Good morning from the wAIve team in London.

Today, Wednesday 4 November 2026: A day for drawing lines and checking sources. Nvidia has told the cloud companies that rent out its chips to keep Chinese business to a minority, and Canva is cutting or moving more than 1,200 jobs. Meanwhile, a Crown Prosecution Service lawyer cited a case in the High Court that turned out not to exist.

  BREAKING  

Nvidia has instructed cloud computing providers that the majority of their business must come from non-Chinese customers, according to a report in the Financial Times. The directive, issued under pressure from Washington, means companies selling cloud access to Nvidia's chips face a stark choice: reduce their Chinese client base or risk losing access to the most sought-after AI hardware on the market.

THE PLAY: This is a supply chain shock masquerading as a policy memo. Any business relying on cloud infrastructure where the provider has significant Chinese exposure now faces potential service disruption.

  INDUSTRY  

Canva is cutting or relocating more than 1,200 jobs, representing 22 out of every 100 employees globally. CEO Melanie Perkins described the decision as "without a doubt the hardest moment in our thirteen-year history." The design platform cited pressure from cheaper AI-native competitors and the need to embed its own AI tools more aggressively.

THE PLAY: Canva is a profitable, privately held company worth tens of billions. That makes this restructuring more revealing, not less.

  LAW  

A senior Crown Prosecution Service barrister cited a wholly fictitious legal precedent during a High Court hearing involving a child rape case, in what the court described as "a serious failure." The imaginary case was traced to an AI-assisted search and adds to a growing catalogue of hallucination incidents affecting legal proceedings in Britain. The CPS has not yet disclosed which tool was used.

THE PLAY: The legal profession's exposure here is a warning for every professional services sector. Lawyers, consultants, and analysts who use AI for research without verification are open to liability and compliance issues.

AI inference startup Fal has raised $1.3 billion in a round led by Bessemer Venture Partners, doubling its valuation to $16 billion. The five-year-old company sits in the infrastructure layer, running compute that powers AI applications.

Why it matters: Fal's rapid rise reflects how much value is accruing in the unglamorous middle of the AI stack.

Palantir reported that its US commercial revenue exceeded its US government revenue for the first time in the company's history. Palantir has been pushing its AI data platforms into enterprise markets aggressively over the past two years.

Why it matters: If a company built on government secrecy is now growing faster in commercial markets, the enterprise AI platform race is moving faster than most tech buyers have planned for.

The owner of Britain's most polluting power plant has struck a £20 billion deal with infrastructure investor Brookfield and AI platform Radiant to convert the Drax site in North Yorkshire into a datacentre with over one gigawatt of capacity, potentially operational by early 2028.

The regional angle: A one-gigawatt facility would significantly expand domestic AI compute capacity and reduce reliance on transatlantic data routing. It also signals that the UK is now actively competing for the infrastructure investment that has been flowing predominantly to the United States and Gulf states.

Closing prices, Tuesday 3 November (US markets). Not investment advice.

Napkin AI converts written text into professional diagrams, charts, and visual frameworks instantly, without requiring any design or technical skills.

THE PLAY: Takes under five minutes and significantly improves comprehension for visual learners around the table.

Fathom records, transcribes, and summarises meetings in real time, then automatically drafts follow-up notes and can push action items directly to CRM tools including Salesforce and HubSpot.

THE PLAY: Yet another note-taking app in this overcrowded market, but worth trying out if you're not settled on which tool to use yet.

Tuesday's revelation that a Crown Prosecution Service barrister cited a wholly invented legal case in a High Court hearing is a watershed moment. Not because AI errors are new, but because this one landed in a formal judicial proceeding involving a child, in a country whose legal system depends on accurate precedent. It joins a lengthening list of professional failures linked to unverified AI output: fabricated academic citations, invented regulatory references, and now a fictitious case name placed before a senior judge.

The instinct is to treat these as individual failures of judgement by people who should have known better. Instead, they're the symptom of a structural gap between tools that are genuinely useful the vast majority of the time, and contexts where even a small error rate is unacceptable.

Stanford researchers published research on Tuesday that should give any senior leader pause. They tested fifty widely cited claims about large language model capabilities against twenty current models. Nearly half of those findings were already obsolete. A further third had weakened significantly. The average lifespan of a published claim about what AI can or cannot do: fifteen months and falling. Claims about what models cannot do expire fastest of all.

The implication is uncomfortable.

AI tools are being deployed faster than the institutional checks. This is visible in legal proceedings. It is likely to be equally visible, if less publicised, in financial analysis, regulatory submissions, and customer-facing communications.

The error rate of any given AI system changes as models are updated, as the tools your suppliers use change, and as the contexts in which AI is deployed expand into areas it was not originally designed for. A verification process that worked adequately six months ago may not be adequate today.

Takeaway: Reliability governance is now a board-level risk issue. Every professional function that has adopted AI for research, drafting, or analysis needs a verification layer: not because AI is unreliable in general, but because the contexts where failures cause most harm are precisely those where checking feels redundant.

Most executives prompt AI as a generalist assistant and receive generalist answers. A sharper technique is to assign the model a specific expert identity before asking your question. This constrains the model's framing and dramatically improves the precision and relevance of its response.

By defining a specialism and presenting a concrete scenario, you force the model to filter its response through a relevant professional lens rather than producing a generic overview. It works because large language models have absorbed enormous volumes of domain-specific professional writing, and assigning a named expert role unlocks that knowledge far more precisely than an open-ended question does.

Thanks for reading, the wAIve Daily Team. Found this useful? Forward it to someone who needs it. Tell us what you need to know. Reply with your questions.

wAIve Daily is published with Centai Strategy