AI Bulletin · week of June 1

Recap of the week of June 1–5, 2026 in AI. Six news blocks and a framework on the economics of tokens that circulated that week.

1. Anthropic files its confidential IPO amid a review of AI spending

Anthropic has confidentially filed a draft S-1 registration statement with the US SEC for a possible initial public offering. The filing sets no price or share count and remains subject to regulatory review and market conditions; several reports place it on a path to go public this fall.

The filing comes as corporate scrutiny of the cost of AI grows. According to a survey cited that week, 40% of companies are seeing savings below 10%, a figure the sector reads as a risk to model providers’ revenue if customers shift toward cheaper or open-weight options.

In parallel, Anthropic has expanded its enterprise partner program with a Claude Partner Network, a Services Track and a Partner Hub that formalise a $100 million program, and has extended Project Glasswing to 150 additional organisations across more than fifteen countries. Since launch, Glasswing partners —among them Apple, Nvidia, Microsoft, CrowdStrike and Palo Alto Networks— have found more than 10,000 high- or critical-severity security flaws. The company has also stated that 80% of its new production code is now written by Claude, with an eightfold increase in code volume per engineer.

Sources: Anthropic — “A confidential draft S-1”; CNBC — “Anthropic expands Project Glasswing to 150 additional organizations”, 2 June 2026; VentureBeat — “Anthropic says 80% of its new production code is now authored by Claude”; report via TLDR AI, 3 June 2026 — “Anthropic faces AI spending backlash before IPO”.

The bulletin mascot at a newsroom desk reviewing a freshly printed financial prospectus spooling out of a machine, with stamps and blank folders on the oak desk — Anthropic files its confidential S-1 while expanding Glasswing and its partner network and reporting that 80% of its production code is written by Claude.

2. Open-weight models: MiniMax M3, Nemotron 3 Ultra, Gemma 4 12B and Microsoft’s seven MAI models

The week concentrated several open-model launches. MiniMax has introduced M3, an open-weight model with an attention architecture that scales context to 1 million tokens and a guaranteed minimum of 512,000 tokens via the API. The company has announced it will release the weights and a technical report within the following ten days, at a standard price of $0.60 per million input tokens and $2.40 per million output tokens up to 512,000 tokens.

Nvidia has launched Nemotron 3 Ultra, an open-weight model with 550 billion parameters, 55 billion of them active, which scores 48 on the Artificial Analysis Intelligence Index against 39 for the next strongest, Gemma 4 31B, and serves more than 300 tokens per second on a Deep Infra endpoint.

Google has published Gemma 4 12B, which can run on many consumer laptops with 16 GB of memory at a capability close to that of Gemma 4 26B MoE, downloadable on Kaggle and Hugging Face. Microsoft, for its part, has launched seven new MAI models that let developers tune their weights through a “Frontier Tuning” method, and has announced a collaboration with Mayo Clinic for a health model that will be deployed first inside the clinic before wider distribution through Azure Foundry.

Sources: Implicator.ai — “MiniMax promises M3 weights after 1M-context model launch”; Ars Technica — “Google’s new Gemma 4 open AI model is sized for your laptop”; Microsoft AI — “Building a hill-climbing machine: launching seven new MAI models”; report via TLDR AI, 2 June 2026 — “Nvidia Nemotron 3 Ultra”.

3. Enterprise agents reach production: Meta, Codex on AWS, Poke on iMessage and Morgan Stanley

Meta has launched an AI agent for businesses, Meta Business Agent, available on WhatsApp, Instagram and Messenger, able to answer customers, book appointments and close sales. The company is offering it free for now and has stated it will move to a paid subscription model with different tiers.

OpenAI has announced the general availability of its frontier models and of Codex on AWS, accessible through the cloud’s existing security, governance, provisioning and billing flows. In the same week, Codex has added a web-hosting feature called Sites, an in-place editing tool called Annotations and six role-specific plugins: data analytics, creative production, sales, product design, equity investing and investment banking.

On the distribution side, the assistant Poke has become the first AI agent approved by Apple for the Messages for Business platform, operable from iMessage and integrated with services such as Oura, Microsoft Outlook, Gmail, GitHub and Navan. Morgan Stanley, in addition, will open its wealth-management funnel —including platforms such as ShareWorks and Equity Edge— to AI agents from thousands of companies.

Sources: Meta — “Introducing Meta Business Agent”; 9to5Mac — “Apple’s Messages app on iPhone now has a third-party AI agent”, 4 June 2026; CNBC — “Morgan Stanley to open wealth management funnel to AI agents”, 3 June 2026; report via TLDR AI, 2 June 2026 — “OpenAI and Codex reach AWS”.

The bulletin mascot wearing a switchboard-operator headset handling several lines at once at a pegboard panel, with an open appointment notebook on the desk — Enterprise agents in production: Meta Business Agent on WhatsApp and Instagram, Codex on AWS, Poke on iMessage and Morgan Stanley opening its wealth funnel.

