Good day, AI enthusiasts!
It was another busy week in AI, but a few themes stood out.
AI agents are getting much closer to real work. Meta launched a personal agent that can handle everyday tasks, while OpenAI introduced a version of ChatGPT built specifically for financial services.
The money going into AI is still huge too. Harvey raised $550 million at a $15.5 billion valuation, Oracle signed more than $30 billion in new AI cloud contracts in a single quarter, and Google committed €13 billion to AI infrastructure in Finland.
There were also some important developments around regulation, surveillance and jobs — including new AI auditing rules in California and fresh research suggesting the impact on employment may be more complicated than “AI replaces workers”.
Here are 10 AI developments worth knowing from this week:
1. 🧭 Meta wants AI to run the to-do list, not answer it
Meta launched Muse, a personal AI agent designed to do things on a user’s behalf rather than simply respond in a chat window. It can send emails, book travel, organise plans and work across connected services, including through WhatsApp.
The more interesting part is the architecture. Muse runs inside a dedicated virtual machine, with a separate “Sentinel” agent reviewing what it tries to do online. Meta says it cannot see users’ passwords or payment details and requires approval before sensitive actions such as purchases or sending emails.
The strategic advantage here may be less about having the best agent and more about having somewhere to put it. Meta already owns several of the interfaces through which billions of people communicate, shop, organise and share information.
The next consumer AI battle may therefore be fought over distribution and permission, not the chat window.
Source: Meta — Introducing Muse
2. 💼 OpenAI moved from office software into the deal room
OpenAI launched ChatGPT for Financial Services, built with input from Morgan Stanley and Evercore and aimed initially at work such as investment banking and equity research.
The product combines GPT-6 Astra with financial datasets from providers including Daloopa, PitchBook and LSEG News. Bankers can use it for research, financial models and client materials, with granular citations back to the underlying data.
That is a meaningful shift. General-purpose AI increasingly looks like the foundation; the commercial battle is moving upwards into specialist data, compliance, institutional workflows and industry-specific interfaces.
In other words, being able to reason is becoming table stakes. Knowing which numbers a banker is allowed to trust is the product.
3. 🕵️ AI is becoming an engineering workforce for surveillance
Anthropic released a new threat-intelligence report documenting how state-linked actors and contractors have used Claude in surveillance operations.
One of the more striking cases involved a consultant working for Malian national-security authorities who, according to Anthropic, used Claude to engineer a platform capable of intercepting communications across the country’s mobile operators and generating dossiers. Iranian actors used it to build a malicious browser extension, while Anthropic also documented China-linked intelligence activity.
The important point is not simply that governments are asking AI to analyse information. Anthropic says AI is increasingly being used in place of an engineering workforce — helping relatively small teams construct systems that previously required considerably more specialist labour.
Cheap intelligence does not just lower the cost of doing business. It can lower the cost of state capacity too.
4. 📊 Anthropic modelled a richer economy where workers receive a smaller share
Anthropic also released an economic scenario model exploring what increasingly capable AI could do to the US economy by 2030.
Its three scenarios range from a relatively modest 1.6% uplift in GDP to an extreme scenario in which GDP is 32.4% higher. But the more interesting figure sits underneath the headline growth: in that extreme scenario, labour’s share of income falls to 45.2%, versus 59.4% in the modest scenario.
5. ⚖️ Legal AI now has a $15.5 billion valuation benchmark
Legal AI company Harvey raised $550 million at a $15.5 billion valuation, in a round co-led by Diffusion and Lightspeed Venture Partners.
Harvey says its software is now used by 80% of the Am Law 100 and by in-house legal teams at five Fortune 10 companies.
That valuation says something larger about vertical AI.
The first wave of enterprise AI largely asked companies to bring their work to a general model. Businesses are now paying — and investors are valuing — systems that already understand the workflow, vocabulary, permissions and source material of a profession.
There may be enormous value in the layer between the frontier model and the person actually doing the job.
6. ⚡ Google is buying power the way cloud companies used to buy servers
Google announced at least €13 billion of investment in Finnish AI and digital infrastructure over the next two years, its largest European investment to date.
The plan includes three new data centres in northern Finland and a 22-year agreement for up to 50% of the output of a Finnish nuclear plant. Google and utility Fortum will also explore additional nuclear and renewable-energy projects.
There is something revealing about the time horizon here.
Software companies used to compete over developers, features and cloud regions. AI companies are increasingly making decisions involving national grids, power stations and energy contracts measured in decades.
Compute is becoming an energy business.
7. 🧾 California decided AI companies need independent examiners
California Governor Gavin Newsom signed SB 813 and AB 1405, creating what the state describes as the first US framework for independent verification organisations and a registry for AI auditors.
The laws establish standards around the independence, transparency and integrity of third-party AI assessment.
The political language was unusually plain: one of the bills’ authors said the industry should not be allowed to “grade its own homework.”
That may prove to be the more consequential regulatory idea. As AI moves into finance, infrastructure, government and other consequential systems, self-published safety reports are unlikely to remain sufficient indefinitely.
AI auditing could become its own industry.
