AI News Weekly Summary — June 1–7, 2026 | Microsoft Build, a White House AI Order, Anthropic’s S-1

This AI news weekly June 1–7, 2026 was about platforms, regulation and the cost of AI. Microsoft showed its own MAI models, the White House signed a new AI order, and Anthropic began its move toward an IPO. In the background, the pressure on infrastructure grew — from chip debt to NVIDIA’s local silicon. (Context: previous week, May 25–31).

1. Microsoft Build 2026 — Its Own MAI Models and Less Reliance on OpenAI

On June 2–3 in San Francisco, Microsoft used Build to show its own MAI models. The headline is MAI-Code-1-Flash — a coding model offered in GitHub Copilot, initially in Visual Studio Code — alongside MAI-Thinking-1 (a reasoning model in private preview via Microsoft Foundry). Microsoft is also leaning harder into a wider choice of models in Foundry, including non-OpenAI ones, while Agent 365 and Entra Agent ID handle identity and oversight for enterprise agents.

The logic is simple: cheaper in-house models ease margins and loosen OpenAI’s exclusivity inside Copilot. Google is playing it the same way by cutting prices on developer tools.

This week Microsoft, Google, NVIDIA and ServiceNow are competing not over the model itself but over the “runtime” an agent runs inside. The key signal: the choice is shifting from a model to a whole platform — chip, runtime and security layer.

2. The White House Signs an AI Order — “Innovation and Security”

On June 2, President Trump signed an executive order, “Promoting Advanced Artificial Intelligence Innovation and Security.” Instead of mandatory pre-release testing (the draft the administration pulled in May), it leans on cyber-defense of government systems and voluntary cooperation with industry. Key points: a Treasury-run AI cybersecurity “clearinghouse” coordinating vulnerability discovery and patching, and a voluntary framework under which developers may give the government access to a “covered frontier model” for up to 30 days before release.

The order explicitly bars any mandatory licensing or permitting to publish models, and directs the Attorney General to prosecute AI-enabled crimes such as system intrusions. It’s an “innovation first, security voluntary” posture.

The new order caps the turn that began in May, when the White House pulled its testing draft. You can read it as an attempt to reconcile two goals — keeping the lead over China and limiting risk — without imposing hard obligations on companies.

3. The Great American AI Act — 269 Pages and a Three-Year State-Law Freeze

On June 4, Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) unveiled a 269-page draft of the Great American AI Act — the most comprehensive federal AI framework yet in the US. The headline provision is a three-year preemption of state laws on frontier-model development. It would hit similar state efforts — including Colorado, where the earlier June 30, 2026 date was replaced by a new regime starting January 1, 2027.

Companies above $500M in revenue would publish “frameworks” for governing their most capable models, report serious safety incidents and admit auditors; the draft also funds a $100M/year federal standards center (CAISI). Reactions are polarized: labor unions (including the AFL-CIO) rejected it as “a giveaway to the AI industry,” while tech groups (ITI, NetChoice) praised it. It’s still a discussion draft, not a law.

4. Anthropic Files a Confidential S-1 — the Road to an IPO Begins

On June 1, Anthropic filed a confidential draft S-1 with the US Securities and Exchange Commission (SEC), starting its path to the public markets. The company hasn’t set a share count or price yet — it’s giving the SEC time to review before a public prospectus. The filing comes days after the close of the Series H ($65B, a $965B valuation) and at a run-rate of roughly $47B.

According to reports, OpenAI is also preparing documents for a potential IPO, which could set up two of the year’s largest AI listings competing for the same investor pool. The cost scale shows in the background: the SpaceX contract alone is $1.25B per month for compute.

Analysts now treat a trillion-dollar debut as the base case — if markets cooperate. That suggests the $965B private valuation is heading quickly for a public-market test, where the growth rate, not the narrative, sets the multiple.

5. Apollo and Blackstone — About $36B of Debt for Anthropic’s TPU Chips

A separate, record transaction is forming around Anthropic’s round. According to reports, Apollo Global Management and Blackstone are structuring about $36B of debt financing to buy Google TPU chips, which would then be leased to Anthropic (data centers in New York, Texas, Louisiana and Indiana). Broadcom, the TPUs’ co-developer, would backstop part of the largest tranches. The debt is being syndicated to investors.

