Own the AI Rails
Nov 14, 2026
24 min Read
Own the AI Rails
*The internet is becoming agentic. Access to it is being handed out for free. Ownership of it is being quietly collected.*
The internet is turning from a place you visit into a system that acts for you. Software reads, decides, buys and builds on your behalf. That is the agentic internet. It is already here in run-rate form.
Everything that follows argues one question. When machines do the work, who owns the data they run on, the compute they run with, and the price they pay to run at all? Two answers exist. Access, on someone else's terms. Or ownership, on yours. They look the same today. They will not look the same in 2028.
Executive summary
- **The consumer story was wrong, and then it was right.** ChatGPT launched in November 2022 as a chat toy and hit 100 million users in two months. [1] The consensus was that AI would handle small consumer tasks and never touch lawyers, doctors or engineers. By 2026, coding is the largest AI revenue line on earth. Anthropic's run-rate went from $1 billion to $44 billion in seventeen months, led by a coding product. [2]
- **The consumer layer is enormous and owned by a handful.** a16z has tracked the top 100 consumer AI apps across seven editions since 2023. ChatGPT has been #1 in every one. [3–9] It holds 79% of consumer chatbot use and 1.2 billion weekly users. [10,11] Only 11 new products made the list in the latest edition, the fewest ever. [9]
- **Every request is a transfer of information.** Prompt, files, corrections, memory, metadata and agent logs all go to the model owner. 73% of consumer messages are now personal. [12] A third of US adults take health questions to a chatbot. [13] Retention has stretched from 30 days to five years. [14]
- **The build-out is historic.** About $416 billion of AI capex in 2025, near $780 billion in 2026, over $1 trillion forecast for 2027. [15] Google processes 3.2 quadrillion tokens a month. [16]
- **Adoption is wide, not deep.** 69% of S&P 500 companies have live AI. About 2% track its effect. 4.5% of US consumers pay. [15,9] Agent scaling within any one function stays near 10%. [17]
- **Price is a tactic, not a floor.** Inference cost falls about 10x a year. [18] OpenAI projects $278 billion of negative cash flow through 2030 and has already started selling ads. [19,11] Someone pays for the free tier.
- **Caution is collapsing while concentration rises.** Enterprise bans on feeding GenAI fell from 28% to 7%. [20] Closed models hold 87% of enterprise workloads. [21] Three clouds hold 63% of spend. [22]
- **Machine payments already work.** x402 went from 58,800 to 138.5 million transactions in four quarters. [23] The agent economy has rails. The question is who lays them.
- **Our position.** The counterweight is ownership. Local and verifiable inference. Data licensed, not surrendered. Compute paid per call. Cluster Protocol is built as the orchestration layer for that internet.
Contents
Part I · Where we are
- 01 · From toy to tool: 2022 to 2026
- 02 · The consumer layer, seven editions
- 03 · The handful
- 04 · How an AI loop actually works
- 05 · What you tell it
- 06 · Where it goes, and how long it stays
Part II · The question
- 07 · The build-out
- 08 · Wide, not deep
- 09 · Price is a tactic, not a floor
- 10 · Concentration, and the quiet switch
- 11 · The quiet surrender of data
Part III · The debate and the answer
- 12 · Access or ownership
- 13 · Agents need a way to pay
- 14 · Where Cluster Protocol stands
- 15 · What to do with this
Part I · Where we are
*Ground evidence. How AI arrived, who captured it, and what it takes from you each time it works.*
01 · From toy to tool: 2022 to 2026
ChatGPT went public on 30 November 2022. By January 2023 it had 100 million monthly users. TikTok took nine months to get there. Instagram took two and a half years.¹
The first reaction was consumer-shaped. People used it to write birthday poems and summarise emails. The serious verdict was that AI would take the small stuff. In March 2023 Goldman Sachs said 300 million jobs were exposed, with administrative work at 46% of tasks and legal at 44%.²⁴ Even that forecast was read as a story about paperwork, not about professions.
Three years later the chart reads differently. Legal is second from the top, and Harvey, a legal AI company, is valued at $11 billion.²⁵ OpenEvidence, a medical assistant, is used by an estimated 50–60% of US physicians.⁹ And the profession everyone said would be last, software engineering, became the largest AI market of all.
Why Coding became the multi-billion-dollar line
Coding is where the loop closes fastest. Code either runs or it does not. The model gets a clean signal on every attempt. Developers correct it constantly, and every correction is training data. So the product improved faster than anywhere else, and the money followed.