8. 💰 Oracle finally put some receipts behind the AI infrastructure boom
Oracle reported $19.3 billion in quarterly revenue, up 30%, while cloud-infrastructure revenue jumped 121% to $7.4 billion.
More strikingly, it booked more than $30 billion of new AI cloud contracts in a single quarter, taking total remaining performance obligations to $664 billion. Oracle says it also delivered more than 300,000 GPUs to AI cloud customers since the end of the previous quarter.
The bill remains enormous. Oracle spent $28.5 billion on capital expenditure during the quarter and still reported negative free cash flow.
But this is an important counterpoint to the “AI infrastructure bubble” discussion: at least some of the extraordinary capital expenditure now has very large customer contracts sitting behind it.
The argument is shifting from “Is anybody buying this compute?” to “Can providers build it economically enough?”
Source: Oracle — Q1 FY27 results
9. 👥 Small businesses using AI appear to be hiring more, not less
New research from payroll platform Gusto produced a useful counterweight to the assumption that AI adoption automatically means fewer employees.
Among businesses in its study, AI adopters had headcounts around 7% higher after one year than comparable firms that knew about AI but had not adopted it. For businesses with fewer than 10 employees, the difference was roughly 10%.
The new jobs were often not AI jobs. Gusto found businesses adding people such as teachers, technicians, cooks, therapists and other hands-on or customer-facing workers.
There is an important caveat: the research is observational and Gusto explicitly says it does not prove AI caused the additional hiring.
Still, it offers an interesting mechanism worth watching. In very small companies, automating admin may not remove the next employee. It may give the owner enough capacity to hire them.
10. 🧠 Amazon’s next AI chip supplier might not be Nvidia
Qualcomm and Amazon announced a multi-generation partnership to develop customised silicon for AWS AI infrastructure, initially focused on inference, alongside optical connectivity capable of reaching 1.6 terabits per second.
The commercial structure is notable. Amazon can purchase up to $60 billion of Qualcomm AI data-centre products under the arrangement, while Qualcomm granted Amazon warrants worth roughly $4 billion that vest alongside purchases.
Qualcomm spent decades being associated primarily with phones. It now expects data-centre revenue to become a major business and is targeting $15 billion annually by 2029.
Nvidia remains extraordinarily dominant, but hyperscalers have every incentive to cultivate alternatives, customise more of their own stack and push down inference economics.
The AI chip race is broadening from who makes the fastest accelerator to who can design the cheapest system around billions of daily AI requests.
So what did this week actually tell us?
The biggest AI stories are increasingly happening outside the models themselves.
There is still plenty of competition around model performance, but the bigger story is becoming what happens once those models are deployed at scale — who pays for the infrastructure, which industries adopt them first, how work changes and where regulation starts to catch up.
And judging by this week, the money — and the consequences — are only getting bigger.
O T H E R__A I _N E W S
OpenAI released ChatGPT Images 2.5, improving image detail, editing precision and consistency across multiple edits while cutting generation latency by up to 50%. It also introduced Sketch and Templates in ChatGPT, plus two new image models for developers.
OpenAI also introduced a new Data agent for ChatGPT Work, designed to connect to company data, investigate business questions and turn the results into interactive dashboards and reports without users needing to write queries.
OpenAI launched the Agents API in public beta, giving developers access to the managed agent harness and infrastructure behind Codex. Agents can work across long-running sessions, use tools and files, and coordinate subagents while running either in OpenAI-hosted sandboxes or external environments.
OpenAI also released GPT-Live-1 through its API, bringing full-duplex voice conversations to developers with improved interruption handling, custom voices and telephony support. The voice layer costs $0.05 per minute and can delegate more complex reasoning to models such as GPT-6 Astra.
Google launched a dedicated Gemini app for Windows, giving users desktop access through an Alt + Space shortcut. The app can work across Google services, help manage tasks and create images and videos without requiring users to leave their current workflow.
Amazon made Amazon Quick generally available on macOS and Windows. Its enterprise AI assistant can work across company systems, draft deliverables, update records and run agents, while Amazon says business data and conversations remain inside the organisation’s controlled AWS environment.
Apple expanded Apple Intelligence into a redesigned Health app, adding personalised health insights and a new Health Age feature that analyses long-term data from Apple Watch and other sources. New vision-based AI assessments can also evaluate movement, strength, balance and flexibility using an iPhone camera.
Mistral raised €3 billion at a valuation of around €21 billion ($24 billion), in what the company called the largest equity funding round by a privately owned European technology company. The French AI company says it is now on track to reach $1 billion in annual recurring revenue by year-end.
Nvidia announced plans for up to 2 gigawatts of new AI data-centre capacity in Australia through partnerships with Firmus, CDC, NEXTDC and AirTrunk. For context, Australia’s current total data-centre computing capacity is estimated at around 1.6 GW.
The U.S. government accused six Chinese AI companies of industrial-scale model distillation, including DeepSeek, Moonshot AI and Alibaba. Officials alleged they used outputs from leading U.S. models to accelerate their own systems; China rejected the claims and described distillation as a widely used, neutral technique.
That's it for this week!
See you soon,
News Digest AI team