It would be the largest chip-financing debt transaction ever. The mechanism lets Anthropic get compute without putting $36B on its own balance sheet — a special-purpose vehicle holds the debt.

When Apollo and Blackstone arrange $36B of debt on AI chips for bond investors, compute is in practice shifting from an R&D expense to an infrastructure asset. That changes the cost of capital and the set of investors around frontier AI.

6. OpenAI Dreaming V3 — ChatGPT With Better Memory

On June 4, OpenAI began rolling out Dreaming V3 — the biggest change to ChatGPT’s memory since its 2024 launch. Instead of waiting for a “remember this” command, the model runs a background process after conversations that synthesizes what matters (preferences, constraints, ongoing projects) and tells current facts from stale ones. It starts with Plus and Pro users in the US, with the rest following in coming weeks.

The key detail is cost: OpenAI cut the memory process’s compute needs roughly fivefold, which makes it viable on the free tier too. Privacy researchers note that in the EU such memory features will be judged through the lens of the GDPR and further AI Act obligations applying from August 2026.

7. NVIDIA RTX Spark — an AI Platform for Windows PCs

On May 31, around Computex / GTC Taipei, NVIDIA revealed RTX Spark — a platform for Windows PCs spanning slim laptops and compact desktops, framed with Microsoft as “reinventing the PC.” The Arm-based chip integrates AI agents, content creation and gaming; Adobe is rebuilding Photoshop and Premiere Pro for it. Devices are expected in autumn 2026, with no price disclosed.

The announcement pushed AMD, Intel and Qualcomm shares lower. Much of the recent investor narrative around NVIDIA has rested on data centers — moving to client devices suggests it sees the next AI bottleneck at the edge, where agents run locally without cloud cost or latency.

It’s really a bet that some AI workloads will move from the cloud onto the device. If it pays off, NVIDIA hedges against the moment cheap local compute undercuts the “everything in the data center” model.

8. GitHub Copilot — Usage-Based Billing and a Developer Backlash

On June 1, GitHub began moving from a flat fee to an AI Credits / usage-based model — felt most sharply for now on individual plans and agentic sessions (Business and Enterprise plans have their own transition periods). The reaction was immediate: “what a joke” spread across Reddit and X, with users posting estimates of bills rising from $29 to as much as $750 a month on heavy sessions.

The logic is simple — usage-based billing passes on the real cost of expensive reasoning models that the flat fee used to hide (classic code completion is treated differently). Developers who manage their prompts carefully report flat or lower costs. The problem is team budgeting, when a single session can cost as much as a month of several others’ usage. The change coincided with Microsoft’s cheaper MAI model.

9. SoftBank — €75B for AI Data Centers in France

Announced May 30 at the Choose France summit hosted by President Macron, and widely discussed across June 1–7: SoftBank will invest up to €75B in 5 GW of AI data-center capacity in France. Phase one is €45B and 3.1 GW in the Hauts-de-France region by 2031, across three planned sites. It’s Europe’s largest single announced AI infrastructure commitment.

SoftBank cited France’s low-carbon grid (about 70% nuclear), industrial land and engineering talent. Caution is warranted — SoftBank’s liquid assets are well below the announced figure, so the investment will almost certainly come from a mix of funds, debt and joint ventures.

10. Sysdig — the First Attack of Its Kind It Has Observed, Run by an Autonomous LLM Agent

Security firm Sysdig documented the first attack it has observed in which an LLM agent autonomously performed post-exploitation actions (the intrusion was seen in May and widely discussed at the turn of the month). Initial access came through CVE-2026-39987 in the marimo tool, and the whole chain — from compromise to exfiltrating an internal Postgres database — took under an hour, with no human directing each step.

A separate but related risk this week was the BadHost flaw (CVE-2026-48710) in the Starlette library, relevant to the FastAPI ecosystem and many AI apps. Patches for both are available and should be applied.

