OpenAI's revenue was $50B, not $68B; an AI agent helped breach seven South Korean banks
The week the numbers moved: OpenAI's reported annualised revenue fell by $18 billion when it was defined precisely, Washington took $2.4 billion of donated compute for federal science, Google shipped one work agent that runs Claude as well as Gemini, Microsoft made Windows a host for agents it does not own, and South Korean police traced AI-assisted intrusions into seven banks.
Between October 2 and October 9, 2026, the figure doing the most work in AI valuations was cut by about a quarter.
OpenAI told investors it reached roughly $50 billion in annualised revenue at the end of September. The $68 billion reported days earlier counted its partners' gross revenue. 1
That afternoon Nvidia fell about 3% and Oracle nearly 6%. CoreWeave slipped nearly 8%.
The same week the White House put donated frontier compute behind federal science agencies. Google Cloud shipped one work agent that can run a rival's model. Microsoft turned Windows into a host for agents it does not own. And in South Korea, AI tooling reached the customer records of seven banks.
| Development | Number to keep | What changed | Why it matters |
|---|---|---|---|
| OpenAI's reported revenue | $50B, not $68B | OpenAI told investors it hit about $50 billion annualised at the end of September; the higher figure counted partners' gross revenue. 1 | The number investors were pricing fell a quarter in one afternoon, and it is not audited. |
| The Genesis Mission | $2.4B pledged, 14 more agencies | Five companies donated compute, model access and cloud credits to federal researchers. 2 | Washington gets frontier capability without buying it, and labs get a standing federal channel. |
| Google's work agent | ~90% of the Fortune 100 | Google Cloud launched one universal agent that routes each job across the Gemini family — and Anthropic's Claude. 3 | Model choice becomes a routing decision, not a vendor decision. |
| Microsoft's Windows | $2,599 laptop, 30M paid Copilot seats | Windows became a host for third-party agents, and Meta's Muse arrives as a native app. 4 | The agent's home turf is now the operating system rather than the browser. |
| South Korea's banks | 68,000 people, 7 firms | Investigators traced AI-assisted intrusions to an open-source pentesting agent driving external models. 5 | Agentic attacks have moved from demonstrations to customer records. |
OpenAI's reported revenue fell $18 billion when the counting changed
OpenAI told investors it reached about $50 billion in annualised revenue at the end of September, below the $68 billion that had been reported days earlier, and AI shares fell on the disclosure. The Financial Times reported the $50 billion figure first, and CNBC confirmed it with a person familiar with the presentation. 1
The gap between the two numbers is a definition. The same person said the $68 billion figure included gross revenue from OpenAI's partners, which makes the comparison with Anthropic more direct. 1
OpenAI also told investors its total run rate grew 77% during the third quarter, and that its enterprise business grew 107%. 1
The market reaction was broad rather than targeted. Nvidia fell 3%, Oracle nearly 6% and CoreWeave nearly 8% on Thursday. AMD and Broadcom each fell 4%, Intel 5% and Super Micro nearly 5%. 1
The context is a company under pressure to justify an $852 billion private valuation ahead of a widely expected listing. OpenAI filed confidentially with regulators in June and executives have signalled a 2027 debut; it is in early talks to raise about $30 billion on top of a $122 billion round closed in March. 1
Anthropic makes the comparison awkward. It reported a $65 billion annualised run rate at the end of July, was reportedly seeking a $2 trillion valuation, and has still not published the prospectus it is listing on. 6
The constraint is that both figures are investor-facing and unaudited. When "annualised revenue" can move by $18 billion on which entities you count, treat any single number as a claim about a definition, not a measurement.
$2.4 billion of donated compute went to 14 more federal agencies
The White House expanded the Genesis Mission to an additional 14 federal agencies on October 8, with five companies pledging $2.4 billion in computing power, model access and cloud credits. The announcement came at the administration's "Science: A New Golden Age" summit. 2
The split is worth reading as the whole of the pledge. Nvidia contributes $1 billion in compute over five years; OpenAI $200 million in token discounts and training for scientists; Anthropic $150 million over three years in access to Claude; and Amazon Web Services $50 million in cloud credits. 2
Those four figures add to $1.4 billion of the $2.4 billion total. Google is named as a contributor in the statements Politico cites, without a published figure.
Nvidia's chief executive framed it in the administration's vocabulary. "With a $1 billion investment, NVIDIA is putting advanced Super Intelligence in the hands of America's scientists to accelerate breakthroughs in medicine, energy and materials," Jensen Huang said. 2
The programme's scope is the science. Genesis Mission is aimed at using AI to improve federal research, and this round extends it from an initial set of agencies to 14 more. 2
The constraint is that these are in-kind pledges, not appropriations. No mechanism for measuring delivery, no stated access terms and no expiry were published with the announcement.
