AI Intelligence Report
Top Stories
Anthropic to pay SpaceX nearly $45 billion for computing power
Anthropic has signed a three-year compute deal with Elon Musk’s SpaceX worth approximately $45 billion — roughly $1.25 billion per month through May 2029. The agreement gives Claude exclusive access to SpaceX’s “Colossus” data center clusters, comprising more than 200,000 Nvidia GPUs. The disclosure landed inside SpaceX’s IPO filing and reframes the AI compute race: SpaceX is now effectively an AI cloud provider, and Anthropic is locking in capacity at a scale that rivals OpenAI’s Microsoft and Oracle commitments.
Andrej Karpathy joins Anthropic’s pre-training team
OpenAI co-founder and former Tesla AI chief Andrej Karpathy started at Anthropic this week, joining the pre-training team under lead Nick Joseph. His mandate is to launch a sub-team that uses Claude itself to accelerate pre-training research. Karpathy is pausing his education startup Eureka Labs to focus on the role. The hire is the year’s most symbolically loaded talent move — Karpathy was the public face of OpenAI’s early research culture, and his decision strengthens Anthropic’s bench at a moment when the company is also doubling down on infrastructure (see the SpaceX deal above).
OpenAI files confidentially for IPO targeting September listing
OpenAI filed a confidential S-1 with the SEC as early as Friday, May 22, targeting a September 2026 listing at a valuation in the $852 billion to $1 trillion range. The filing arrives just as critics note that ChatGPT user growth has plateaued — pivot-to-ai called it “running out of other people’s money.” SpaceX’s own IPO is set for June 12, raising questions about whether the market can absorb two mega-listings of this size in the same quarter.
Alibaba’s Qwen 3.7-Max runs autonomously for 35 hours straight
Alibaba Cloud unveiled Qwen 3.7-Max on May 20. In a demonstration that has the AI engineering community buzzing, the model ran on its own for 35 continuous hours — executing 1,158 distinct tool calls, performing 432 kernel evaluations, debugging compilation errors, and ultimately writing optimized software for a chip architecture (the T-Head ZW-M890) it had never seen during training. The model boasts a 1-million-token context window and scored 44.5 on the Apex Math Reasoning benchmark, ahead of Claude Opus 4.6 Max (34.5) and DeepSeek V4-Pro Max (38.3). Notably, Alibaba is keeping the weights closed and offering API-only access — a departure from previous Qwen releases.
Tencent unveils sweeping Hunyuan ecosystem upgrade
At a Tencent Cloud industry summit on May 21, Tencent rolled out major upgrades to its Hunyuan family: a fast model (Hunyuan Turbo S), a deep-reasoning model (Hunyuan T1), a visual reasoning model (T1-Vision), low-latency voice (Hunyuan Voice), and multimodal releases including Hunyuan Image 2.0, Hunyuan 3D 2.5, and Hunyuan-Game. The T1 update delivers an 8% lift in competitive math, 8% in commonsense Q&A, and a 13% jump in agent task performance. The release is Tencent’s clearest agentic-AI push of the year and lines up with President Martin Lau’s pledge to more than double Hunyuan spend to over RMB 36 billion in 2026.
AI News Roundup
Newsletter Highlights
Note: dedicated Substack tools were not connected this morning. The summaries below were drawn from your Inoreader newsletter feeds, which mirror your Substack inbox.
Inoreader AI Folder
14 unread items in your AI folder from the past 24 hours. Items already covered in Top Stories or Newsletter Highlights are referenced rather than duplicated.
OpenAI user numbers go flat — just in time for the IPO
Gerard’s characteristically skeptical read on OpenAI’s S-1 filing: weekly active user growth has flattened and the company is timing a public listing precisely as the easy growth runs out. He notes that SpaceX’s June 12 IPO will absorb investor cash that OpenAI also needs.
