the SiliconANGLE articleAI Weekly, Nov 15
AT&T, earlierAn average of 45 billion tokens per day; its AI gateway cuts costs by as much as 90%; its telecom model OTel 2.0 runs through Microsoft FoundryAT&T blog, “The Tokenomics Equation”, Jul 23
AT&T, earlier8 billion tokens a day in the headline and text, and 27 billion in the article’s narration, a more than threefold increase in months; up to 90% cost savings after the orchestration layer was rebuilt; more than 100,000 employees with access to Ask AT&T WorkflowsVentureBeat, Feb 26
AT&T, earlierAsk AT&T processes roughly 27 billion tokens per day (Mark Austin)SiliconANGLE (theCUBE at MWC), Mar 5
AT&T, earlierAbout 8 billion tokens a day last year, now upwards of 45 billion (Mark Austin); about 150 AI use cases in production; open models at about 25% of tokens, 25–40% expected this year, and an aim of 65–70%Fierce Network, 2026 (undated)
AT&T, earlierOTel 2.0 is a post-trained version of Gemma 4 31BGSMA Newsroom, Jul 24
AT&T, earlierA two-times return on AI investmentCIO.com, Jan 28
AT&T, earlierA fivefold return within the year; about 45 billion tokens a day; a plan to run 70–80% of AI workloads on open models eventuallyThe AI Innovator, Aug 13
AT&T, scaleThird quarter of 2026: revenue $32.3B; operating expenses $28.2B, including a $2.6B non-cash impairment; capital expenditures $5.1B; 2026 capital investment guided to $23–24BAT&T release, Oct 21, as carried by Finviz
AT&T, scaleAn additional $4B of annual cost savings expected by 2028AT&T Q4 2025 call, Jan 28, transcript via The Motley Fool
PricesGemma 4 31B: $0.10 input, $0.34 output per million tokensOpenRouter, read Nov 16
PricesGemini 3.8 Flash: $0.75 input, $3.75 output, shown as a discounted priceOpenRouter, read Nov 16
PricesGPT-6 Luna: $0.10 and $0.50; GPT-6 Astra: $10 and $50, for prompts up to 272,000 tokensOpenAI pricing, read Nov 16
PricesClaude Sonnet 5.5: $2 and $10; Claude Fable 5.2: $8 and $40Anthropic models overview, read Nov 16
Other volumesGoogle’s first-party model APIs: nearly 32 billion tokens a minute; more than 2,600 customers used at least 100 billion tokens in 12 months and nearly 75 at least 10 trillion on Gemini Enterprise (Thomas Kurian)Alphabet Q3 2026 call transcript, Oct 28
Other volumesOver 300 Foundry customers on track to process over one trillion tokens this yearMicrosoft FY2026 Q3 call, Apr 29
Other volumesByteDance’s Doubao passed 180 trillion tokens of daily usage, stated at a June 23 conferenceLeon Liao, Jul 4, relaying Volcano Engine

What we read. The SiliconANGLE articles, AT&T’s July blog post, the GSMA’s, VentureBeat’s, Fierce Network’s, CIO.com’s and The AI Innovator’s articles, Alphabet’s and Microsoft’s call transcripts and the price pages were read directly. AT&T’s third-quarter figures are from its release as carried by Finviz, not from the 10-Q. The Doubao figure is from a secondary account. Growth. 45 ÷ 8 = 5.625; 45 ÷ 27 = 1.667. March 5 to July 23 is 140 days; July 23 to November 14 is 114. Nearly 500 × 3 = nearly 1,500. Volume. 45 billion × 365 = 16.425 trillion; × 30.4 = 1.368 trillion. Cost. 45 billion tokens is 45,000 million, so $1 per million tokens is $45,000 a day and $45,000 × 365 = $16.425M a year. Blended price = 0.8 × input + 0.2 × output: Gemma 4 31B 0.8 × 0.1 + 0.2 × 0.34 = 0.148, × 16.425 = $2.4M; GPT-6 Luna 0.8 × 0.1 + 0.2 × 0.5 = 0.180, × 16.425 = $3.0M; Gemini 3.8 Flash 0.8 × 0.75 + 0.2 × 3.75 = 1.350, × 16.425 = $22.2M; Claude Sonnet 5.5 0.8 × 2 + 0.2 × 10 = 3.600, × 16.425 = $59.1M; Claude Fable 5.2 0.8 × 8 + 0.2 × 40 = 14.400, × 16.425 = $236.5M; GPT-6 Astra 0.8 × 10 + 0.2 × 50 = 18.000, × 16.425 = $295.7M. 18.0 ÷ 0.148 = 121.6. $295.7M ÷ $4,000M = 7.39%. $28.2B × 4 = $112.8B; $295.7M ÷ $112,800M = 0.262%. Comparisons. 32 billion × 1,440 minutes = 46.08 trillion a day; 45 ÷ 32 = 1.41 minutes. 45 billion ÷ 180 trillion = 0.025%. Caveats: the 80/20 split is assumed, and agent workloads that reread long contexts can be more input-heavy still, which would lower the blended price. List prices overstate what a customer of this size pays. The Gemini price was shown as discounted; Gemma is also offered free by some hosts. The definitions of a token count differ between companies, and AT&T does not say whether cached tokens are included. The 2025 figure has no exact date, and the March figure was given for the Ask AT&T platform, so the series may not be like for like. Not used: a 1 billion tokens-a-day figure for September 2024 that appears only in an unattributed secondary source; counts of AT&T use cases, which follow different definitions; headcount.

Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.