AI So Whats
Edition

Wednesday, September 16, 2026

19 issues 23h window ~14 min read

Running through today’s issues

The UI is no longer the moat, and computer use is what killed it

3 sources

Thompson reads Salesforce's AIforce as Benioff conceding that burying functionality in a UI stops being defensible once Astra's computer use is both accurate and fast — and charging $550 a seat for the concession before it becomes free. Every's team supplies the ground truth: Mike Taylor flipped Astra from a yellow rating to his daily driver purely on computer use, and staffers are now handing over Google Slides builds, Verizon support chats and link audits. The same issue reports Anthropic shipping a Salesforce plugin with 37 sales skills, which is exactly the headless pattern arriving from the other direction. If the UI is reachable anyway, every enterprise SaaS pricing model built on opacity has months, not years.

Anthropic is now priced above OpenAI, on numbers worth interrogating

3 sources

The Information has OpenAI in early talks at $1.2T after deferring its IPO, while Anthropic's planned listing is pegged at $1.5T or higher; Forge confirms both have confidentially filed draft S-1s. Superhuman flags that Anthropic's two reported profitable quarters exclude training costs and revenue share, and that its $30T revenue opportunity claim is a quarter of global GDP against $65B annualized today. Meanwhile Anthropic is losing Palantir and Booz Allen workloads to OpenAI over zero data retention. The valuation ordering and the operational facts are pointing in opposite directions.

The slowdown debate has become a market-structure fight, not a philosophy one

4 sources

The Information reports SMB customers already balking at AI purchases after catastrophic-risk coverage went mainstream, and Anthropic, OpenAI and Google meeting since July on a standards body that would advantage incumbents over neolabs. At All-In, Musk pitched cross-lab peer review as faster than regulation and Trump called the alarm a hoax on Jensen Huang's speakerphone; Zuckerberg countered in The Information that alignment is a competitive distinction labs should handle unilaterally. Superhuman frames it as doomer versus utopian, but the actual contest is over who writes the standards. Whoever does gets to define the compliance cost that keeps new entrants out.

Cheap calibrated judgment is becoming its own model category

2 sources

TypeSafe's Jev returns probabilities rather than prose, priced per billion tokens instead of per million, with output tokens unmetered. Every's test ran 1,709 judgments for under a cent and clocked 0.35 seconds per passage against Fable 5.1's 8.83, with only a one-defect accuracy gap out of seven. Superhuman notes the catch: it can't generate text at all. That constraint is the point — it makes continuous mid-task verification economically trivial, which is the missing piece for letting agents run longer without supervision.

Superhuman AI 06:15 PT

🤯 ChatGPT's co-inventor drops a model that can't hallucinate

So what

Discount the Anthropic profitability headline: excluding training costs and rev-share makes it a marketing number, not a pre-IPO fact.

A promo-heavy issue whose one real item is TypeSafe's Jev, a non-generative "System One" model from InstructGPT co-author Diogo Almeida, plus fresh scrutiny of Anthropic's pre-IPO numbers.

  • Jev claims 20-200x faster inference, 40-400x cheaper, no hallucination — but it is code-output only, no prose
  • Anthropic pressing ahead with a 2026 listing at $2T; reported two profitable quarters exclude training costs and revenue share
  • Anthropic told investors its addressable revenue opportunity is $30T, about 25% of world GDP, against $65B annualized today
  • Anthropic shipped a Salesforce plugin with 37 sales skills, landing the same week as Dreamforce
  • Zach Yadegari (Cal AI) launched Persona, a wrist-worn agent band with three-day battery
typesafe jev ×2 anthropic ipo ×3 ai safety debate ×4 wearables ×2 Open in Gmail →
The Information 06:00 PT

Exclusive: Apple Considers Return to Server Market, Has Talked With Nvidia to Use Network Tech

So what

Apple silicon becomes a credible inference-side alternative for enterprise buyers who want capacity without queueing for Nvidia GPUs.

