Revenue doubts as AI agents move in to stay── OpenAI revenue report, always-on agents, reasoning research
Download the video (MP4, 12 MB)
An AI developer may be earning less than we have been told. That news arrived today. At the same time, the range of work we hand to AI is set to grow again. When hopes and money are shaking, how far can we trust AI with real work?
Today we carry that single question through money, through ways of putting AI to work, and through research that checks how AI thinks. What must pharma professionals confirm before they let AI take a task on?
What concerns me is not whether the figures are big or small. It is what it means that they moved.
01Reports on OpenAI's revenue
Source CNBC ／ OpenAI公式 ／ Reuters
We start with a plain question. How much does an AI developer really earn? If that answer shifts, the logic behind the money poured into AI is open to doubt. For those who pick the tools, the maker's finances matter too.
| Item | Detail |
|---|---|
| Annualized revenue (reported) | About $50 billion |
| Gap vs. earlier signals | About $20 billion less |
| Reported by | Financial Times, CNBC |
| Same-day official blog | Oracle cuts days of work to minutes with ChatGPT and Codex |
| Same-day lawsuit | USA Today sues over copyright in training data |
Revenue figure comes from press reports, not an official OpenAI announcement
The reported figure fell well short of what had been expected. And it came from the press, not from the developer. That same day, the company promoted a big client whose workload shrank sharply. Meanwhile a newspaper took it to court for using articles without permission. Earning, selling and being sued all landed at once. I read this as a sign. The time of being valued on hopes alone is ending.
For pharma professionals, the weight here is the weight of a contract. Bringing a tool in-house means reshaping procedures and training around it. If the maker falls short, prices or policies may change. This is my own guess, but relying on a single vendor gets riskier. Depending on the court case, users may also be asked how source material was handled.
As earnings wobble, how is money flowing out elsewhere? Next, we look at another developer's spending.
02Spending by a company eyeing an IPO
Source Bloomberg.com ／ Benzinga
From income, we turn to outgoings. For a company preparing to list its shares, every dollar spent is a question put to investors. How far will it go, and what is it bracing for?
Compute plan
Anthropic reportedly plans to spend $518 billion on compute. The scale and the demand for cash stand out.
Funding for AI science
Anthropic pledged $150 million to an AI science initiative backed by the White House.
IPO and rogue-AI risk
Bloomberg reported that investors eyeing an Anthropic IPO struggle to price the risk of rogue AI.
The screen shows a huge spending plan and money for state-backed research. The plan's size could not be confirmed in the original reports gathered today. Then there is a market still unable to price machines that stop doing what we intend. That last point caught my eye. We can compare sums of money. We have no yardstick yet for comparing danger. What has no yardstick tends to drop out of investment decisions.
Pharma professionals know what it means to lack a measure. Medicines have long had harms weighed as seriously as benefits, with records kept. For AI, that way of weighing is not yet settled. So users should list the dangers themselves and keep steps to check them. Adding one line for risk next to benefit in an approval request already changes the decision.
With no yardstick for danger, AI is now set to work longer, further from human hands.
03Gemini's always-on agents
Source blog.google ／ VentureBeat ／ TechCrunch ／ SiliconANGLE
Here the story moves from money to how AI is put to work. It is changing from a tool that answers each question once into something that waits and works for hours. That changes how we decide what to hand over.
What matters most is that the AI gets its own mail, schedule and document store. It does not wait for each human instruction. It pushes long jobs forward by itself. In effect, it starts receiving messages and keeping records for us.
| Company | Announced |
|---|---|
| Goodfire | "Inside-out" monitor of internal states; cheap rogue-AI detection |
| Darwinium | Two new features to detect fraud by AI agents |
That same day, several products to check AI from the inside or by its behavior were launched. As delegation widens, demand for checking it grows too.
Meanwhile, fewer firms now let AI fix the systems they run live. My view is that more companies tried it and learned to be wary.
Safety reports and quality records must show who did what, and when. If an AI works from its own inbox, its actions must be kept in a form people can trace. Decide how to check before you delegate. Never reverse that order.
As delegation widens, big money is also flowing to firms that use AI to find medicines. Next, what lies behind those sums.
04Money in AI drug discovery
Source Bloomberg.com ／ Axios ／ 36Kr
When we see a big figure, we tend to assume that much money really changed hands. But a figure can be an estimate, a sum still to be raised, or a number kept hidden. Before reading investment news in this field, sort out what kind of figure it is.
| Company | Amount | Nature of amount |
|---|---|---|
| Isomorphic Labs | Over $40 billion | Valuation; fundraising in talks or being explored |
| Iambic | $150M | Planned IPO raise |
| NVIDIA (to 14 AI drug discovery firms) | Withheld as "$X" | 3-year investment; size unknown |
Amounts as reported. No currency conversion.
