Wednesday afternoon a message notification from the AI enthusiasts WhatsApp group appeared on my screen. It was from a good friend, and it said:
“when you start seeing stuff like this floating around, every projection is nonsense”
The culprit was a headline in the Wall Street Journal from Tuesday: “Anthropic Expected to Tell Investors It Sees Over $30 Trillion in Potential Revenue”. A whopping amount, as any enthusiastic AI model would have written it.
We have been writing The Compute for nearly four months, and every time we publish, a couple of readers tell us that we must have a typo in the numbers. The same happens whenever we talk about AI in conversations, meetings or events, we get raised eyebrows and jaw drops.
Tuesday’s $30tn figure is an order of magnitude or two larger than anything we have mentioned so far. So we set out to make this issue about understanding the numbers rather than reporting them.
What the WSJ headline describes is a total addressable market, or TAM: the revenue Anthropic would collect if it captured the WHOLE market and left no customers for any of its competitors or peers. As Reuters put it more precisely, the TAM is the annual revenue opportunity available at 100 per cent market share.
That is neither a target nor a forecast nor an expectation of revenue. It is a ceiling, probably unreachable and certainly lofty, and while this explains what the number is, it does nothing to tell us how big it is, or why it is that big.
How much is Thirty Trillion?
To fathom the scale, take time and start with a second.
A thousand seconds is about 17 minutes.
A million seconds is 11 and a half days, just shy of a fortnight.
Go up one step and a billion seconds is 31 years and eight months, most of a career.
The next step leaves history behind: a trillion seconds is 31,700 years, nearly twice as far back as the cave paintings at Lascaux.
Anthropic’s figure is 30 of those. Thirty trillion seconds is 951,000 years, more than three times as long as humans have been around.
Now back to the figure itself. $30tn is roughly the size of the US economy, which the International Monetary Fund’s April 2026 World Economic Outlook estimates at $32.4tn this year. The Outlook also estimates the world output at about $126tn, so Anthropic’s TAM is also a quarter of everything produced on earth.
To put things in perspective, $30tn is more than 12 and a half times the combined revenue of the 191 technology companies in the S&P 1500 index, which made around $2.4tn between them last year.
The Going Rate
Looking at AI frontier labs’ run rates, their annualised revenue calculated by multiplying a recent month or quarter, OpenAI just passed the $40bn mark this month (roughly double its end-2025 pace) and Anthropic exceeded $65bn at the end of July, with roughly 80 per cent of that contributed by enterprise customers. This means the two pure-plays together are running at around $105bn, which computes to a mere 0.35 per cent of the $30tn TAM we started with.
Google does not report Gemini figures separately, the number sits somewhere within the Google Cloud revenue, which came out at nearly $25bn last quarter, and we know that Gemini crossed the 1bn monthly active users threshold on August 11, 2026. Anthropic on the other hand does not disclose user count and most of its revenue comes from enterprise customers.
Which leaves us with OpenAI, where by dividing the $40bn annual run rate by the more than 1bn active users the company announced its models had reached on July 31, 2026, we get a figure of $40 per user per year. Now that is a far cry from the $10,000 per year we would need per user to reach $30tn, assuming 3bn users; that is ChatGPT, Gemini and another 1bn between Anthropic and everything else, and it is generous, because the same people are counted more than once across those services. Even if the user base is double, we are talking about generating a revenue of $5,000 per user per year, still more than 100 times today’s average revenue per user.
If the consumer number looks small, the enterprise sector holds bigger amounts and a much wider spread. According to Ramp’s spending data, the top 1 per cent of US companies spend about $7,500 per employee per month on AI, roughly $90,000 per year. That is well above the $5,000 a year the $30tn assumes. The top 10 per cent spend $611 a month, which annualises to $7,332, and clears the bar too. At first glance, the number seems attainable now, until we look at the median firm spend of $11.38 a month, or about $137 a year.
Some companies got a bill shock earlier this year and started capping their staff’s AI spend. Uber set the limit at $1,500 per employee per month, per tool, Tesla capped it at $200 per week, and SemiAnalysis surveyed per-employee caps ranging from $250 at an aerospace and defence manufacturer all the way up to $2,000 at Workday and Stripe. According to developer and writer Simon Willison, who ran the arithmetic on the Uber cap, an engineer running two tools at the ceiling will cost about $36,000 a year, roughly 11 per cent of a typical Uber engineer’s total compensation and well above the $5,000 level.