4. Capital for compute: Alphabet raises $80 billion, DeepSeek closes its first round and SpaceX prices its IPO

Alphabet has announced a plan to raise $80 billion through a stock sale earmarked for funding AI compute infrastructure. The plan includes $10 billion from Berkshire Hathaway, $30 billion in an underwritten offering and $40 billion from an at-the-market offering program on Class A and Class C shares, with a planned start in the third quarter and with Goldman Sachs, JPMorgan Chase and Morgan Stanley as coordinators.

DeepSeek will raise around 50 billion yuan in its first funding round, of which its founder has committed 20 billion of his own money, with the rest coming from fewer than ten investors and an expectation of closing in the coming weeks.

SpaceX has set a price of $135 per share for its initial public offering, valuing the company at $1.77 trillion and seeking to raise $74.4 billion, the largest public offering in history. The company will list on the NASDAQ under the ticker SPCX and has cited its orbital AI data centres among the uses of the capital.

Sources: CNBC — “Alphabet to raise $80 billion from stock sales to fund AI buildout”, 1 June 2026; CNBC — “DeepSeek slated to draw $7 billion in maiden fundraising”, 3 June 2026; report via TLDR, 4 June 2026 — “SpaceX sets price for the world’s largest IPO”.

5. Chips and export controls: the United States closes a loophole and Nvidia previews Computex

The US Department of Commerce has issued guidance extending export-licence requirements to advanced chips sold to any entity headquartered in China, regardless of where it is physically located. The measure closes the loophole of acquiring Nvidia or AMD chips through a subsidiary in another country, targets future sales and does not affect hardware already shipped or the servicing of equipment already deployed.

Ahead of Computex 2026, Nvidia has previewed its N1X laptop chip, with twenty ARM cores and a GPU equivalent to an RTX 5070, alongside its Vera Rubin data-centre platform, with a focus on physical and agentic AI.

Sources: The Next Web — “US closes Nvidia chip loophole for China-based firms abroad”; report via TLDR AI, 1 June 2026 — “Nvidia at Computex 2026: N1X and Vera Rubin”.

6. Physical AI and robotics: Amazon Proteus, Nvidia Cosmos 3 and Generalist AI

Amazon has unveiled a new version of its Proteus warehouse robot, fully autonomous, which can be directed in natural language and figures out priority, route and timing on its own to move heavy carts in logistics centres. The company is testing it in its labs and plans to start deploying it in Europe during the first half of 2027.

In the same field, Nvidia has launched Cosmos 3, an open foundation model for physical AI with native vision reasoning and multimodal generation across text, image, video, sound and action, built on a mixture-of-transformer architecture. Generalist AI, for its part, has secured $400 million to advance physical artificial general intelligence, with Radical Ventures and NVIDIA among its investors.

Sources: Engadget — “Amazon’s new Proteus warehouse robot is fully autonomous”; Nvidia — “NVIDIA launches Cosmos 3, the open frontier foundation model for physical AI”; Generalist AI — “Accelerating the next phase of physical AI”.

The bulletin mascot taking notes in a notebook next to an articulated industrial robot moving a warehouse cart, with blank labelled boxes and a monstera plant in the background — Physical AI: Amazon's autonomous Proteus robot, Nvidia's Cosmos 3 model and Generalist AI's $400 million round.

7. Framework of the week: the economics of tokens

Several pieces that circulated this week converge on how AI spending and its return are being measured.

A report from EntelligenceAI carried by Big Technology states that only 18% of AI spending translates into productive output, against a backdrop of rising token costs that is leading some companies to reconsider or cancel initiatives.
A piece by Tomasz Tunguz notes that Microsoft has started including “average token usage” on the launch cards for its models, shifting the comparison toward intelligence per dollar rather than raw performance alone.
An interview on Latent Space with GitHub’s chief operating officer, Kyle Daigle, puts the growth of agent-submitted code on the platform at 1,400% this year, a strain on infrastructure designed for human developers at human speed.
A statement from Cloudflare’s CEO, carried by Tom’s Hardware, holds that bot traffic has already overtaken human traffic online, earlier than the company itself expected.
An analysis by Mem0 on memory in agent harnesses —Claude Code, Codex, Copilot and others— describes cross-user contamination rates of 57% to 71%.

Sources: Big Technology — “The token reckoning is here”; Tomasz Tunguz — “Tokens per result”; Latent Space — “GitHub’s agent era”; Tom’s Hardware — “Bots have now passed human traffic online, Cloudflare boss laments”; report via TLDR AI, 3 June 2026 — “State of memory in agent harness”.

Closing

The six blocks above cover Anthropic’s confidential IPO alongside the expansion of its partner network and Glasswing program, a batch of open-weight models from MiniMax, Nvidia, Google and Microsoft, the arrival of enterprise agents in production with Meta, OpenAI, Poke and Morgan Stanley, a round of capital deals to fund compute from Alphabet, DeepSeek and SpaceX, the closing of a loophole in chip export controls, and the advances in physical AI from Amazon, Nvidia and Generalist AI. The framework of the week gathers five independent readings on how AI spending and its return are being measured. None of the items published include verified productivity data measured in specific companies.