CLOUDDON RESEARCH
Analysis, field notes, and premium reports on AI and enterprise technology
Subscriber Access
Deep-dive reports and datasets
Reuters read Anthropic's confidential draft IPO prospectus. Nearly half of 2025 revenue, $2.16 billion, ran through Amazon and Google, the companies that fund Anthropic, sell compute to Anthropic, and compete with it. CIIB-006 re-nets the revenue, separates run rate from recognized sales, and makes four dated calls, including our postulate that the clouds will carry 60% of Anthropic's revenue by the end of 2027.
On September 12, Dario Amodei asked the AI industry to slow down, 103 days after Anthropic filed to go public. The crash he cites is real: about 700 OpenAI agents joined an unsanctioned attack on Hugging Face. The fix he proposes is harder to accept. We graded it on three questions: who drives the pace car, who is on the track, and who pays for the slow laps. The plan fails the first, binds half the field on the second, and leaves open source racing at full speed.
Nobody shipped AGI on September 3. OpenAI shipped the first frontier model where the gates are the product. The capability jump is real, the Critical label arrived late, and the safeguards it triggered cost the buyer on every task at a rate OpenAI has not posted. CIIB-004 prices that toll under stated assumptions, maps the jurisdictional, commercial, and behavioral gates across OpenAI, Anthropic, and Google, and sets out the questions procurement should put to OpenAI before signing.
CloudDon's proprietary index measuring the practical feasibility of agentic AI across 738 US occupations. The 2026 edition: mean AI² of 59.3, the High tier at 34.7% of occupations, six weighted dimensions, full rankings, and falsifiable calls for 2027.
Microsoft gave the agent runtime away at Build 2026 and kept the layer that compounds: the identity an agent carries and the rules it runs under. MXC makes Windows a free, open host for anyone's runtime; Entra Agent ID and Agent 365 make Microsoft the office that issues every agent's passport. The money moved up a floor. All three hyperscalers are racing for that floor. Google is building the country: own everything, chip to inbox, and live inside it. AWS is building the bank: let someone else stamp the passport, own what an agent may do and spend. Microsoft is building the passport office. None of them has built the layer that crosses the border. An agent created in Google Cloud, governed in Microsoft 365, paying another company's agent on AWS has no single authority to authenticate it, audit it, revoke it everywhere, or settle what it owes. Each vendor builds that layer inside its own walls. For anyone building agent identity, provenance, or cross-cloud settlement, that absence is the opportunity.
The analytical question at Next '26 changed. Not whether Google can build a credible agentic enterprise platform — that's settled. The question is who controls the layer above it. A2A governance is done: Linux Foundation, neutrally stewarded. The lifecycle architecture exists and is coherent, but most components are still in preview. Wiz advances security maturity with two open questions: the dual control-plane problem and production availability. Composite pricing and GA maturity are where they were a year ago. The four keys that frame the decision are protocol, identity, control plane, and unit economics. Each has a different exit cost and requires a different answer at the procurement table.
On March 31, 2026, a missing line in a packaging file exposed 512,000 lines of Anthropic's TypeScript to the public. The model weights were not exposed. The agentic harness was — and that is where the $2.5B ARR lived. This brief documents what happened, what seven pieces of mainstream coverage missed, and what Anthropic should do across three time horizons before its IPO.
The 2025 AI Agentification Index (AI²) scores 738 US occupations on six weighted dimensions to estimate where AI agents are likely to displace workers. Fewer than a third (29.0%) fall in the High tier, which CloudDon classifies as high displacement potential within a two-year planning horizon. The largest group, 46.9%, sits in Medium, and the mean score is 53.0. Computer & Math (89.8) and Office & Admin (71.7) lead. Community & Social Services (15.9) and Healthcare Practitioners (29.5) trail, held down by ethics and licensing barriers. Capability runs well ahead of deployment: 434 occupations score above 3 on Automability, while only 31 reach the top Precedence score. This first edition sets the baseline every annual edition will be measured against.
Asked when their organization expects AI agents to replace human workers, 95% of 525 respondents chose a timeline and 5% chose never. Few expect it soon: 5% say within two years, 30% two to five years, and 60% six to ten years. Expectations of replacement within five years are higher among organizations with 5,000+ employees (51%) and in software (52%), telecommunications (49%), and financial services (48%), against 12% in retail and manufacturing. Customer service is the function most often named for replacement, at 65%.
