June 9th, 2026

Tier 1: Palantir CEO weaponises the Anthropic/OpenAI cost backlash — your CourseBuilds wedge just got sharper

[Verified shipped] Karp went onstage outside SF last week and told customers they’re “foolish to do business directly with AI firms rather than with intermediaries.” His best line: “You’ll go to a large language model company and learn that they don’t care about you at all… you’re gonna pay a lot in tokens, and it’s gonna be very hard to understand how it helps you.” The kicker: Uber CTO Praveen Neppalli Naga told The Information Uber blew its full-year AI budget in the first few months of 2026 — pure Claude Code overrun. Palantir’s commercial head Ted Mabrey says there’s been an “explosion” in customer cost concern in the six weeks since. Mitigations spreading: model routers and per-employee usage caps.

Action: Build the “cost transparency dashboard + per-user router” demo into the CourseBuilds Aria wedge. Aria’s lease team is exactly the kind of buyer Karp is signalling to — high-volume, low-AI-literacy, terrified of bill shock. Lead the Zaicek conversation with “here’s how I stop your firm from being the next Uber CTO quote.”

Tier 1: Apple Siri AI ships on Gemini + Extensions API opens Claude as default — the iPhone is now an agent surface

[Verified shipped — dev beta live, fall 2026 GA] Apple rebuilt Siri ground-up on a reported ~$1B/year Google Gemini deal feeding Apple Foundation Models, with on-device + Private Cloud Compute. The Siri AI app lives in the Dynamic Island, reads screen context, acts across apps. The bigger story for you: iOS 27 Extensions let users swap ChatGPT, Claude, or Gemini as default assistant, and Xcode 27 ships first-class Anthropic, Google, and OpenAI coding agents in the IDE. No EU/China launch. Requires iPhone 15 Pro+.

Action: Two project hits. (1) Fillarup — when iOS 27 GA’s, fuel-cost queries become voice-native (“hey Claude, when should I fill the tank this week?”). Map a Claude-Extension surface for Fillarup’s core calc before the App Store gets crowded. (2) Always-On Reeve — see PAY ATTENTION.

Tier 1: Apple Shortcuts becomes natural-language buildable — automation gets commoditised on hundreds of millions of phones

[Verified shipped — iOS 27 fall] Apple’s Shortcuts app now accepts a plain-language description (“ping my partner when I leave work with my ETA”) and Apple Intelligence wires up the leave-work trigger, stored address, Apple Maps ETA, and Messages send. Editing works the same way — describe the change in natural language. Cecilia Dantas conceded on stage that Shortcuts “can feel complicated.” That layer just got generic.

Action: Anxiety-flip — most consumers will discover automation through Apple, not your stuff. The defensible PaperClip + Reeve play is multi-account orchestration, persistent memory, and cost discipline — things Apple Intelligence cannot offer because it lives in one user’s keychain. Update the Always-On Reeve positioning doc to drop any framing that competes with iOS-native Shortcuts.

You already built what Karp is selling — MACA has a per-run cost dashboard that most CIOs haven’t shipped

Palantir’s pitch is that Foundry/AIP plus FDEs give customers “dashboards showing per-task spend by model so customers can pinpoint cost overruns.” You shipped that into MACA two months ago: api/lib/costs.ts logs per-run cost across 14 agents in 4 waves, public/cost-dashboard.html renders it, and scripts/photo-costs.json tracks the photo pipeline. The Uber CTO blew his budget because no one wired that in. When you walk into Aria with CourseBuilds, you’re not pitching Palantir’s promise — you’re showing the working artefact. Receipts beat slides.

“The Model Is No Longer the Bottleneck” — k-dense.ai

https://www.k-dense.ai/blog/the-model-is-no-longer-the-bottleneck

The argument: the bottleneck has moved from model capability to the workflow around it. This is the exact failure pattern in MACA right now — the open issues are copy quality (human review pass) and UX/flow for a gym owner, not “Opus isn’t smart enough.” 30 minutes here will reframe how you triage MACA’s next sprint and what you tell Anil for the business-naming pipeline. Worth it before your next MACA session.

“Built to Benefit Everyone: Our Plan” — Sam Altman + Jakub Pachocki

https://openai.com/index/built-to-benefit-everyone-our-plan/

Altman’s framing of OpenAI’s “third phase” — automated AI researcher, AI-shaped economy, “personal AGI for everyone” — and a global coordination body that could pause frontier work. Skim it. Direct hit on AI Edge positioning: when Aria or RT asks “what’s OpenAI’s strategy now,” you need to be able to answer in 30 seconds without sounding like a TLDR summary. The post IS the positioning.

“Uber’s CTO Praveen Naga came out last week saying they blew their full-year AI budget in the first few months of 2026 — pure Claude Code overrun. Palantir’s commercial head says there’s been an explosion in customer cost concern in the six weeks since that quote. The mitigations spreading are model routers and per-employee usage caps. That’s the actual story under the Anthropic-projected-eleven-billion-this-quarter headline — enterprises are eating consumption-pricing alive.”

Use case: Drop this on Zaicek when CourseBuilds comes up, or on the RT AI team when someone asks “should we go all-in on Claude Code?” Signals: you read primary sources (The Information’s Bratton), you know the named CTO, you understand the consumption-pricing mechanics, and you’ve already shipped cost dashboards into MACA. Expert positioning without name-dropping models.

