Two AI companies filed to go public, one major coding tool got acquired for $60 billion, and four separate pricing structures changed. All in seven days. The through-line is consolidation: tools that were cheap are getting metered, platforms that were independent are getting bought, companies that stayed private longer than anyone expected are suddenly racing to list. The builders who feel it first are the ones who built workflows on the old economics.

THE WEEK AT A GLANCE

WHERE TO START

SpaceX, which absorbed xAI in February and owns Grok, has an option to buy Cursor for $60 billion roughly 30 days after its June 12 IPO. Cursor routes to Claude and GPT-5 today because those are the best models for coding. After acquisition, those routing decisions belong to people whose parent company has a competing model and a Colossus cluster equivalent to 1 million H100 GPUs providing financial incentive to use it. I give it 18-24 months before the changelog shows Grok's share of completions rising without a capability announcement to justify it — and I'd genuinely like to be wrong.

→ READ THE FULL TAKE

Anthropic filed a confidential S-1 on June 1, valued at $965 billion, with a $47 billion annualized revenue run rate. Up from $10 billion annually just one year ago. That growth rate is the argument for going public now. What changes for Claude users isn't the product the morning after but the incentive structure underneath it over the next 12-18 months. OpenAI filed two weeks earlier, added ads to ChatGPT within weeks of that, and the trajectory for what public-market pressure does to an AI product is already visible.

→ READ THE FULL TAKE

Starting June 15, Agent SDK usage, claude -p, Claude Code GitHub Actions, and third-party apps authenticating via Agent SDK move to a separate credit pool billed at full API prices: $20/month for Pro, $200/month for Max 20×. The subsidy gap closing was 15-30× at the median; on heavy long-context workloads, independent analysis has measured it as high as 175×. The builders most at risk are running automated CI/CD pipelines and won't find out until a workflow stops at 2am on June 16. Audit now and claim credits before the deadline.

→ READ THE FULL TAKE

Gemini CLI (100K GitHub stars, Apache 2.0, 6,000 merged community pull requests) retires June 18. The replacement, Antigravity CLI, is technically better: Go rewrite, faster startup, five parallel subagents. What it isn't: open-source. Google accepted years of community contributions, then replaced the tool with a closed binary and an 11-day migration window. Migrate anyway; the feature migration is clean and the deadline is real. The complaint stands independently of the migration advice.

→ READ THE FULL TAKE

GitHub Copilot moved to AI Credits on June 1: 1 credit = $0.01, monthly pools equal to your plan price, code completions still unlimited. Agentic workflows, code review, and multi-model chat sessions now draw from the pool. There's a promotional higher allotment through September 1 that's masking what this actually costs for heavy users. September is the first honest bill. If you run heavy agentic pipelines on Copilot infrastructure, model what September looks like while the cushion is still in place.

→ READ THE FULL TAKE

Microsoft unveiled Project Polaris at Build 2026: an in-house mixture-of-experts coding model targeting GitHub Copilot GA in August, running on Maia accelerators inside Azure's own infrastructure. The benchmark claims (beats GPT-4 Turbo on HumanEval and MBPP) are Microsoft's own, unverified externally; don't make architecture decisions on those numbers yet. What doesn't need verification: the inference stack is entirely Microsoft's. You don't build a first-party coding model and a first-party reasoning model (MAI-Thinking-1 also shipped) the same quarter your key vendor files to go public by accident.

→ READ THE FULL TAKE

Anthropic expanded Project Glasswing to 150 organizations across 15 countries on June 2: power utilities, water systems, hospitals. Initial partners have found 10,000+ high-severity security flaws. Cloudflare alone: 2,000 bugs, 400 rated critical, false-positive rate better than their own human security team. Eight weeks, fifteen countries, codebases for power grids — and public visibility into the access controls for those systems has not kept pace. Net positive for infrastructure security. The thing I'd want to see is an independent audit of the program doing the finding, not just the bugs it found.