How Virgin Atlantic ships faster with Codex
OpenAI customer story: Virgin Atlantic used Codex to ship its revamped mobile app for a fixed holiday-travel deadline, reaching near-total unit-test coverage and zero P1 defects. A useful enterprise-deployment data point for any comms team writing about AI in operations.
OpenAI named a Leader in enterprise coding agents by Gartner
OpenAI was placed in the Leaders quadrant of the inaugural 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents. Codex is specifically cited for innovation and enterprise-scale deployment — an external validator that complements the Virgin Atlantic case study above.
Google I/O: AI for Everyone
Recap of Google’s I/O keynote and the “AI for everyone” framing around Gemini 3.5 Flash, Gemini Spark, and Gemini Omni. Pairs the launch news with concerns about Big Tech versus human workers.
OpenAI and SpaceX IPO era begins
Same SpaceX–Anthropic compute story but framed around the IPO calendar. Adds: Anthropic’s $45 billion three-year contract is roughly the entire 2025 Anthropic revenue forecast multiplied several times over.
Codex hit 4 million weekly developers
Headline data point for the week: Codex now has approximately 4 million weekly developers — a number disclosed quietly at an OpenAI enterprise event. The note frames AI buyers as “absorbing the gap” between what humans can ship and what agents can deliver overnight.
The Prompt That Builds Your Audience From Zero
Tactical prompt template aimed at solo creators: structured persona-pillar-proof framework for posting consistently to social platforms. Worth a skim for any communications team experimenting with AI-assisted social production.
Qwen model delivers 35 hours of continuous autonomous reasoning
Coverage of Alibaba’s Qwen 3.7-Max demo with the kernel-optimization breakdown and benchmark comparison versus Claude Opus 4.6 Max and DeepSeek V4-Pro Max.
Six AI search engines ranked for privacy, speed, and minimal AI chaos
Ranked comparison of six AI search tools (Perplexity, You.com, Kagi-AI, Andi, Phind, Brave Leo). Practical if you’re auditing which AI search tool to standardize the comms team on.
Lore Issue #185: Google Throws Everything at the AI Race
Lore’s roundup of the week: Google’s product blitz, the Karpathy-to-Anthropic move, Gemini 3.5 Flash’s pricing edge, and Spark’s “agents in the background” framing.
How OpenAI found a missing math proof
The Neuron pairs the OpenAI math claim with Qwen’s 35-hour agent run as two examples of AI doing sustained intellectual work. Includes Sam Altman’s offer of free OpenAI access to every Y Combinator startup.
OpenAI disproved a famous math equation
Aggregated coverage of the same OpenAI math story — useful if you want a second angle, but the substance is identical to The Neuron item above.
OpenAI claims it solved an 80-year-old math problem — for real this time
Skeptical framing of the same math story, plus Sam Altman’s “mic drop” offer to every Y Combinator startup — essentially free GPT credits and an inside track to enterprise distribution. The “for real this time” headline reflects skepticism after earlier overstated claims.
AI Workflows & Tool Watch
Anthropic ships MCP Tunnels and self-hosted sandboxes for Claude
At Code with Claude London this week, Anthropic released two features that directly affect anyone using Claude with internal tools. MCP Tunnels create a secure, outbound-only connection from your environment to Anthropic, so Claude can reach a private MCP server without you opening any inbound firewall holes or exposing a public endpoint. Self-hosted sandboxes keep your sensitive files and packages on your own infrastructure while Claude’s orchestration loop stays on Anthropic. For a global comms team running internal automations, this removes the biggest objection IT usually raises about AI agents touching internal systems.
Perplexity Personal Computer is now open to all Mac users globally
Perplexity’s “Personal Computer” — a desktop agent that can edit local files, drive native Mac apps, browse the web, and orchestrate voice tasks — opened to all Pro and Max subscribers earlier this month. GPT-5.5 is now the default orchestration model, and Claude Opus 4.7 powers background agents and Spaces collaboration. This is the closest competitor to Claude’s Cowork mode, and worth a side-by-side comparison for a comms workflow (drafting, file triage, light research).