Apple is building an enterprise server around clustered M8 Ultra chips and has talked to Nvidia about using its networking gear.

  • Two configurations under consideration: two M8 Ultras clustered, or four
  • Target buyers are AI developers, businesses and governments — Apple's first real server push in two decades
  • Nvidia networking inside an Apple box would be a first, and an admission Apple can't do the interconnect alone
  • Driven by surging interest in Macs for AI inference workloads
ai infrastructure apple silicon nvidia Open in Gmail →
The Information 04:00 PT

OpenAI in Early Talks for New Funding Round at $1.2 Trillion Valuation

So what

Anthropic at $1.5T versus OpenAI at $1.2T inverts the old ordering — reprice any secondary exposure you hold in either.

OpenAI is in early talks at a $1.2T valuation after pushing its IPO to next year or later, while Meta monetizes AI through subscriptions and Musk floats peer-review as an alternative to regulation.

  • Last round was $122B at $852B post in March; Caplight secondaries already trade past $1T on Codex and GPT-6 Astra strength
  • Anthropic's planned IPO is pegged at $1.5T or higher; Altman says an offering this year doesn't make sense amid safety concerns
  • Meta One goes global: $2.99 for WhatsApp Plus up to a $499 Max tier; 15 million subscribers already
  • SK Hynix in talks to lease Intel's Ohio fab or form a JV — Seoul's sensitivity to advanced memory export is the main obstacle
  • Microsoft raised its dividend 8% to 98 cents while guiding $50B capex this quarter; Salesforce unveiled Koa, built with Nvidia on Nemotron
openai valuation ×2 anthropic ipo ×3 ai safety debate ×4 meta subscriptions ×3 Open in Gmail →
Ben Thompson (Stratechery) 03:20 PT

Salesforce AI Force, Agents as UI, The Race to Headless (Stratechery Update 9-16-2026)

So what

If you hold enterprise SaaS, test whether the API exposes everything the UI does — opacity stops being a defensible pricing strategy this year.

Benioff introduced AIforce as an interface revolution; Thompson reads it as Salesforce conceding that the UI was never the moat and racing to headless before someone forces it.

  • Claudeforce, Slackforce and Agentforce Coworker ship Salesforce logic, skills and workflows into Claude and Slack rather than pulling users into Salesforce
  • Thompson's own anecdote: after vibe-coding a Channels management UI, he stopped using any UI and now just tells Codex what to do
  • The bet is that Salesforce's defensibility is being the system of record, not the front end
  • Astra's computer use is as accurate as Sol's but massively faster — meaning any functionality buried in a UI is reachable anyway
  • Benioff is charging $550 a seat for abandoning the moat, which Thompson calls clever because it will be free soon enough
agents as ui ×2 computer use ×2 enterprise saas salesforce aiforce Open in Gmail →
Noahpinion 23:08 PT

The anti-American Right

So what

H-1B and Indian-founder talent pipelines face political risk from a faction that ignores the merit arguments your recruiting deck relies on.

Noah Smith uses Texas GOP figure Bo French's anti-Indian tweet about Longhorns fans to argue the New Right is structurally incapable of coalition-building and therefore self-limiting.

  • French, the GOP nominee for Texas Railroad Commissioner, trails his Democratic opponent and has called for deporting 100 million people when only 50 million foreign-born residents exist
  • Cornyn, Cruz and Karl Rove denounced him; Rove said he'd vote Democrat; Abbott, who endorsed him, was mild
  • Smith's case: legality, skill level, assimilation, low incarceration and even voting Republican change nothing for this faction
  • Indian Americans are the highest-earning group in the country and incarcerated at 25% the rate of non-Hispanic whites
  • Polling cited puts 73% of Americans saying immigration is good, making the New Right a loud but numerically fringe bloc
us politics ×2 immigration talent Open in Gmail →
The Information 18:51 PT

Exclusive: AI Agent Startup Instinct in Talks for $10 Billion Valuation

So what

A $10B mark on 100,000 users sets the comp for every consumer agent in your pipeline — price accordingly or pass now.