→ READ THE FULL TAKE

DeepSeek built R1 and V3 without a dollar of venture capital. Now it's raising $7.4 billion for the first time, at a $52-59 billion valuation, with Tencent as anchor investor and founder Liang Wenfeng contributing 40% of the round from his own balance sheet. That 40% is the tell: this is a strategic partner selection, not a capital raise in the conventional sense. The Tencent position is the one I keep coming back to. If DeepSeek models end up running inside WeChat's 1.4 billion monthly active users, the conversation about Chinese AI reach changes considerably.

→ READ THE FULL TAKE

GPT-5.5, GPT-5.4, and Codex went generally available on Amazon Bedrock on June 1. The models didn't change. Every API call now inherits IAM, VPC isolation, KMS encryption, and CloudTrail audit logging. Organizations with existing SOC 2 or FedRAMP approval on Bedrock don't need a separate security review for OpenAI models — the compliance framework is inherited. That paperwork was the actual constraint for banks, health systems, and government contractors. Not the model quality.

→ READ THE FULL TAKE

OpenAI shipped Lockdown Mode on June 6: a free toggle on all ChatGPT tiers that disables Agent Mode, Deep Research, live web, Canvas networking, and file downloads. More interesting than the toggle: Elevated Risk labels deployed the same day on Agent Mode, Codex codebase access, and autonomous email sending. Informational only, not blocking. But they're OpenAI publicly acknowledging these features carry risk they haven't fully fixed yet, with a commitment to remove the labels when security improves. Watch whether those labels accumulate or actually get retired.

→ READ THE FULL TAKE

OpenAI shipped Dreaming V3 on June 4: a background memory system that reads across all your past conversations and synthesizes structured facts without you prompting it. Temporal awareness means stored memories update as circumstances change — "you're going to Singapore" becomes "you went to Singapore" after the date passes. No raw transcripts retained, just conclusions. The most useful thing you can do right now: Settings → Personalization → Memory, read what's actually listed, correct anything wrong.

→ READ THE FULL TAKE

Trump signed an executive order on June 2 creating a 30-day pre-release review for the most powerful AI models. Labs can decline. No penalty for declining. The order explicitly bars its language from creating mandatory requirements. What's actually binding: an AI cybersecurity clearinghouse (Treasury, NSA, CISA) forming within 30 days, and CISA Binding Operational Directives for federal civilian systems. For most builders, nothing changes this week. For anyone selling AI tools into federal agencies, watch July.

→ READ THE FULL TAKE

Colorado's original AI Act was repealed May 14, replaced with CADMA, effective January 1, 2027. Gone: governance frameworks, annual bias testing, the requirement to report discovered algorithmic discrimination to the AG. Kept: consumer notice at the point of interaction, 30-day adverse-outcome notices, right to request human review. Dropping bias testing is not a regulatory efficiency improvement. It removes a check on documented failure modes in hiring, lending, and housing. CADMA is still real; if your AI is a non-de-minimis factor in consequential decisions affecting Colorado residents, January 1 applies.

→ READ THE FULL TAKE

OpenAI added live job listings (from Indeed, Upwork, and Appcast, free in the US across all tiers) and a resume tailoring tool to ChatGPT this month. The listings are useful as a filter. The resume tool is the underrated half: upload a resume, pick a job posting, get a language-aligned version in 45 seconds. The barrier to tailoring applications isn't knowing you should. It's the hour it takes per application. Turn memory on first (Settings → Personalization → Memory) or the personalization doesn't work.

→ READ THE FULL TAKE

NVIDIA put the DGX Spark chip (128GB unified memory, 6,144 CUDA cores) into Windows laptops at Computex 2026. Runs 120-billion-parameter models locally. Price floor reportedly can't go below ~$2,900. The "agentic AI OS" framing is software vision, not a shipping product. The buying question is simple: do you regularly run large models locally where privacy or air-gapping requires it? If no, you're paying 100% for a chip you'll run at 10%.

→ READ THE FULL TAKE

Next week: June 12 is the SpaceX IPO. How it prices determines whether the Cursor acquisition closes at $60 billion or falls back to the $10 billion collaboration path. Watch the listing.

— SAMWISE