Cursor's revenue run-rate went from about $100 million in early 2025 to $500 million in June, $1 billion in November, and more than $2 billion by February 2026.²⁶ Claude Code, launched in May 2025, reached a $1 billion run-rate in six months and about $2.5 billion by February 2026. [7,2]
The parent company tells the same story at larger scale. Anthropic's run-rate was about $1 billion in December 2024. It was $9 billion in December 2025, $19 billion in March 2026, and passed $44 billion in May 2026, according to SemiAnalysis. [2] Enterprise use accounts for more than half of Claude Code revenue, and the product is estimated to touch about 4% of public GitHub commits. [2]
The developers themselves confirm both halves. Stack Overflow's 2025 survey of 49,000 developers found 84% use or plan to use AI tools, up from 76%. In the same survey 46% said they do not trust the accuracy of the output, up from 31%. [28] Use went up. Trust went down. People kept using it anyway.
That is the shape of the whole era in one chart. Dependence rises faster than trust. The gap between the two is where the data goes.
We said AI would never be the lawyer, the doctor or the engineer. We were wrong on all three. We were right that it would need everything we know.
02 · The consumer layer, seven editions
a16z has published the Top 100 Gen AI Consumer Apps seven times since September 2023. Read in sequence, the seven editions are the best public record of how the consumer layer formed and who it formed around. [3–9]
Three lines run through all seven.
The top never changes. ChatGPT has been #1 on web and mobile in every edition. Its multiple over #2 was 5x in 2024 and is still 2x to 2.7x in 2026, with the runner-up now a Google product. [4,8,9]
The list is closing. New entrants fell from 22 to 12 to 17 to 11 to 11. [4–9] Fewer new products can break in. The ones that do are often features of the incumbents: Google alone added four entries in one edition. [7]
The money is moving to the model owners. In 2023, 90% of the list monetised by subscription at an average $21 a month. [3] In 2026, 84% of AI-native products still sell subscriptions, 64% sell usage credits, and 14% run ads. [9] But the big shift is at the top. OpenAI reported a $1 billion annualised advertising run-rate in August 2026 on a base of 1.2 billion weekly users. [9,11] Advertising was 97.6% of Meta's 2025 revenue and 73.2% of Alphabet's. [9] The largest consumer AI product on earth has started to adopt the business model of the two largest advertising companies on earth.
The user base that funds all this grew as fast as anything ever built. ChatGPT had 100 million monthly users in January 2023. By February 2025 it had 400 million weekly users. By February 2026, 900 million. By August 2026, 1.2 billion. [1,6,8,11] More than one in ten humans uses one product every week. [8]
The consumer internet did not get a hundred AI companies. It got one, a runner-up owned by Google, and a long tail fighting for the last 10%.
03 · The handful
So how many companies are we actually talking about?
On the consumer surface, StatCounter's August 2026 data puts ChatGPT at 79.4% of chatbot use. Gemini has 10.9%. Perplexity 4.3%. Microsoft 2.8%. Claude 2.6%. Everything else rounds to zero. [10] a16z's panel agrees from a different angle: ChatGPT has about 3x the US paid subscribers of Claude or Gemini, and only 8% of ChatGPT subscribers also pay for Claude. [9]
Behind the products sits the capital. OpenAI filed for IPO in June 2026 after a $122 billion round at an $852 billion valuation. Anthropic's filing implies $965 billion. xAI, merged into SpaceX in February 2026, is carried near $250 billion. Perplexity is near $20 billion and Mistral near $14 billion. [29] Add Google and Meta, which fund models from advertising, and the list of companies that can afford a frontier model fits on one hand and a thumb.
Paid adoption is small and concentrating. 4.5% of eligible US consumers pay for ChatGPT, Gemini or Claude, up from 2.1% a year earlier. The top 1% of payers spend $903 a month and account for 19.5% of all consumer AI spend. The top 10% account for about half. The median payer spends $25. [9]
So 95% of users pay nothing. Their contribution is data. A small group pays heavily, and they are the ones delegating the most. Either way the value flows to the same six boxes.
````If you are not paying, you are the input. If you are paying, you are the input with a receipt.````
04 How an AI loop actually works
Now the basics. No jargon.
An AI assistant is not a search box. A search box takes a few words and returns links. An assistant takes everything you give it and returns a finished thing. To do that, it needs more from you. It needs context. It needs your files. It needs your corrections when it gets something wrong.
An agent needs even more. You are not asking a question. You are delegating a task. The agent reads your inbox, your calendar, your documents, your code. It acts for you. Every step it takes is logged somewhere.
Look at the right-hand column. The prompt is the smallest part. The corrections tell the model what you actually wanted. The re-asks tell it what you care about. The files tell it what you own. The metadata tells it when and where you work. The memory ties all of it to one identity across months.
