A house newsletter · daily · researched and written by the Kurzia AI

The story that matters most

Crunchbase News · yesterday

Anthropic IPO

Takeaway: $83 a share

Anthropic priced its IPO at $83 a share on Wednesday, above the $70–$78 range in its filing, valuing it near $1.5 trillion, per Crunchbase News citing CNBC, Reuters and the WSJ; trading starts Thursday on Nasdaq as ANT. The company plans to sell just over 800 million shares for about $66.4 billion, plus $26 billion of non-voting stock to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC — more than $92 billion in total, against SpaceX's $75 billion June debut. Per Crunchbase News, from the filing: $33.2 billion of nine-month revenue, up from $2.5 billion a year earlier, and a net loss that widened to $104 billion.

The first pure frontier lab to list sets the public mark every private AI valuation now gets tested against — good for funds holding Anthropic-adjacent positions, a hard benchmark for anyone raising above it. Anthropic itself has not confirmed the price; expect the Thursday print to move every comparable mark on your books.

Read it at Crunchbase News

Rounds and exits

  • Payhawk — Payhawk, spend management rebuilt around AI agents for finance, raised a $280m Series C led by Goldman Sachs at a $2bn valuation, reported by Tech.eu.

  • MegaRobo — MegaRobo, Suzhou lab automation and AI-for-science infrastructure, raised nearly US$200m joined by Oman Investment Authority; stage not stated.

  • Boltz — Boltz, open-source molecular design models deployed at Pfizer, GSK, Takeda and BMS, raised a $110M Series A with a16z investing.

  • Humyn Labs — Humyn Labs, real-world training data for robotics, raised $47M: $26M equity led by Accel plus $21M venture debt, reported by TechCrunch.

  • Vero Security — Vero Security, runtime policy enforcement for AI tools and agents, raised an $11M seed led by TLV Partners.

  • Unsiloed AI — Unsiloed AI, turning complex enterprise documents into structured data for agents, raised a $7.5M seed led by Felicis, reported by TechCrunch.

  • each::labs — each::labs, generative media workflow platform, raised a $7M seed co-led by Revo Capital and the EBRD, reported by Tech.eu.

  • Linq — Linq, messaging infrastructure connecting AI agents to customers, raised a Series B led by a16z; amount not disclosed.

  • Voice AI — Voice AI startups raised at least $9.5B in 2026 across 154 deals, up 67% on 2025, per Crunchbase data.

What the AI labs and big platforms did

  • Gemini 4 Neon — Google's third Gemini 4 model, built for production coding and long analytical work, shipped Wednesday in the Gemini API, Antigravity and the Gemini app; Google ranks it #3 on the Artificial Analysis Intelligence Index. At $3/$15 per million tokens with a 1M context, it costs 32% less per task than Carbon at its default setting, pushing near-frontier coding into everyday agent workloads — thin coding-agent wrappers get squeezed. Cached input runs $0.15 per million tokens, a 95% discount. Google blog

  • GPT-6.2 Astra and Sol — OpenAI shipped two frontier models Tuesday: Astra at $8 input/$40 output per million tokens and Sol at $2/$10, both with 1M context. Astra scores 62 on the Artificial Analysis Intelligence Index, behind only Anthropic's Fable 5.2. Sol's price flattens the workhorse tier and squeezes wrapper margins. Testers split — Matt Shumer called the pair essentially unusable — and in its preparedness write-up OpenAI gave Astra its top rating for self-improvement, the first such grade on a model it has released. Latent Space

  • MAI Licensed Weights — Alongside the seven-model MAI family announced at Ignite, Microsoft will let trusted customers take a frontier MAI model's weights, fine-tune them and run them inside their own Azure environment. Licensed weights open a self-hosted frontier path for regulated enterprises — pressure on fine-tuning platforms, opportunity for compliance-heavy startups. Microsoft also claims MAI Cyber 2 Flash beats GPT-6 Astra's quality at under 2% of the cost per task on CyberGym-E2E, a claim that would reprice security agents. Microsoft AI

For your portfolio

European Business Magazine · yesterday

Project Exponential

Takeaway: more than 100 AI agents

KPMG UK put more than 100 AI agents to work on a FTSE 250 company's financial statements alongside its regular audit team, the Financial Times reported Wednesday, via European Business Magazine — and in several cases the agents dug deeper and raised questions the human team had not. KPMG plans to widen the trial to as many as 20 larger audits.

This is the Big Four moving agents from helper to first-draft labour in the most junior-heavy workflow in professional services. It hits accounting-automation and audit-support startups from both sides: buyers will expect agent-drafted, traceable workpapers, and the firms may build rather than buy. One agent mis-matched regulatory requirements to company documentation and needed a person to step in; the findings were not used to sign off the accounts. Ask any founder selling into audit or compliance: when KPMG runs agents on 20 of its largest engagements next year, is your product inside that workflow or priced against it?

Read it at European Business Magazine

TechCrunch · yesterday

Nvidia's backstops now evaporate on delay

Takeaway: $105 billion backstop

In quarterly CFO commentary Tuesday, Nvidia spelled out that its data-center loan guarantees — including the $105 billion backstop behind OpenAI's PORTS-Pike campus in Ohio — only take effect once the customer has raised its own financing and installed Nvidia's hardware, and interim CEO Colette Kress told analysts Nvidia will be more selective about which companies it helps finance. Reported by TechCrunch.

The buildout has run partly on Nvidia paper; a pickier Nvidia means portfolio companies signing GPU or neocloud capacity cannot assume vendor money closes the capex behind it. Diligence now runs to who funds the data center, not just who sells the tokens. Question for any founder with a big compute contract on the roadmap: if Nvidia's guarantee lapses the moment the builder slips, who actually holds the obligation to deliver your capacity?

Read it at TechCrunch

“our guarantee doesn't exist anymore”

TechCrunch · today

Cowork

Nadella explained at the DealBook Summit on Wednesday why he offered buyouts to 60,000 employees in October: Microsoft is reorganising so that people manage AI rather than do the work directly, and he framed the offer as giving staff a choice rather than a layoff. The team building Cowork, the AI agent inside Microsoft 365 apps, is 11 developers whose job is running the AI that builds the product.

Those are the two clearest public numbers yet on agent-first engineering: 60,000 buyouts at a 51-year-old incumbent, and an 11-person team shipping a core product. Nadella also conceded the transformation will take longer than anyone expects. For board work, a team still hiring like it is 2023 is the tell. Ask each founder: what is revenue per engineer this quarter versus a year ago — and are the open reqs for doers or for agent supervisors?

Read it at TechCrunch

Share this issue

How this issue was made. Researched and written by an agent, produced on 19 November 2026, 05:35. 94 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us