AI So Whats
Edition

Wednesday, September 23, 2026

20 issues 23h window ~15 min read

Running through today’s issues

Both frontier labs cut price on the same day because their customers' bills got scary

5 sources

Opus 5.5 lands at 40% below Opus 5 and GPT-6 Sol and Luna at half their predecessors' API price, shipped within hours of each other. The Information's Applied AI data explains why: AI natives went from 1.4% to 8% of enterprise software spend while Pegasystems' Copilot bill went from $20,000 to $260,000 a month, and Microsoft responded with 30–50% Copilot seat discounts. Every's testing shows the cuts are real capability-per-dollar, not marketing — Sol matched Astra on 17 of 18 computer-use runs. Price, not benchmarks, is now the axis of competition.

Anthropic's China accusation is reshaping the Chinese AI capital markets

3 sources

Anthropic's report alleging Chinese labs rerouted government and military data to Claude triggered a CAC probe that knocked Z.ai down 12.4%, MiniMax 4% and Alibaba 4.4%, with DeepSeek and Moonshot — both pre-IPO — the focus. Peers points out the genuinely undercovered story is the fraught relationship between Chinese AI firms and their own government, and notes Xi is arriving in Washington without corporate executives. Meanwhile Founders Fund and Khosla partners are quietly touring Beijing, Shanghai and Shenzhen precisely because DeepSeek matters. The accusation created regulatory risk in Beijing that U.S. investors are now diligencing in person.

Owning the asset has replaced writing the check

4 sources

Anthropic is negotiating for 1GW of Apollo-controlled Stream capacity to stop renting from Amazon, Google and SpaceX. a16z is turning 1,600 portfolio companies and Bay Area concentration into a school that functions as a hiring funnel. General Catalyst bought a hospital and an asset manager so it sees problems first and lowers portfolio cost of capital. All three concluded that in a market where capital is abundant, the only durable moat is a physical or proprietary asset nobody can replicate.

Muse's week-one numbers make agentic commerce a supply-side question

3 sources

Muse posted 500,000 triers, 250,000 DAUs and the top App Store spot in seven days. Thompson argues the decisive variable is not consumer pull but whether Walmart signs on — it joined OpenAI's Instant Checkout instantly despite one-third conversion and smaller carts, and Muse charges no transaction fee. Expedia already jumped aboard. The demand side is proven; whether retailers accept smaller carts and no ad surface decides whether Amazon actually bleeds.

Superhuman AI 06:13 PT

🦾 Anthropic's Opus 5.5 matches Fable, but cheaper

So what

Renegotiate your inference budget now: both frontier vendors just cut effective prices roughly in half on the same afternoon.

Anthropic and OpenAI shipped price-cutting models on the same day and both sweetened usage limits, turning a model launch into a customer-retention fight.

  • Opus 5.5 hits Fable-level performance at 40% less than Opus 5, with raised five-hour limits and an on-demand rate limit reset on paid plans
  • GPT-6 Sol and Luna roll out in ChatGPT Work and Codex with a banked usage reset and 50% API price cut
  • Mirage's Tesseract gives agents real video editing primitives — compositions, keyframes, audio, local render — inside ChatGPT and Claude
  • Anthropic's Opus 5.5 prompting guide took 5K bookmarks in two hours; Meta's Muse won a skeptic over by clawing a $250 flight credit
  • FT claims SpaceX, OpenAI and Anthropic listings could outweigh every tech IPO since 1980
opus 5.5 ×4 gpt-6 family ×2 model pricing ×6 meta muse ×3 Open in Gmail →
The Information 06:00 PT

Exclusive: Founders Fund, Khosla Visit China as Country’s AI Prowess Rises

So what

Your China diligence gap is now a competitive gap; Thiel's partners are on the ground while LPs assume the door is closed.

U.S. venture firms have publicly retreated from Chinese startups while quietly sending partners on weeks-long fact-finding trips.

  • Founders Fund partners Sean Liu, John Luttig and Joey Krug toured Beijing, Shanghai and Shenzhen last month
  • Khosla Ventures also made the trip; the stated purpose is learning, not investing
  • Second China visit in a year for Liu, first ever for Luttig
  • DeepSeek's influence is the draw — Chinese labs now set the agenda in the hottest parts of the stack
china ai ×4 us-china tech ×4 venture capital ×3 Open in Gmail →
The Information 04:00 PT

Exclusive: Meta’s Muse Surpassed 500,000 Users After First Week

So what

Microsoft just told the market a Copilot seat is worth half list price; reprice any enterprise AI seat contract you're signing this quarter.

A dense AM briefing where the through-line is cost: Anthropic chasing owned compute, Microsoft slashing Copilot prices, and both frontier labs cutting model prices the same day.