This is how a knowledge graph of you gets built. Not from one chat. From the pattern across all of them. It is the same mechanism that made coding products improve so fast. There, the signal was "did the code run". Here, the signal is "did the person come back".
````You think you are asking a question. You are also answering a hundred that were never asked.````
05 What you tell it
People do not treat an assistant like a public forum. They treat it like a private one. That changes what they type.
OpenAI's study with NBER covered 1.58 million messages from May 2024 to July 2025. In June 2024, work and non-work were about even. By July 2025, 73% of messages were non-work. Non-work messages grew about 8x. Work messages grew about 3x. 12
"Practical guidance" is the largest category. That is life advice. Relationships, money, parenting, health. Users aged 18 to 25 sent 46% of all messages.12 The youngest adults are the heaviest contributors to the corpus.
Health is the sharpest case. Pew surveyed 3,488 US adults in June 2026. 34% use AI chatbots for health. 25% use them to work out what is causing symptoms. 18% use them for information they are uncomfortable discussing with a person. 71% of those users are at least somewhat comfortable sharing personal health information with the chatbot. 13
Read the blue bars together. One in three adults takes health questions to a chatbot. Nearly one in five takes the questions they would not say out loud. Those are the most sensitive vectors a person has. They are going to a server.
This is why the vulnerability is not about being a power user. The casual user asks the private question once, with no idea where it goes. The power user delegates whole workflows and feeds in everything. Both leave a trail. The trails differ in volume, not in kind.
````The more useful the assistant becomes, the more of you it needs. That is the design, not a flaw.````
06 Where it goes, and how long it stays
Follow the data after you press send. Three documented events from the last eighteen months show the pattern.
Retention got longer. In August 2025, Anthropic changed its consumer policy. Free, Pro and Max users were asked to allow training on their chats. The toggle defaulted to on. Users who allowed it moved from 30-day retention to up to five years. The deadline was 28 September 2025.14 Business and API customers were exempt. The consumer was the one asked to give more.
Deletion stopped being deletion. In May 2025, a US federal court ordered OpenAI to preserve all ChatGPT output logs that would otherwise be deleted, including chats users had asked to delete, as part of the New York Times copyright case. OpenAI objected that users often share sensitive information in chats they expect to vanish.32 The objection was correct. The order stood.
Private became public. In July 2025, thousands of shared ChatGPT conversations were found indexed by Google and searchable by anyone. Many contained personal details. OpenAI's security chief said the feature created too many chances for people to share things they did not intend to, and withdrew it. 33
Memory makes it stickier, by design. In April 2025 OpenAI rolled out memory that references all your past conversations, on by default for users who already had memory enabled.³¹ The feature is useful. That is the point. It is useful because it builds a durable model of you. The model lives on their side.
Part II
The question
An AI loop needs your data to work. It stores it. It learns from it. It gets better at you specifically. The agentic version needs even more: your inbox, your code, your decisions, your wallet. Should the control of all of that sit with a handful of companies? Should the price of running your own agents be set by them? Should the only option be access, on their terms, with ownership nowhere in the picture? The next five sections show how fast the money is moving while that question goes unasked.
07 The build-out
The money is not ambiguous.
Microsoft, Amazon, Google and Meta spent about $416 billion on AI infrastructure in 2025. a16z's State of Markets II, published 30 September 2026, puts 2026 near $780 billion. It relays Morgan Stanley's projection of more than $1 trillion in 2027. ¹⁵
The quarterly filings show the slope in real time. The same four reported $97 billion of capex in Q3 2025, $119 billion in Q4, and $130 billion in Q1 2026.³⁴ Alphabet alone guided to $175–185 billion for the year.³⁴
Demand is louder than supply. At I/O in May 2026, Google reported 3.2 quadrillion tokens processed per month. A year earlier, 480 trillion. Two years earlier, 9.7 trillion. More than 375 Google Cloud customers each processed over a trillion tokens in a year.¹⁶
08 Wide, not deep
Here the picture gets uncomfortable for anyone selling certainty.
Adoption is broad. 69% of S&P 500 companies report live AI deployments.¹⁵ Stanford's 2026 AI Index notes generative AI reached 53% of the global population within three years, faster than any prior technology.³⁵ Pew finds nearly half of US consumers use AI, but only 25% daily.⁹
Depth is thin. About 2% of those S&P 500 companies disclose tracked metrics on AI's effect.¹⁵ McKinsey's survey finds 39% of organisations experimenting with agents, 23% scaling an agentic system in at least one function, and no more than 10% scaling agents within any single function.¹⁷ Stanford puts agent adoption in the single digits in nearly every department.³⁵
Gartner adds the forecast. More than 40% of agentic AI projects will be cancelled by the end of 2027. Of thousands of vendors claiming agents, about 130 are real. Yet by 2028 Gartner expects 33% of enterprise software to include agents and 15% of day-to-day work decisions to be made autonomously.³⁶
````Read the two Gartner numbers together. Most projects fail. The ones that survive will make your decisions.````
That is the next three years. Shallow, expensive, uneven. Then suddenly structural. The terms get set now, while adoption is shallow. Once agents make 15% of decisions, the terms are not up for discussion.
