AI So Whats
Edition

Saturday, September 26, 2026

6 issues 23h window ~5 min read

Running through today’s issues

Agents are now governed by permissions, not capability

2 sources

Microsoft built Autopilot to slot into the org chart with its own credentials, and Every's reporter watched it refuse an ordinary Word-to-Excel handoff on permissions. Stratechery's week is the same story one level up: Amazon simply revoked Meta's Muse access, and the block holds because Amazon's moat is physical. In both cases the binding constraint on an agent is who lets it in, not what it can reason about.

The industry is retreating from scoreable public tests just as it demands better private ones

2 sources

Ginkgo postponed its OpenAI-versus-Jewett protein contest and then killed the man-versus-machine framing, abandoning the Deep Blue moment it had hired a film crew for. Meanwhile Microsoft spent its launch event telling customers to build company-specific evals, with 16,000 internal ones of its own, because general benchmarks say nothing about your process. The credible measurement is moving inside the enterprise, and public head-to-heads are being quietly withdrawn.

Capital is flowing to the layer beneath the labs

2 sources

Fal is raising at $15B and possibly $20B and Fireworks is following, purely on the volume of tokens developers push through them. Thompson's Frontier Overhangs argues labs have strategic reasons to pace model releases — which makes the serving tier, whose economics depend on usage rather than capability jumps, the more durable place to sit. Inference demand keeps compounding whether or not the frontier moves.

The Information 15:59 PT

Exclusive: Fireworks, Fal Consider New Rounds as Inference Demand Soars

So what

Price any inference-layer position off Fal's $15–20B mark; the serving tier is now repricing faster than the labs it depends on.

Inference providers are the hot trade: Fal is talking to investors at a $15B valuation, with some pushing toward $20B, and Fireworks is weighing a round too.

  • Fal specializes in video and image inference — Google's Nano Banana models among them — and talks are early at $15B, possibly $17–20B
  • Fireworks AI is also considering new funding as developer inference spend climbs
  • The pitch is serving speed and model breadth, not model ownership: pure picks-and-shovels on top of open and hosted weights
  • Reported by Valida Pau, Stephanie Palazzolo and Aaron Holmes
ai funding fal / fireworks inference infrastructure Open in Gmail →
McKinsey & Company 13:42 PT

Here’s why some US industrial companies achieve higher multiples

So what

Useful only as a framing check on your industrial holdings: sector tailwind is already in the price, execution discipline is what re-rates.

A thin dispatch arguing that industrials in fast-growing markets don't automatically earn premium multiples — four repeatable management actions separate the re-raters from the rest.

  • The claim: market growth is necessary but not sufficient; the multiple comes from what management does inside it
  • Four actions define the "accelerators," per the linked research
  • Bundled with Inspired for business growth: How five companies beat the market and the Global Private Markets Report 2026
  • From Andy West and Sébastien Lacroix, who lead McKinsey's Strategy & Corporate Finance Practice
corporate strategy industrials valuation multiples Open in Gmail →
Every 11:01 PT

Copilot Gets a Seat in the Org Chart

So what

Before you buy Copilot seats, budget for eval-building and a permissions audit — that configuration work, not the model, determines day-one usefulness.

Ryan Sloan reports from Microsoft's Redmond launch of a unified Copilot with a persistent agent called Autopilot — and finds the permissions that make it enterprise-safe are the same thing that breaks it on the demo floor.

  • New app replaces separate consumer and work chat apps; three tabs — Home (chat plus Word/Excel/PowerPoint with shared context), Code (non-developers ship shareable apps), Autopilot (persistent goal-driven agent)
  • Autopilot draws on 30 days of work, org chart and calendar, gets a user-given name like "Dot (Ryan's Autopilot)," and is shown reporting to its owner in the org chart; runs on usage-based credits and is off by default
  • Sloan's off-script request — move a Word table into Excel — failed on permissions, as did a Chat-to-Code handoff, the exact handoffs the "one app" pitch depends on
  • Charles Lamanna says Microsoft runs roughly 16,000 internal evals, builds them with customers, and that Autopilot must "clear a higher bar" than on-demand tools; Nadella called for "evals specific to your company"
  • Nadella's line at the event: "The new massive insider risk is the AI agent"
ai agents ×2 agent permissions enterprise evals microsoft copilot Open in Gmail →
Ben Thompson (Stratechery) 10:01 PT

Begun, the Aggregator Wars Have (This Week in Stratechery)

So what

If you hold commerce-adjacent agent plays, note that distribution blocks now decide outcomes: Walmart's posture is the swing variable, not Meta's model quality.

The week's frame is aggregator-versus-aggregator: Amazon blocked Meta's Muse agent, and Thompson argues Amazon's physical-world investment is the moat that makes the block stick.

  • Muse is better when it can use Amazon, but Muse doesn't show ads — so Amazon blocked it; Thompson says Meta needs Walmart to wait Amazon out
  • Expedia is maneuvering to hold its middleware position; the open question is where Google is in all this
  • GM reversed course and is re-integrating CarPlay after three years of customer and dealership backlash — Andrew Sharp's told-you-so
  • Frontier Overhangs argues pacing the frontier is strategically convenient for labs regardless of sincerity; the Sharp Tech video asks why Anthropic or OpenAI can't unilaterally slow down
  • Stratechery Interview with Colossus EIC Jeremy Stern on profiling Zuckerberg, Scott Wu, Shyam Sankar, Josh Kushner, Neil Mehta and Jake Sullivan
ai agents ×2 agentic commerce aggregation theory meta vs amazon Open in Gmail →
The Information 10:00 PT

Contest Involving OpenAI, Human Scientists Goes Awry

So what

Read the retreat as a signal: nobody involved wanted a scoreable public result, so discount vendor claims about agentic science until someone submits to one.

Ginkgo Bioworks' planned human-versus-OpenAI protein design contest quietly fell apart as a spectacle — postponed, then revived with the man-versus-machine framing stripped out.

  • The setup: Stanford's Michael Jewett, an expert in cell-free protein synthesis, against a team of OpenAI agents in Ginkgo's 15,000-square-foot robotic lab, with Ginkgo executives judging three rounds
  • Jewett could use any commercial model; OpenAI would run its most advanced unreleased models
  • Branded "Mike Versus the Machines," with a documentary crew hired; postponed days before the Sept. 14 start and now back on with that concept scrapped
  • Explicitly pitched as biology's Kasparov–Deep Blue or Lee Sedol–AlphaGo moment — both of which the humans lost
ai benchmarks ai in biology openai Open in Gmail →
AI Market Fit 07:38 PT

Why Peter Thiel Wants AI to Move Faster

So what

Thiel is now openly pricing political regime risk into 2028 — factor that into fund life and exit timing for US-domiciled positions.

A paywalled unpacking of Peter Thiel's 65-minute conversation with Axel Springer's Mathias Döpfner, in which he accepts real extinction risk and argues for acceleration anyway.

  • Six major LLMs asked point blank put p(doom) somewhere between 1% and 10%; Thiel agrees the risk is real
  • His argument: slowing AI loses even at 10% extinction risk, because Chinese labs convert Western caution into advantage
  • AI is the proof case for his long-running stagnation thesis; Europe is falling further behind than it did in the Minitel era
  • Political tail: Weimar-style extremism, a competence crisis among Western elites, and 50% odds on a democratic socialist winning the White House in 2028
  • Argentina's decline offered as the template for the West; health framed as the only progress with no downside
ai policy ai safety peter thiel tech stagnation Open in Gmail →