OpenAI unveiled its first custom silicon — a chip called Jalapeño, designed with Broadcom and built specifically to run large language models (the “inference” step, meaning the work of actually answering your prompts rather than training the model). The pitch is faster, cheaper, more reliable AI, and less dependence on Nvidia, whose chips have powered nearly the entire industry. It is OpenAI’s clearest move yet to control its own costs and supply chain. The day was paired with a quieter rollout of an updated GPT-5.5 Instant. For communications watchers, the read is simple: the AI giants are now becoming hardware companies, and that reshapes the competitive map.
Anthropic (maker of Claude) told the U.S. Senate Banking Committee that operators linked to Alibaba’s Qwen lab ran the largest “distillation attack” it has ever seen — using roughly 25,000 fraudulent accounts to harvest 28.8 million conversations with Claude between April 22 and June 5. Distillation means quietly training your own cheaper model on the outputs of a more capable rival, effectively copying its skills. Anthropic says the campaign targeted Claude’s most prized abilities in software engineering and agentic reasoning, and that it dwarfs three earlier campaigns it attributed in February to DeepSeek, Moonshot and MiniMax. This is now squarely a U.S.–China policy story, and a sensitive one for any China-headquartered tech company to track.
Anthropic launched Claude Tag, which replaces its old “Claude in Slack” app with a persistent AI teammate that lives in a channel under its own identity — building memory, monitoring conversations, taking initiative and working asynchronously alongside the humans on a team, rather than answering one person at a time. It is in beta for Enterprise and Team customers now; organizations that do nothing will be auto-migrated on August 3. Anthropic says its own product team now has 65% of its code written by an internal version of this agent. For a comms team coordinating across time zones in Slack, this is the most directly usable launch of the week — see the Workflows section below.
Azeem Azhar’s Exponential View published what it calls the first bottom-up, deduplicated measure of consumer and enterprise AI spending: roughly $110 billion in sales over the past 12 months, with an annualized run rate above $175 billion. The argument is that AI revenue has reached a “tipping point” suggesting the hundreds of billions being spent on it may be economically sustainable. Two important caveats for anyone quoting it: the figures are modeled from public statements, leaks and internal estimates rather than audited filings, and version one deliberately excludes all China activity, internal efficiency gains and professional services — so it captures only a defined slice of the market.
Noam Shazeer — a co-author of the foundational 2017 “Attention Is All You Need” paper that made modern AI possible, and a co-lead of Google’s Gemini models — is leaving for OpenAI, a move that surfaced widely across this week’s newsletters. Google had paid about $2.7 billion to bring him back from Character.AI less than two years ago. Coming alongside DeepMind’s John Jumper departing for Anthropic, it feeds a growing storyline that Google’s research bench is being raided as rivals race ahead. A clean example of how AI’s most valuable asset right now is people.
“Office workers are spending way too much on AI too”Critic David Gerard on the shift from flat subscriptions to per-token billing — and the surprise costs hitting enterprise AI users. See also · Inoreader AI Folder.
Colorado AI Act takes effect June 30, 2026One of the first comprehensive U.S. state AI laws goes live next week, after being postponed from February — a signal of America’s state-led patchwork approach.
A dedicated Substack inbox tool was not available for this run, so these highlights are drawn from the newsletter editions that landed in your Inoreader “Newsletters on AI” folder and Readwise Reader feed in the past 24 hours. Coverage of your AI-focused newsletters is below.
First bottom-up, deduplicated estimate of AI spending: about $110B in sales over 12 months, with an annualized run rate above $175B.
Azhar’s thesis: AI revenue has hit a “tipping point” that makes the industry’s enormous capital spending look potentially sustainable.
Key caveat for quoting: figures are modeled from public statements and leaks (not audited), and version one excludes all China activity, internal efficiency gains and professional services.
Makes the case that letting AI agents do real work requires purpose-built tooling — not just bolting chat onto existing apps.
A more technical read, but the throughline (workflows must be redesigned around agents) applies to any team adopting AI seriously.
From Your AI Feeds
Inoreader AI Folder.
Nine fresh items from your “AI” folder in the last 24 hours, dominated by OpenAI’s first chip, Anthropic’s move into Slack, and the China-distillation dispute. Duplicates from mirror feeds have been collapsed.
A skeptical look at how vendors (Microsoft, Anthropic, OpenAI) are moving customers from predictable monthly subscriptions to per-token billing, where costs can balloon. Worth a scan if you ever brief leadership on AI budgets.
Newsletter coverage of the OpenAI–Broadcom Jalapeño inference chip, plus the new GPT-5.5 Instant. The clearest sign yet that the model labs are becoming chip companies.