Personal AI assistant Instinct is raising $1 billion at roughly a $10 billion valuation, with Sequoia and Benchmark in talks to lead.

  • User base passed 100,000 as the invitation-only service opened up, and it is capacity constrained
  • The raise follows earlier reporting this month that it was seeking funding to buy compute
  • Both Sequoia and Benchmark circling the same round signals a consumer-agent category bet, not a single-company one
consumer agents ×3 venture funding ×3 instinct Open in Gmail →
The Information Briefing 17:01 PT

The Briefing: Meta’s AI Subscriptions

So what

Subscription revenue won't cover AI capex at any of these companies — model Meta One as optionality, not as a funding source.

Martin Peers argues AI is finally making consumer subscriptions matter for big tech, but the evidence that they can offset AI capex is thin.

  • Meta One launched Tuesday with 15 million existing subscribers; Google reported 350 million paid subscriptions yet buried the revenue in a $12.9B line item that is 11% of the total
  • Apple services were 28% of June-quarter revenue at nearly double hardware gross margin; Snap's subscriptions drove 11% growth against 5.8% ad growth
  • The structural problem: every AI service has a free tier and users can hop between Anthropic, OpenAI, Google, Meta and Grok
  • On the DoorDash–Wonder deal: Wonder bought Grubhub for $650M last year and just sold its campus dining unit for $300M, recovering nearly half the price
  • Wonder has raised $3B at a $9.65B valuation and keeps raising, which Peers reads as evidence it isn't profitable
meta subscriptions ×3 consumer ai monetization doordash wonder Open in Gmail →
The Information 16:25 PT

Anthropic, OpenAI ‘Pacing’ Spooks Some Startup Customers

So what

Rotate toward safety tooling and app-layer names now; neolab bets get harder the moment the three big labs write the standards.

Julia Hornstein reports that Anthropic's and OpenAI's pledges to slow down are already costing startups deals, while reshaping where VC money goes.

  • One VC describes an SMB AI app where customers turned wary after catastrophic-risk coverage went mainstream
  • Anthropic, OpenAI and Google have met regularly since July about a joint safety standards body; OpenAI policy chief Chris Lehane confirmed it Tuesday
  • Standards bodies favor incumbents with compliance resources and squeeze neolabs pursuing novel architectures
  • Counter-thesis from another investor: money rotates to model security and infra — Neo raised $100M from a16z and Bessemer, Gimlet Labs took three tranches
  • A third view: a frontier pause gives app-layer companies like Instinct room to close the capability gap
ai safety debate ×4 venture funding ×3 consumer agents ×3 frontier labs ×2 Open in Gmail →
The Information 15:22 PT

Exclusive: Meta to Launch Camera-Free Smart Glasses Amid Mounting Privacy Concerns

So what

The wearable fight is shifting from capture to always-available voice agents; camera-centric hardware theses need a rethink.

Meta will ship camera-free smart glasses this fall, code-named Luna, as a direct answer to the privacy backlash against its Ray-Ban line.

  • Six microphones for talking to Meta AI and Muse, Meta's new consumer agent, plus a dedicated side button
  • Open-ear speakers and slimmer temple arms make them read as ordinary eyewear
  • Removing the camera trades capability for social acceptability — a bet that voice is enough for agent access
wearables ×2 meta subscriptions ×3 consumer agents ×3 Open in Gmail →
The Information 14:45 PT

Join us live: The Information’s 50 Most Influential People in Tech on Thursday, 9/17

So what

Worth an hour only if you want The Information's read on which investors and lab leaders they think compound influence from here.

A live event promo for The Information's inaugural 50 Most Influential People in Tech list, airing Thursday.

  • Subscriber-only list ranking who will hold sway over the next decade, not just who is biggest today
  • Lessin, Efrati, McEvoy and Brown discuss who made it, who didn't and what the omissions reveal
  • Thursday, September 17, 12pm PT
media ×2 events ×2 Open in Gmail →
Forge Global 14:16 PT

OpenAI, Anthropic revenues climb as both aim for IPO

So what

Both S-1s are already filed, so your window to accumulate OpenAI or Anthropic secondaries at pre-listing marks is measured in months.

