AI So Whats
Edition

Saturday, October 17, 2026

9 issues 23h window ~7 min read

Running through today’s issues

The AI trade has moved from models to owned cash flows

3 sources

AI Market Fit documents venture money buying operating businesses outright: Thrive at a $12B valuation with OpenAI on the cap table, General Catalyst's Long Lake taking Amex GBT private for $6.3B, Dwelly rolling up 10 UK letting agencies. The Information shows the same capital logic at the infrastructure layer, where DayOne is taking a data center business public on ByteDance's expansion. OPEXEngine supplies the measurement problem that follows: when AI resets gross margin and headcount, the benchmark set that underwrites those deals is already stale. The returns case now rests on operating metrics nobody has a clean comparison for.

Single-counterparty dependence is the quiet risk in AI infrastructure

2 sources

DayOne draws nearly 70 percent of revenue from ByteDance and has acknowledged the concentration in its own filings ahead of a New York listing. Stratechery describes the mirror image: mobile carriers and airlines signing deals specifically to avoid relying on SpaceX as it buys spectrum and moves from emergency satellite backup toward primary cellular coverage. One party is selling the dependence to public markets; the other is paying to escape it.

Practitioners cannot agree whether agents should remember anything

2 sources

Mike Taylor keeps memory switched off in Claude and ChatGPT, starts a fresh thread per task, and catalogues seven ways accumulated context degrades output, from Katie Parrott's 3,855-word style guide to Gemini's Pokémon agent chasing a non-existent item for 100 turns. Allie Miller is selling the opposite four weeks later: a personal AI operating system of agents that hold ongoing responsibilities and know how you work. Persistence is the product in one and the failure mode in the other, and neither side is arguing from capability limits.

Every 13:04 PT

Watch the replay: How Codex Happened, behind the scenes

So what

Publishing your source interviews alongside an AI-assisted piece is a cheap credibility move you can adopt before anyone demands it.

A promo for the replay of How Codex Happened: Behind the Scenes, where Dan Shipper, Kate Lee and Jack Cheng unpick reporting on OpenAI's coding agent while using that agent to write the piece.

  • The editorial tension flagged is reporting on people you know: changing "worried" to "wondered" is the example that surfaced it
  • Shipper published the underlying interviews so readers can check context, find what he missed, and build a different story from the same material
  • Shipper's test for AI-assisted writing is specificity to your own experience; Cheng's is whether the work shows someone paid attention
  • The companion piece, How Codex Happened, covers the people and decisions behind the product's history
ai writing ×2 editorial process every openai codex Open in Gmail →
Every 11:01 PT

Context Engineering Is Knowing When to Start Over

So what

Edit the offending message instead of correcting the model in-thread; your error and your complaint both stay in context and compound.

Mike Taylor argues context rot, not model capability, is what degrades long AI conversations, and the fix is starting a new thread rather than arguing with the one you have.

  • Seven failure modes from his forthcoming book Context Engineering with DSPy: distraction, confusion, neglect, drift, clash, poisoning, underspecification
  • Adding a handful of "i-" examples pushed ChatGPT from 21 percent to 100 percent "i-" product names, killing every suggestion he actually liked
  • Katie Parrott's style guide grew from 759 to 3,855 words and the writing collapsed: 127 drafts, over a million words, fixed by cutting back to 513 words
  • Stanford's lost-in-the-middle finding: a fact buried mid-context scores worse than giving the model no documents at all
  • Microsoft and Salesforce found splitting one prompt into step-by-step messages cost 39 percent accuracy; Gemini's Pokémon agent spent 100-plus turns hunting an item that did not exist
agent memory ×2 ai writing ×2 context engineering llm limitations Open in Gmail →
Ben Thompson (Stratechery) 10:45 PT

Home At Last (This Week in Stratechery)

So what

Price SpaceX as a cellular competitor, not a satellite supplier: the carriers buying around it are telling you where the threat lands.

The week's roundup pairs Apple's return to home appliances, which Gruber calls pointless and Thompson defends, with a SpaceX spectrum story that has outgrown its original framing.

  • Apple HomeView is official; the open question is the addressable market for expensive devices running subpar AI
  • SpaceX's latest spectrum purchase moves it from de-facto emergency satellite coverage toward primary cellular; carriers and airlines are cutting deals specifically to avoid depending on it
  • Ari Emanuel sat for a conversation on his memoir, AI's impact on Hollywood, and the power of a phone call
  • Sharp China covers the Brussels–Beijing deal, weak holiday spending data, and reports of Xi taking rest breaks
  • Asianometry on ASML's new multibeam inspection tool; this week's video is Apple and a Hacker's Future
ai in hollywood apple spacex spectrum Open in Gmail →
McKinsey & Company 09:30 PT

As the world order turns, leading companies are keeping pace. Here’s how

So what

You get a framing device for the board-level resilience conversation and nothing operational until someone reads the underlying six-pattern report.

Shubham Singhal's Geopolitics Practice pitches six emerging patterns of geopolitical change that leading companies are using to realign for resilience and growth.

