AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #54

Owning the Stack

Thursday, June 25, 2026 · 3 Developments · 2 Perspectives · 2 Case Studies · 1 Resource

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

The frontier labs spent today integrating downward, into silicon and scarce talent, while the work that actually pays you keeps moving upward, into judgment and governance. (Thursday edition; 24-hour lookback.)

The through-line for operators: as compute and talent consolidate at the bottom of the stack, your edge is everything at the top — deciding what to build, getting chosen, and governing the agents that execute.

Developments

Technology

OpenAI builds its own silicon: the Broadcom-made Jalapeño chip (June 24, 2026)

On June 24 OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom chip: an inference-only accelerator co-developed from design to tape-out in nine months, with OpenAI's own models helping optimize the work. OpenAI says early testing shows performance per watt "substantially better than current state-of-the-art," with initial deployment by the end of 2026 as the first step in a multi-generation platform aimed at gigawatt-scale data centers. The move pushes OpenAI down the stack into chip architecture, kernels, memory, and networking, and reduces its reliance on Nvidia.

So what

The labs are vertically integrating to attack the one cost they cannot prompt away: inference. If OpenAI's perf-per-watt claims hold, cheaper served tokens reach everyone building on its models, and the gap between frontier quality and frontier cost keeps narrowing. The action: treat your AI cost-of-goods as a moving target, revisit pricing and gross-margin assumptions anchored to today's inference rates, and avoid architecting your product around any single provider's current price.

Confidence: Established Fact — announced by OpenAI and Broadcom and corroborated by CNBC, TechCrunch, and Tom's Hardware; the perf-per-watt figure is OpenAI's own early claim.

Go deeper OpenAI (announcement) · Curated: CNBC

People & Organization

Two more Gemini researchers leave Google for Anthropic (June 24, 2026)

Bloomberg reported on June 24 that Jonas Adler and Alexander Pritzel, both viewed internally as key Gemini contributors, are leaving Google for Anthropic, the second senior pair to make that move in a week. They follow DeepMind director and Nobel laureate John Jumper to Anthropic and Noam Shazeer's exit to OpenAI, part of a steady drain of frontier talent from Google. Adler worked on Google's AI coding efforts and Pritzel on model training. TechCrunch framed the moves as an intensifying talent war among the labs.

So what

Research talent, not compute, is the scarce input the leading labs are now fighting hardest over, and where it concentrates shapes which model you will be building on in eighteen months. For operators this is a supplier-stability signal, not industry gossip. The action: track where the people behind your core model are moving, keep a tested fallback model in your stack, and weight provider roadmaps by the strength of the teams still shipping them.

Confidence: Established Fact — reported by Bloomberg and corroborated by TechCrunch and Neowin; the departures are described as planned. Developing: extends the Jumper and Shazeer moves covered earlier this week.

Go deeper Bloomberg · Curated: TechCrunch

Technology

OpenAI pushes Codex beyond code into knowledge work

OpenAI is repositioning Codex from a coding tool into a general knowledge-work agent, adding role-specific plugins, a "Sites" output layer, and a Record & Replay feature that watches you complete a Mac workflow once and drafts an editable, natural-language "skill" it can rerun. Knowledge workers are now roughly 20% of Codex users and reportedly growing more than three times faster than developers. Record & Replay shipped in mid-June for macOS on paid and business plans, outside the EU, UK, and Switzerland. Every documented the shift in its Context Window column and adoption guide.

So what

This lowers the barrier to automation from "can write code" to "can do the task once on screen," putting agent-building in reach of operations, finance, and marketing staff. That is where most repetitive white-collar work actually lives. The action: have one non-engineering team pick a recurring manual workflow, record it, and measure time saved against error rate, then decide whether to standardize the resulting skill or keep a human checkpoint.

Confidence: Established Fact — features have shipped and are documented; adoption and reliability at scale are still early, and the rollout excludes the EU, UK, and Switzerland.

Go deeper OpenAI Developers (Record & Replay) · Curated: Every

Perspectives

People & Organization

KPMG: worker AI-agent use jumps to 56%, but pragmatism sets in

KPMG's Q2 2026 AI Quarterly Pulse finds employee adoption of AI agents rose to 56% from 23% the prior quarter, even as leaders shift from experimentation toward ROI discipline and pragmatism. Where workers resist, the drivers are increased workload or complexity (58%) and job-security fears (58%), not technophobia. Roughly a third of organizations are now deploying and scaling agents, with many more orchestrating several at once, and most still require human review of agent outputs. The Deep View summarized the quarter's read on enterprise adoption.

So what

Adoption is no longer the constraint; absorption is. The gap is now between employees using agents and organizations redesigning work, governance, and incentives around them. The action: pair every agent rollout with an explicit answer to the workload and job-security concerns your own people will raise, keep a human-review step on consequential outputs, and measure ROI per workflow rather than declaring victory on usage counts.

