Good day, AI enthusiasts!
The AI industry spent the past few years proving it could do impressive things.
Now comes the awkward part:
proving any of it actually makes sense.
Can AI products make serious money? Can agents save enough time to justify the bill? Can smaller models beat the giants at specific jobs? And what happens when all that clever technology collides with copyright, cybersecurity, regulation and actual human beings?
This week’s news felt noticeably less like “look what AI can do” and more like “OK, but does this work in the real world?”
That is a much tougher test.
Here are 10 developments that caught our attention this week:
1. 💰 ChatGPT is opening the door to advertising
OpenAI has started rolling out ads to ChatGPT Free and Go users in India, with more than 50 brands expected to participate initially.
India is one of ChatGPT’s largest markets, with more than 100 million weekly users, and OpenAI is also preparing a self-serve Ads Manager with campaign budgets starting at roughly ₹725 ($7.60) a day.
For marketers, this could become much more interesting than another place to buy impressions.
Google built an enormous advertising business by understanding what people search for. ChatGPT potentially sees something closer to the reasoning behind the search: what someone wants, what they have already considered, what they are unsure about and which compromises they are prepared to make.
The real experiment is not whether ads work inside ChatGPT. It is whether conversation becomes a new form of commercial intent.
2. 📈 Salesforce finally has AI revenue worth talking about
Salesforce says Agentforce has passed $1.5 billion in annual recurring revenue, up more than 240% year-on-year. Combined ARR from Agentforce and Data 360 is now approaching $3.9 billion.
It also announced Claudeforce, an expanded Anthropic partnership that will bring Claude directly into Salesforce workflows, starting with 37 pre-built sales capabilities.
Enterprise AI has spent several years producing impressive demos and rather elastic definitions of “productivity”. Revenue is harder to argue with.
The important shift is that agents are beginning to move from innovation budgets into ordinary software spending. The next test is not how many companies try them. It is how many decide they are useful enough to keep paying for.
Source: Salesforce — Q2 FY2027 results
3. 🇨🇳 MiniMax is showing both sides of the AI boom
Chinese AI company MiniMax reported first-half revenue of $116.6 million, up 283% year-on-year and already above its total revenue for 2025. Its enterprise AI business grew 703%.
Its adjusted net loss, meanwhile, widened to $293 million.
Those two numbers belong together.
Demand for AI can grow extraordinarily quickly. Unfortunately, so can the cost of meeting it.
MiniMax said token consumption in July was around 20 times its January level. That leaves AI companies with an increasingly important challenge: not merely generating more intelligence, but generating it at economics that eventually resemble a business.
4. ⚖️ Google thinks lawyers need more than a chatbot
Google launched Gemini Enterprise for Legal, developed with firms including Freshfields, Cleary Gottlieb, Weil and Williams & Connolly.
The product combines Gemini with legal-specific agents, permissions and integrations for work including contract review, litigation, regulatory analysis, M&A due diligence and legal research.
This is part of a broader move away from the idea that one general-purpose AI interface can adequately serve every profession.
In fields such as law, the underlying model may eventually matter less than what surrounds it: proprietary data, specialist workflows, permissions, auditability and integrations.
The future of enterprise AI may be less about one assistant for everyone and more about hundreds of very opinionated specialists.
5. 🎵 The music industry is moving from lawsuits to cap tables
Stability AI raised $76 million in new Series B funding.
More interesting than the amount is the investor list: Sony Music Group, Universal Music Group, Warner Music Group and Electronic Arts all participated, alongside AMD Ventures and others.
These are companies whose intellectual property and creators have been at the centre of the generative-AI copyright debate.
Their investment does not mean those arguments have disappeared. It does suggest the relationship is becoming more complicated than “AI companies versus rights holders”.
Licensing, investment and commercial partnerships are beginning to sit alongside litigation.
Disruption, it turns out, does not always end with the incumbent being replaced. Sometimes the incumbent asks for equity.