09 Price is a tactic, not a floor
The economics of this era rest on one fact. Intelligence is getting cheaper faster than anything in computing history.
a16z found that for a fixed capability, inference cost falls about 10x a year. A GPT-3-class model cost $60 per million tokens in November 2021. By November 2024 it cost $0.06.¹⁸ Stanford measured a different model class: GPT-3.5-level capability fell from $20 to $0.07 between November 2022 and October 2024. 280x in 18 months.³⁷
Two things follow. They pull in opposite directions.
First, the curve that makes hosted inference cheap also makes local inference possible. A capability that needed a data centre in 2021 runs on a workstation in 2026. The open-weight gap on human-preference leaderboards narrowed to 1.7 points by early 2025.37 Running your own model is no longer a compromise on quality. It is a choice about control.
Second, the prices you see are not the costs. OpenAI's own projections, reported by the FT in September 2026, show about $36 billion of revenue this year and $350 billion in 2030. Alongside that: $278 billion of negative free cash flow across 2026–2030.19 xAI took $3.2 billion of revenue in 2025 against a $6.4 billion operating loss, and is spending at a $30 billion annual capex run-rate.³⁰
Alphabet and Meta together earned close to $500 billion in advertising in 2025.¹⁵ That is the business model that paid for the last free internet. OpenAI's $1 billion ad run-rate says it is already being wired in to pay for this one.¹¹
````A subsidised price is a loan. The collateral is your behaviour.````
10 Concentration, and the quiet switch
Platforms behave one way while they win users. They behave another way once they have them. Every operator knows this. The question is where we are in the cycle.
Infrastructure is already concentrated. Synergy Research puts Q2 2026 cloud spend at $143 billion for the quarter, up 43% year on year. AWS 28%. Microsoft 20%. Google a record 15%. Three companies, 63%.²²
The model layer is concentrating too, in the direction that matters. Menlo Ventures' mid-2025 survey of enterprise LLM API spend found closed models at 87% of workloads, up from 81% six months earlier. Open-source share fell from 19% to 13%.²¹
Now watch what leaders do with that position.
On 13 February 2026 OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini from ChatGPT. The stated reason: only 0.1% of users still chose GPT-4o daily. It had retired the model once before, in 2025, then restored it after paying users said they needed more time.³⁸ None of that is villainy. It is what a vendor does when it owns the model and you rent it. Your workflow was built on a surface someone else can remove.
Transparency moved the same way. Stanford's Foundation Model Transparency Index scores how much the major labs disclose about data, compute and deployment. It fell from 58 to 40.³⁵
````Not how they must operate. How they want to. That gap is where your leverage used to be.````
11 The quiet surrender of data
Cisco's 2026 Data Privacy Benchmark surveyed 5,200 security and privacy professionals in twelve countries. One number stands above the rest. Organisations with outright bans or hard limits on putting data into GenAI fell from 28% in 2025 to 7% in 2026.²⁰
11 The quiet surrender of datawledge risk from using proprietary or customer data in AI. 29% use unedited, unredacted customer data. 81% report higher demand for data localisation because of AI. 86% associate local storage with greater security.²⁰
Read both halves as one sentence. The people responsible for data know it is at risk. They believe local storage is safer. They are being asked for more localisation because of AI. They dropped their bans anyway. Convenience won. It usually does, until the bill arrives.
The data that flows in during this phase is not inert. It is the training signal for the next model. The individual contributes phrasing and preference. The enterprise contributes its codebase, its contracts, its customer logs. Both are teaching a system they do not own how to do what they do.
````You are not the customer of the free tier. You are its curriculum.````
Part III
The debate, and the answer
Everything above is how it is for you right now. The rest of this report is about what you do with that. You can stay a consumer and rely on the handful. Or you can own what you use and keep your integrity in place. Here is the case for the second, with the objections answered.
12 Access or ownership
This is the question the report exists to pose. Everything above is evidence. Here is the argument.
The strongest objections to the ownership model deserve a straight answer. This report is only useful if it survives them. Each objection comes with the data behind it.
