AI So Whats
Edition

Saturday, September 12, 2026

6 issues 23h window ~5 min read

Running through today’s issues

Agent adoption has decoupled from agent returns

3 sources

McKinsey has 40 percent of large organizations scaling agents, up from 27 percent, with EBIT impact frozen at 37 percent and only 6 percent reporting significant impact. Its Quarterly piece blames the operating model rather than the technology. Every supplies the micro-version of the same failure: Jack Cheng noticing that cheap, fast models pushed him toward snacking on low-impact work because nothing forced him to ask whether it was worth doing. The constraint that used to do the prioritizing is gone at both the individual and enterprise level.

The money story is now the AI story

2 sources

The Information's Midha profile turns on an Anthropic IPO valued at up to $2 trillion and on compute access as the thing a VC actually sells. Stratechery's week includes "Anthropic and the Revival of (P)Doom Angst" alongside Meta giving away agent hardware to consumers. Capital structure and compute allocation, not model benchmarks, are what separate the players.

Oil no longer prices geopolitical risk — data centers might

2 sources

Noah Smith notes that with Hormuz closed and Iranian shipments blockaded, oil still hasn't reached its early-2010s highs in real terms, and concedes some of the remaining inflation may come from data center demand. Set against McKinsey's evidence of AI spend that isn't converting to earnings, the macro risk is migrating from energy supply shocks to unrecovered compute capex.

The Information 06:16 PT

Early Anthropic Investor Seeks VC Glory With Cash and Compute

So what

Compute allocation is now the differentiated term sheet item; if your fund can't offer it, you are bidding on price alone.

A profile of Anjney Midha, the a16z partner whose 2021 personal check into Anthropic at sub-$1B is about to become the defining venture return of the cycle.

  • Midha, 34, invested a substantial share of his and his wife's savings when Anthropic was valued under $1 billion
  • The piece pegs Anthropic's approaching IPO at a valuation of up to $2 trillion
  • His pitch to founders now combines cash with compute access, not just capital
  • Side bets include a neural sensor startup whose hardware he wears daily under a monogrammed Stetson
  • Framing: a new generation of investors is winning on access to scarce GPUs rather than on price or brand
anthropic anthropic ipo frontier-lab capital venture capital Open in Gmail →
Noahpinion 01:44 PT

Get the Middle East out of my politics!

So what

Your energy-shock hedges are mispriced: Hormuz closed and oil still sits below 2010s highs in real terms.

Noah Smith argues on the 25th anniversary of 9/11 that America should disengage from the Middle East entirely — militarily, diplomatically, and in its domestic politics.

  • Harvard Harris polling puts El-Sayed, Mamdani and Netanyahu below almost any American politician on net approval
  • The Iran War has failed to raise oil prices to even their early-2010s highs; in hours-of-work terms the spike is negligible
  • Blanchard and Gali: the 2000s economy was a third to half as oil-sensitive as the 1970s; the US is now a net oil exporter
  • He concedes some inflation may be coming from data center demand rather than energy
  • CAIR Action and AIPAC are both funding candidates on Israel/Palestine, which he calls fuel for attacks on Muslims and Jews alike
energy markets geopolitics us politics Open in Gmail →
McKinsey & Company 13:24 PT

AI adoption is scaling. Here’s how to make it pay off.

So what

Adoption metrics in your portfolio companies' board decks are no longer evidence of anything; ask for EBIT attribution or ignore the slide.

McKinsey's State of AI update finds agent deployment scaling fast while enterprise profit impact sits exactly where it was a year ago.

  • 40 percent of large-organization respondents report scaling AI agents, up from 27 percent last year
  • Nearly a third now use agentic coding tools to build products or features in-house
  • Share reporting enterprise-level EBIT impact from AI: flat at 37 percent
  • A companion McKinsey Live session cites just 6 percent of companies seeing significant EBIT impact
  • The framing is now a spend–value gap and a governance problem: weigh agentic AI against the true cost of the work it replaces
enterprise ai ×2 ai agents ×3 ai roi ×2 Open in Gmail →
Every 12:23 PT

What I Learn When I Run Out of AI

So what

Cheap inference removes the friction that used to filter your roadmap; build the pause back in deliberately or ship features nobody needed.

Jack Cheng argues that Anthropic's weekly usage caps are accidentally useful: the wait forces him to decide whether a question is worth asking.

  • He batches his hardest bugs and one-shot builds into weekly sessions he calls consulting the oracle
  • As Fable 5.1 got token-efficient enough to run as a daily driver, his weekly allocation now stretches three to four days instead of one
  • The result is more of what Will Larson calls snacking — easy low-impact work crowding out hard high-impact work
  • Historical frame: Annie Easley hauling punch cards, Alan Kay calling time-sharing "a very institutional way of thinking about computing"
  • Turkle's 1984 line lands hardest: it is hard to walk away from a program with an undiscovered bug
ai agents ×3 developer workflow inference economics Open in Gmail →
Ben Thompson (Stratechery) 10:02 PT

Duo Threats (This Week in Stratechery)

So what

Meta is giving away the agent hardware layer to consumers while Apple bets on apps — watch which one owns the default personal agent.

Thompson's week pairs Apple's $2,000–$3,200 foldable iPhone Duo with a sharper argument: Meta's unglamorous Muse agent may matter more for ordinary users than OpenAI's Navier-Stokes result.

  • The Duo arrives at $2,000 to $3,200; Gruber called the name back in April
  • Thompson's read on Apple: hardware-software integration still works, but its biggest AI blindspot is believing in the primacy of apps
  • OpenAI solved one of the most famous open math problems — impressive, of little impact, and possibly chilling to advanced math research
  • Meta's Muse bundles free hardware and software so consumers can run agents without buying a Mac Mini
  • Sharp Tech covers Astra and AGI, plus "Anthropic and the Revival of (P)Doom Angst"; Ballmer's Clippers drew the harshest penalty in NBA history
ai agents ×3 apple meta openai Open in Gmail →
McKinsey & Company 10:00 PT

The key to AI value is hiding in plain sight: Your operating model

So what

If your operating model still routes work through the same approval layers, agent deployments will keep landing as cost without showing up in EBIT.

A McKinsey Quarterly piece relocates the AI value problem from technology to org design: there is no best operating model, only intentional choices well executed.

  • Five lessons drawn from companies it labels reinventors
  • Same drumbeat as the State of AI note: only 6 percent see significant EBIT impact from AI
  • Companion framing pieces: From adoption to impact: Three horizons of AI transformation, and A new operating model for a new world
  • The spend–value gap is positioned as a governance responsibility, not a procurement one
  • Recommendation: assess agentic AI against the true cost of the work it replaces, and pull back where it doesn't clear
enterprise ai ×2 ai roi ×2 operating model Open in Gmail →