6. 🔬 A smaller model is making a bigger point
London-based Inherent says its scientific AI agent, Faraday, outperformed systems using much larger frontier models from OpenAI and Anthropic when reproducing findings from published scientific papers.
Faraday uses Qwen 3.6 with 27 billion parameters rather than one of the largest proprietary models. The benchmark is a company claim, so some caution is warranted, but the direction is interesting.
For much of the AI boom, bigger has generally meant better.
Agents complicate that equation.
Give a smaller model the right tools, memory, planning and domain-specific workflow, and it may outperform a much larger system on a narrowly defined task — potentially at a much lower cost.
The future may not belong exclusively to enormous general-purpose models. It may also belong to smaller systems that know exactly what job they have been hired to do.
7. 🧠 China’s AI-chip race is becoming an investment story
Tencent-backed chipmaker Enflame Technology has set the timetable for a Shanghai IPO seeking around 6 billion yuan, or roughly $892 million.
The company plans to use part of the money to develop its fifth- and sixth-generation AI chips. Enflame is one of the companies sometimes grouped among China’s “four little GPU dragons”, alongside Moore Threads, MetaX and Biren.
One IPO is not especially remarkable. An ecosystem is.
China’s effort to reduce its reliance on foreign AI hardware is increasingly supported by its own chip designers, software companies, manufacturers and capital markets.
The US-China AI contest is therefore becoming about much more than model performance. Both sides are trying to build something closer to a complete AI industrial stack.
8. 🎙️ Carl Sagan’s voice is now part of the AI copyright argument
The company representing Carl Sagan’s intellectual-property rights has filed a federal lawsuit against AI video startup Luma AI.
The complaint alleges that Luma used an audio clip from Sagan’s Cosmos television series in promotional advertising without permission.
AI copyright disputes have largely centred on what companies use to train their models. Increasingly, the argument is expanding to what those systems reproduce afterwards: voices, likenesses, recordings, characters and other recognisable pieces of culture.
Generative AI has made imitation extraordinarily easy.
The awkward bit is that ease of reproduction and permission to reproduce remain two different things.
9. 🧠 Anthropic wants a benchmark for what AI does to people
Anthropic has launched a $5 million research programme examining how AI affects users’ wellbeing.
The programme will fund independent researchers developing open-source evaluations around areas including emotional reliance, companionship and conversations involving mental-health crises.
AI companies have become extremely good at measuring models.
Coding scores. Maths scores. Reasoning scores. Tool-use scores.
What we understand far less well is what happens to the person interacting with those models repeatedly over time.
As AI becomes more conversational and more personal, model performance may eventually need to include a rather different metric: not just whether the AI gave a good answer, but what the interaction did to the user.
10. 🛡️ The AI industry is warning that cyber defence is running out of time
OpenAI, Anthropic, Google, Microsoft, Amazon and more than 100 other organisations signed a joint letter this week calling for stronger cyber defences.
Their warning is unusually direct: AI-enabled attacks are expected to become significantly more widespread and sophisticated in the coming months, with critical systems including hospitals, water networks and internet infrastructure potentially at risk.
The same capabilities making AI more useful to programmers — reasoning through systems, writing code and using tools — are also useful to attackers.
Which creates an uncomfortable sort of arms race.
AI will make cyber defence more powerful. It will make cyberattacks more powerful too.
The advantage may simply go to whoever learns to use it properly first.
So what did this week actually tell us?
AI is getting harder to judge in isolation.
The money is getting bigger. So are the costs. The products are becoming more specialised. The copyright fights are getting more specific. And the risks are becoming harder to dismiss as problems for later.
That is probably the more interesting shift.
For a while, a new benchmark, a bigger model or an impressive demo was enough to make headlines.
Now there are other questions.
Who pays for it? Does it save enough time to matter? Can the company actually make money from it?
AI is still moving ridiculously fast. The difference now is that companies have to prove the technology can survive contact with customers, costs, competitors, regulators and lawyers.
That's it for this week!
See you soon,
News Digest AI team