Divide $30tn by the world’s roughly 3.6bn workers and it comes to about $8,300 per worker per year. The average user generates $40 of that today, while the top 10 per cent of US companies pay $7,332 per employee and the median firm pays $137. Reaching $30tn requires every employer on earth to spend more than the top tenth of American firms do today.
Gone with the Token
A token is a fragment of text, roughly one word or a part of one, and the basic unit that AI counts. Put plainly, it is the currency of AI.
Google chief executive Sundar Pichai said at Google I/O on May 19, 2026 that Google processes more than 3.2 quadrillion tokens a month across its products. A year earlier the figure was 480 trillion. Two years earlier, in May 2024, it was 9.7 trillion. If 30 trillion seemed like a big number, 3.2 quadrillion is more than 100 times bigger. Qualcomm chief executive Cristiano Amon estimated global demand at 31.7 billion tokens every 10 seconds in 2026, and projected a 40-fold increase to 1.27 trillion tokens for the same 10 seconds by 2030.
Thankfully, AI is billed by the million tokens, which makes the numbers more palatable, and the price at a similar level of model quality has fallen roughly 1,000-fold in the past three years, according to Andreessen Horowitz. The Tokenando Token Price Index, which we have been tracking since early last week and whose methodology is published here, puts the average price of a million tokens from OpenAI at $1.76 today against $8.04 from Anthropic, more than four times as much for two US frontier labs. DeepSeek sits at around $0.25 and Mistral at around $0.17.
Falling prices have not produced falling bills, which is the Jevons paradox at work; William Stanley Jevons noticed in 1865 that more efficient steam engines increased Britain’s coal consumption rather than reducing it.
Bain found that token costs halved between December 2024 and December 2025 while tokens consumed grew four and a half times, so spending still more than doubled. Reaching a $30tn market requires that gap to persist, with consumption outrunning the price falls every year, which if we extrapolate from the two top AI labs’ current $105bn run rate, reaches $30tn within seven years… around 2033.
All these tokens carry a cost, and it is incurred at the inference level: the running of a trained model to answer a prompt, as distinct from the training runs where the model is built. That is why Anthropic is set to fundraise $100bn at a $2tn valuation, while SpaceX already raised $86bn at a $1.77tn valuation, having stated its AI TAM at $26.5tn in its IPO prospectus earlier this spring (out of an overall TAM of $28.5tn). Between them, both companies are using these ceilings to attract investment at unprecedented scale, however that money is not going into marketing, sales, customer service or salaries. Most of the funds go into token factories, the data centres hosting the machines that AI labs lease to produce the intelligence; the funds cover the cost of the land, the buildings and the machines themselves, as well as powering and cooling them.
Racks and Macs
Starting with the machines, these are housed in racks, fridge-sized cabinets where AI inference actually happens. Nvidia’s current flagship, the GB300 NVL72, contains 72 Blackwell Ultra accelerators (the graphics processing units, or GPUs, that run AI models) and about 20TB of high-bandwidth memory, stacked next to the GPUs for speed. The GB300 is sold as a single unit rather than chip by chip, and while Nvidia has never published a list price, estimates range from $3.7m per rack up to $6.5m. That is a spread of nearly 1.75 times on a cabinet bought by the tens of thousands every year, and it is due to the fact that each deal is negotiated privately, bundled with networking, software and support, and increasingly wrapped in financing (we covered that aspect in last week’s Issue XVI of The Compute).
On the opposite side of the machine spectrum is the latest product from Apple: the M5 Ultra Mac Studio, announced August 25, 2026, comes with 256GB of unified memory and a 16TB drive in its maximum available configuration today, and costs $18,299 as per Apple’s website (a 512GB version is planned for late October, at a yet to be disclosed price). For the sake of reference, one GB300 NVL72 features roughly 78 times as much memory (20TB against the Mac’s 256GB) and costs somewhere between 200 and 355 times the Mac’s price.
Going back to the racks, these machines are typically rented by the GPU-hour, though the biggest customers sign multi-year leases, and the reported terms of those leases give us the best perspective of what the capacity costs at scale. Anthropic pays about $1.25bn a month to SpaceX for Colossus 1, the data centre near Memphis, Tennessee that holds more than 220,000 GPUs, and for its successor Colossus 2 as it comes online: roughly 325,000 GPUs in all, which at around 4,500 racks would cost between $17bn and $29bn, assuming they were all Nvidia GB300s. The lease runs to May 2029 and is worth roughly $45bn if it runs its course; Google separately signed a similar lease at $920m a month for about 110,000 GPUs.