Objection 1. Enterprises are moving toward closed models, not away.
True. Closed share rose from 81% to 87% in six months.²¹
That is a measure of today's convenience, not tomorrow's preference. 89% of teams did not switch vendor in a year. That is not loyalty. It is friction. The ownership model attacks the friction directly. One API surface. Hundreds of models. Your history stays with you. When switching costs nothing, closed share stops being destiny.
Objection 2. Decentralised infrastructure is tiny.
True. Messari counts about $72 million of on-chain DePIN revenue in 2025 against a $10 billion sector valuation.³⁹
Revenue is small. The slope is not. Funding rose 43% in a year. The leading networks grew real usage 1.7x to 8x while their tokens fell.³⁹ Usage decoupled from price. Nobody judges 2026 hyperscaler capex by 2006 AWS revenue.
Objection 3. Most agent projects will fail anyway.
True. Gartner expects more than 40% cancelled by the end of 2027.³⁶
Failure is why ownership matters. A failed project on rented rails leaves you a bill and a dataset in someone else's warehouse. A failed project on owned rails leaves you the weights, the data and the pipeline. You keep the learning either way. You only keep the assets one way.
Objection 4. Consumers do not care. They chose the free tier.
True. 95.5% of US consumers pay nothing, and bans on GenAI inside companies fell to 7%.⁹,²⁰ Consumers chose convenience when the trade was invisible. The trade is now visible. Ads arrived in the product in 2026.¹¹ Retention terms went to five years.14 Deleted chats were preserved by court order.³² Cisco's respondents still say local storage is safer at 86%.²⁰ Nobody stopped caring. They stopped having an option that was as easy as the default. Ownership only wins if it is as easy as access. That is a product requirement, not a philosophy, and it is the one Cluster is built against.
````Access is what they sell you while they still need you. Ownership is what you keep when they no longer do.````
13 Agents need a way to pay
An agent that cannot pay cannot act alone. It can only act on a human's card, inside a human's subscription, under a human's rate limit. Autonomy without settlement is theatre.
That constraint is lifting fast. x402 is an open standard from Coinbase's Developer Platform. It revives HTTP status 402, Payment Required. A server answers a request with 402 and a price. The client pays on-chain. The response follows. No account. No key. No invoice.⁴⁰
Chainalysis' October 2026 report counts about 138.5 million x402 transactions on Base by the end of Q3 2026, up from roughly 58,800 in Q3 2025. Over 87% of payments were between $0.0001 and $0.10. The median time between a wallet's first funding and its first x402 payment fell from 73 days to 18 minutes.²³
The consumer side is moving the same way. a16z's 7th edition reports that Instinct, a personal purchasing agent, passed 100,000 users in about three weeks and claims over $1 billion in annualised transaction volume. 40% of its users connected a credit card within three weeks.⁹ People are already letting software spend for them. The only question is whose rails the spending runs on.
````The loop is financial in the back. It is seamless in the front. That is what good rails feel like.````
14 Where Cluster Protocol stands
Cluster Protocol is built as the orchestration layer for the ownership model. Not a wrapper on one vendor's API. Not a marketplace that takes the vendor's margin. A set of rails an agent can run on without a human in the loop and without a single party in the middle.
What is Live
We built Cluster as the orchestration layer for autonomous workflows, with three components behind a single endpoint. Inference gives developers serverless access to more than 500 models. Compute provides GPU capacity, billed per job rather than per hour. Data is a marketplace where owners list and price their own datasets.
Our x402 payment rails have processed more than 100,000 transactions, with over 6,000 weekly transacting users.⁴² On top of this sits CodeXero, our application layer, which turns a plain-language prompt into an on-chain app or agent that uses inference, compute and data directly. No code. Free to start. Private, public or tokenised on the builder’s terms.⁴¹
So that's where Cluster Protocol, as a true decentralized AI infrastructure, comes into picture for general purpose agentic utilities like, commerce, day to day errands, or whatever you make your GrokBot, Muse, Chat.
15 What to do with this
If you build agents
Design for the 402. Give every agent a wallet and let it pay per call. Build against one API surface that spans many models, so your product does not inherit one vendor's retirement schedule.
If you run an enterprise
Reinstate the question the 7% still ask. Where does the prompt go? Route sensitive inference to local nodes or TEEs. License data out under terms you set, instead of uploading it under terms you accept.
If you hold data
Tokenise it. The agent economy will pay for data per use, in fractions of a cent, hundreds of millions of times. Owners who license will earn. Owners who upload will train their replacement for free.
If you are reading this on a free tier
Every correction you type is a labelled example. Every file you drop is a document in someone else's corpus. Every health question is a vector. That is not paranoia. It is the business model, stated in their own filings.