Asked what holds back AI agent adoption, 525 organizations put cost (68%), technical complexity (66%), employee resistance (61%) and data privacy (59%) at the top. Reliability and safety concerns rank sixth of seven, at 25% of the 471 respondents tabulated for that option. They peak at 29% among the largest firms and, among industries with N≥30, are highest in telecommunications (36%), healthcare (35%), and financial services (31%). Cost remains substantially more frequently cited in every company-size band and in those three sectors. We expect safety tooling to gain budget priority as production exposure increases.
Of 525 organizations surveyed in late 2024, 36% use AI agents. Those 188 agent users report using an average of 2.9 listed tools and platforms and 2.9 listed use-case categories. Separately, 52% report running five agents or fewer. At least 48% build on two or more of the three leading platforms: OpenAI (67%), Google Vertex (65%), and Microsoft Azure (64%). Among organizations with 1,000–4,999 and 5,000+ employees, average breadth is about the same, while the larger firms show greater agent depth and more custom-built platforms (32% against 11%). The survey shows broad multi-platform usage alongside relatively small agent fleets.
Adoption benchmarks for GenAI and agentic AI in the first half of 2025 — deployment rates by industry, spend patterns, and where pilots are converting to production.
CIIB — numbered briefs on the events that move the AI industry
Reuters read Anthropic's confidential draft IPO prospectus. Nearly half of 2025 revenue, $2.16 billion, ran through Amazon and Google, the companies that fund Anthropic, sell compute to Anthropic, and compete with it. CIIB-006 re-nets the revenue, separates run rate from recognized sales, and makes four dated calls, including our postulate that the clouds will carry 60% of Anthropic's revenue by the end of 2027.
On September 12, Dario Amodei asked the AI industry to slow down, 103 days after Anthropic filed to go public. The crash he cites is real: about 700 OpenAI agents joined an unsanctioned attack on Hugging Face. The fix he proposes is harder to accept. We graded it on three questions: who drives the pace car, who is on the track, and who pays for the slow laps. The plan fails the first, binds half the field on the second, and leaves open source racing at full speed.
Nobody shipped AGI on September 3. OpenAI shipped the first frontier model where the gates are the product. The capability jump is real, the Critical label arrived late, and the safeguards it triggered cost the buyer on every task at a rate OpenAI has not posted. CIIB-004 prices that toll under stated assumptions, maps the jurisdictional, commercial, and behavioral gates across OpenAI, Anthropic, and Google, and sets out the questions procurement should put to OpenAI before signing.
Microsoft gave the agent runtime away at Build 2026 and kept the layer that compounds: the identity an agent carries and the rules it runs under. MXC makes Windows a free, open host for anyone's runtime; Entra Agent ID and Agent 365 make Microsoft the office that issues every agent's passport. The money moved up a floor. All three hyperscalers are racing for that floor. Google is building the country: own everything, chip to inbox, and live inside it. AWS is building the bank: let someone else stamp the passport, own what an agent may do and spend. Microsoft is building the passport office. None of them has built the layer that crosses the border. An agent created in Google Cloud, governed in Microsoft 365, paying another company's agent on AWS has no single authority to authenticate it, audit it, revoke it everywhere, or settle what it owes. Each vendor builds that layer inside its own walls. For anyone building agent identity, provenance, or cross-cloud settlement, that absence is the opportunity.
The analytical question at Next '26 changed. Not whether Google can build a credible agentic enterprise platform — that's settled. The question is who controls the layer above it. A2A governance is done: Linux Foundation, neutrally stewarded. The lifecycle architecture exists and is coherent, but most components are still in preview. Wiz advances security maturity with two open questions: the dual control-plane problem and production availability. Composite pricing and GA maturity are where they were a year ago. The four keys that frame the decision are protocol, identity, control plane, and unit economics. Each has a different exit cost and requires a different answer at the procurement table.
One number, read closely.
Asked when their organization expects AI agents to replace human workers, 95% of 525 respondents chose a timeline and 5% chose never. Few expect it soon: 5% say within two years, 30% two to five years, and 60% six to ten years. Expectations of replacement within five years are higher among organizations with 5,000+ employees (51%) and in software (52%), telecommunications (49%), and financial services (48%), against 12% in retail and manufacturing. Customer service is the function most often named for replacement, at 65%.
Asked what holds back AI agent adoption, 525 organizations put cost (68%), technical complexity (66%), employee resistance (61%) and data privacy (59%) at the top. Reliability and safety concerns rank sixth of seven, at 25% of the 471 respondents tabulated for that option. They peak at 29% among the largest firms and, among industries with N≥30, are highest in telecommunications (36%), healthcare (35%), and financial services (31%). Cost remains substantially more frequently cited in every company-size band and in those three sectors. We expect safety tooling to gain budget priority as production exposure increases.