Nvidia claims the GB300 delivers inference at $0.12 per million tokens, a benchmark computed on the vendor’s own utilisation assumptions. At that rate, Anthropic’s monthly capacity leased at Colossus alone would buy around 10 quadrillion tokens, three times what Google says it processes each month across all its products.
From Wafers to Planes
Nvidia, which is trending towards a $400bn annualised run rate, provides most of the AI chips currently, but it is not the only company in the space. Amazon Web Services has more than $225bn of customer commitments for Trainium, its own AI chip, named initially for its training use but now sold for inference too. Google runs much of its AI workload on TPUs (tensor processing units, its in-house AI chips), and OpenAI showcased Jalapeño this week, an inference chip it developed in collaboration with Broadcom. AMD sells accelerators of its own, and Intel supplies CPUs, which agentic AI (models that carry out multi-step tasks rather than answer once) has put firmly back in demand.
Memory is a separate industry by itself. SK Hynix, Samsung and Micron make nearly all the world’s DRAM (dynamic random-access memory, the kind found in everyday computers), and they have shifted capacity towards the high-bandwidth memory that GPUs need, so prices for standard DRAM have risen sharply through 2026. That cost increase reaches well beyond AI, with Apple raising consumer device prices by up to $300 overnight on June 25, 2026, after having absorbed the cost increases for months. Nvidia’s financials are impacted too, and it expects its gross margin will drop from 75 per cent today to a floor of 71 to 72 per cent by the fourth quarter of fiscal 2027, with the drop not arriving immediately because memory is bought in advance and held in inventory.
One layer down, the company that builds nearly all of Nvidia’s chips is Taiwan Semiconductor Manufacturing Company, or TSMC, the world’s largest contract chipmaker, headquartered in Hsinchu, Taiwan. TSMC has forecast its 2026 capital expenditure at between $60bn and $64bn, and one of its leading-edge fabs turning out roughly 20,000 wafers a month (the silicon discs chips are cut from) costs $25bn to $35bn to build and equip. Those wafers convert to racks: each of them yields roughly 30 finished Blackwell Ultra GPUs, and since a GB300 rack holds 72 accelerators, a fab producing only Nvidia’s Blackwell chips would supply the GPUs for around 8,000 racks a month, about two and a half Colossus 1s. In July 2026 TSMC committed another $100bn to its site in Arizona, where the first fab has been producing since late 2024, with Nvidia’s Blackwell chips among its output, and the rest are under construction or planned; the commitment takes TSMC’s total investment in the United States to $265bn towards a planned 10 fabs in the Arizona site.
Inside every one of those cutting-edge fabs stands manufacturing equipment from ASML, of Veldhoven in the Netherlands, the only company on earth that makes extreme ultraviolet lithography machines, EUV for short: the machines that print the circuits onto the silicon. A standard EUV unit costs about $200m; the latest High-NA version, which draws even finer lines, costs about $400m, travels in 250 crates on seven planes, and roughly 10 of them will be produced in all of 2026. A single leading-edge fab would need around 10 standard EUV machines, of which ASML shipped 48 units in all of 2025, so equipping one new fab absorbs about a fifth of the world’s yearly supply; the 10 new TSMC fabs in Arizona would absorb the full output of two years at ASML.
Adding It Up
Assume the ceiling is reached and the market pays $30tn a year for work delivered by AI. At the 40 per cent gross margin Anthropic now projects, that revenue implies about $18tn a year in inference costs. At the SpaceX rent of $1.25bn a month, $18tn a year pays for 1,200 leases the size of Anthropic’s, which would hold around 390m GPUs in 5.4m GB300 racks: that is $20tn to $35tn of hardware. One leading-edge fab, supplying 8,000 racks a month, would need 56 years to produce the silicon for them. TSMC’s Arizona site, at its planned 10 fabs, would need nearly six years, and equipping those 10 fabs would absorb two years of ASML’s entire production. Note that every figure here is estimated at today’s rates, and token prices have already fallen 1,000-fold in three years, so today’s rates will not last. The power, the grid connections and the copper below the ground carry numbers of their own, and we are keeping those for future issues.