The largest is an estimate of a company's worth, not cash in hand. Talks are still going on. Another is what a first share sale is meant to bring in. For the last deal, we know how many firms and for how long, but the key sum is hidden. So none of this is settled money. If only headline figures stick, we overrate what the market expects.
This bears directly on pharma professionals' own decisions. When choosing partners or vendors, judging momentum by the size of a figure misleads. An estimate of worth falls fast when hopes fade. It is the same pattern as the developer's earnings we saw earlier. Separating settled sums from hoped-for ones is the first check on a partner's real finances. Asking about it is not rude.
Just as we separate settled sums from hopes, can we check what AI is really thinking? Next, a piece of research.
05Research on shortened AI reasoning
Source Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability
If we hand AI long jobs, we must read not just its answers but the thinking behind them. So if we train it to write that thinking more briefly, does it become unreadable? That is what this study asks.
The results split. How well the written thinking reflects the true reasons for an answer fell in many cases. Yet when someone tampered with the question, traces of that still showed in the reasoning. Shorter means less honest, but tampering stays visible. I see value in being able to separate those two.
| Method | How pressure is applied |
|---|---|
| Fixed cap | Sets a limit on generation length |
| Target length | Gives a target length per example |
| Relative reward | Rewards outputs shorter than others in the group |
| Models, size of drop and per-method gaps not given in abstract |
However, the summary does not say how much the approaches differed. Nor does it show the range of results. Other researchers have not yet checked the work, so it is too early to lean on it heavily.
This distinction helps pharma professionals who use AI for review or record keeping. Reading an AI's explanation as the true reason for its judgment is risky. As a clue for spotting tampering, though, it may be useful. Do not simply believe the explanation. Treat it as material for review.
So, in settings close to pharma, in what forms is AI already at work today?
06In pharma practice
Source Morningstar ／ Yahoo Finance ／ CNN
Let us bring all this inside a pharma company. Seeing where AI has started to work makes clear what to check. Where it enters changes both how to check and who is responsible.
AI for production data
Seeq announced a collaboration with Cognizant to speed up adoption of industrial AI in the life sciences.
Building QSP models
Simulations Plus announced a partnership funded by a global pharma company, developing AI-enabled QSP models with Turin.
Claude-led discovery
CNN reported on Anthropic's "Claude-led biological discovery". Scientists are divided in their assessment.
It is entering plant records analysis, prediction by equations of how a medicine behaves in the body, and research into how living things work. The analysis and the prediction move forward through a tie-up between outside firms and a deal paid for by a drugmaker. In the research, scientists disagree about a finding said to be led by AI. In these reports, each seems to support human judgment rather than replace it.
To me, this split in views captures today's question well. How do people check what AI produces, and who answers for it? Quality records and medicine forecasts face the same question. For record analysis, quality staff would set the checks. For forecasts, clinical staff would. As AI spreads, checking changes shape rather than disappears. Here the careful habits of pharma pay off directly.
So where is society trying to place that duty to check? Finally, how safety gets decided.
07What is being overlooked
Source 36Kr ／ The Economist ／ myMotherLode.com
Should AI safety be confirmed before release, or should blame come after a problem? For people in pharma, this is a familiar question. The difference lies in who bears responsibility, and when.
- Position
- AI researcher, Nobel laureate
- Statement
- Calls for FDA-style safety approval before model release
- Position
- Editorial
- Statement
- Liability laws alone cannot control AI
Same day: an AP-NORC poll reported that many Americans see AI developing too fast
One side argues that, like a medicine, AI should face a regulator's check before it goes out. The other, an opinion piece, says that laws assigning blame afterward cannot rein AI in by themselves. I read them as overlapping. Both say asking questions later is not enough. And in America, many people think AI is moving too quickly. When public opinion moves, rules follow. The case for binding AI in advance seems to be gaining ground.
Pharma professionals have worked inside a system with both pre-approval review and post-launch monitoring. The AI debate is searching for exactly that shape. So this industry's experience applies directly when setting AI rules in-house. Check before you delegate, and keep tracking records afterward. I recommend writing both into department rules now.
Check before delegating, and keep following the record after. With that in mind, let us close.
How the OpenAI revenue report affects fundraising, compute contracts and IPO valuations.
Open the full transcript
Intro
An AI developer may be earning less than we have been told. That news arrived today. At the same time, the range of work we hand to AI is set to grow again. When hopes and money are shaking, how far can we trust AI with real work?
Today we carry that single question through money, through ways of putting AI to work, and through research that checks how AI thinks. What must pharma professionals confirm before they let AI take a task on?