Confidence: Informed Perspective — survey data self-reported by respondents; a curated newsletter summary cited different headline figures, so numbers here are anchored to KPMG's published Q2 release.

Go deeper KPMG (Q2 2026 AI Pulse) · Curated: The Deep View

Go-to-Market

Cloudflare's CEO: AI agents could "destroy small businesses" (June 23, 2026)

At an Axios event in Cannes on June 23, Cloudflare CEO Matthew Prince warned that "AI is going to destroy small businesses," because shopping agents compare thousands of products with no brand loyalty, making it nearly impossible for lesser-known companies to get chosen. Without a way for new entrants to be discovered and survive, he argued, agent-mediated commerce drives "massive consolidation." The remarks extend this briefing's running thread on AI eroding click-through and brand discovery.

So what

As buyers delegate choice to agents, the contest shifts from winning human attention to being legible and selectable by machines, and incumbents with structured data and integrations start with the advantage. Discoverability is becoming an engineering problem. The action: make your product and pricing machine-readable, get into the data sources and catalogs that agents query, and start testing how leading shopping and research agents represent you versus larger rivals.

Confidence: Informed Perspective — a CEO's argument, directionally important but not a measured outcome.

Go deeper Axios · Curated: The Rundown AI

Case Studies

AI Strategy & Transformation

How Every is fine-tuning a model to clone its editor's judgment

At media-and-software company Every, CEO Dan Shipper fine-tuned models on roughly 30,000 of editor-in-chief Kate Lee's real copy-edit decisions, using Tinker, a fine-tuning API from Thinking Machines, after prompts and style guides failed to capture her judgment when rules collided. The bar for success is exacting: the model must catch nine of ten of Kate's edits, Kate must accept nine of ten of its suggestions, and it must make no serious errors. The lesson the team draws is that tacit judgment is learned from examples, not written down in instructions.

So what

This is a concrete template for capturing expert judgment that lives in someone's head rather than in a document, the kind of knowledge that walks out the door when a key person leaves. Fine-tuning on real decisions is becoming a practical alternative to ever-longer prompts. The action: identify one high-value judgment your business depends on, start logging the expert's actual decisions now, and treat that dataset as an asset whether or not you fine-tune this year.

Confidence: Informed Perspective — a firsthand company account; Tinker is a real product, but the results are self-reported and the experiment is ongoing.

Go deeper Every (Context Window) · Curated: Thinking Machines (Tinker)

Operations & Finance

HP cut PC refreshes about 28% using AI fleet management

On The Neuron's podcast, HP's Larry Meadows described how HP's Workforce Experience Platform shifts enterprise IT from reactive ticketing to prediction: agents on Windows, macOS, and Chromebook devices feed a data lake spanning tens of millions of machines that flags issues early and recommends whether to fix, upgrade, or replace each device. By matching hardware to how employees actually use it rather than refreshing on a fixed cycle, HP says it cut its own PC refreshes by roughly 28%. HP's Imagine 2026 update expanded the platform's AI-driven insights and reporting.

So what

This is a near-term, measurable AI ROI story in an unglamorous place, hardware lifecycle, where data-driven decisions replace calendar-based capex. The pattern generalizes to any asset you refresh on a schedule rather than on evidence. The action: ask whoever owns IT to pull real utilization data before the next hardware refresh and right-size the spend, then look for other "replace on a timer" budgets where usage data could defer or avoid cost.

⚠ Hype check: The ~28% reduction is HP's own internal result for an HP product, shared on HP's podcast. Treat it as a vendor benchmark to test against your own fleet, not an independently audited figure.

Confidence: Established Fact — HP's product and its own reported results; the 28% figure is company-reported.

Go deeper HP (Imagine 2026) · Curated: The Neuron

Resources

Security

A planted agent "skill" hijacked 26,000 agents and passed every scanner

Security firm AIR built a malicious agent skill called "brand-landingpage," gave it credibility with GitHub stars, a marketplace merge, and an Instagram ad, and reached roughly 26,000 agents belonging to designers, marketers, and their companies. The skill carried no malicious code itself; it pointed agents to setup instructions hosted on a domain AIR controlled, which it could rewrite after the security scan cleared. Scanners from Cisco, Nvidia, and skills.sh all marked it safe, because they inspect a skill's local files, not the external content it fetches at runtime.

So what

Agent skills and MCP connectors are a software supply chain, and a clean one-time scan is not a safety guarantee when a skill reaches out to mutable external instructions. Anyone letting agents install community skills has just inherited a new attack surface. The action: allowlist agent skills and MCP servers, pin or vendor the external resources they call, and require review of any skill that fetches instructions from a domain you do not control.

Confidence: Established Fact — a documented proof-of-concept by AIR, corroborated by CSO Online and The Next Web.

Go deeper AIR (The Story of Skills) · Curated: CSO Online

Sources in this edition: The Deep View, The Rundown AI, The Neuron, Every, Superhuman AI, TLDR AI, AINews

← Latest Editions and Search