````The future of the internet is agentic. The rails are already being laid. The only decision left is whether you own the ones your agents run on.````
Next step
Build on rails you do not have to trust.
Run inference across 500+ models. License data on your terms. Let your agents pay their own way with x402. Start your journey with hub.clusterprotocol.ai🔗
Method and sources.
All figures come from the cited primary reports or contemporaneous reporting of them, as noted. Where a figure comes from secondary coverage of a paywalled or long report, the entry says "as reported" or names the outlet. Exhibits are drawn to scale from the figures listed. Survey populations differ across sources; comparisons across surveys are directional. Run-rate figures are annualised and unaudited.
- UBS analyst note citing Similarweb and Sensor Tower, Feb 2023, via [Reuters / Inquirer](https://business.inquirer.net/384799/chatgpt-sets-record-for-fastest-growing-user-base-analyst-note) (100M monthly users in ~2 months; TikTok ~9 months; Instagram ~2.5 years).
- SemiAnalysis estimates, as reported by [iTiger](https://www.itiger.com/news/1183910360), 4 May 2026 (Anthropic run-rate $1B Dec 2024 → $7B Sept 2025 → $9B Dec 2025 → $14B Feb → $19B Mar → $30B Apr → $44B+ May 2026; Claude Code ~$2.5B Feb 2026; ~4% of public GitHub commits; enterprise >50% of Claude Code revenue).
- a16z, ["How Are Consumers Using Generative AI?"](https://a16z.com/how-are-consumers-using-generative-ai/), 1st edition, 13 Sept 2023 (ChatGPT 60% of list traffic; 80% of sites new; 15 of 50 with mobile apps; 90% monetising; avg $21/month).
- a16z, [Top 100 Gen AI Consumer Apps](https://a16z.com/100-gen-ai-apps/), 2nd edition, 13 Mar 2024 (~2B monthly visits, 5x #2; 22 web newcomers; companion apps 2 → 8; [Character.AI](https://character.ai/) 2 hours/day).
- a16z, [3rd edition](https://a16z.com/100-gen-ai-apps-3/), 21 Aug 2024 (nearly 30% new; 12 web newcomers, 58% creative; Suno #36 → #5; ByteDance six spots).
- a16z, [4th edition](https://a16z.com/100-gen-ai-apps-4/), 6 Mar 2025 (17 newcomers; ChatGPT ~400M weekly; DeepSeek #2; Cursor #41, Bolt #48).
- a16z, [5th edition](https://a16z.com/100-gen-ai-apps-5/), 27 Aug 2025 (11 web newcomers; Gemini ~12% of ChatGPT web visits; Grok #4; Lovable #22; Replit debut; revenue retention >100% at one vibe-coding platform).
- a16z, [6th edition](https://a16z.com/100-gen-ai-apps-6/), 9 Mar 2026 (900M weekly users; 2.7x Gemini on web; >10% of global population weekly; Claude Code $1B run-rate in six months; Codex 2M weekly; Notion AI attach 20% → 50%+).
- a16z, Top 100 Gen AI Consumer Apps, 7th edition, 5 Oct 2026, via [aVenture](https://aventure.vc/news/2026-10-05-the-top-100-gen-ai-consumer-apps-7th-edition) and [The Neuron](https://theneuron.ai/news/ai-top-1-percent-spend-903-a-month-a16z-data/) (11 newcomers; 4.5% paid from 2.1%; top 1% = $903/month, 19.5%; top 10% ≈ half; 2x Gemini, 6x Claude web, 14x Claude mobile; 84% subscriptions, 64% credits, 14% ads, 2% fees; Meta 97.6% / Alphabet 73.2% ads; OpenAI $1B ad run-rate on 1.2B weekly users; OpenEvidence 50–60% of physicians; Instinct 100K users, $1B+ volume, 40% card connect; NSFW >20% excluded; Pew ~half use AI, 25% daily).
- StatCounter, Aug 2026, and OpenAI disclosures compiled by [DemandSage](https://www.demandsage.com/chatgpt-statistics/) (79.4 / 10.9 / 4.31 / 2.79 / 2.57%; 800M weekly Oct 2025; 900M Feb 2026; 50M+ paying consumers).
- [Adweek](https://www.adweek.com/programmatic/chatgpt-openai-billion-advertising-revenue/), 31 Aug 2026 (ChatGPT >$1B annualised ad revenue).
