OpenAI released the first performance results for Jalapeño, the custom chip it designed with Broadcom to run AI models rather than train them. In plain terms: instead of renting Nvidia hardware for every answer ChatGPT gives, OpenAI now has silicon built for its own workloads. Tested against GPT-OSS 120B, DeepSeek R1 and Kimi K2.5, Jalapeño delivered 1.5 to 1.9 times more work per watt and answers 1.7 to 3.6 times faster end-to-end than comparison systems. Sam Altman summed it up as “we made a chip and it is fast.” OpenAI says Jalapeño is the first generation of a long-running platform where chips, memory, models and products are designed together — a direct challenge to Nvidia’s pricing power and a signal that the cost of running AI is about to fall.
A powerful free AI model called Ox Alpha appeared on OpenRouter, OpenCode, Klio and Nousportal with no announcement, no marketing and no named owner. OpenRouter states only that it is “developed and operated by a third-party provider who has chosen to remain anonymous,” and that OpenRouter itself neither builds nor operates it. Developers report it outperforming paid frontier models on coding work. Independent researchers fingerprinting its tokenizer say it matches Zhipu AI’s GLM-5.3 family on every normalised test, though Google, Xiaomi, DeepSeek and Microsoft’s MAI have all been floated. The communications angle is the interesting one: thousands of engineers are sending proprietary code to servers whose jurisdiction, data-retention policy and owner are all unknown — a governance story waiting to break.
Elon Musk said SpaceX’s first AI satellites — running exclusively on Nvidia chips under the Starmind programme — will launch in the fourth quarter of 2027 and reach “significant scale” in 2028. The pitch is that space offers unlimited solar power and free cooling, the two constraints throttling data centres on Earth. Nvidia’s contribution is the Vera Rubin Space Module, built around its IGX Thor and Jetson Orin platforms and engineered for the severe size, weight and power limits of a satellite; Nvidia claims up to 25 times more AI compute for orbital inference. Musk projects SpaceX AI compute capacity above 2GW by the end of this year and close to 10GW by the end of 2027. Treat the timeline with the usual Musk discount, but the Nvidia exclusivity is real and signed.
Stanford’s latest “Canaries in the Coal Mine” update finds that employment among 22-to-25-year-olds in the occupations most exposed to AI now sits roughly 19% below where it would be had it tracked their peers in less-exposed jobs. That gap was 15% a year ago. Since 2022, employment for young workers in the 40% of occupations most exposed to AI has fallen about 11%, while the least-exposed 60% has risen about 10%. Crucially, the researchers find no evidence of broad economy-wide displacement: the adjustment is happening through reduced hiring of juniors, not layoffs, and experienced workers show flat or rising employment. This is the single most quotable AI-and-jobs dataset in circulation right now, and it will be cited at you in interviews.
The US Federal Trade Commission issued a proposed policy statement for public comment setting out its intention to enforce against certain forms of personalised pricing enabled by AI — that is, algorithms that quietly charge different customers different prices for the same product based on inferred willingness to pay. Separately, Representative Suhas Subramanyam (D-VA) said he will amend the FRONTIER Act at a September markup to address the containment of autonomous AI agents. Both moves land against a US backdrop that Andreessen Horowitz executives this week described as a “patchwork” of state laws with “ever-shifting goalposts,” and roughly three weeks after the European Commission began actively enforcing the EU AI Act on 2 August.
OpenAI launches an Admin plugin for ChatGPT Work and CodexWorkspace administrators can now review usage and credits, manage members and permissions, adjust limits and action requests through conversation rather than a settings panel. Relevant for any comms team running ChatGPT at scale.
Codex can now hand routine tasks to cheaper modelsOpenAI’s coding agent will delegate simpler sub-tasks to lower-cost models — the same cost-control pattern now appearing across agent platforms, where an expensive model supervises cheap ones.
xAI’s Grok Bot gives every task its own cloud computerIn beta on macOS, iOS, Windows and Linux, with Android to follow. Each task gets a persistent cloud machine that keeps working after you close your device — but access starts at USD 300 per month via SuperGrok Heavy.
Anthropic IPO odds climb toward an October listingPrediction-market pricing on an Anthropic IPO rose sharply in late August. The company closed its Series H in May at a USD 965 billion valuation on USD 65 billion raised, with Morgan Stanley, Goldman Sachs and JPMorgan underwriting.
Forge · Aug 25
08
OpenAI’s S-1 is expected on EDGAR imminentlyThe filing would give the first look at audited financials, Microsoft revenue-share terms and offering structure ahead of a September listing target — reportedly USD 2 billion a month in revenue against a projected USD 14 billion loss for 2026.