  • The framing is realignment for resilience, growth and competitive advantage rather than risk avoidance
  • Paired with The global balance sheet 2026, subtitled imbalance and divergence
  • Also linked: Geopolitics and the geometry of global trade, 2026 update
  • The email itself carries no numbers; the six patterns sit behind the report
mckinsey research ×2 geopolitics supply chain resilience Open in Gmail →
The Information 09:00 PT

Exclusive: Singapore’s DayOne Pushes Ahead With Data Center IPO Tied to TikTok

So what

Any DayOne allocation is a leveraged bet on ByteDance capex and TikTok's regulatory survival, sold to you as diversified infrastructure.

DayOne Data Centers is pushing toward a multibillion-dollar New York IPO on a revenue base that is almost entirely ByteDance.

  • Two people with knowledge of operations put ByteDance at nearly 70 percent of DayOne's revenue
  • The Singapore-based operator began four years ago as the overseas business unit of a Chinese data center company
  • Its growth tracks TikTok's global expansion directly
  • DayOne has acknowledged that the single-customer dependence is a significant business risk
bytedance customer concentration data centers ipo Open in Gmail →
AI Market Fit 08:50 PT

The AI Rollup Playbook: Why Buying a Boring Business and Running It With AI Agents Is the Best AI Bet of 2026

So what

Your competition for boring services businesses now carries permanent capital and an OpenAI-grade engineering bench; expect entry multiples to move against you.

The issue lays out the AI rollup playbook: stop building AI companies, buy cash-flowing boring businesses and rewire the operations with agents.

  • Thrive Holdings raised $2B at a $12B valuation in August from SoftBank, D1 and Altimeter, with OpenAI a shareholder since December 2025, and embeds its own engineers inside acquired accounting firms and MSPs
  • General Catalyst has co-created Crescendo in support, Eudia in legal, Titan in IT services, and Long Lake, which closed the $6.3B all-cash take-private of American Express Global Business Travel in September and calls itself the world's first AI holding company
  • Dwelly raised $263M in under six months and made 10 UK letting-agency acquisitions in 2026
  • CNBC named the category on June 8, 2026: AI rollup, Silicon Valley's new buyout playbook hits Wall Street
  • The rest of the issue covers target industries, three financing routes, the automation stack and downside scenarios
ai operating leverage ×2 ai rollups general catalyst private equity Open in Gmail →
AI with Allie 08:32 PT

ONLY 3 DAYS LEFT: our last AI Mastermind of 2026 🎉

So what

If your team buys agent training, demand a working agent as the output; four weeks of sessions is not a deliverable.

A closing-window pitch for the last AI Agent Mastermind of 2026, framed around hiring one more teammate rather than learning a tool.

  • Four weeks, starting Tuesday; enrolment closes Monday, October 19 at 3pm ET
  • Deliverables are a personal AI operating system, agents that hold ongoing responsibilities, and proactive agents that act without being asked each time
  • Includes setup help, office hours, and a private group session with OpenAI's Dominik Kundel
  • Social proof is April-vintage student testimonials, which the email itself notes predate current capability
agent memory ×2 ai agents training programs Open in Gmail →
McKinsey & Company 07:44 PT

On World Food Day, making every dollar count

So what

Pitch food and agriculture buyers on cost-to-serve, not growth; the 24-point sentiment drop sets the budget conversation for two planning cycles.

A World Food Day package arguing that farmers and shoppers are both buying more selectively, and that value has become the operating problem across the food chain.

  • Global farmer net spending sentiment has fallen 24 percentage points since 2024, squeezed by falling commodity prices, volatile input costs and trade disruption
  • Farmers cutting back on fertiliser to protect margin risks smaller harvests later
  • 61 percent of consumers say price matters more now than two years ago; food prices have outpaced general inflation
  • The State of Grocery North America 2026 runs 64 pages and argues pricing, promotions, loyalty and experience must operate as one system
  • Fermented novel proteins are framed as a $100B to $150B opportunity needing new business models
mckinsey research ×2 consumer spending food and agriculture Open in Gmail →
OPEXEngine by Bain 07:36 PT

Planning for 2027? Put the Latest SaaS Benchmarks to Work

So what

Stop anchoring 2027 targets on 2026 ratios; gross margin and headcount are moving faster than the benchmark set refreshes twice a year.

The Q4 benchmark refresh is live, with the pitch that AI has moved SaaS cost structures far enough that last year's baseline is the wrong anchor for 2027 planning.

  • Proprietary database updated twice a year; Q4 refresh adds broader coverage and a larger sample
  • The 2027 survey opens in January with new questions on AI investment, adoption and monetisation, plus a streamlined format and a redesigned BenchmarkEngine platform
  • Relativity's Senior Director of FP&A describes connecting long-range strategy to the annual plan on the SaaS Conversations podcast
  • An on-demand panel with Bain, Insight Partners and Gorgias asks how AI changes moats, revenue durability, growth economics and operating leverage
ai operating leverage ×2 fp&a saas benchmarks Open in Gmail →