Good day, AI enthusiasts!
Nvidia is pushing further up the AI stack with Nemotron 4, an open-weight model family whose largest version is expected to reach at least one trillion parameters. It comes as more of the AI market shifts towards ownership and control: startups are raising heavily around trainable models, while infrastructure providers are selling capacity before it’s even built. The interesting question is becoming less about which chatbot is best, and more about who controls the models, compute and workflows underneath it.
In today’s digest:
Nvidia is building a trillion-parameter Nemotron 4 model family
River AI raises $1.1bn for personally trainable AI assistants
IBM and Together AI sign a $240m AI inference deal
CoreWeave says its near-term AI capacity is effectively sold out
Anthropic adds invisible watermarks to Claude-generated text
A I__M O D E L S
Nvidia is building a 1-trillion-parameter Nemotron 4 family
News Digest: Nvidia is developing Nemotron 4, a new open-weight model family designed to compete more directly with leading open models from the US and China. The largest version is expected to contain at least one trillion parameters. Training is still under way and Nvidia hasn’t announced a release date, although people working on the project told Reuters it could be ready as early as late autumn.
What changed
The largest Nemotron 4 model is expected to contain at least one trillion parameters.
Final training is still under way, with no confirmed launch date.
Nvidia also unveiled Nemotron 3.5 Lightning and NeMo Switchyard, an open-source model-routing library.
Why it matters: Nvidia increasingly wants a role in the model and software layers, not just the chips underneath them. A competitive open-weight Nemotron family could give businesses another serious deployment option while making Nvidia’s wider infrastructure and software ecosystem harder to avoid.
Source: Read the original Reuters report
A I__A G E N T S
River AI raises $1.1bn to build personally trainable AI assistants
Image source: New Digest AI
News Digest: River AI, founded by xAI co-founder Igor Babuschkin, has raised $1.1 billion across seed and Series A funding. Its first product lets developers fine-tune and reinforce open models through an API. The larger ambition is more interesting: AI agents that organisations or individuals can train around their own needs rather than relying entirely on assistants controlled by another platform.
Key points:
Investors include General Catalyst, AMP, Nvidia, AMD Ventures, Y Combinator and Temasek.
River already supports reinforcement learning and LoRA fine-tuning of open models through its API.
Why it matters: For businesses, the next AI buying decision may increasingly involve ownership as well as model quality. If customised open models become easier to train, companies could keep more of their workflows, data and differentiation under their own control.
Source: TechCrunch
A I__+__I N F R A S T R U C T U R E
IBM and Together AI sign a $240m Nvidia-powered inference deal
Image source: Verdict.co.uk
News Digest: IBM and Together AI have signed a $240 million multi-year agreement to build a large AI inference cluster on IBM Cloud. The initial US deployment will use roughly 2,000 Nvidia Blackwell 300 chips. Together AI expects much of the capacity to be committed before it becomes available, as companies look for infrastructure to run open models without building it themselves.
Key Points:
The initial cluster will use around 2,000 Nvidia Blackwell 300 chips.
Together AI expects capacity to sell out two to three months before availability.
Why it matters: As AI moves into production, inference is becoming a major infrastructure market in its own right. Specialist providers can increasingly compete by giving companies flexible access to open models and scarce GPUs without requiring them to build their own stack.
Source: Reuters
A I__+__I N F R A S T R U C T U R E
CoreWeave says its near-term AI capacity is effectively sold out
Image source: qz.com
News Digest: CoreWeave reported second-quarter revenue of $2.58 billion as demand for AI infrastructure continued to exceed available capacity. CEO Michael Intrator said near-term capacity is effectively sold out, while its revenue backlog reached $104.2 billion. The company also disclosed roughly $25 billion of additional early-Q3 customer commitments that aren’t yet included in that figure.
Key points:
Quarterly revenue more than doubled year on year to $2.58 billion.
Backlog reached $104.2 billion, excluding roughly $25 billion in newer commitments.
Why it matters: There’s still little evidence that demand for top-tier AI compute is cooling. The harder question is financial: whether long-term revenue will justify the debt, data-centre investment and hardware spending needed to satisfy it.
