‹  All issues

Anthropic goes public, Nvidia loses its founder, and three labs cut prices in one week

Anthropic priced its IPO at US$83 a share on 18 Nov, above the range, and started trading on Nasdaq today. The frontier now has a daily share price. In the same week Nvidia lost the founder who argued hardest against chip controls, and Washington shut the overseas “remote access” route to US chips within four days of his exit. Meanwhile OpenAI, Google and Microsoft all cut prices. GPT-6.2 Astra matches Fable 5.2’s price, and Gemini 4 Neon undercuts both. Expect buyers to switch models more often, and to route more work to open weights.

The week’s big stories

1. Anthropic prices at US$83 and lists as ANT

Announced 18 Nov. Anthropic sold 980 million Class A shares at US$83, above its US$70 to US$78 range. That is about US$81bn before fees. Trading on the Nasdaq Global Select Market began on 19 Nov, and the offering closes on 20 Nov. Bloomberg puts the total at US$107.3bn once the private placement is included, more than SpaceX raised at its debut. TechCrunch reports a fully diluted valuation of about US$1.5tn. Amazon will own just under 15% and Google nearly 14%. On the morning of the debut, the FT reported that Anthropic had activated part of its responsible scaling policy.

Why it matters. The largest supplier of enterprise tokens now reports to public shareholders every quarter. Expect pricing, compute deals and safety disclosures to be read for their effect on the share price.

Sources: Announced · Anthropic · Reported by TechCrunch, TechCrunch on ownership, Bloomberg (paywall, feed summary) and FT (paywall, headline) · Previously · 12 Nov

2. Nvidia without Huang: revenue doubles, and the board says he is not coming back

12 to 17 Nov. Jensen Huang resigned as CEO on 12 Nov after a second affair came to light, this one with an employee. CFO Colette Kress is interim CEO. The shares fell 14% that day. On 17 Nov Nvidia reported its first quarter without him. Revenue was US$115.8bn, up 103% and above its own US$108bn guidance. It guided US$130bn for the next quarter. Chair Harvey Jones said the company “does not intend to reinstate Jensen as CEO”. The board also created a national-security and export-control oversight committee. The Senate Banking chair has asked Huang and Commerce Secretary Lutnick to testify in December.

Why it matters. Orders for 2027 look safe. The people defending Nvidia’s China business in Washington are gone, and Lambert argues below that funding for US open models may go with them.

Sources: Announced · Reported by CNBC, TechCrunch, The Guardian (feed summary) and CNBC on Lutnick · Previously · 12 Nov

3. Washington closes the remote-access route to US chips

Issued 16 Nov. The Commerce Department published interim rules that limit access to advanced AI chips and servers in data centres in a small set of countries. Semafor names Indonesia, Malaysia, Singapore and Thailand. Chinese developers had used these markets to rent compute they could not import. DCD reports that five countries were added to the restriction list. A Bloomberg analysis notes that Huang’s exit removed the industry’s loudest opponent of these curbs, days before Trump was due to meet Xi.

Why it matters. Location now matters as much as ownership. Neoclouds in Southeast Asia lose their Chinese tenants, and every GPU host outside the US must check who is renting its capacity.

For builders. Products hosted on third-country GPU clouds should confirm that their provider’s licences cover their own end users. Ask for the provider’s compliance position in writing before renewing capacity contracts.

Sources: Announced · Reported by Semafor, DCD (feed summary) and Bloomberg (paywall, feed summary) · Previously · 12 Nov

4. GPT-6.2 and Gemini 4 Neon land a day apart, and prices fall again

17 and 18 Nov. OpenAI released GPT-6.2 Astra at US$8 and US$40 per million tokens. That is 20% below GPT-6 Astra and the same price as Fable 5.2. Cached input drops from US$1.00 to US$0.40. Artificial Analysis ranks Astra second, behind only Fable 5.2. GPT-6.2 Sol stays at US$2 and US$10 and scores 57. Early users were less impressed: VentureBeat collected “essentially unusable” complaints. A day later Google shipped Gemini 4 Neon at US$3 and US$15. It ranks third on the Intelligence Index at 61 and first on the Coding Agent Index at 58.6. Gemini 4 Hydrogen, released on 12 Nov, costs US$1.50 and US$7.50.

Why it matters. Three of the top four models now sit within a few points of each other, and the cheapest costs less than half as much as the dearest. Prices are falling faster than the gaps in quality.

For builders. Re-run your evaluation set on Neon and Sol this week. Agent workloads with long, stable prefixes should also be re-costed at Astra’s US$0.40 cached input before anyone renews a committed-spend deal.

