Europe unveils a trillion-parameter AI rival

Good day, AI enthusiasts!

It was an interesting week in AI, particularly when it came to the money behind the industry.

OpenAI's reported annualised revenue figure for September came in at around $50 billion, rather than the near-$70 billion previously signalled. Meanwhile, Nvidia-backed data-centre company Firmus abandoned a planned $5 billion Australian IPO after investors questioned its valuation and expansion plans.

There was plenty of investment activity, too. The White House announced $2.4 billion in technology-company commitments for AI-enabled scientific research, while Mistral unveiled a trillion-parameter model developed in Europe.

And on the product side, ChatGPT introduced interactive interfaces, Google announced a new workplace agent, and Microsoft made a bigger push into AI running directly on PCs.

Here are 10 AI developments worth knowing from this week:

Reuters reported on Thursday that OpenAI had told investors its annualised September revenue was approaching $50 billion, compared with an earlier indication that it was approaching $70 billion.

There is an important distinction here. According to Reuters, much of the difference comes down to how revenue is counted, particularly sales through cloud partners. It does not necessarily represent a $20 billion collapse in underlying business performance.

The episode illustrates a problem investors will increasingly face as large AI companies prepare for public listings: comparing businesses whose revenue definitions, cloud arrangements and costs are not necessarily equivalent.

Annualised revenue is also a run-rate calculation, rather than money actually earned over a full year. With numbers this large, the accounting details deserve just as much attention as the growth rates.

Source: Reuters

Australia's Firmus, backed by Nvidia and several major investors, shelved its planned $5 billion initial public offering on Friday after the deal failed to attract sufficient investor confidence.

The proposed listing valued Firmus at approximately $30.6 billion, nearly three times its valuation following an August funding round. Investors were concerned about debt, the scale of its expansion plans and reports that a major development partnership was no longer proceeding.

Firmus intends to pursue private fundraising and may consider another public-market route.

The AI infrastructure industry still needs enormous amounts of capital, but the willingness to fund every ambitious expansion plan at an equally ambitious valuation is clearly being tested.

Source: Reuters

The White House announced $2.4 billion in industry commitments for the Genesis Mission, an initiative designed to use AI and advanced computing to accelerate scientific research across US federal agencies.

According to the administration, Nvidia committed $1 billion, AMD $500 million, OpenAI $200 million, and Google and Anthropic $150 million each. The commitments include computing resources and tools, so they should not be interpreted as $2.4 billion in new cash funding.

The programme is intended to support research in areas including energy, health and space.

For the companies involved, scientific computing is becoming an increasingly important application for powerful AI systems. For researchers, the potential benefit is access to computing resources that would otherwise be difficult or expensive to obtain.

Source: White House fact sheet

French AI company Mistral introduced Mistral Large 4, a model with one trillion parameters, placing it among the largest open-weight AI systems announced to date.

The model entered public API preview on Tuesday, with the company planning to release its weights later in October. Mistral says it performs competitively across coding, multimodal tasks and specialist enterprise applications, including cybersecurity.

The model was trained using 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centres.

Its significance extends beyond model benchmarks. European businesses and public institutions increasingly want strong AI capabilities without relying entirely on US or Chinese providers. Mistral is offering a technically ambitious alternative, although its performance claims will need continued independent scrutiny.

Source: Mistral

OpenAI introduced Intelligent UI alongside a broader rollout of GPT-6 in ChatGPT this week.

Instead of relying exclusively on text responses, ChatGPT can now generate interactive charts, comparisons, calculators, visual explanations and other interfaces based on what someone is trying to accomplish.

A financial question might become a calculator. A complicated explanation might become an interactive diagram. A comparison might appear as a structured interface rather than several paragraphs of text.

OpenAI says the new experience is rolling out globally across paid and free tiers, with timing varying by plan.

This is an interesting product development because it changes how people interact with the information an AI system produces. For software businesses, it also raises practical questions about which everyday tools people will continue to open separately.

Source: OpenAI

The US government, Meta, Google DeepMind and Isomorphic Labs are supporting Biohub's Virtual Biology Initiative, bringing total commitments associated with the effort to approximately $1.8 billion.