OpenAI disrupts a Russian covert influence campaignOpenAI banned Russia-origin accounts that used its models to promote a fabricated Israel-based think tank and a “sovereignty index” praising Russia and criticising the West. A useful case study in AI-enabled information operations.
EU AI Act transparency rules are now actively enforcedSince 2 August the Commission’s AI Office and national authorities have been enforcing general-purpose AI obligations and new transparency requirements — the compliance clock is running for anyone shipping AI features into Europe.
European Commission
18
Taiwan indicts nine in an AI server diversion caseReported by Superpower Daily: prosecutors allege AI servers were diverted in breach of export controls — the enforcement side of the chip-restriction story that usually only appears as policy.
Nvidia launches its Space Computing platform for orbital inferenceThe Vera Rubin Space Module, IGX Thor and Jetson Orin are now positioned as a product line for size-, weight- and power-constrained environments — space first, but the same constraints apply to vehicles and edge deployments.
Nvidia
22
Google’s A2A protocol joins the Agentic AI Foundation alongside Anthropic’s MCPBoth major agent-interoperability standards now sit under neutral Linux Foundation governance, with 250-plus members including AWS, Anthropic, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. In practice: fewer proprietary lock-ins for enterprise agent deployments.
The dedicated Substack tools were not connected for this run, so newsletter coverage below was assembled from the same newsletters as they arrive in your Inoreader feeds (via kill-the-newsletter and the Newsletters on AI stream). Coverage is complete for the past 24 hours; paid-tier full text was not accessible without the Substack connection.
The week’s theme is agents moving off the desktop and onto phones, with OpenClaw arriving free on iOS and Android and xAI’s Grok Bot shipping paid always-on agents.
The architectural shift that matters: agents now run on a persistent cloud machine rather than inside the app, so work continues after you close your phone.
Ben’s practical read is that mobile agents are still best for monitoring and retrieval rather than execution — useful for a comms lead wanting alerts, not autonomous posting.
Azhar’s weekly data roundup across AI, energy and markets, drawn from Exponential View’s own tracking of the AI economy.
Headline finding: the share of inference running on open-weight models has doubled in the last twelve months — meaning more AI work is now done on models anyone can download and self-host rather than closed APIs.
The implication for corporate policy is direct: “which model are we using” is no longer a question with one answer per vendor.
Argues that running several frontier models in parallel produces less diversity of thought than expected, because the leading models increasingly converge on similar answers.
Draws on Exponential View’s own research into model homogeneity — “great minds think (a little too much) alike.”
Worth reading before commissioning any multi-model review process: the redundancy may be buying less assurance than it appears.
Leads on the Nvidia-SpaceX orbital compute story and Nvidia’s move into CPUs, which puts it in direct competition with Intel and AMD rather than just complementing them.
Secondary item: reports that Anthropic’s flagship model is losing benchmark ground to cheaper rivals, feeding the “is frontier pricing sustainable” debate.
The Neuron remains the most efficient daily read for a non-technical audience — plain-language framing throughout.
Covers OpenAI’s task-delegation feature in Codex alongside the Stanford employment findings and Taiwan’s indictment of nine people in an AI server diversion case.
An earlier edition the same week covered OpenAI’s decision to pause frontier training while AI control research catches up.
Consistently the densest newsletter in the stack for hard news items with links.
A technical but readable history of the “harness” — the scaffolding around a model that manages tools, memory and retries, and which increasingly determines whether an agent is useful.
Core claim: the harness, not the model, is now where most real-world performance differences come from.
Useful background for understanding why Claude Code, Codex and Cursor produce different results from similar underlying models.
A quarterly catch-up edition rather than a weekly one — the best single document for reconstructing the summer’s model releases and policy moves in order.
Recommended if you need to brief someone who has been away.
Reads OpenAI’s frontier-training pause as partly performative positioning ahead of its IPO.
Covers Anthropic’s debt financing and argues Chinese labs are deliberately using open-weight releases as competitive strategy — directly relevant background for China tech coverage.
Argues Amazon’s AI position rests on many pillars — chips, cloud infrastructure, enterprise inference, commerce and robotics — so it profits regardless of which model wins.
A useful counter-narrative to the “Amazon is losing AI” framing common in tech press.
Analyses how companies handled the USD 168 billion in US tariff refunds distributed from April after the Supreme Court invalidated the tariffs in February.
The reputational question: whether firms passed savings to customers or retained them — and how media coverage tracked the difference.
The closest thing in your stack to a pure corporate-communications case study this week.