Google shipped one work agent that can run a rival's model
Google Cloud launched the Gemini agent on October 8: a single universal agent for work that answers questions, runs knowledge work, generates media and writes code from one prompt box, and picks its model per job from the Gemini family and Anthropic's Claude. 3
What changed is the shape of the product. Earlier Gemini Enterprise releases sold models and assistants; this one sells an agent that owns an objective, runs in the cloud across sessions and devices, and spawns temporary sub-agents for multi-step work. 3
Two features matter for anyone buying. Gemini can act as a "coworker agent" with a persistent role, its own
@agents.company.com address, its own storage and directory presence, so colleagues interact with it as they would a person. And model choice is deliberately open: Google says the agent routes across its own models and Claude today, with other private and open models later. 3Google's governance answer is identity plus a network boundary. Every agent gets a cryptographically attested identity with least-privilege permissions; every action is written to an audit trail attributed to the agent rather than a person; agents run in an Agent Sandbox and all traffic passes through Agent Gateway, which enforces company policy such as "agents may not open documents classified Need to Know". 3
The scale claims are the adoption signal. Google says nearly 500 Google Cloud customers each processed more than a trillion tokens in the past year, nearly 80% of its cloud customers now use its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise. 3
Its named customers give the numbers a baseline. Japanese insurer SOMPO built more than 10,000 custom agents across 34,000 employees; Bradesco cut document review from an hour to five minutes; BNP Paribas is deploying to more than 65,000 employees; Orange Spain built over 1,000 custom agents. 3
The constraint is that Google's governance layer enforces the policy the customer writes. Identity, permissions and the gateway are shipped; the rules an agent must never break are still the buyer's to specify.
Microsoft made Windows a host for agents it does not own
At an event in San Francisco on October 7, Microsoft rebuilt Windows around agents that run on the PC — and opened the platform to a competitor, with Meta's Muse arriving as a native Windows app. 4
Chief executive Satya Nadella described the design as "hybrid intelligence": cloud AI and on-device AI presented as one system. Copilot can act on local files with permission, run models locally and reach for cloud compute only when a task needs it. 7
Jacob Andreou, who runs Copilot, gave the worked example: having Copilot pull the files needed to file taxes from across a laptop without touching the cloud, which makes the task cheaper and faster. "Copilot now truly understands your PC," he said. 7
A competitor's agent is part of the offering. Muse was already finding an audience on its own; on Windows it becomes a native app alongside Perplexity's Portable Computer, which Microsoft says can run locally without spending Perplexity cloud credits. 4
The hardware is the other half of the pitch. The Surface Laptop Ultra runs Nvidia's Arm-based RTX Spark system-on-chip, with up to 128GB of unified memory and a stated draw of no more than 80W, from $2,599, available October 16; the Surface RTX Spark Dev Box starts at $5,999. 4
Scale gives Microsoft a distribution advantage. Microsoft 365 Copilot passed 30 million paid seats by July, with net seats added doubling quarter over quarter. 7
The constraint is trust, and Microsoft set the bar itself. It removed Copilot buttons from Windows 11 and pulled back the Recall feature after user backlash, so the agents announced this week arrive to an audience that has already rejected one version of the idea. 4
An AI pentesting agent was used to break into seven banks
South Korea's National Police Agency opened an investigation after intrusions at seven financial firms exposed the personal data of about 68,000 people, with investigators pointing to an AI agent used as the attack tool. 5
The tool is the news. Investigators found traces of ARTEX, an open-source "autonomous penetration-testing" agent built by a Chinese developer, Li Puhua, who works under the alias Autumn. Artex is not a model itself: it drives external AI models — among them Anthropic's Opus, OpenAI's GPT and China's DeepSeek — to find vulnerabilities and plan attacks with little human input. 5
The firms are named. The Financial Services Commission held an urgent session after the attacks, which were first detected last week; Shinhan Bank, KB Kookmin Bank and Hana Bank were the worst affected according to South Korean media. 8
The stolen fields are what make this a fraud story rather than a disclosure story. Names, phone numbers, annual income figures, loan limits and loan products were accessed, plus a small number of national identification numbers. 8
Per-bank figures circulating do not reconcile with the national total: BleepingComputer reported Shinhan's breach at 25,000 customers and Kookmin card data for 119,000, against a nationwide figure of about 68,000. Treat the breakdown as preliminary. 5
