The day Grok's running costs exceeded revenue── Costs and usage limits at Grok, Anthropic and OpenAI
Download the video (MP4, 11 MB)
When we choose an AI tool, we usually compare how smart it is and what it costs. Today's news raises another question. Is there really enough money to keep these tools running?
Many makers seem to lose more the more their products are used. Who pays for that gap has rarely been discussed. Now the strain is showing. Allowances are shrinking, and some firms are cutting staff. If pharma teams hand work to AI, they cannot ignore this strain. Where does it appear, and what should users prepare?
Every story today leads back to that question.
01Grok tightened its usage limits
Source The Times of India ／ Gizmodo ／ Anadolu Ajansı ／ Startup Fortune
We start with paying customers. They expected to get what they paid for. One day, that changed for the maker's own reasons.
The reason is what matters. It was not a rush of demand. The boss himself admitted the price lost money on every sale. So a monthly fee never guaranteed a fixed amount of use. Same price, less service. That is a price rise in disguise.
- Role
- xAI founder
- Said
- Running costs “unsustainably high” relative to revenue
- Criticism
- Contradicts “AI makes saving money pointless” days earlier (Gizmodo)
- Role
- Grok Bot subscribers
- Said
- Asked for an early reset of weekly limits
- Criticism
- Called an “indirect rug pull”
I think customers were angry about more than the smaller allowance. Days earlier, the same man said cutting costs did not matter. Now he admits the costs are heavy. When a maker's words and its books drift apart, user trust breaks first. Users who lose trust start preparing to switch.
The same thing could happen in a pharma office. If one tool drafts your documents and runs your searches, work could stop mid-week. For regulatory filings with deadlines, that delay is serious. So decide on a backup tool and a fallback routine in advance. This is my guess, but other makers may follow the same path.
Is this pricing strain unique to one maker? The books of a much larger company give an answer.
02Anthropic's losses and compute
Source The New York Times
Does the same loss-making pattern hold for an industry leader? The clue lies in papers it filed ahead of selling shares to the public.
| Item | Reported | Source |
|---|---|---|
| Revenue | About $17 billion | Amended IPO filing |
| Loss | About $23 billion | Amended IPO filing |
| AI spending commitments | On the order of $900 billion | Forbes |
| Compute | Racing to secure it | NYT |
| Tone | Narrowing loss, surging revenue / slowing growth, rising costs | Reuters / Fortune |
Coverage of the same filing differed in tone
According to reports, its losses are larger than its income. Its promises to pay in the future are reportedly bigger again, by more than an order of magnitude. I see today's prices as backed by investors, not as true costs.
New money from the markets fills this gap. If that money dries up, prices will rise or allowances will shrink. What happened in the earlier story could repeat on a far larger scale.
Another point stands out. Outlets reading the same document came out upbeat or gloomy. Read only the headlines, and you could reach the opposite judgment. The safest way is to check the original document yourself.
If a pharma company picks a long-term partner, it should check the partner's finances, not just performance. Check contract length and the terms for price changes. Then sudden changes are easier to absorb. Not relying on a single maker is part of choosing well.
So what are makers who want smaller losses starting to ask of users? The next change is not about price. It is about how use is counted.
03One usage limit for ChatGPT
Source TechCrunch ／ Yahoo Finance
A maker that wants to limit losses may change how it counts use before it raises prices. One major firm chose exactly that.
The chat screen and the work screen became one. Time spent talking and time spent coding now come out of one pool. It looks more convenient. I read it as a change that helps the maker control total use. For users, every task now drains toward one ceiling.
Writing code takes far more computing than conversation. So under one contract, a few heavy users could push others to the ceiling early. The ceiling itself is also said to be lower. Users share a pool that also drains faster.
Pharma teams often run summaries and automated reporting on the same contract. You need a way to measure who used how much. Each department needs its own share. Otherwise, tools may stop working at the end of the month. Managing use is no longer just a job for IT.