- Chatterji et al., *How People Use ChatGPT*, NBER w34255, Sept 2025, via [DeepLearning.AI](https://www.deeplearning.ai/the-batch/chatgpt-users-now-more-likely-to-be-young-female-and-seeking-info-study-shows) (73% non-work; 8x vs 3x; 28.3 / 28.1 / 21.3%; 46% from ages 18–25; ~700M weekly users at time of study).
- Pew Research Center, [*From Diagnoses to Treatments*](https://www.pewresearch.org/science/2026/08/25/from-diagnoses-to-treatments-why-americans-use-ai-chatbots-for-health/), 25 Aug 2026, n = 3,488 (34%; 28 / 25 / 22 / 22 / 20 / 18 / 15%; 71% at least somewhat comfortable).
- Anthropic consumer policy change, Aug 2025, via [TechCrunch](https://www.techcrunch.com/2025/08/28/anthropic-users-face-a-new-choice-opt-out-or-share-your-data-for-ai-training) and [Implicator](https://www.implicator.ai/anthropics-five-year-data-plan-for-claude-consumer-users/) (toggle default on; 30 days → 5 years; deadline 28 Sept 2025; business and API exempt).
- a16z, *State of Markets II*, 30 Sept 2026, via [GPTS24](https://www.gpts24.com/en/news/a16z-state-of-markets-ii-finds-ai-adoption-shallow-while-capex-approaches-780-billion) (capex $416B / ~$780B / >$1T; 69% / ~2% / ~2%; Alphabet + Meta ~$500B ads).
- Google, [I/O 2026 keynote](https://blog.google/innovation-and-ai/sundar-pichai-io-2026/), 19 May 2026 (9.7T → 480T → 3.2 quadrillion tokens/month; 375+ customers over 1T tokens).
- McKinsey, *State of AI 2025*, via [Forbes](https://www.forbes.com/sites/josipamajic/2026/03/22/10-of-enterprise-functions-use-ai-agents-mckinsey-finds/), 22 Mar 2026 (39% / 23% / ≤10%).
- a16z, ["Welcome to LLMflation"](https://www.a16z.com/llmflation-llm-inference-cost), Nov 2024 (10x/year; $60 → $0.06).
- Financial Times via [Reuters / The Star](https://www.thestar.com.my/tech/tech-news/2026/09/19/openai-forecasts-cash-burn-near-280-billion-by-2030-ft-reports), 18–19 Sept 2026 (OpenAI ~$36B 2026; $350B 2030; $278B negative FCF).
- Cisco, [*2026 Data Privacy Benchmark Study*](https://www.cisco.com/c/dam/en_us/about/doing_business/trust-center/docs/cisco-privacy-benchmark-study-2026.pdf) (28% → 7%; 70%; 29%; 81%; 86%; n = 5,200+).
- Menlo Ventures, *2025 Mid-Year LLM Market Update*, via [BigDATAwire](https://www.hpcwire.com/bigdatawire/this-just-in/menlo-ventures-report-enterprise-llm-spend-reaches-8-4b-as-anthropic-overtakes-openai/), 1 Aug 2025 ($8.4B; 32 / 25 / 20 / 9 / 1%; closed 87% from 81%; 11 / 66 / 23%).
- Synergy Research Group, Q2 2026, via [Shattered](https://shattered.io/aws-falls-28-percent-google-cloud-record-15-2026/) ($143.4B; +43%; 28 / 20 / 15%).
- Chainalysis, *North America: Inside the Growth of Agentic Commerce*, 14 Oct 2026, via [Brave New Coin](https://bravenewcoin.com/insights/coinbases-x402-records-138-5m-transactions-chainalysis-shows-ai-agents-reshaping-crypto-payments) (58,800 → 138.5M; 87% under $0.10; 73 days → 18 min; 70.6% → 89.3%).
- Goldman Sachs, *The Potentially Large Effects of Artificial Intelligence on Economic Growth*, 26 Mar 2023, via [SiliconANGLE](https://siliconangle.com/2023/03/29/goldman-sachs-report-says-ai-put-300-million-jobs-risk/) (300M jobs; admin 46%, legal 44%, architecture/engineering 37%, business/finance 35%, community 33%, management 32%, arts/media 27%, construction 6%, maintenance 4%; 7% GDP).
- Harvey valuation ($11B, Mar 2026) per [ValueAdd VC](https://valueaddvc.com/ai-valuations).
- Cursor/Anysphere run-rate milestones compiled by [Panto](https://www.getpanto.ai/blog/cursor-ai-statistics) from company disclosures, Bloomberg and TechCrunch (~$100M early 2025; >$500M June 2025; $1B Nov 2025; $2B+ Feb 2026; $29.3B valuation Nov 2025).