In practice, it lowers the barrier for sophisticated attacks — from “a skilled human team” to “model API access plus knowledge of a working CVE.” Warnings about offensive AI just gained one of their first documented examples.

11. Cognition — $1B at a $26B Valuation, Devin’s Revenue Up 1,230%

Cognition, maker of the autonomous engineer Devin, raised (announced May 28, widely discussed in early June) over $1B at a $26B post-money valuation, co-led by Lux Capital, General Catalyst and 8VC. Devin’s annualized revenue grew from $37M in May 2025 to roughly $492M in May 2026 — about 1,230% in a year. More than 90% of the company’s own code is now written by Devin.

Customers include Goldman Sachs, Mercedes-Benz, NASA and Santander. CEO Scott Wu stresses the round lets the company stay independent — which matters when every major player (OpenAI with Codex, Anthropic with Claude Code, Google with Jules) has its own coding agent.

12. Anthropic Project Glasswing — Claude Moves Into Critical Infrastructure

On June 2, Anthropic expanded Project Glasswing — the program giving vetted organizations Claude Mythos Preview for cybersecurity — into new sectors: power grids, water systems, healthcare, communications and hardware makers. Per Anthropic, for most of the new partners a serious attack could affect systems that matter to more than 100 million people.

It’s complemented by connecting Claude to 28 security and compliance platforms (including CrowdStrike, Palo Alto Networks, Okta and Zscaler) via the Claude Compliance API, embedding the model directly in enterprise security stacks.

13. GPT-5.5-Cyber in Europe — Access for Vetted Defenders

Back in May (announced May 11), OpenAI laid out an “EU Cyber Action Plan,” expanding limited access to GPT-5.5 and the GPT-5.5-Cyber variant for vetted European businesses, governments, agencies and EU institutions, including the EU AI Office. The topic resurfaced across June 1–7 as part of the wider contest for European cybersecurity contracts — partly because, from June 1, members of the Trusted Access for Cyber program must enable stronger account security.

The contrast with Anthropic is softer than it looked: Anthropic is expanding Glasswing globally and, per reports, opened its Mythos model to the EU’s cybersecurity agency (ENISA) around the same time. Both are chasing the same European budgets.

14. Foundation Phantom — Humanoid Robots Tested in Ukraine (a Continuing Story)

In reports over recent months, Foundation has described testing its Phantom humanoid robots in Ukraine, mainly for logistics and reconnaissance — carrying supplies, assessing damage, working in places too dangerous for people. By those accounts, the robots are not meant for offensive operations and run under military supervision and Foundation engineers.

Even in a purely logistics role, humanoids near a war zone push the policy conversation about autonomous weapons closer. The distance between support and offensive capability is real but not infinite.

15. Demis Hassabis: AGI Around 2030, “Plus or Minus a Year”

In interviews in early June, Google DeepMind CEO Demis Hassabis pointed to around 2030, give or take a year, as a realistic window for AGI — human-level general intelligence across a wide range of tasks. Hassabis has been among the more cautious voices, so naming a specific window is a notable shift.

Read it with context: Jack Clark (Anthropic) put 60%+ odds on recursive self-improvement by the end of 2028. The most prominent lab CEOs are converging on a 2028–2030 window for something qualitatively different from today’s systems — close enough to plan decisions around.

16. Wikipedia — Editors Threaten a Strike After the Community Tech Team Is Dissolved

Volunteer Wikipedia editors are threatening a strike after the Wikimedia Foundation dissolved its Community Tech team (a May 21 decision that laid off several engineers). Editors worry it will weaken tools, technical fixes and the day-to-day work of the community that has maintained the site’s quality for years.

There’s a separate issue here: Wikipedia’s quality matters for AI models trained on human knowledge. If the editorial backbone weakens, over time that may show up in the training data for future models too.