So far, makers have been rationing and recounting use. Meanwhile, some makers are growing their sales.
04Tempus AI's earnings
Source TradingView ／ BigGo Finance ／ TIKR.com ／ CoinCodex
Next is a company built on medical data. Unlike chat tools for everyone, it serves focused clients. Is that business making money?
| Item | Figure | Nature |
|---|---|---|
| Revenue | $411.5 million (up 23%) | Actual |
| Net loss | $14.9 million | Actual. Headline: “first profit” |
| Full-year revenue guidance | $1.60–1.605 billion | Company outlook. Raised |
| Share price | Jumped 10% | TIKR.com report |
| Ark Invest | Bought $12 million of shares | CoinCodex report |
Headlines do not show which profit measure turned positive
The answer: still in the red, but the gap is small, and sales are growing. The company also lifted its outlook for the year. Platforms that own data and have a clear use find buyers more easily, I think. General tools are rationing use, while focused firms grow.
But one caution. Headlines said the firm made money. The figures show a loss. Headlines alone cannot tell you how the gain was counted. Before using a figure in a decision, check what it actually measures.
The stock rose sharply, and a well-known investor bought in. The market is paying for the company's forecast. But that forecast is an estimate, not a promise. Expectations can move the other way when they miss.
Pharma companies are key buyers of such platforms. When a partner is growing, check how your data is used and what happens when the contract ends. The partner's results shape where your data goes.
Once such platforms reach the workplace, where does the AI itself go wrong? The question shifts from money to use.
05Where proactive AI fails
Source HALO: Human-Grounded Granular Evaluation of Proactive Agents
AI that acts without being asked is convenient, but it brings new kinds of mistakes. One study had people check those mistakes one by one.
| Item | Reported result |
|---|---|
| Acceptance rate | Best commercial model beat the best open-source model |
| Most common failure | Over-proactivity: too early, or too strong an intervention |
| Share of over-proactivity | Majority of failed responses |
| Scale | 10 models × 400 scenarios = 8,000 responses |
Judged by 20 experts. Preprint, not yet peer-reviewed
The most frequent mistake was doing too much. The AI stepped in when no help was needed, or acted forcefully too soon. Making AI smarter does not help if it cannot tell when to hold back. Overstepping is harder to manage than being unhelpful.
Four types of need
Dependencies, dead ends, risks and opportunities. Scenarios built in 10 everyday domains and 50 personal environments.
Gather evidence first
Before judging whether help is warranted, the agent must gather evidence from databases and documents.
Diagnosing failures
Failures were split into timing, intensity and other issues. Reasoning tokens and search counts were also logged.
Here, the AI had to collect material before deciding whether to help. Acting without it means intervening on assumptions. People work the same way. We check first, then give our view. AI is no different.
However, these findings come from work not yet reviewed by experts. Details may change. Still, I think the idea that overstepping is the core mistake matches what practitioners see.
In pharma, AI must not start editing or deciding on its own in safety reports or material reviews. People need to set in advance when and how far it may act. A good start is writing human checkpoints into the SOPs.
If people set the limits, who builds the foundation? Moves by makers close to our industry point to that question.
06In the pharma workplace
Source FinancialContent ／ TradingView ／ Yahoo Finance
Drug companies rarely build AI from scratch. Most buy foundations made by specialist firms. So where did money flow among the firms that build them?
nPhase raises funds
$30 million Series D to expand globally its AI-native clinical data platform for life sciences.
RWS acquires Acolad
Deal completed on November 2. Adds interpreting, transcription and regulated-industry expertise to AI content.
Trademark data into Copilot
Clarivate will bring CompuMark trademark data and insights into Microsoft Copilot via IPOne.
It went to firms that sell trustworthy content. Trial records, translation for heavily regulated sectors, and brand records. I read this as value moving from the chat tool to the reliable content inside it. Anyone can build chat, so it tends toward a price war.