Forge's weekly digest frames OpenAI versus Anthropic as a race to be AI's first public-market giant, with both having confidentially filed draft S-1s.

  • Both companies have submitted draft registration statements to the SEC, setting up two of the largest tech IPOs ever
  • Figure AI raised $4.19B Series C; Crusoe raised $3B Series F at $30B post
  • Stoke Space raised another billion; PsiQuantum finalized a $100M Commerce Department award
  • August commentary: cap-weighted recovery alongside wide dispersion across individual private names
  • OpenAI, Databricks, Hugging Face and Stripe top the most-viewed list on the secondary marketplace
anthropic ipo ×3 openai valuation ×2 venture funding ×3 secondaries Open in Gmail →
McKinsey & Company 13:48 PT

The AI economy: Interconnected forces, feedback loops, and speeds of change

So what

Useful only as board-deck scaffolding for explaining why compute buildout outruns grid and institutional capacity.

MGI publishes a systems framework for the AI economy, arguing the interesting questions are about the couplings between capability, power, institutions and labor rather than the models.

  • Authors Shubham Singhal, Arvind Govindarajan and Tanguy Catlin map capabilities, physical resources, institutions, businesses, geopolitics and people as one system
  • Core claim: components fall out of step with each other, and the gaps are where bottlenecks and risk show up
  • Positioned as a diagnostic for where a change in one part ripples through the rest
research frameworks ×2 macro ai economy Open in Gmail →
Every 11:36 PT

Mini-Vibe Check: TypeSafe's Jev Judged Everything I’ve Written in 0.7 Seconds

So what

Swap your LLM-as-judge eval layer to a calibrated classifier — you get 500x cost reduction and can grade every step, not samples.

Mike Taylor's hands-on with TypeSafe's Jev: it returns calibrated probabilities instead of prose, cheap and fast enough to check an agent's work mid-task.

  • 37 documents against 21 questions each — 777 judgments in under 0.7 seconds for roughly a quarter of a cent
  • Priced at $42 per billion tokens versus per-million pricing everywhere else; output tokens aren't metered at all
  • Head-to-head with Fable 5.1: median 0.35s per passage versus 8.83s, about 580x cheaper, but caught six of seven planted defects to Fable's seven
  • 11 experiments, 1,709 judgments, total cost under a cent — routing, grading support replies, flagging risky agent actions, sorting pitches
  • The framing that matters: a code linter for knowledge work, run continuously during generation rather than once at the end
typesafe jev ×2 evals ×2 ai agents ×2 inference economics Open in Gmail →
The Information 11:29 PT

Why Anthropic’s Data Policy Drama Is Good for OpenAI

So what

Make zero data retention a hard gate in your next model procurement — it's currently deciding enterprise deals faster than benchmarks are.

Anthropic's refusal to offer blanket zero data retention on Fable models is pushing regulated enterprise customers to OpenAI.

  • Palantir and Booz Allen are expanding Astra use because they secured zero data retention; both restricted Fable after Anthropic's June 30-day retention policy
  • Anthropic's fix, Enterprise Frontier Safeguards, only lands later this fall with unclear eligibility criteria; Salesforce, Comcast and Visa helped design it
  • Palantir began distributing Astra to its customers this month, giving enterprises weeks of Astra exposure before they can test Fable
  • A cybersecurity vendor serving US government and large financial clients switched its customer-facing apps to OpenAI over the same issue
  • Palantir CEO Alex Karp has publicly accused labs of using customer data to compete with customers
frontier labs ×2 data retention enterprise ai procurement Open in Gmail →
AI Market Fit 09:37 PT

The All-In Summit 2026: What Silicon Valley’s Most Powerful People Actually Said

So what

The White House is now explicitly against a slowdown, which caps how much regulatory downside you should price into US frontier exposure.