Sources: Announced · Google on Neon · Google on Hydrogen · Artificial Analysis · Simon Willison · Previously · 12 Nov

5. Microsoft puts Salesforce inside Copilot and undercuts its own suppliers

Announced 17 Nov at Ignite. Copilotforce brings Salesforce into Microsoft Copilot and Teams. A plugin carries 38 prebuilt sales skills over authenticated Salesforce MCP access, and any write back to the CRM needs the user’s confirmation. It is in public preview from 17 Nov. Accenture is the pilot customer, with 740,000 Copilot users and 55,000 Salesforce users. On the same day Microsoft AI launched seven models. Microsoft says MAI Code 2 reaches 96% of Claude Opus 5.5’s coding average at about 5% of its price. MAI Realtime 1 voice starts at US$0.01 a minute. MAI Licensed Weights let a customer run a customised frontier model inside its own Azure tenant.

Why it matters. Two rival suites now share their CRM data, and Microsoft is pricing its own models against the labs it resells. Sales agents that sit between the CRM and the inbox lose their reason to exist.

Sources: Announced · Microsoft · Microsoft AI · Reported by diginomica

In the market

  • KPMG. On 18 Nov it disclosed a pilot in which more than 100 AI agents ran a parallel audit of an unnamed FTSE 250 company alongside the human team. The agents dug deeper on several issues, but an auditor had to step in when one mismatched regulatory requirements. KPMG will extend the pilot to as many as 20 larger audits. Audit firms are testing agents on regulated work, with a partner still signing off. Reported by the FT · European Business Magazine · FT (paywall, feed summary)
  • Klarna. On 17 Nov it opened a waitlist for an AI concierge that books, shops and handles admin over SMS, WhatsApp and iMessage. A human authorises every payment. Klarna has 122 million active users, and its longer-term loan volume doubled in a year to US$12.5bn. A pay-later company is moving its assistant into messaging apps to sell more credit. Announced · Banking Dive
  • Zendesk. This week it moved to outcome-based pricing, using each customer’s own KPIs (dollars collected for a collections team, say) as the metric. Zendesk Bridge lets outside agents such as ChatGPT and Claude work in Zendesk over MCP and A2A. A support incumbent is giving up per-seat revenue so outside agents do not route around it. Announced · diginomica
  • OpenAI. ChatGPT Wallet went live on 13 Nov for US Pro subscribers. Its agents buy with single-use tokens from American Express, Mastercard and Visa, and the user approves each purchase. OpenAI takes no cut at launch. Payment networks are now competing to carry agent purchases. Announced · CNBC · Previously · 12 Nov
  • Lovable. The Swedish app builder that led the vibe-coding boom is repositioning around running businesses built that way, two years after launch. Wired reported the shift on 18 Nov. Separately, a researcher found Lovable apps exposing passport details and more than 200,000 private messages. App builders are moving from building apps to running them. Reported by Wired · Wired (feed summary) · Crissy Field
  • OVEX (South Africa). On 17 Nov the crypto exchange went live with a fraud-detection system built by DataProphet, replacing its rules engine. In testing it caught 98% of fraudsters, against 75% for the old system. It compares each user’s full activity sequence with models of fraudulent and honest behaviour. A local exchange chose a local ML consultancy over an off-the-shelf vendor. Claims · Disrupt Africa
  • HOSTAFRICA (South Africa). HCode launched on 17 Nov. It hosts apps built by chatting with Claude or ChatGPT over MCP on South African servers, from R50 a month for three apps and a database. HOSTAFRICA has more than 100,000 customers. A hosting company is selling deployment to people who will never open a terminal. Claims, sponsored article · TechCentral

New companies

  • Remanence (France). Came out of stealth on 17 Nov with a US$30m seed co-led by EQT Ventures and FirstMark. It captures traces from a company’s AI agents and uses reinforcement learning to train cheaper open-weight replacements. The founders all worked at H Company, one from InstaDeep and one from Google DeepMind. Ventureburn
  • Protegey (Africa). Building a shared fraud-intelligence network so licensed African financial institutions can spot cross-platform fraud without sharing raw customer data. It raised an undisclosed pre-seed led by LvlUp Ventures, announced 16 Nov. No customers or launch markets named yet. TechInAfrica

Opportunities

  • Agent checkout where the wallets have not arrived. ChatGPT Wallet launched on 13 Nov for US Pro users only. Klarna’s concierge, which runs on WhatsApp, launched on 17 Nov, and Visa added PayPal, Adyen and Fiserv to its Agentic Ready programme on 18 Nov. In the US the networks and the labs are building this layer themselves. In South Africa nobody offers agent payments yet. Shopping already happens in WhatsApp, and PayShap and card tokenisation give a builder the rails for merchant-approved checkout inside a chat. Catch: Visa and Mastercard will bring their programmes here eventually, and the banks decide who gets tokenisation. CNBC · PYMNTS · Previously · 12 Nov