Claude Code 2.1.147 ships pinned background sessions and /code-review
The May 21 release of Claude Code keeps “pinned” background sessions alive while idle, restarts them in place to apply updates, and only sheds them under memory pressure. The /simplify command has been renamed to /code-review and can post findings directly as inline GitHub PR comments. A regression in the same release was fixed on May 22 (the Bash tool was returning exit code 127 for some users). If you’ve been running Claude Code for podcast-prep or content scripts, update.
Anthropic adds a Compliance API for IT and security teams
Claude now integrates with security and compliance tools via a dedicated Compliance API. For an enterprise comms team operating in a tightly governed environment (Tencent’s compliance posture being a clear example), this opens the door for IT to govern Claude the same way they govern Microsoft 365 or Slack — central logging, retention policies, DLP rules. Not a workflow itself, but it unblocks the rest of them.
Anthropic restructures programmatic-Claude pricing starting June 15
From June 15, programmatic usage of Claude (Agent SDK, GitHub Actions, third-party frameworks) will be metered separately from your standard Pro or Max chat subscription. Each tier comes with a monthly credit roughly mirroring the subscription price (Pro $20, Max 5x $100, Max 20x $200). Worth budgeting for now if you run automation against the API; otherwise you could be surprised at month-end.
Practical pattern: Obsidian → n8n → Airtable, with AI in the middle
An n8n template worth bookmarking: highlight a question in an Obsidian note, fire it to a webhook via the Post Webhook plugin, let an OpenAI node interpret the query, pull the answer from Airtable, and write it back into your note. For a knowledge-management workflow built on Obsidian and DEVONthink, this is the cleanest “ask my own notes” pattern that doesn’t require a custom RAG pipeline.
Comms-relevant data point: 82% of journalists now use AI
Muck Rack’s 2026 State of Journalism report (covered this week) finds AI adoption among journalists at 82%, up from 77% in 2024. ChatGPT leads at 47%, Gemini at 22%, Claude at 12%. Forty-three percent of journalists say they rarely receive pitches that match what they actually cover — a hard prompt to revisit PR pitching workflows and to consider AI-assisted personalization at the reporter level.
Tencent Mentions
Tencent doubles down on agentic AI with latest Hunyuan updates
The most consequential Tencent AI story of the week. Hunyuan Turbo S, Hunyuan T1, T1-Vision, Hunyuan Voice, Hunyuan Image 2.0, Hunyuan 3D 2.5 and Hunyuan-Game all announced on May 21 at the Tencent Cloud industry summit. T1 benchmarks improved meaningfully on agent tasks (+13%). Likely to drive a wave of “China AI is catching up” coverage in Western outlets through the weekend.
Tencent unveils Hunyuan Hy3 preview model
Companion announcement focused on the Hy3 preview’s enhanced multimodal capabilities and the integration story across Tencent applications.
Tencent’s Q1 revenue miss heightens pressure for faster AI payoff
Earlier in May but worth flagging as media context. Q1 revenue rose 9% to RMB 196.5 billion — the slowest growth in six quarters — while net income climbed to RMB 58.1 billion. President Martin Lau confirmed Hunyuan + Yuanbao spend will more than double to over RMB 36 billion in 2026. Expect to see this framing recycled in every story about this week’s Hunyuan launch.
Tencent pledges higher AI investment after chip curbs hit capex plans
Reuters via Investing.com: Tencent’s leadership reiterating the 2026 AI investment commitment, with explicit reference to the impact of US chip export restrictions on capex planning. Useful background context for any comms outreach on Hunyuan in Western markets.
Generated by Claude · AI Intelligence Report for Cam MacMurchy
Sources include web search, 143 Substack subscriptions, Inoreader AI folder, Reddit, Discourse forums, Hacker News, and Product Hunt.