Goldman and JPMorgan are exploring compute futures trading as an AI financing hedge. Both banks (per The Information’s Yang/Campbell exclusive in today’s Applied AI) are scoping a derivatives market for GPU/compute capacity — treating compute the way they treat oil, gas, and grain. If this lands in the next 12 months, every enterprise AI consumption-pricing conversation gets a hedging layer attached to it, and CFOs start asking vendors about compute-cost certainty the same way they ask about FX risk.

Why it matters for you: queue, don’t act. Too early to build into CourseBuilds, but worth keeping a note in the AI Edge vault. When you eventually pitch Aria, “do you want a hedge against AI cost volatility?” is going to be a CFO-level question — and you’ll already have the frame loaded.

• [TLDR — “OpenAI S-1 🇺🇸”]: OpenAI filed confidential draft S-1, no timing — just IPO posturing, not actionable.
• [Rundown — “Third phase of OpenAI”]: Altman manifesto — better as deeper-look link, not a Tier 1 item.
• [TLDR — “Xiaomi MiMo UltraSpeed 15x faster”]: 1T-param FP4 inference at 1000 tok/sec — interesting infra benchmark, no Roy hook.
• [Rundown — “Argentina non-human corporations”]: Milei novelty law, AI-owned companies — zero project relevance.
• [Rundown — “NotebookLM agentic chat”]: Google added sandboxed code exec to NotebookLM — not in Roy’s daily stack.
• [Rundown — “Devin Desktop / Windsurf rebrand”]: Cognition rebrand and desktop wrapper — not a stack change.
• [Rundown — “Moonshot $30B valuation”]: China model lab fundraising — geopolitical noise.
• [Rundown — “UK £1.1B AI Hardware Plan”]: National compute spend — landscape, not actionable.
• [Rundown — “Google/Nvidia consider Intel backup”]: TSMC hedging story — landscape.
• [Rundown — “Cut recurring meetings with Claude + Granola”]: Tutorial, fine to bookmark but not a brief item.
• [Rundown — “Lovable campus menu app”]: Reader demo, charming but not signal.
• [Information — “SpaceX Mega IPO event”]: Event marketing, not AI news.
• [Information — “Anthropic vs OpenAI IPO race”]: Marketing for paid IPO research — same theme already covered.
• [Information — “Databricks $165B talks”]: Valuation news, no project hook.
• [Information — “Anthropic OpenAI VC fundraising pace”]: Fundraising landscape, not actionable.
• [TLDR — “FrontierCode benchmark”]: Cognition code-mergeability benchmark — interesting but not a tool you’d adopt yet.
• [TLDR — “DX engineering velocity 10-15%”]: AI lifts PR throughput modestly — could be RT narrative material later.
• [TLDR — “Perplexity Computer 87/94% cuts”]: Marketing-framed numbers from Perplexity’s own research.
• [TLDR — “xAI looks like datacentre REIT”]: Alderson analysis — niche, skip unless infra play surfaces.
• [TLDR — “Perplexity IPO 2028”]: IPO date noise.
• [TLDR — “Model is no longer the bottleneck”]: Surfaced to Section 3 instead.
• [TLDR — “Agent harness should repair itself”]: Interesting agent-ops post, no immediate Roy hook.
• [TLDR — “OpenAI Economic Research Exchange”]: Researcher program announcement.
• [Practicaly — Apple WWDC roundup]: Same material as Tier 1 #2 and #3, no new signal.
• [Practicaly — “Reframe, Extend, Enhance in Photos”]: Worth noting for UBX sale info-memo photography, but tactical not strategic.
• [Practicaly — “Tim Cook → John Ternus Sept 1”]: Leadership transition, not AI-actionable.
• [Practicaly — “Spatial Panoramas in visionOS 27”]: Vision Pro feature, no Roy surface.
• [TheTip — “AI Money Group Wolf of Wall Street clone”]: Sponsored creator pitch — skip.
• [TheTip — “Workshop facilitation prompt”]: Generic prompt template — not Roy’s voice.
• [TheTip — “Amazon AI Custom Merch”]: E-commerce print-on-demand, not Roy’s lane.
• [TheTip — “Beehive audio AI”]: Did-you-know trivia.
• [BagelBots — “You’re probably fixing the wrong problem”]: Generic AI-projects-fail content marketing.
• [BagelBots — “Prompt That Builds Your Competitive Advantage”]: Lead-gen prompt drop, no signal.
• [Neil Patel — “robots.txt mistake crushing SEO”]: SEO 101, no AI angle.
• [a16z — “Month-End Is Now Just Another Day”]: Modern-ERP post — tangentially relevant to Ben/XeroAgent but more vendor narrative than insight.
• [Information — “Goldman/JPM compute futures”]: Surfaced to Section 5 instead.

• Sources scanned: 9 newsletters (TLDR, Rundown, The Information, Practicaly, TheTip, BagelBots, Neil Patel, a16z; aiwithkyle/superhuman/theaireport/aiwithallie/agentai inboxes returned no recent hits)
• Items extracted: 46
• Items surfaced: 9 (1 PAY ATTENTION, 3 Tier 1, 1 Section 2, 2 Section 3, 1 Section 4, 1 Section 5)
• Items skipped: 37
• Read time: ~7 min