Firms at a later funding stage are entering a phase of global selling. For drug company buyers, that means more options to compare and more places to switch. More leverage in talks is good news for buyers. It also demands sharper judgment.
Still, there is a catch. When outside data enters a chat tool, it gets harder to see where an answer came from. Stale records could pass straight into wrong decisions. For filings or labeling decisions, demand that makers let you trace the evidence.
Once tools and data are in place, staffing comes next. The final story looks at it from the workers' side.
07AI job cuts and China's guidelines
Source Caixin Global ／ The Register ／ GitHub
The heavier AI costs become, the more firms look for savings elsewhere. Increasingly, they look at headcount.
| Actor | Action | Source |
|---|---|---|
| China State Council | Asks high-impact AI projects to assess job risks | Caixin Global |
| China State Council | Urges firms to give advance notice and re-employment help | Caixin Global |
| SUSE | To cut one in ten staff, citing AI | The Register |
| Dropbox | Disclosed ~200 job cuts (~9% of staff) after adding Copilot link | GitHub |
China's plan has 18 points
One firm announced it would cut a visible share of its workforce and pointed to AI. Another cut roles right after adding an AI link. Buying tools and cutting people now come up together. It looks as if firms pay for AI with jobs. That is my reading, not the companies' explanation.
In response, a major government asked firms to act early. For big deployments, they should estimate the effect on work, warn employees early, and support their next roles. The state is now shaping how AI is introduced.
In our industry, putting AI into document work or inquiry handling changes the job. Show early whose work will change and how, and offer retraining. I think that is what makes adoption last. Cutting people alone breeds anxiety among those who stay. It can also lower the quality of AI use.
Costs, then allowances, then people. That is the order in which the strain has surfaced. So what should users prepare?
We will watch the timing and price of Anthropic's IPO, and whether usage limits spread to other makers.
Open the full transcript
Intro
When we choose an AI tool, we usually compare how smart it is and what it costs. Today's news raises another question. Is there really enough money to keep these tools running?
Many makers seem to lose more the more their products are used. Who pays for that gap has rarely been discussed. Now the strain is showing. Allowances are shrinking, and some firms are cutting staff. If pharma teams hand work to AI, they cannot ignore this strain. Where does it appear, and what should users prepare?
Every story today leads back to that question.
CH 01 Grok tightened its usage limits
We start with paying customers. They expected to get what they paid for. One day, that changed for the maker's own reasons.The reason is what matters. It was not a rush of demand. The boss himself admitted the price lost money on every sale. So a monthly fee never guaranteed a fixed amount of use. Same price, less service. That is a price rise in disguise.I think customers were angry about more than the smaller allowance. Days earlier, the same man said cutting costs did not matter. Now he admits the costs are heavy. When a maker's words and its books drift apart, user trust breaks first. Users who lose trust start preparing to switch.The same thing could happen in a pharma office. If one tool drafts your documents and runs your searches, work could stop mid-week. For regulatory filings with deadlines, that delay is serious. So decide on a backup tool and a fallback routine in advance. This is my guess, but other makers may follow the same path. Is this pricing strain unique to one maker?
The books of a much larger company give an answer.
CH 02 Anthropic's losses and compute
Does the same loss-making pattern hold for an industry leader?The clue lies in papers it filed ahead of selling shares to the public.According to reports, its losses are larger than its income. Its promises to pay in the future are reportedly bigger again, by more than an order of magnitude. I see today's prices as backed by investors, not as true costs.New money from the markets fills this gap. If that money dries up, prices will rise or allowances will shrink. What happened in the earlier story could repeat on a far larger scale.Another point stands out. Outlets reading the same document came out upbeat or gloomy. Read only the headlines, and you could reach the opposite judgment. The safest way is to check the original document yourself.If a pharma company picks a long-term partner, it should check the partner's finances, not just performance. Check contract length and the terms for price changes. Then sudden changes are easier to absorb. Not relying on a single maker is part of choosing well. So what are makers who want smaller losses starting to ask of users?