- [Gadget Scout](https://www.gadgetscout.co.uk/articles/anthropic-overtakes-openai-on-revenue-in-race-to-ipo.html) citing CNBC, July 2026 (Anthropic $47B run-rate, 28 May 2026; flagged as pending review by the outlet).
- Stack Overflow, [2025 Developer Survey](https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/), 29 July 2025, n = 49,000+ (84% from 76%; 46% distrust from 31%; 45% say debugging AI code is time-consuming).
- [ValueAdd VC, AI valuations tracker](https://valueaddvc.com/ai-valuations), mid-2026 (OpenAI $852B; Anthropic $965B; xAI ~$250B; Databricks $188B; Perplexity ~$20B; Mistral ~$14B; Cohere $7B).
- SpaceX IPO filing via [Yahoo Finance](https://finance.yahoo.com/sectors/technology/articles/xai-burned-6-4b-last-222608682.html), 20 May 2026 (xAI $3.2B revenue, $6.4B operating loss 2025; $12.7B capex 2025; $7.7B Q1 2026; Anthropic first operating profit expected Q2 2026).
- OpenAI memory update via [TechCrunch](https://techcrunch.com/2025/04/10/openai-updates-chatgpt-to-reference-your-other-chats/), 10 Apr 2025.
- US District Court (S.D.N.Y.) preservation order in *NYT v. OpenAI*, 13 May 2025, via [Malwarebytes](https://www.malwarebytes.com/blog/news/2025/06/openai-forced-to-preserve-chatgpt-chats).
- Shared-chat indexing, July–Aug 2025, via [UPI / Yahoo](https://www.yahoo.com/news/articles/openai-ends-chatgpt-users-option-193718975.html), first reported by Fast Company.
- SEC 10-Q/10-K filings, Q3 2025 – Q1 2026, compiled by [Silicon Analysts](https://siliconanalysts.com/market-data/hyperscaler-ai-capex); Alphabet 2026 guidance via CNBC, 10 Feb 2026.
- Stanford HAI, *AI Index 2026*, 13 Apr 2026, via [Unite.AI](https://www.unite.ai/stanford-ai-index-2026-reveals-a-field-racing-ahead-of-its-guardrails/) (53% of global population; agents single digits; Transparency Index 58 → 40).
- Gartner press release, 25 June 2025, via [SD Times](https://sdtimes.com/agentic-ai/gartner-more-than-40-of-agentic-ai-projects-will-be-canceled-in-next-few-years/) (>40% cancelled; 33% of apps; 15% of decisions; ~130 real vendors).
- Stanford HAI, *AI Index 2025*, via [R&D World](https://www.rdworldonline.com/?p=77809) ($20 → $0.07, 280x; open–closed gap 1.7%).
- OpenAI, ["Retiring GPT-4o and older models"](https://openai.com/index/retiring-gpt-4o-and-older-models), 29 Jan 2026.
- Messari & Escape Velocity, *State of DePIN 2025*, via [Cointelegraph](https://cointelegraph.com/news/depin-revenues-grow-despite-token-losses), 29 Jan 2026 ($72M; $10B; $698M → ~$1B; Helium ~8x / −77%; GEODNET ~1.7x / −41%).
- Coinbase Developer Platform, x402 documentation, as referenced in [KuCoin's listing note](https://www.kucoin.com/blog/cluster-protocol-cp-kucoin-listing-tokenomics).
- [clusterprotocol.ai](https://www.clusterprotocol.ai/) and [KuCoin Learn](https://www.kucoin.com/learn-and-earn/one-api-infinite-ai-learn-earn-with-cp) (500+ models; OpenAI-compatible API; TEE inference; CodeXero; x402).
- [yellow.com asset profile](https://yellow.com/ko/asset/cp) (100,000+ x402 transactions; 6,000+ weekly transacting users; CCIP; ERC-8004).
- [FXEmpire](https://ar.fxempire.com/crypto/cluster-protocol) (5B fixed supply on Base; April 2026 round led by DAO5).
- a16z Top 100 series index: [editions 2–6 on a16z.com](https://a16z.com/100-gen-ai-apps/); 1st edition at source 3; 7th at source 9.
- Bloomberg Law, ["OpenAI Has More Than 1 Billion Users, 2 Million Businesses"](https://news.bloomberglaw.com/artificial-intelligence/openai-has-more-than-1-billion-users-2-million-businesses), 2026 (cross-check for the 1B+ user figure).
- Pew Research as cited within source 9 (nearly half of US consumers use AI; 25% daily).
- Cisco prior-year bans figure cross-checked via [HCA Magazine](https://www.hcamag.com/au/news/general/nearly-half-of-organisations-entering-employee-info-into-genai-survey/530859).