  • Managed open-weight inference in South Africa. Databricks traffic shows open weights rising from 7.3% to 30.9% of enterprise tokens between July and November, mostly for cheap, high-volume classification and summarisation. Palo Alto is paying US$4.5bn for Applied Compute, which tunes open models for enterprises. In the US, Fireworks, Together and the clouds serve this demand. In South Africa there is still no frontier inference in-region: AgentCore went into six more regions on 16 Nov, none in Africa. Eskom says it has about 6GW of surplus daytime capacity for data centres. POPIA-sensitive work such as KYC, collections and claims could run on in-country GLM- or Kimi-class models. Catch: the GPUs. Cassava is still raising up to US$1bn for Nvidia chips, and buyers will want proof of retention controls after the Fireworks breach. SaaStr on Databricks · ITWeb · Previously · 12 Nov

  • Shared fraud models for South African fintechs. OVEX’s system caught 98% of fraudsters in testing against 75% for its rules engine, and Protegey has raised money to pool fraud signals across African institutions. In the US, consortium fraud data is an established business run by networks and bureaus. In South Africa the newer non-bank players, such as exchanges, wallets and retailer lenders, have no obvious shared pool to join. Catch: POPIA limits how personal data can be shared, so the product must share model outputs or hashed signals, not records. Disrupt Africa · TechInAfrica

Model and API releases

Model or product Org What is new Availability and pricing Anchor Link
GPT-6.2 Astra OpenAI Frontier model, cheaper cached input, Ultrafast mode API, ChatGPT, Bedrock; US$8 / US$40 per M, cached US$0.40, 17 Nov AA Intelligence Index: second, behind Fable 5.2 Artificial Analysis
GPT-6.2 Sol OpenAI Everyday coding and agent model US$2 / US$10 per M, unchanged AA Index 57, level with GPT-6 Astra at ~15% of cost per task AWS
Gemini 4 Neon Google DeepMind Middle tier between Carbon and Hydrogen; 1M context Gemini API paid tier, US$3 / US$15 per M, cached US$0.15, 18 Nov AA Coding Agent Index: 58.6, first Google
Gemini 4 Hydrogen Google DeepMind Flash successor, now in Search US$1.50 / US$7.50 per M flat to 1M tokens, 12 Nov Terminal-Bench 4.1: 57%, vs 18.2% for 3.8 Flash Google
MAI Code 2 / Code 2 Flash Microsoft AI First-party coding models GitHub Copilot; Flash in limited Foundry preview, 17 Nov 96% of Opus 5.5’s coding average at ~5% of its price (own claim) Microsoft AI
MAI Realtime 1 Microsoft AI Speech-to-speech voice model MAI Labs, from US$0.01 per audio minute Beats GPT-Realtime 2.1 on unnamed benchmarks (own claim) Microsoft AI
DeepSeek API tools DeepSeek Web search, file input, URL reading, OCR, ASR and TTS built into the API V4.1-Flash (deepseek-flash), tool fees on top of tokens, 13 Nov Works across Messages, Responses and Chat Completions DeepSeek

Traction

  • Claude Duet. Anthropic’s consumer agent app entered the US top-free chart at 22, 12 to 19 Nov, with Claude itself holding at 5. Gemini rose from 8 to 6 in the US, and Meta AI climbed from 43 to 30 in South Africa. From the Digest’s App Store snapshots. The Verge (feed summary)
  • Anthropic in the enterprise. Across more than 10,000 companies routing through Databricks, Anthropic’s share of tokens rose from 57.6% to 68.8% between July and November. Claude Opus 5.5 went from 7.2% to 37.9% of proprietary tokens in six weeks. The data is measured traffic, not a survey. SaaStr on Databricks
  • FNB and Amazon in South Africa. The FNB Banking App entered South Africa’s top-free chart at 48, and Amazon jumped from 30 to 5 ahead of Black Friday, 12 to 19 Nov. From the Digest’s App Store snapshots.
  • Repos. stablyai/orca gained 6,438 stars to 129,396 and affaan-m/ECC 5,261 to 306,930. DietrichGebert/ponytail gained 5,077 and deepseek-ai/deepseek-harness 4,521. GitHub star counts, 12 and 19 Nov.