Source: Fortune
A I__+__R E G U L A T I O N
Anthropic says Claude-generated text will carry invisible watermarks
Image source: News Digest AI
News Digest: Anthropic says newer Claude models will add machine-readable watermarking to generated text as it adapts to EU AI Act transparency requirements. The signal is embedded at model level and designed to follow text when it’s copied and pasted. The system will apply across Claude products including its API, Claude Code and Cowork, while generated files will use C2PA.
Key points:
Models released after 2 August will automatically watermark generated text and files.
Anthropic says the text watermark may remain detectable after some editing.
Why it matters: AI provenance is moving closer to being a standard product requirement rather than an optional trust feature. Businesses may gain better ways to identify AI-generated material, although the usefulness of the system will depend on how difficult those signals are to remove.
Source: TechCrunch
U S E__T H I S__T O D A Y
Build a reusable company context file for AI
If you regularly use ChatGPT, Claude or other AI tools for business work, stop explaining your company from scratch every time.
Create one compact context file containing the information AI repeatedly needs to understand your business, then attach or reference it when starting important work.
Help me create a reusable Company Context File that I can give to AI tools before asking them to work on my business.
Here is what I can provide:
Company: [ADD COMPANY]
Website: [ADD WEBSITE]
What we sell: [ADD PRODUCTS/SERVICES]
Target customers: [ADD CUSTOMER TYPES]
Positioning: [ADD POSITIONING]
Main competitors: [ADD COMPETITORS]
Tone of voice: [ADD TONE]
Important terminology: [ADD TERMS]
Things AI should avoid: [ADD RULES OR CONSTRAINTS]
Turn this into a concise reference document covering:
what the company does;
who we serve;
our main customer problems;
positioning and differentiators;
products or services;
brand voice and writing preferences;
important terminology;
competitors and alternatives;
claims or assumptions AI should not make;
any information that should be confirmed before use.
Keep it factual, easy for another AI to understand and short enough to reuse across different tasks.
At the end, identify any important context I haven’t provided and ask me for it.
Best for: Founders, marketers, sales teams, agencies and operations teams using AI regularly.
O T H E R__N E W S
Worth Knowing
Ryanair plans to use Gemini, Google DeepMind models and custom AI agents across areas including decision-making, crew scheduling, disruption management, fleet planning and maintenance. The deal is a useful example of enterprise AI moving beyond employee chatbots and into operational systems.
French media groups have asked the country’s competition authority to intervene over Google’s AI-generated search summaries, arguing they reduce traffic to publishers while using their content. The dispute adds to the growing commercial tension between AI search products and the sites supplying their underlying information.
Researchers at A Security say they discovered and developed an exploit for a serious Zoom vulnerability in a day using fewer than 20 prompts to publicly available AI models. Zoom has patched the flaw, but the case shows how AI can dramatically reduce the cost and expertise needed for vulnerability research.
A I__T O O L S
🛠️ Top AI Tools & App Releases1.
Tines 3B brings agents, apps and automations into one secure environment. A particularly useful detail is how it handles credentials: agents can take actions through integrations without directly seeing the underlying secrets.
Worth watching for: Operations, security and IT teams building agents that need access to real business systems.
Bullet is a coding agent that tries to reduce the time AI spends searching and reasoning before making changes. It routes work between models, uses targeted code search and can execute tasks in parallel.
Worth watching for: Developers and small technical teams using coding agents heavily enough for latency to matter.
bb is a desktop interface for orchestrating coding agents including Claude Code, Codex and OpenCode from one place. Rather than committing teams to a single model provider, it provides a layer for managing different agents and can even extend parts of its own interface through prompts.
Worth watching for: Engineering teams experimenting with several AI coding tools rather than standardising on one.
Vizard Agent takes a broader approach to AI video creation. It can start from footage, a URL, script, image or idea, then handle editing, generation, repurposing and localisation through a conversational workflow.
Worth watching for: Marketing and content teams producing short-form video across several channels or markets.
Kubit launched a product analytics capability designed to connect what an AI agent does with what the user does afterwards. Teams can analyse agent traces alongside funnels and user behaviour, making it easier to see whether higher latency, token use or specific agent actions actually improve outcomes.
Worth watching for: Product and AI teams that need to measure agents by business outcomes rather than model metrics alone.
That's it for today!
See you soon,
News Digest AI team