The next change is not about price. It is about how use is counted.
CH 03 One usage limit for ChatGPT
A maker that wants to limit losses may change how it counts use before it raises prices. One major firm chose exactly that.The chat screen and the work screen became one. Time spent talking and time spent coding now come out of one pool. It looks more convenient. I read it as a change that helps the maker control total use. For users, every task now drains toward one ceiling.Writing code takes far more computing than conversation. So under one contract, a few heavy users could push others to the ceiling early. The ceiling itself is also said to be lower. Users share a pool that also drains faster.Pharma teams often run summaries and automated reporting on the same contract. You need a way to measure who used how much. Each department needs its own share. Otherwise, tools may stop working at the end of the month. Managing use is no longer just a job for IT. So far, makers have been rationing and recounting use. Meanwhile, some makers are growing their sales.
CH 04 Tempus AI's earnings
Next is a company built on medical data. Unlike chat tools for everyone, it serves focused clients. Is that business making money?The answer: still in the red, but the gap is small, and sales are growing. The company also lifted its outlook for the year. Platforms that own data and have a clear use find buyers more easily, I think. General tools are rationing use, while focused firms grow.But one caution. Headlines said the firm made money. The figures show a loss. Headlines alone cannot tell you how the gain was counted. Before using a figure in a decision, check what it actually measures.The stock rose sharply, and a well-known investor bought in. The market is paying for the company's forecast. But that forecast is an estimate, not a promise. Expectations can move the other way when they miss.Pharma companies are key buyers of such platforms. When a partner is growing, check how your data is used and what happens when the contract ends. The partner's results shape where your data goes. Once such platforms reach the workplace, where does the AI itself go wrong?
The question shifts from money to use.
CH 05 Where proactive AI fails
AI that acts without being asked is convenient, but it brings new kinds of mistakes. One study had people check those mistakes one by one.The most frequent mistake was doing too much. The AI stepped in when no help was needed, or acted forcefully too soon. Making AI smarter does not help if it cannot tell when to hold back. Overstepping is harder to manage than being unhelpful.Here, the AI had to collect material before deciding whether to help. Acting without it means intervening on assumptions. People work the same way. We check first, then give our view. AI is no different.However, these findings come from work not yet reviewed by experts. Details may change. Still, I think the idea that overstepping is the core mistake matches what practitioners see.In pharma, AI must not start editing or deciding on its own in safety reports or material reviews. People need to set in advance when and how far it may act. A good start is writing human checkpoints into the SOPs. If people set the limits, who builds the foundation?
Moves by makers close to our industry point to that question.
CH 06 In the pharma workplace
Drug companies rarely build AI from scratch. Most buy foundations made by specialist firms. So where did money flow among the firms that build them?It went to firms that sell trustworthy content. Trial records, translation for heavily regulated sectors, and brand records. I read this as value moving from the chat tool to the reliable content inside it. Anyone can build chat, so it tends toward a price war.Firms at a later funding stage are entering a phase of global selling. For drug company buyers, that means more options to compare and more places to switch. More leverage in talks is good news for buyers. It also demands sharper judgment.Still, there is a catch. When outside data enters a chat tool, it gets harder to see where an answer came from. Stale records could pass straight into wrong decisions. For filings or labeling decisions, demand that makers let you trace the evidence. Once tools and data are in place, staffing comes next. The final story looks at it from the workers' side.
