AI Bulletin · week of May 18

Recap of the week of May 18–22, 2026 in AI. Six news items and a framework on the AI labs’ P&L and IPO pressure.

1. Google I/O 2026: Gemini 3.5 Flash, an open agent runtime and a consolidation of coding tools

Google used I/O 2026 to launch Gemini 3.5 Flash, a new model focused on agentic workflows, coding and long-horizon tasks, and started rolling it out the same day across Search, Workspace, Android XR and its developer platform. The company put monthly token usage across its AI systems at more than 3.2 quadrillion.

In parallel, Google Cloud published Agent Executor, an open and distributed runtime for executing long-running agent workflows with durable execution, secure isolation and connection recovery, integrated with Kubernetes Engine via Agent Substrate. The developer tooling block is also being unified under a new brand, Antigravity, which bundles IDE, CLI, SDK and agent tooling.

Search added an agentic search box with personal context from Gmail and Photos, and Chrome added an llms.txt check to Lighthouse under a new “Agentic Browsing” category.

Sources: Google — “Gemini 3.5 Flash” and “Google I/O 2026: Sundar Pichai keynote”, 19 and 20 May 2026; Google Cloud — “Introducing Agent Executor”; InfoWorld — “Google to unify AI coding tools under Antigravity”; ITPro — “Google adds AI to the search box”; Search Engine Land — “Google adds llms.txt check to Chrome Lighthouse”.

The bulletin's mascot reading a freshly printed newspaper at a newsroom desk with a typewriter, a coffee mug and an open notebook — Google I/O 2026: Gemini 3.5 Flash, Agent Executor and the consolidation of Google's coding tools under Antigravity.

2. Anthropic locks in a $45 billion, three-year compute commitment with SpaceX, talks Maia chips with Microsoft and projects $10.9 billion in Q2 revenue

Anthropic has agreed to pay SpaceX nearly $45 billion over three years for compute capacity, at a rate of $1.25 billion per month through May 2029, with either party able to end the agreement on 90 days’ notice. The commitment puts a financial figure on the access to more than 220,000 GPUs at SpaceX’s Colossus 1 datacenter in Memphis reported earlier in May, and coverage is extended to a second SpaceX datacenter.

In parallel, Anthropic is in talks with Microsoft to be supplied with its Maia 200 AI chips, an alternative to compute from Amazon and Google that the company is pursuing to ease its capacity constraints. Microsoft already invested $5 billion in Anthropic last November.

Anthropic’s Q2 revenue is projected at $10.9 billion, according to figures disclosed to investors as part of the ongoing funding round. Quarterly growth is outpacing the previous run rates of Google and Facebook in their respective years before going public, although the report adds that planned compute spend may prevent Anthropic from closing the full year in profit. Andrej Karpathy has also joined the company’s pre-training team, and the company’s new consulting venture has closed its first acquisition, Fractional AI.

Sources: report via TLDR AI, 21 May 2026 — “Anthropic to pay SpaceX nearly $45 billion for computing deal”; CNBC — “Anthropic, Microsoft in talks for AI chip deal after $5 billion investment”, 21 May 2026; TechCrunch — “OpenAI co-founder Andrej Karpathy joins Anthropic’s pre-training team”, 19 May 2026; Yahoo Finance — “Anthropic’s new consulting venture makes its first acquisition”; Q2 projection reported via TLDR AI, 21 May 2026 — “Mind-blowing growth is about to propel Anthropic into its first profitable quarter”.

3. A jury dismisses Musk’s claims against Altman, and OpenAI prepares a possible IPO in September

A US jury has dismissed all of Elon Musk’s claims against Sam Altman and OpenAI on the grounds that the plaintiff waited too long to file them. Musk has said he will appeal. With the ruling out of the way, specialist coverage points to OpenAI preparing for a possible IPO as early as September.

OpenAI has also published an offering called Guaranteed Capacity that lets corporate customers reserve compute for one, two or three years, with discounts based on the length of the commitment. The company reported $5.7 billion in Q1 revenue, ahead of Anthropic for the same period, and closed a partnership with Dell to deploy Codex in hybrid and on-premises environments, announced at Dell Technologies World.

A separate analysis published by CNBC argues that falling prices and increased competition, including from Chinese labs, could complicate both OpenAI’s and Anthropic’s valuations in their respective IPOs.

Sources: NPR — “Jury dismisses all claims in Elon Musk’s lawsuit against OpenAI CEO Sam Altman”, 18 May 2026; TechCrunch — “OpenAI barrels toward IPO that may happen in September”, 20 May 2026; CNBC — “OpenAI announces new Guaranteed Capacity offering for customers to secure compute”, 19 May 2026; Sherwood News — “Report: OpenAI’s Q1 revenue was $5.7 billion, beating Anthropic”; CNBC — “Cheap AI could derail OpenAI and Anthropic’s IPOs”.

The bulletin's mascot standing in front of a ticker board with lines rising and falling, on a desk with two stacks of folders and a wax-sealed envelope — OpenAI: ruling in Altman's favour, a path toward an IPO in September and a warning about pricing pressure from cheaper models.

4. Cursor hits $3 billion ARR, SpaceX has an option to buy at $60 billion and Microsoft pulls Claude Code licences

Cursor reached an annualised revenue rate of $3 billion in late April, with more than 3,000 customers paying at least $100,000 per year for its software. SpaceX holds an option to buy Cursor for $60 billion, exercisable during a 30-day window that opens shortly after SpaceX’s IPO, scheduled for 12 June.