17. Other Important Signals of the Week

• OpenAI Rosalind — a biodefense program (announced May 29, with further communication on June 3) giving vetted teams and select US government agencies and allies access to the GPT-Rosalind model; partners include Lawrence Livermore, Johns Hopkins APL and CEPI.
• OpenAI Codex for enterprise — new plugins for specific roles and workflows (analytics, sales, product design, finance and more) plus “annotations” and “Sites” features.
• OpenAI Ads — separately, the company is building out its own advertising layer; worth treating as a distinct business signal.
• Cursor and SpaceX — an earlier deal (from April) giving SpaceX an option to acquire Cursor for $60B resurfaced in commentary as an example of how strategic coding tools have become.
• Colorado AI Act — the earlier June 30, 2026 effective date was replaced by a new regime starting January 1, 2027; it’s exactly the kind of state rule the federal Great American AI Act could freeze.
• Unconfirmed “Claude Sonnet 4.8” rumor — developer communities increasingly expect a mid-June mid-tier Anthropic model; for now there’s no model card or API ID.

Weekly Takeaway

The week of June 1–7, 2026 showed the center of gravity shifting from individual models to the platforms that run them. At Build, Microsoft revealed its own MAI models and leaned harder into a wider choice of models in Azure AI Foundry, Google is cutting prices, and NVIDIA is moving onto Windows PCs with RTX Spark — everyone is fighting over the runtime an agent lives in, and over its cost. The fight over GitHub Copilot’s new billing is the same economic battle, seen from the developer’s side.

The second thread is regulation and money. The White House signed an order favoring innovation and voluntary security, while Congress unveiled the Great American AI Act with a three-year freeze on state rules — two different answers to the same question. In the background, Anthropic filed a confidential S-1, and Apollo and Blackstone are arranging about $36B of debt for chips, nudging compute toward a capital-markets product. The attack documented by Sysdig was a reminder of what these security programs are for: offensive AI is less and less only theoretical.

Źródła / Sources (wspólne dla obu wersji):

• Microsoft Build / wojny o kodowanie: Microsoft and Google take on Anthropic and OpenAI in AI coding — CNBC, Microsoft unveils new AI models to lessen reliance on OpenAI — CNBC
• Rozporządzenie / Great American AI Act: Promoting Advanced AI Innovation and Security — The White House, Fact Sheet — The White House, Obernolte–Trahan — Great American AI Act draft — house.gov, 3-year preemption of state laws — Roll Call, Great American AI Act framework — FedScoop
• Anthropic S-1 / dług na chipy: Confidential draft S-1 — Anthropic, Anthropic files to go public — TechCrunch, Apollo $36B debt deal for Google chips — Bloomberg
• OpenAI Dreaming V3: Dreaming: better memory for ChatGPT — OpenAI, ChatGPT memory, especially free tier — Engadget
• NVIDIA RTX Spark: NVIDIA’s new PC chips — CNBC
• GitHub Copilot — rozliczenie: Copilot is moving to usage-based billing — GitHub Blog, Updates to Copilot billing and plans — GitHub Changelog, „What a joke”: billing backlash — TechCrunch
• SoftBank Francja: SoftBank up to €75B in French data centers — TechCrunch
• Sysdig / atak agenta LLM: AI agent at the wheel — Sysdig, Attackers use LLM agent after marimo CVE-2026-39987 — The Hacker News
• Cognition / Devin: Cognition raises $1B at $26B — The Next Web
• Anthropic Project Glasswing: Project Glasswing — initial update — Anthropic, Mythos expansion — Investing.com
• OpenAI Rosalind / GPT-5.5-Cyber: Rosalind Biodefense — OpenAI, OpenAI biodefense program — Axios, Trusted Access for Cyber / GPT-5.5-Cyber — OpenAI, OpenAI EU cyber model vs Anthropic Mythos — CNBC
• Cursor–SpaceX: SpaceX can buy Cursor for $60B — CNBC, SpaceX option to acquire Cursor for $60B — Reuters/Investing.com
• Wikipedia / Community Tech: Editors plot strike after Wikimedia layoffs — The Register, Editors threaten strike — Cybernews
• Hassabis / kontekst tygodnia: AI News June 5, 2026 — Build Fast with AI, AI News June 1, 2026 — Build Fast with AI