From Your AI Feeds
Inoreader AI Folder.
Twelve articles in the last 24 hours from your “AI” folder, after removing duplicates where the same newsletter arrives through both a direct feed and the Newsletters on AI stream. Coverage is dominated by two stories: OpenAI’s Jalapeño chip results and Nvidia’s move into orbital compute.
OpenAI’s own write-up of the Jalapeño benchmarks, tested against GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. The company claims 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than comparison systems, breaking the usual trade-off between throughput and speed.
OpenAI CFO Sarah Friar explains how gains across chips, compute, models and products compound to deliver more capability at greater scale and lower cost. Read alongside the expected S-1 filing, this functions as investor-facing narrative as much as technical explanation.
Administrators can review workspace activity and credit usage, manage members and permissions, adjust limits and act on requests through a single conversation rather than navigating a settings interface. For enterprise deployments this makes AI governance considerably less laborious.
OpenAI banned a network of Russia-origin accounts that used its models to build out a fake Israel-based think tank and publish a “sovereignty index” flattering Russia and criticising Western countries. A concrete example of generative AI being used to manufacture the appearance of independent expert opinion.
Newsletter coverage of the Nvidia-SpaceX orbital data centre programme, with a secondary theme that AI is increasingly moving into physical systems rather than staying in software. Arrived twice in your folder, via both the AI Valley feed and Newsletters on AI.
Covers Nvidia’s entry into CPUs — putting it head-to-head with Intel and AMD — alongside the orbital compute launch. The secondary item reports that Anthropic’s flagship model is under pressure from cheaper competitors on benchmark performance.
Ben Tossell on the shift of AI agents from desktop to phone, prompted by OpenClaw’s free iOS and Android release and xAI’s paid Grok Bot. The technical change that matters is persistence: agents now run on cloud machines that keep working after you put the phone down.
Codex can now route simpler sub-tasks to lower-cost models, cutting the bill for long agentic runs. The same edition covers the Stanford finding on AI exposure and young workers, and Taiwan’s indictment of nine people over diverted AI servers.
A quarterly consolidation issue covering the summer’s model releases, funding rounds and regulatory developments in sequence. The most efficient way to reconstruct what happened between June and August if you need to brief someone quickly.
Argues that elaborate prompting compensates for a poorly built system. As models get better context and tooling, instructions should shrink rather than grow — a useful corrective if your team maintains long prompt templates.
A consumer-safety flag on a rapidly growing AI application, bundled with a promotional segment on free AI assistants for small businesses. Skim the first section; the rest is marketing.
On Ox Alpha, with the sharpest framing of the risk: the model retains everything sent to it and there is no published retention or jurisdiction policy. If anyone on a comms or legal team asks why an anonymous free model is a problem, this is the paragraph to send them.
See also · Top Stories · Substack Highlights
Discoveries
Workflows & Tool Watch.
DEVONthink now speaks MCP — and supports Claude Fable 5
DEVONtechnologies’ latest maintenance release adds support for Anthropic’s Claude Fable 5 and for image diffusion models running locally through Ollama. More significant for your setup: DEVONthink can now connect to any AI application that supports the Model Context Protocol — Claude and Codex included — with granular controls to restrict which databases are reachable, redact private information and exclude sensitive documents entirely. There are also four new AppleScript commands for locating and applying bibliographic metadata, and AI actions can now be embedded directly into DEVONthink automations using new triggers and placeholders. In plain terms: you can now ask Claude a question and have it search your DEVONthink archive to answer, without handing over the whole database.
RelevantDEVONthink · Claude · Codex · Ollama · AppleScript
Claude Code’s August release adds Remote Control and a plugin marketplace
Anthropic shipped a substantial Claude Code update this month. Remote Control lets you resume the most recent session from another machine with claude remote-control –continue, which matters if you start something on the Mac mini at home and want to pick it up on the MacBook Pro. A plugin marketplace command source now re-resolves the plugin directory each session and applies changes without a restart. On the reliability side, MCP OAuth was fixed for strict authorization servers by using 127.0.0.1 rather than localhost in the redirect URI — worth knowing if any of your MCP connections have been failing to authenticate. Streaming also gained keepalive pings to stop long thinking pauses from timing out.
RelevantClaude Code · MCP servers · Cowork · Mac mini · MacBook Pro
The Admin plugin makes AI governance a conversation instead of a spreadsheet
OpenAI’s new Admin plugin for ChatGPT Work and Codex, released yesterday, lets a workspace administrator ask a question, explore the detail, make an authorised change and confirm the result in one conversation — reviewing activity and credit usage across the organisation and identifying teams that need more enablement. For a global comms function running AI across multiple time zones, this is the difference between quarterly usage audits and continuous oversight. It also creates an audit trail in natural language, which is easier to hand to legal than a log export.