CH 07 AI job cuts and China's guidelines
The heavier AI costs become, the more firms look for savings elsewhere. Increasingly, they look at headcount.One firm announced it would cut a visible share of its workforce and pointed to AI. Another cut roles right after adding an AI link. Buying tools and cutting people now come up together. It looks as if firms pay for AI with jobs. That is my reading, not the companies' explanation.In response, a major government asked firms to act early. For big deployments, they should estimate the effect on work, warn employees early, and support their next roles. The state is now shaping how AI is introduced.In our industry, putting AI into document work or inquiry handling changes the job. Show early whose work will change and how, and offer retraining. I think that is what makes adoption last. Cutting people alone breeds anxiety among those who stay. It can also lower the quality of AI use. Costs, then allowances, then people. That is the order in which the strain has surfaced. So what should users prepare?
Wrap-up
AI is neither cheap nor unlimited. That is today's answer. Its costs come back to users as smaller allowances and staffing changes. So keep a backup, measure your use, and let people decide how far AI may act. You can start all of this today. Tomorrow, we watch how the market prices a maker with deep losses. That price will shape how much we can use.
- CH 01The Times of India「Elon Musk just admitted for his company's chatbot what is also the 'biggest problem' of all AI companies;」 timesofindia.indiatimes.com
- CH 01Gizmodo「Elon Musk Needs to Save Money on Grok, Days After Saying AI Makes Saving Money Pointless」 gizmodo.com
- CH 01Anadolu Ajansı「Musk says cost of running Grok Bot 'unsustainably high' relative to revenue」 aa.com.tr
- CH 01Startup Fortune「Elon Musk admits Grok Bot costs too much to run as users hit weekly caps」 startupfortune.com
- CH 02The New York Times「Inside Anthropic’s Urgent Hunt for A.I. Computing Power」 nytimes.com
- CH 03TechCrunch「OpenAI is finally merging ChatGPT’s Chat and Work tabs」 techcrunch.com
- CH 03Yahoo Finance「OpenAI Tightens ChatGPT, Codex Usage Limits As Fundraising Plans Reportedly Shrink」 finance.yahoo.com
- CH 04TradingView「Tempus AI 3Q 2026: Revenue $411.5M, Net loss $14.9M — 10-Q Summary」 tradingview.com
- CH 04BigGo Finance「[Tempus AI Q3 2026 Earnings Call] Tempus AI Turns First Profit as Revenue Jumps 23% to $411.5M; Raises Full-Year Guidance to $1.60B-$1.605B」 finance.biggo.com
- CH 04TIKR.com「Tempus AI Soars 10% on Q3 Beat and Raised Guidance as Oncology Sees “Accelerating Unit Growth”」 tikr.com
- CH 04CoinCodex「Cathie Wood's Ark Invest Buys $12M in Tempus AI Stock Amid Bullish Wall Street Calls」 coincodex.com
- CH 05HALO: Human-Grounded Granular Evaluation of Proactive Agents（「中心の発見は、出過ぎ（over-proactivity）が観察された中で最も多い失敗の型だったことである。」）
- CH 06FinancialContent「nPhase Secures $30 Million in Series D Funding to Drive Global Expansion of AI-Native Clinical Data Platform for Life Sciences」 financialcontent.com
- CH 06TradingView「RWS completes acquisition of Acolad, giving enterprise customers more ways to scale with AI and specialist expertise」 tradingview.com
- CH 06Yahoo Finance「Clarivate to Bring Trusted Trademark Data into Microsoft Copilot Through IPOne」 finance.yahoo.com
- CH 07Caixin Global「China Issues 18-Point Plan to Protect Workers From AI Job Losses」 caixinglobal.com
- CH 07The Register「SUSE to cut one in ten employees because of... AI」 theregister.com
- CH 07GitHub「At Dropbox, a New Copilot Link and 200 Fewer Jobs」 digidai.github.io
Articles used
- AI Daily News 2026-11-03
- AI and the Pharmaceutical Industry 2026-11-03
- AI Daily News: key points of each article (Nov 3, morning)
- AI Daily News: key points of each article (Nov 2, evening)
- HALO: over-proactive AI agents
Today's related reports
- AI Daily News November 3, 2026
- AI & Economy News November 3, 2026
- AI & Finance News November 3, 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.