A recap of the All-In Summit, where Jensen Huang took a live call from Trump onstage and the AI safety split inside Silicon Valley played out in public.

  • Trump, on speaker, called AI alarm a hoax and argued slowing US development mainly helps China
  • Musk proposed labs peer-review each other's models pre-release, framed as faster than regulation and open to Chinese participation
  • Nadella pushed human control, independent evaluators and companies owning their own AI learning loops
  • Huang covered doom, regulation, open source, China and Nvidia's capital strategy beyond chips
  • Shotwell on direct-to-cell Starlink, reusability and orbital data centers; Musk teased a SpaceX–Tesla combination
ai safety debate ×4 us politics ×2 all-in summit Open in Gmail →
Every 09:05 PT

You're Probably Sleeping On Computer Use

So what

Audit which internal workflows you kept human because no API existed — computer use just removed that constraint at usable speed.

Every's team argues computer use is the underrated capability of the moment: Astra flipped from a yellow rating to Mike Taylor's daily driver purely on the strength of it.

  • Taylor now defaults to Astra in Codex for work he never delegated before — hundreds of edits across six Google Slides decks, checking every link in his book's PDF proofs into a defect spreadsheet
  • Staff examples: Codex maintaining a calendar of kids' school events, negotiating in Verizon's support chat, clearing WhatsApp storage via iPhone Mirroring, listing items for sale from photos
  • Codex runs computer use in the background; Claude Code pulls the window to the front and blocks you
  • Dan Shipper built Hands, an internal Mac app letting the Every Agent drive apps and hand tasks to Codex or Claude from Slack
  • Douglas Brundage's deck workflow: build two or three reference slides, generate copy with Codex, let it build the rest in Google Slides, then convert your edits into saved rules
computer use ×2 ai agents ×2 agents as ui ×2 evals ×2 Open in Gmail →
McKinsey & Company 08:00 PT

Winning: What the best athletes, teams, and coaches do differently

So what

Skip unless you need fresh language for a performance-culture conversation with a portfolio company CEO.

Seth Goldstrom launches a Winning video series extracting eight repeatable behaviors from elite athletes, coaches and teams for corporate performance.

  • Built on three decades of performance consulting plus interviews with championship coaches, Olympians and team executives
  • The premise is that repeat winners share identifiable behaviors, not just talent
  • First episode introduces the eight behaviors and their corporate translation
research frameworks ×2 leadership Open in Gmail →
Every 07:39 PT

Thesis: 2027 ticket price increases by $100 on Wednesday + the After Party is free

So what

Decide by Wednesday midnight if the operator-heavy speaker list is worth $1,400 to you; the after party gets you the room for free.

Ticket logistics for Thesis: 2027, Every's first conference, on November 5 at Pioneer Works in Brooklyn.

  • Full ticket price rises $100 to $1,500 at midnight Wednesday; applications reviewed on a rolling basis
  • New speakers: Josh Miller (The Browser Company), Ben Cera (Polsia), Kieran Klaassen (Cora), Emmett Shine (Little Plains)
  • Joining Ivan Zhao of Notion, Cat de Jong of Anthropic, Cristóbal Valenzuela of Runway, Lauren Reeder of Sequoia and OpenAI's Andrew Ambrosino and Ed Bayes
  • The 21+ after party is included with a ticket and free to Every subscribers
events ×2 media ×2 Open in Gmail →
The Information 07:31 PT

How Erebor Bank’s Sweet Stablecoin Deal Backfired

So what

A reminder that novel-charter banks underwrite crypto flow risk badly at first — relevant if you hold or are underwriting Erebor.

Palmer Luckey's Erebor Bank pulled its free stablecoin-to-cash conversion offer after crypto trading firms arbitraged it.

  • The offer was a customer-acquisition pitch at launch earlier this year; it lasted a few months
  • Sophisticated trading firms found a way to profit from the conversion at Erebor's expense
  • Reads as a young bank still learning to price risk in crypto flows
crypto fintech risk Open in Gmail →