Pairs the Jalapeño chip with the day’s other big story: Anthropic’s accusation that Alibaba ran the largest-ever attempt to extract Claude’s capabilities. A handy two-in-one summary of the day.
Covers Anthropic’s Claude Tag — the persistent Claude agent now living inside Slack channels. The single most directly useful launch of the week for a Slack-based comms team.
A second take on the same launch, emphasizing that Anthropic has replaced its old Slack app with an agent that monitors the channel and acts autonomously.
The week’s most-cited analysis piece, putting hard numbers on AI revenue ($110B in sales; $175B run-rate) while flagging that China and internal-efficiency value are excluded.
OpenAI’s own research write-up on AI agents handling longer, more complex tasks across roles. Useful source material if you ever need OpenAI’s framing on AI and jobs.
Media-intelligence vendor Signal AI launches a meaning-based search that moves beyond Boolean keywords — directly relevant to how comms teams monitor coverage and reputation.
Discoveries
Workflows & Tool Watch.
Claude Tag: a standing AI teammate inside Slack
Anthropic’s Claude Tag puts Claude permanently in a Slack channel under its own name. Unlike the old app you had to summon, it watches the conversation, remembers context, and can take on tasks — drafting, summarizing threads, chasing action items — across everyone in the channel. For a comms team running fast-moving issues across time zones, this is worth piloting in one channel first (start read-only) before giving it room to act. It is in beta for Enterprise/Team; auto-migration from the old Slack app lands August 3.
RelevantClaude · Slack · Team coordination · Issues management
“Record a skill” — teach Claude a task by demonstrating it
This week’s Ben’s Bites highlights a pattern that fits your “automate without coding” approach perfectly: instead of writing instructions, you record yourself doing a repetitive task once, and Claude turns that into a reusable skill it can run later. Think of it as recording a macro, but Claude figures out the steps. A natural fit for recurring comms chores — formatting a media list, prepping a standard briefing, tidying transcripts from MacWhisper.
Obsidian + Claude: the safe way to let AI into your vault
A widely shared 2026 guide pulls together the community’s hard-won lessons on connecting Claude to an Obsidian vault. Two stand out for you: first, always put the vault under git version control before giving Claude write access — one cautionary Reddit thread (“the dark side, rampant destruction”) shows how an agent can corrupt notes, and git lets you instantly undo. Second, a short USER.md “identity” file describing who you are improved the quality of Claude’s suggestions more than any fancy retrieval setup. There’s also an MCP server that tunnels your local vault to Claude on mobile.
Reliably capturing system audio on macOS for transcription
A new technical explainer from Recall.ai walks through how apps can tap your Mac’s system audio — the sound coming out of your speakers, like a Zoom or VooV call — and the real-world limits of Apple’s newer ScreenCaptureKit and Core Audio Taps APIs. If you ever wrestle with getting clean call or interview audio into MacWhisper for transcription, this is the clearest current map of what works and what doesn’t.
RelevantMacWhisper · Zoom · VooV Meeting · Podcast production
Product Hunt trend: AI that embeds where you already work
The week’s notable Product Hunt launches share one move — they don’t ask you to open a new app, they live inside surfaces you already touch. Examples include Dune Keypad (a physical keypad with Claude built in), the Wispr Flow voice-dictation tool, and the Granola meeting notetaker, alongside automation staples like n8n. The takeaway for your stack: the best new tools reduce friction by meeting you in your keyboard, your calls, or your existing automations rather than adding another tab.
Tencent confirmed it is testing “Xiaowei,” a native AI assistant, on a small scale within Weixin (the Chinese version of WeChat), letting users call up and operate mini-programs through natural-language conversation. CNBC frames it as Tencent stepping up to challenge domestic rivals — a story likely to keep generating reporter questions about WeChat’s AI roadmap and timing.
Coverage of Tencent’s plans to significantly increase AI investment underscores the capital intensity of the race, and pairs with reporting that Hunyuan ranks among the top tier in domestic benchmarks while Yuanbao’s monthly active users have passed 100 million.
Per Tencent’s published schedule, older Hunyuan model versions such as HY 2.0 are slated to begin retiring starting June 26, 2026, as newer Hy3-generation models take over. A small but useful date to have on hand if developers or partners ask about model continuity.
No direct Tencent mention surfaced in today’s AI news, but the Anthropic accusation against Alibaba’s Qwen lab is the China-AI story most likely to come up in reporter conversations this week. The framing — U.S. labs alleging Chinese rivals are illicitly copying model capabilities — is the relevant backdrop any China-headquartered tech communicator should be ready to address.