What concerns me is not whether the figures are big or small. It is what it means that they moved.
CH 01 Reports on OpenAI's revenue
We start with a plain question. How much does an AI developer really earn?If that answer shifts, the logic behind the money poured into AI is open to doubt. For those who pick the tools, the maker's finances matter too.The reported figure fell well short of what had been expected. And it came from the press, not from the developer. That same day, the company promoted a big client whose workload shrank sharply. Meanwhile a newspaper took it to court for using articles without permission. Earning, selling and being sued all landed at once. I read this as a sign. The time of being valued on hopes alone is ending.For pharma professionals, the weight here is the weight of a contract. Bringing a tool in-house means reshaping procedures and training around it. If the maker falls short, prices or policies may change. This is my own guess, but relying on a single vendor gets riskier. Depending on the court case, users may also be asked how source material was handled. As earnings wobble, how is money flowing out elsewhere?
Next, we look at another developer's spending.
CH 02 Spending by a company eyeing an IPO
From income, we turn to outgoings. For a company preparing to list its shares, every dollar spent is a question put to investors. How far will it go, and what is it bracing for?
The screen shows a huge spending plan and money for state-backed research. The plan's size could not be confirmed in the original reports gathered today. Then there is a market still unable to price machines that stop doing what we intend. That last point caught my eye. We can compare sums of money. We have no yardstick yet for comparing danger. What has no yardstick tends to drop out of investment decisions.Pharma professionals know what it means to lack a measure. Medicines have long had harms weighed as seriously as benefits, with records kept. For AI, that way of weighing is not yet settled. So users should list the dangers themselves and keep steps to check them. Adding one line for risk next to benefit in an approval request already changes the decision. With no yardstick for danger, AI is now set to work longer, further from human hands.
CH 03 Gemini's always-on agents
Here the story moves from money to how AI is put to work. It is changing from a tool that answers each question once into something that waits and works for hours. That changes how we decide what to hand over.What matters most is that the AI gets its own mail, schedule and document store. It does not wait for each human instruction. It pushes long jobs forward by itself. In effect, it starts receiving messages and keeping records for us.That same day, several products to check AI from the inside or by its behavior were launched. As delegation widens, demand for checking it grows too.Meanwhile, fewer firms now let AI fix the systems they run live. My view is that more companies tried it and learned to be wary.Safety reports and quality records must show who did what, and when. If an AI works from its own inbox, its actions must be kept in a form people can trace. Decide how to check before you delegate. Never reverse that order. As delegation widens, big money is also flowing to firms that use AI to find medicines. Next, what lies behind those sums.
CH 04 Money in AI drug discovery
When we see a big figure, we tend to assume that much money really changed hands. But a figure can be an estimate, a sum still to be raised, or a number kept hidden. Before reading investment news in this field, sort out what kind of figure it is.The largest is an estimate of a company's worth, not cash in hand. Talks are still going on. Another is what a first share sale is meant to bring in. For the last deal, we know how many firms and for how long, but the key sum is hidden. So none of this is settled money. If only headline figures stick, we overrate what the market expects.This bears directly on pharma professionals' own decisions. When choosing partners or vendors, judging momentum by the size of a figure misleads. An estimate of worth falls fast when hopes fade. It is the same pattern as the developer's earnings we saw earlier. Separating settled sums from hoped-for ones is the first check on a partner's real finances. Asking about it is not rude. Just as we separate settled sums from hopes, can we check what AI is really thinking?
Next, a piece of research.
CH 05 Research on shortened AI reasoning
If we hand AI long jobs, we must read not just its answers but the thinking behind them. So if we train it to write that thinking more briefly, does it become unreadable?
That is what this study asks.The results split. How well the written thinking reflects the true reasons for an answer fell in many cases. Yet when someone tampered with the question, traces of that still showed in the reasoning. Shorter means less honest, but tampering stays visible. I see value in being able to separate those two.However, the summary does not say how much the approaches differed. Nor does it show the range of results. Other researchers have not yet checked the work, so it is too early to lean on it heavily.This distinction helps pharma professionals who use AI for review or record keeping. Reading an AI's explanation as the true reason for its judgment is risky. As a clue for spotting tampering, though, it may be useful. Do not simply believe the explanation. Treat it as material for review. So, in settings close to pharma, in what forms is AI already at work today?
CH 06 In pharma practice
Let us bring all this inside a pharma company. Seeing where AI has started to work makes clear what to check. Where it enters changes both how to check and who is responsible.It is entering plant records analysis, prediction by equations of how a medicine behaves in the body, and research into how living things work. The analysis and the prediction move forward through a tie-up between outside firms and a deal paid for by a drugmaker. In the research, scientists disagree about a finding said to be led by AI. In these reports, each seems to support human judgment rather than replace it.To me, this split in views captures today's question well. How do people check what AI produces, and who answers for it?