Back in the group chat, the conversation has moved on. A few members are discussing pooling money for one or more M5 Ultra Mac Studios, at $18,299 each: the price of a decent car. The plans range from running open-weight models and agents on their own hardware, all the way up to building a small data centre and renting it out.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
This week’s issue was fuelled by live albums or concert recordings of some old favourites. Started with Paolo Nutini a the 2014 iTunes Festival and his delivery of Candy and ended with R.E.M. bringing back old memories with their Live at the Olympia in Dublin. There is something about live performances that a studio version never quite gives you… the stage dynamics, the crowd, the small imperfections.
On the screen, we finally finished season three of Borgen, the Danish political drama. It took a while, but it was definitely worth it.
Elsewhere: we “built” a kick-ass club sandwich on Saturday, a lazy beach day on Sunday (I am tempted to coin the term “beachmaxxing”), and now with the weather cooling down… the running shoes are coming back out.
References
“Anthropic Expected to Tell Investors It Sees Over $30 Trillion in Potential Revenue”, Corrie Driebusch, The Wall Street Journal, August 25, 2026. https://www.wsj.com/tech/ai/anthropic-expected-to-tell-investors-it-sees-over-30-trillion-in-potential-revenue-a611efea (also the source, via FactSet, for the S&P 1500 technology revenue figure)
“Anthropic expected to tell investors it sees over $30 trillion in potential revenue”, Reuters, August 25, 2026. https://www.reuters.com/business/media-telecom/anthropic-expected-tell-investors-it-sees-over-30-trillion-potential-revenue-wsj-2026-08-25/
World Economic Outlook, International Monetary Fund, April 2026 (US and world output). https://www.imf.org/external/datamapper/NGDPD@WEO/OEMDC/ADVEC/WEOWORLD/USA
“Anthropic revenue run rate tops $65 billion”, Reuters, August 17, 2026. https://www.reuters.com/technology/anthropic-revenue-run-rate-tops-65-billion-source-says-2026-08-17/
“OpenAI’s Revenue Run Rate Tops $40 Billion Ahead of IPO”, Bloomberg, August 13, 2026. https://www.bloomberg.com/news/articles/2026-08-13/openai-s-revenue-run-rate-tops-40-billion-ahead-of-ipo
Alphabet second quarter 2026 results (Google Cloud revenue), SEC exhibit. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm
“One billion monthly users”, Google, August 11, 2026. https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/
“Building abundant intelligence”, OpenAI, July 31, 2026. https://openai.com/index/building-abundant-intelligence/
“How much does it cost to be AI-pilled?”, Ramp AI Index, June 26, 2026. https://ramp.com/data/ai-index-june-2026
Uber and the end of tokenmaxxing, Fortune, August 7, 2026 (Uber’s cap and exhausted 2026 budget). https://fortune.com/2026/08/07/uber-ai-spending-tokenmaxxing-is-over-cto/
“Tesla caps employee AI spending at $200/week”, Electrek, July 2, 2026, reporting The Information. https://electrek.co/2026/07/02/tesla-caps-employee-ai-spending-200-week/
“Tokenbudgeting: our conversations”, SemiAnalysis.
“Uber Caps Usage of AI Tools Like Claude Code to Manage Costs”, Simon Willison, June 3, 2026. https://simonwillison.net/2026/Jun/3/uber-caps-usage/
World Employment and Social Outlook, International Labour Organization, 2026 (global workforce). https://www.ilo.org/publications/flagship-reports/employment-and-social-trends-2026
“Google I/O 2026: Sundar Pichai’s opening keynote”, Google, May 19, 2026. https://blog.google/innovation-and-ai/sundar-pichai-io-2026/
Qualcomm Computex 2026 keynote coverage (Cristiano Amon token estimates), ServeTheHome, June 1, 2026. https://www.servethehome.com/qualcomm-computex-2026-live-coverage/
“LLMflation: LLM inference cost”, Andreessen Horowitz. https://a16z.com/llmflation-llm-inference-cost/
Tokenando Token Price Index. https://tokenando.ai/indices/ttpi
“How Token Economics Will Change Opex”, Bain & Company, June 10, 2026. https://www.bain.com/insights/how-token-economics-will-change-opex/
“Anthropic Lowers Gross Margin Projection as Revenue Skyrockets”, The Information, January 22, 2026. https://www.theinformation.com/articles/anthropic-lowers-profit-margin-projection-revenue-skyrockets (free summary: https://www.investing.com/news/stock-market-news/anthropic-trims-profit-margin-outlook-as-ai-operating-costs-rise--the-information-4459316)