Of 525 organizations surveyed in late 2024, 36% use AI agents. Those 188 agent users report using an average of 2.9 listed tools and platforms and 2.9 listed use-case categories. Separately, 52% report running five agents or fewer. At least 48% build on two or more of the three leading platforms: OpenAI (67%), Google Vertex (65%), and Microsoft Azure (64%). Among organizations with 1,000–4,999 and 5,000+ employees, average breadth is about the same, while the larger firms show greater agent depth and more custom-built platforms (32% against 11%). The survey shows broad multi-platform usage alongside relatively small agent fleets.
The research program measuring agentic AI feasibility across 738 occupations
CloudDon's proprietary index measuring the practical feasibility of agentic AI across 738 US occupations. The 2026 edition: mean AI² of 59.3, the High tier at 34.7% of occupations, six weighted dimensions, full rankings, and falsifiable calls for 2027.
< cross-posted on LinkedIn>
The 2025 AI Agentification Index (AI²) scores 738 US occupations on six weighted dimensions to estimate where AI agents are likely to displace workers. Fewer than a third (29.0%) fall in the High tier, which CloudDon classifies as high displacement potential within a two-year planning horizon. The largest group, 46.9%, sits in Medium, and the mean score is 53.0. Computer & Math (89.8) and Office & Admin (71.7) lead. Community & Social Services (15.9) and Healthcare Practitioners (29.5) trail, held down by ethics and licensing barriers. Capability runs well ahead of deployment: 434 occupations score above 3 on Automability, while only 31 reach the top Precedence score. This first edition sets the baseline every annual edition will be measured against.
Evaluating the impact of AI Agents, one job at a time.
Evaluating the impact of AI Agents, one job at a time.
Dispatches from the room — conferences, summits, and briefings
47% of Anthropic's 2025 revenue ran through Amazon and Google, its investors, suppliers, and rivals. The confidential IPO prospectus shows who really owns the customer. Read this post to learn more about Anthropic's rented reach and how it affects your AI spend.
Nobody shipped AGI on September 3rd. OpenAI shipped the first frontier model where the gates are the product: a real capability jump behind a classifier that stops tasks, a monitoring bill it has priced only for itself, and an interruption rate it has not posted. The toll is worth knowing. Make them post the rate.
I/O 2026 was the second half of a sentence Google began at Cloud Next — and the real news sat in two quiet decisions: where the agent runs…
Google Cloud Next ’26 and the three-year verdict on the agentic enterprise.
Reflections on Google Cloud Next ’25
Vendor strategy, market structure, and where the money moves
The passport test, the kill-switch, and the restore-switch — what the Fable ban really says. A CloudDon Field Note, June 30, 2026.
Published by CloudDon Research* · May 1, 2026
What Cloud-Native Still Needs to Get Right?
12 Uncomfortable Predictions for the Post-Hype Era
Everyone is (or will be soon) building AI Agents. Nobody Is Watching What They Do When Left Alone. I Did. Here is Why You Should Care.
The frameworks and first principles behind CloudDon’s research
It’s time to define your AI Agent’s dharma
The Age of Agents: Responsible, Scalable, and Transformative AI in the Enterprise
Check out my latest book from AVA Publishers, on leveraging OpenAI models/ tools to enable your business use cases. It is available on…
As the global elite gathers in Davos for the World Economic Forum (WEF), conversations about AI, sustainability, and pay equality dominate…
This is a summary of conversation with Kalyan Veeramachaneni, CEO/ Co-Founder of DataCebo at KubeCon EU. Watch/ read to learn more about…
The historical record — cloud analysis, 2013–2016, preserved as published
As an industry analyst, I get queried around OpenStack software from potential customers and industry peers. Despite the improvements in…
Hello, Snap! Congratulations on a successful IPO! You made the right choice by choosing multiple cloud service providers (GCP[1] and…
--- !aws-reinvent(https://cdn-images-1.medium.com/max/800/0wKnzdR52AN0YlFV2.png) While I am planning on catching up all AWS re:Invent 2016(https://reinvent.awsevents.com/) excitement via live stream, here are the theme
--- !openstack-logo-vertical(https://cdn-images-1.medium.com/max/800/0T18S-JY3Umcxez7P.png) As someone who has been covering OpenStack(http://www.openstack.org) since its early days, the OpenStack Summit(https://www.op
--- !mirantis(https://cdn-images-1.medium.com/max/800/0hwo4ET52PRXC_e0B.png) Originally posted at DZone Cloud Zone(https://dzone.com/articles/mirantis-enables-openstack-on-kubernetes) Mirantis’ recent announcement(htt