Quality records and medicine forecasts face the same question. For record analysis, quality staff would set the checks. For forecasts, clinical staff would. As AI spreads, checking changes shape rather than disappears. Here the careful habits of pharma pay off directly. So where is society trying to place that duty to check?
Finally, how safety gets decided.
CH 07 What is being overlooked
Should AI safety be confirmed before release, or should blame come after a problem?For people in pharma, this is a familiar question. The difference lies in who bears responsibility, and when.One side argues that, like a medicine, AI should face a regulator's check before it goes out. The other, an opinion piece, says that laws assigning blame afterward cannot rein AI in by themselves. I read them as overlapping. Both say asking questions later is not enough. And in America, many people think AI is moving too quickly. When public opinion moves, rules follow. The case for binding AI in advance seems to be gaining ground.Pharma professionals have worked inside a system with both pre-approval review and post-launch monitoring. The AI debate is searching for exactly that shape. So this industry's experience applies directly when setting AI rules in-house. Check before you delegate, and keep tracking records afterward. I recommend writing both into department rules now. Check before delegating, and keep following the record after. With that in mind, let us close.
Wrap-up
Even as earnings outlooks waver, the work handed to AI keeps growing. So check what lies behind a figure. Treat what AI says about its thinking as material for review. Keep tracking records after you delegate. The careful habits of pharma professionals already answer today's question. Tomorrow, we look at how the earnings news touches the flow of money.
- CH 01CNBC「OpenAI annualized revenues $20 billion less than previously signaled: Report」 cnbc.com
- CH 01OpenAI公式「How Oracle turns days of work into minutes with ChatGPT and Codex」 openai.com
- CH 01Reuters「USA Today sues OpenAI for copyright infringement over AI training」 reuters.com
- CH 02Bloomberg.com「Anthropic IPO Investors Struggle to Put a Price on Rogue AI Risk」 bloomberg.com
- CH 02Benzinga「Anthropic Commits $150 Million To White House-Backed AI Science Mission」 benzinga.com
- CH 03blog.google「Google Cloud introduces the Gemini agent.」 blog.google
- CH 03VentureBeat「Share of enterprises trusting AI agents to make production changes falls from 75% to 56% in latest VB Intelligence survey」 venturebeat.com
- CH 03TechCrunch「Goodfire says its new ‘inside-out’ monitors catch rogue AI agents at a fraction of the cost」 techcrunch.com
- CH 03SiliconANGLE「Darwinium launches two intent intelligence capabilities to catch fraud by AI agents」 siliconangle.com
- CH 04Bloomberg.com「Alphabet’s Isomorphic Labs in Funding Talks for at Least $40 Billion Value」 bloomberg.com
- CH 04Axios「AI drug discovery startup Iambic seeks $150M in IPO」 axios.com
- CH 0436Kr「NVIDIA's 3-Year $X Investment in 14 AI Pharma Startups: Is a Wave of Star-Studded AI Pharmaceutical IPOs On the Horizon?」 eu.36kr.com
- CH 05Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability（「忠実性は多くの設定で下がったが、モニタ可能性は CoT がずっと短くなっても持ちこたえた。」）
- CH 06Morningstar「Seeq Collaborates with Cognizant to Accelerate Industrial AI Adoption in Life Sciences」 morningstar.com
- CH 06Yahoo Finance「Simulations Plus Announces Funded Partnership with Global Pharmaceutical Company to Advance AI-Enabled QSP Model Development with Turin」 finance.yahoo.com
- CH 06CNN「What to know about Anthropic’s ‘Claude-led’ biological discovery — and why scientists aren’t convinced」 cnn.com
- CH 0736Kr「AI Self-Creation Era Begins: "Godfather of AI" Hinton Warns the Most Dangerous Stage Is Approaching」 eu.36kr.com
- CH 07The Economist「Liability law will not tame AI」 economist.com
- CH 07myMotherLode.com「Most Americans think artificial intelligence is developing too fast, a new AP-NORC poll finds」 mymotherlode.com
Articles used
- AI Daily News 2026-10-09
- AI and the Pharmaceutical Industry — 2026-10-09
- AI and the Latest Technology: If reasoning gets shorter, can we still monitor it?
Today's related reports
- AI Daily News October 9, 2026
- AI & Economy News October 9, 2026
- AI & Finance News October 9, 2026
The narration is synthetic speech. The content draws only on this site's articles from the same day. Each edition passes seven plain-language checks before publication. That means known obstacles to comprehension are held below threshold — it is not a guarantee of comprehension.