Attribution is unresolved and officials are saying so. CrowdStrike's early findings point to China, while the Financial Supervisory Service identified 28 IP addresses in the United States, Japan, Germany and at least ten other countries. Officials have explicitly said a Chinese-built tool does not mean Chinese attackers. 8
The entry point was the unglamorous one. Attackers exploited weak authentication in portals used by external loan recruiters, employees' mobile tools and sales-support systems — auxiliary systems that receive less security attention than internet and mobile banking. 8
The fallout extended past finance. Two of South Korea's largest churches and Korea Electric Power Corp. confirmed their systems had been illegally accessed the same week, though it is not clear one actor was responsible for all of it. 8
The impact reached the market. South Korean cybersecurity stocks jumped by as much as 30% on Tuesday. 5
Prime Minister Han Seong-sook framed the risk in a cabinet meeting. "This is a serious situation because this incident is believed to have taken advantage of artificial intelligence, and if AI is used in phishing attacks, it could lead to secondary damage," she said. President Lee Jae Myung's instruction was simply: "Speed is of the essence." 58
The constraint for defenders is that Artex was legitimate software. It was built to find vulnerabilities in your own network, placed first in a September competition for agentic security platforms, and only after the bank reports did it revise its usage terms to bar unauthorised access, theft of data or any other malicious use. 5
Also this week
- The State of AI Report put the frontier at three labs and the price of capability on a steep curve. Its ninth edition, published October 8, describes the frontier as a race between Anthropic, OpenAI and Google, with Anthropic leading Artificial Analysis's Intelligence Index and Google leading Arena's preference ranking. Epoch AI work cited in the report estimates that holding a fixed benchmark score has become 47% cheaper each quarter since 2023, roughly 13-fold a year, and it puts OpenAI's and Anthropic's combined reported annualised run rates at about $105 billion by late summer, up from about $30 billion at the start of the year. 9 It also makes nine predictions for the next twelve months, including a regulator attributing an abnormal stock move to correlated retail AI-agent orders and, outright, "AGI 2027". 10
- OpenAI put GPT-6 and its new Intelligent UI in front of every ChatGPT user. The release went to Plus, Pro, Business and Enterprise tiers on October 7 and to Free and Go tiers the next day, powered by GPT-6 Sol on paid tiers and GPT-6 Luna on the free ones. Intelligent UI lets a response compose itself as charts, forms and interactive tools rather than text alone, and OpenAI says answers requiring web search begin 44% sooner on average than GPT-5.6 Instant. 11
- A Reuters/Ipsos poll found voters well ahead of Washington on AI risk. Fifty-seven percent said the Trump administration is not taking AI risks seriously enough and 54% said the same of Congress; 84% see AI as a threat to American workers; 56% support limits on new data centre construction; and 62% believe AI could eventually spiral beyond human control. Stricter regulation was backed by 80% of Democrats and 61% of Republicans. 12
- An independent research firm put a number on the Anthropic sceptics. New Constructs valued the company at about $150 billion against the $2 trillion it is reportedly seeking on the Nasdaq, called the offering the "most ridiculous IPO of 2026", and argued that rising operating losses plus open-source competition leave it without a viable business. 6
- Finland ordered Google to stop work on two data centre sites. The Licensing and Supervision Agency, LVV, told Google subsidiary Tuike Finland Oy on October 6 to suspend tree removal, excavation, quarrying, ditching and road building at Muhos and Kajaani until environmental impact assessments are complete, with a halt deadline of October 23. "We understand the concerns and have fallen short of our own high standards in this instance," a Google spokesperson said. 13
- AI took $102 billion of a $159 billion quarter. Crunchbase data for Q3 2026 shows AI-driven startups raising 64% of global venture funding, down from the two previous quarters but 14 points above the same quarter last year, alongside a record 27 rounds of $1 billion or more. 14
Three checks will show what this week's paper is worth. Whether OpenAI's next disclosure reconciles its two revenue figures. Whether the Genesis Mission's pledges appear as delivered compute in a federal budget line. And whether South Korean investigators ever name an attacker, rather than a tool.
References
- 1
- 2
- 3Gemini at Work 2026: Introducing Gemini agent
cloud.google.com
- 4
- 5
- 6
- 7
- 8
- 9The State of AI Report 2026
nathanbenaich.substack.com
- 10
- 11GPT-6 and Intelligent UI for everyone
openai.com
- 12
- 13
- 14Q3 2026 Posted A Record Count Of Billion-Dollar Rounds
news.crunchbase.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.