The project aims to create large, standardised biological datasets that AI systems can use to predict how cells respond to different conditions.

There is an important funding distinction: the total includes commitments made before this week, as well as newly announced support. It is not $1.8 billion of entirely new investment.

The scientific ambition is substantial. Drug discovery and biological research often require expensive experiments and lengthy testing. Better predictive models could help researchers decide which experiments are most promising, although the datasets and models will take time to develop and validate.

Source: Reuters

Google Cloud announced a new Gemini agent for work capable of answering questions, carrying out tasks, creating content and writing code across business applications.

It connects with Google Workspace, Microsoft 365 and Slack, and Google says it can select between Gemini and Anthropic's Claude models depending on the task.

One particularly interesting feature is the ability to create specialised coworker agents with their own email addresses and access permissions. Google also announced industry-focused versions, including previews for financial services and legal work.

For businesses, the practical issue will be less about whether an agent can demonstrate a useful task and more about how reliably it operates across company systems, what information it can access and how its work is supervised.

Source: Reuters

Microsoft unveiled an AI coding model designed to run directly on personal computers, alongside new security technology intended to restrict what autonomous agents can access and do.

The company also presented a high-powered Surface Laptop Ultra using Nvidia technology.

Running more AI locally could reduce dependence on cloud processing for certain workloads, potentially improving privacy, responsiveness and control over sensitive business information.

There is an obvious commercial complication: powerful local AI requires capable hardware, and premium computers are not cheap. Microsoft's challenge is to make the benefits compelling enough to justify the equipment costs.

Source: Reuters

Anthropic released Claude Haiku 5.5, its latest small model designed for high-volume applications where speed and operating costs matter.

According to Anthropic, Haiku 5.5 costs approximately 75% less to run on average than Haiku 4.5. The company also reduced the price of Sonnet 5.5 cache reads, estimating that this will make many agentic workloads around 20% cheaper.

These are company-reported comparisons, and actual savings will depend on the application.

For businesses deploying AI at scale, however, such changes matter. A model used occasionally by an individual has a very different cost profile from one handling thousands of customer interactions, database queries or repetitive business processes each day.

Source: Anthropic

A newly published AP-NORC poll found that 64% of US adults believe AI is developing too quickly, compared with 27% who think the pace is about right and 8% who believe it is moving too slowly.

Around eight in ten respondents also said it was very or extremely important for government to ensure that AI remains under human control.

The survey was conducted in late September and published this week. It measures public opinion, rather than demonstrating whether AI is objectively advancing too quickly.

Still, the results are worth paying attention to. Technology companies are investing heavily in making AI more capable and widely available, while significant parts of the public remain concerned about oversight, safety and control. That gap is likely to influence how AI products are received and how governments respond.

Source: Associated Press

Overall, it was a week of impressive technical ambition, serious capital commitments and some increasingly practical questions about value, reliability and trust.

O T H E R__N E W S

Arena raised $200 million at a $3.1 billion valuation, as the company behind a widely followed AI model leaderboard expands its commercial evaluation business.

Manus secured more than $500 million in funding, with parent company Butterfly Effect announcing its first financing round since its proposed acquisition by Meta collapsed. The company did not disclose its new valuation.

Anthropic published updated usage policies, clarifying restrictions on deceptive campaigns, election interference, surveillance and high-risk autonomous actions. The revised policy takes effect on 12 November.

Amazon introduced three new Alexa+ tablets, starting at $229.99, with built-in AI assistance, Android and access to the Google Play Store. Shipping begins on 14 October in North America and 19 October in several European markets.

Liquid AI released two open-weight decision models, including a 3-billion-parameter model designed for fast classification and decision tasks. The company reports response times as low as eight milliseconds on an Nvidia RTX 4090 in its tests.

Google released AI Edge Foresight for Mac, an experimental AI meeting assistant that can work entirely offline, using on-device models to process notes, transcripts and documents.

China issued new guidelines for AI and advanced technology investment, calling for stronger risk monitoring, reliable AI systems and greater accountability for reckless investment in emerging industries.

OpenAI published details of two AI-assisted influence operations linked to Russia and Iran, involving fabricated institutional identities and journalist personas. The company said it had banned the operations.

See you soon,

News Digest AI team