Anthropic fundraise reporting, Financial Times. https://www.ft.com/content/840ac156-af1c-4a82-b260-ae791072fcfa
“Musk’s SpaceX prices record $75 billion IPO”, Reuters, June 11, 2026. https://www.reuters.com/world/musks-spacex-prices-record-75-billion-ipo-135-share-2026-06-11/
SpaceX S-1 (AI total addressable market), SEC. https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm
“NVIDIA Blackwell Ultra AI Factory Platform”, Nvidia, 2025 (GB300 NVL72 specifications). https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-Blackwell-Ultra-AI-Factory-Platform-Paves-Way-for-Age-of-AI-Reasoning/default.aspx
Loop Capital GB300 NVL72 estimate ($3.7m to $4m a rack): client note by analyst Ananda Baruah, March 2025, first reported by Investor’s Business Daily; as covered by 9to5Mac. https://9to5mac.com/2025/03/25/apple-is-about-to-spend-1-billion-on-nvidia-servers-for-ai-analyst/
Tom’s Hardware GB300 NVL72 reporting ($6m to $6.5m for an inference-optimised system), Anton Shilov, March 24, 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/price-of-nvidias-vera-rubin-nvl72-racks-skyrockets-to-as-much-as-usd8-8-million-apiece-but-server-makers-margins-will-be-tight-nvidia-is-moving-closer-to-shipping-entire-full-scale-systems
“Apple introduces new Mac Studio with M5 Max and M5 Ultra”, Apple, August 25, 2026. https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/
“New Compute Partnership with Anthropic”, SpaceXAI, May 6, 2026. https://x.ai/news/anthropic-compute-partnership
SpaceX-Anthropic compute agreement disclosure, SEC. https://www.sec.gov/Archives/edgar/data/1181412/000162828026040874/spacexukfwp.htm
“Beyond rockets and satellites, SpaceX is quietly building an AI compute business”, Fortune, July 19, 2026 (325,000 GPUs across the Colossus data centres). https://fortune.com/2026/07/19/spacex-ai-compute-renting-business-google-anthropic-pentagon-deals-revenue-valuation-elon-musk-colossus-data-centers/
SpaceX-Google compute agreement disclosure, SEC. https://www.sec.gov/Archives/edgar/data/1181412/000162828026041150/spacexagreementfwp.htm
Nvidia inference cost benchmark ($0.12 per million tokens). https://www.nvidia.com/en-us/solutions/ai/inference/
“NVIDIA Announces Financial Results for Second Quarter Fiscal 2027”, Nvidia, August 26, 2026 (revenue, gross margin guidance). https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx
Andy Jassy on Trainium customer commitments, Amazon. https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-amazon-chips-business-q1-2026-earnings
“OpenAI and Broadcom: the Jalapeño inference chip”, OpenAI. https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
“AI demand reshapes DRAM rankings in Q2 2026”, Counterpoint Research. https://counterpointresearch.com/en/insights/ai-demand-reshapes-dram-rankings-in-q2-2026
Apple price increases of June 25, 2026, CBS News. https://www.cbsnews.com/news/apple-price-hikes-macbook-ipad-2026/
TSMC 2026 capital expenditure guidance, Financial Times, July 2026. https://www.ft.com/content/491927e1-1532-486d-94ec-8d6ee2de7bcd
TSMC additional $100bn US commitment, US Department of Commerce, July 2026. https://www.commerce.gov/news/press-releases/2026/07/trump-administration-secures-additional-100-billion-us-semiconductor
“TSMC begins Blackwell manufacturing in Arizona”, Nvidia blog. https://blogs.nvidia.com/blog/tsmc-blackwell-manufacturing/
“The $250 million ASML printer behind Nvidia’s chips”, Reuters, July 28, 2026 (EUV and High-NA pricing). https://www.reuters.com/world/asia-pacific/250-million-asml-printer-behind-nvidias-chips-2026-07-28/
ASML Annual Report 2025 (48 EUV systems shipped). https://www.asml.com/en/investors/annual-report/2025
“Rapidus to install 10 EUV chipmaking tools”, Tom’s Hardware (EUV machines per leading-edge fab). https://www.tomshardware.com/tech-industry/rapidus-to-reportedly-install-10-euv-litho-tools-into-its-fab-in-japan
The fab construction cost range, the 30-GPUs-per-wafer yield and all descent arithmetic in “Adding It Up” are industry estimates and Tokenando calculations from the figures above, at the stated assumptions.