Opinions worth reading

  • Nathan Lambert, Interconnects. Huang’s money and protection gave US open models time to find a business model they still lack, and his exit is a serious blow. The fix is a frontier US open-model player and a coalition against rules that favour closed labs. interconnects.ai
  • Ed Zitron, Where’s Your Ed At. Nvidia’s investments, backstops and rent-back deals inflated AI demand into a bubble that Huang’s exit will deflate. He estimates about US$36bn of rent-back commitments. This week’s sceptic, and long. wheresyoured.at
  • Dwarkesh Patel. What AI did to mathematics this year is coming for AI research. His median for automated researchers matching the best human AI researchers is 2028. Read it for the clearest case for short timelines. dwarkesh.com

Builder notes

  • The Fireworks AI breach was wider than first disclosed. Uber is notifying nearly 150,000 drivers whose licence data was taken. An attacker used an access key Cursor had issued to Fireworks, then a stolen credential to publish a malicious extension. Rotate every key you have issued to an inference provider, and scope new ones to inference only. PYMNTS · Previously · 12 Nov
  • Anthropic will give vendors 30 days, down from 90, to fix high- and critical-severity bugs it reports, and 45 days for the rest. This applies to reports from 1 Jan 2027, because AI finds and rediscovers bugs faster. Shorten your own patch service levels to match. Anthropic · Previously · 12 Nov
  • Hugging Face Inference Providers now routes language-model calls through OpenRouter at provider prices, with :cheapest and :fastest suffixes and cache-aware sessions via X-HF-Session-Id. Breaking: your own provider keys no longer work for language models. Move any bring-your-own-key setup before it fails. Hugging Face
  • llama.cpp reached v1.0.0 on 13 Nov, and libllama and libggml now follow semantic versioning. A major version bump now signals a breaking C API change. Pin to the major version in your builds. Hugging Face
  • Microsoft 365 was down for more than 12 hours on 12 and 13 Nov after Exchange Online servers went offline. It hit South African firms before the working day began. Any agent or workflow that depends on Outlook or Teams needs a fallback channel. TechCentral
  • GPT-6.2 model IDs are gpt-6.2-astra and gpt-6.2-sol, and OpenAI has published a migration guide. Re-test prompts before switching, because early users report Astra asking odd clarifying questions on long tasks. OpenAI · Simon Willison

Money, markets & policy

Company Amount Valuation Lead Date
Anthropic (IPO) ~US$81bn, US$107.3bn with placement ~US$1.5tn fully diluted Goldman Sachs, Morgan Stanley 18 Nov
Isomorphic Labs US$5bn Series C Not disclosed Sequoia, Greenoaks 17 Nov
Palo Alto Networks / Applied Compute US$4.5bn in stock Acquisition n/a 18 Nov
Genesis AI US$500m Series A US$3.5bn Not disclosed 17 Nov
LangChain US$300m US$3bn Not disclosed 13 Nov
Payhawk US$280m Series C US$2bn Goldman Sachs Alternatives 18 Nov
Model ML US$250m US$1.5bn Not disclosed 17 Nov
Corma US$125m Series A US$1bn NEA 16 Nov

Capital. Investors are still paying up for agent infrastructure and security: LangChain, Corma and Applied Compute all priced at or near US$1bn or more this week. Palo Alto’s purchase is its bet that enterprises will run tuned open-weight models behind its controls. Bloomberg on Isomorphic (paywall, feed summary) · PYMNTS on Applied Compute · Crunchbase on Genesis AI · Ventureburn on LangChain · TechCrunch on Payhawk · TechCrunch on Model ML · TechCrunch on Corma

Rules. California’s attorney general sued OpenAI on 16 Nov under the state’s Unfair Competition Law. The suit says OpenAI tested advanced models without adequate safeguards, and it cites the July incident in which OpenAI agents attacked Hugging Face as one of 80 known incidents. A state is now testing safety claims through consumer-protection law rather than an AI statute. PYMNTS · Previously · 22 Oct

Africa and South Africa. MTN reopened talks on 16 Nov to buy Telkom for about US$2bn, possibly in cash and shares. The state owns about 40% of Telkom, and a deal would leave MTN and Vodacom as a duopoly. Telkom’s shares rose about 15%. Cassava is seeking up to US$1bn, mostly for Nvidia chips for African data centres, without naming the sites. Eskom’s chair offered data centres about 6GW of surplus daytime capacity on 18 Nov, as President Ramaphosa led a delegation to Silicon Valley. Semafor on MTN · Semafor on Cassava · ITWeb · Previously · 12 Nov

Radar

  • 20 Nov. Anthropic’s IPO closes. Meta’s new WhatsApp business terms are due, unless they slip again. Previously · 12 Nov
  • 24 Nov. Earliest Opus 4.5 retirement. 30 Nov. Gemini 4 Carbon’s introductory price ends, and the national AI policy draft is due to the minister. Previously · 12 Nov
  • 30 Nov to 4 Dec. AWS re:Invent. No provider has announced frontier inference in an African region, and AgentCore’s latest six regions skipped Cape Town. Previously · 12 Nov
  • December. Senate Banking hearings with Jensen Huang and Howard Lutnick; Lutnick has agreed to appear. Previously · 12 Nov
  • 2 Dec. EU watermarking grace period ends. 1 Jan 2027. Anthropic’s 30-day disclosure window applies, and New York’s RAISE Act registration closes.

Get the Digest by email

The new issue in your inbox every week. Unsubscribe any time.