RelevantChatGPT Work · Codex · team coordination · media relations governance
OpenClaw brings a free, open-source agent to iPhone and iPad
OpenClaw is now available as a native app on iOS and Android. You pair your phone with the OpenClaw Gateway, a routing layer that connects your requests to AI agents and the tools those agents can reach. Because it is open source and self-hostable, it is the only mobile agent option here that does not require sending your workflow through a vendor’s cloud — relevant if you want to experiment with agentic monitoring of media coverage without a procurement conversation. Start it in read-only mode: agents with write access to your accounts are not something to trial casually.
RelevantiPad Pro · iPhone · media monitoring · MCP
Task delegation is becoming the standard way to control agent costs
OpenAI’s decision to let Codex hand simpler sub-tasks to cheaper models is part of a wider pattern: an expensive model plans and supervises, cheap models execute. If you are budgeting for any agentic workflow — transcribing and summarising a week of podcast recordings, for instance, or sweeping coverage across dozens of outlets — this is the architectural choice that determines whether the bill is manageable. Ask any vendor pitching you an agent whether it supports model tiering; if the answer is no, the running cost will be several times higher than the demo suggests.
RelevantCodex · Claude Code · n8n · Zapier · Make · podcast production
Agent standards consolidate: A2A joins MCP under the Linux Foundation
Google’s Agent2Agent protocol has formally joined the Linux Foundation’s Agentic AI Foundation, the same neutral body that governs Anthropic’s Model Context Protocol. The foundation now has more than 250 members including AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. The practical consequence for anyone building on MCP — which increasingly means anyone connecting DEVONthink, Obsidian, Notion or Things 3 to an AI assistant — is that the protocol is now less likely to be changed unilaterally by a single vendor. That is a meaningful stability signal for the tooling you have been assembling.
agent-device: an MCP server for driving mobile apps
Callstack has published agent-device, an open-source MCP server plus CLI and Node.js API that lets an AI agent drive and verify apps across iOS, Android, HarmonyOS, TV, web, macOS and Linux. It is built for developers testing their own apps, but the underlying capability — an agent that can operate a phone app and confirm what it sees — is the missing piece for anyone wanting to automate checks on how content actually renders in WeChat, LinkedIn or X on a real device rather than in a preview.
RelevantMCP · WeChat · social publishing QA · Claude Code
Ox Alpha is free, fast and available through OpenRouter, and it will be tempting. It is also operated by an anonymous provider with no published data-retention policy, no named jurisdiction and no accountable owner. WhatPlugin.ai reports that it retains everything sent to it. For a corporate communications function handling embargoed announcements, draft statements and issues material, this is a straightforward no — and it is worth a note to your team before someone pastes a holding statement into it because a colleague recommended it.
Relevantissues management · information security · team policy
Nothing in today’s news cycle, your Inoreader AI folder or the newsletter stack names Tencent directly. The items below are the most recent Tencent AI lines a reporter is likely to raise, with the relevant numbers to hand.
Tencent Hy3 has been available free to users worldwide through WorkBuddy since the global rollout, and that free period runs to 31 August 2026 Pacific Time — this Monday. Hy3 is also accessible via Tencent Design Miora and Tencent Cloud TokenHub. Expect questions about what happens on 1 September, and about conversion from free users to paid.
Second-quarter revenue rose 11% year-on-year to RMB 204.8 billion, ahead of the RMB 202.8 billion consensus, with non-IFRS net profit of RMB 68.4 billion up 9%. Domestic games rose 17% to RMB 47.3 billion and marketing services rose 22% to RMB 43.6 billion. Capital expenditure rose 176% to RMB 52.8 billion, pushing free cash flow to negative RMB 13.8 billion, and losses from new AI products widened to RMB 10.5 billion. The ADR fell 5.34% on the day. The framing to expect from reporters is “is the AI spending justified” — the counter-data point is that Hunyuan 3 preview has held the top ranking on OpenRouter by token usage since 28 April, and that WorkBuddy is described by management as the most widely used productivity AI agent in China.
Hunyuan 3 uses 295 billion total parameters with 21 billion active, reports 90% task completion on internal agent benchmarks, and cut API input pricing to RMB 1 per million tokens. Given today’s Ox Alpha story — where fingerprinting points at Zhipu AI’s GLM-5.3 — it is worth being ready for questions about whether any Tencent model has appeared anonymously on routing platforms. It has not.