In the same coding block, Microsoft has cancelled its Claude Code licences and shifted its developers to GitHub Copilot CLI. Anthropic also announced it has acquired Stainless, the SDK automation startup used by OpenAI, Google and Cloudflare among others. Cursor introduced Composer 2.5 the previous week, a coding agent trained with targeted reinforcement learning, synthetic data and new distributed training techniques.

Sources: report via TLDR AI, 22 May 2026 — “Cursor hits $3 billion annual sales rate ahead of SpaceX deal”; Windows Central — “Microsoft cancels Claude Code licenses, shifting developers to GitHub Copilot CLI”; Anthropic — “Anthropic acquires Stainless”; Cursor — “Composer 2.5”.

5. The enterprise agent “control plane” is consolidating: Microsoft 38.6%, OpenAI 25.7%, Anthropic 5.7%

An analysis by VentureBeat puts enterprise agent orchestration share at 38.6% for Microsoft, 25.7% for OpenAI and 5.7% for Anthropic, and argues that the next phase of competition in the enterprise segment will be decided at the control layer: permissions, audit and workflow orchestration above the underlying model.

Along the same line, Forrester has published a post arguing that SAP is trying to position itself as a mandatory gateway between external agents and SAP data, and explicitly recommends that CIOs push back. The critique points directly to SAP’s 2026 API policy that this series covered last week.

Concrete moves arrived in the same consolidation block. Zscaler announced the acquisition of Symmetry Systems to add data discovery and classification to its Zero Trust platform, with a focus on how agents access sensitive information. 1Password integrated runtime credentials with OpenAI Codex so coding agents can access secrets without exposing them in prompts or the repository. Hitachi and Anthropic signed a strategic partnership within the Lumada 3.0 programme focused on deploying AI in physical environments.

Sources: VentureBeat — “Claude’s next enterprise battle is not models: it’s the agent control plane”; Forrester — “SAP is attempting to become the gatekeeper of enterprise AI: CIOs should push back”; SiliconANGLE — “Zscaler acquires Symmetry Systems to extend AI agent security capabilities”, 21 May 2026; Business Wire — “Hitachi announces strategic partnership with Anthropic”, 18 May 2026.

The bulletin's mascot holding a notebook and a microphone in front of three coloured glass panels, while a human hand moves a geometric token between them — The enterprise agent "control plane": orchestration share, SAP's gatekeeper positioning and new moves in identity and data security.

6. Attacks on the developer supply chain: GitHub, npm packages and OAuth phishing in Microsoft 365

GitHub has confirmed that attackers stole data from around 3,800 internal repositories after an employee device was compromised through a malicious VS Code extension. The company says there is no evidence that customer repositories or enterprise data were impacted.

An analysis published by Wiz describes a new variant of the Shai-Hulud family, a self-propagating worm that spread through widely used npm packages including TanStack, UiPath and Mistral AI, and installs a daemon that deletes the user’s home directory if the stolen GitHub tokens are revoked.

The Hacker News reports that a phishing-as-a-service platform called EvilTokens compromised more than 340 organisations using Microsoft 365 in five weeks by abusing the OAuth consent flow, and that identity teams are starting to treat application authorisation and delegated access with the same level of scrutiny as passwords or MFA.

Sources: TechCrunch — “GitHub says hackers stole data from thousands of internal repositories”, 20 May 2026; Wiz — “Mini Shai-Hulud strikes again: TanStack + more npm packages compromised”; The Hacker News — “The new phishing click: how OAuth is being abused”.

7. Framework of the week: the AI labs’ P&L and IPO pressure

Several pieces that circulated this week converge on the economics of serving AI in the run-up to a possible OpenAI IPO.

CNBC publishes an analysis arguing that falling prices and competition from both the US and China could complicate the valuations of OpenAI and Anthropic in their respective IPOs, as enterprises adopt cost-saving measures and look for cheaper alternatives.
Anthropic’s Q2 projection reported this week, $10.9 billion, comes with an explicit caveat that planned compute spend may prevent the company from closing the year in profit.
A piece distributed by TLDR Founders on the AI P&L in enterprise argues that every current AI lab is operating at a loss, and that there is a wide gap between what companies pay for AI subscriptions and what it actually costs to serve them, pointing to a price reset when subsidies are pulled.
Y Combinator and OpenAI have struck a programme under which every startup in the winter batch receives $2 million in OpenAI API credits; across 199 startups that adds up to roughly $400 million in credits, according to the report. OpenAI receives in exchange a 2% stake in each company at a $100 million valuation.
An analysis published by The Next Web notes that unemployment among entry-level workers has widened relative to experienced workers specifically in occupations most exposed to AI substitution, based on post-pandemic US data.

Sources: CNBC — “Cheap AI could derail OpenAI and Anthropic’s IPOs”; Anthropic’s Q2 projection via report from TLDR AI, 21 May 2026; The State of Brand — “Every AI subscription is a ticking time bomb for enterprise”; report via TLDR Founders, 20 May 2026, on the $2 million per startup in OpenAI credits to Y Combinator; The Next Web — “Gen Z is not booing AI. It is booing its own job market”.

Closing

The six items above cover the largest single day of Google updates of the year, an escalation in Anthropic’s compute commitments and revenue, a ruling that clears the way for an OpenAI IPO in September, the financial maturation of Cursor and an internal pivot at Microsoft on its choice of coding tools, the consolidation of the enterprise agent “control layer” and three attacks on the developer supply chain. None of the items published include verified productivity data measured in specific companies, and the framework gathers five independent readings on how revenue, prices and the cost of serving AI line up today.