  • Meta's Muse: 500,000+ triers, 250,000+ DAUs, 2M+ prompts in week one, number one on the App Store with 31,000 ratings
  • Anthropic in early talks to lease up to 1GW from Apollo-controlled Stream Data Centers, possibly with a Google credit guarantee, filled with Broadcom/Google TPUs
  • Z.ai fell 12.4%, MiniMax 4%, Alibaba 4.4% after the CAC probe into data leaks to Anthropic; CAC zeroing in on DeepSeek and Moonshot, both pre-IPO
  • Microsoft authorizing 30% discounts above 1,000 Copilot seats and 50% above 10,000, timed to the Copilot super app launch and usage-based add-ons
  • CoreWeave upsized its convertible to $4.2B at 2.875% due 2033, above the $3B target
model pricing ×6 meta muse ×3 anthropic compute ×2 china ai ×4 enterprise ai spend ×4 Open in Gmail →
AI Market Fit 03:51 PT

BREAKING: Andreessen Horowitz is building a University

So what

If you back talent-pipeline companies, a16z just vertically integrated recruiting into its fund and invited everyone to try copying assets they don't have.

a16z is launching the Horowitz Andreessen Academy, and the argument is that it is not a college competitor but a hiring and dealflow funnel built on assets no university can buy.

  • 1,600 portfolio companies, 80%+ concentrated in the Bay Area, is the stated moat; Databricks, NVIDIA and Stripe named as placement partners
  • Horowitz's diagnosis of prior college alternatives: they shipped one layer, not the bundle — training, credential, brand, status, fun
  • Non-negotiables per Biyani: residential San Francisco, elite demand, direct employer access; he claims only two or three attempts at that bundle in 50 years
  • For-profit and legally separate on purpose, so the student is the only customer
  • Pitch to teenagers without a billion-dollar idea: join someone who has one, citing Zuckerberg's Harvard roommates who didn't
a16z ×2 venture capital ×3 talent pipeline Open in Gmail →
Ben Thompson (Stratechery) 03:01 PT

More on Muse, Amazon, and Walmart; Muse and Expedia; Whither Google? (Stratechery Update 9-23-2026)

So what

Watch for a Walmart–Muse deal as the signal that agentic commerce has a real supply side, not just a demo.

Thompson argues Muse's real leverage is Walmart, that travel middleware like Expedia is more fragile than its Google-era survival suggests, and that Google's agent hesitation is a business-model problem, not an execution one.

  • Walmart bailed on OpenAI Instant Checkout after it converted at one third the rate of Walmart's own site with smaller carts, but signed on fast — and Muse charges no per-transaction fee, so Thompson expects Walmart to join
  • Every non-Amazon purchase raises the depreciation burden on Amazon's fixed physical assets; Meta's play is attrition, not a knockout
  • Expedia benefits today because Muse saves compute by hitting one site, but Meta will build an index and could make aggregators bid for the Muse slot
  • Agents remove the human from the loop, which is structurally bad for Google's ad business — hence slow Gemini Spark releases and competing products
  • Fallback thesis: Google protects search as long as possible while becoming The Google Capital Company selling the world's compute
meta muse ×3 agentic commerce google strategy Open in Gmail →
The Information 19:22 PT

Exclusive: Anthropic in Talks to Lease Apollo-Backed Data Centers

So what

Anthropic's cost curve improves structurally if this closes, which is how it sustains a 40% price cut without gutting gross margin.

Anthropic is moving from cloud tenant to data center tenant, in talks for up to 1GW of Apollo-owned capacity.

  • Stream Data Centers, 27 years old, majority-owned by Apollo Global Management, with U.S. sites
  • Fill would be Broadcom- and Google-designed TPUs, with Nvidia GPUs possible
  • Motive is reducing dependence on Amazon, Google and SpaceX clouds and cutting unit costs
  • Follows smaller leases with ex-bitcoin miners Hut 8 and TeraWulf
anthropic compute ×2 model pricing ×6 Open in Gmail →
Claude Code 19:21 PT

Introducing Claude Opus 5.5

So what

Your existing Claude integrations break if they set thinking budgets manually — run the claude-api migrate skill before you switch model names.

Anthropic's developer launch note for Opus 5.5 prices it at $4/$20 per MTok and makes thinking mandatory.

  • 40% cheaper to run than Opus 5 on typical workloads, 30%+ faster output, cache reads down 60% to $0.20/MTok
  • Fable 5.1-level performance on most work, including codebase-wide migrations
  • Thinking is always on — requests disabling it or setting a manual budget now error; use effort to control depth
  • Higher-risk cyber and bio requests refuse outright, with an opt-in fallback to another model on the Messages API
  • Fast mode research preview: up to 2.5x faster at 2x price; available on AWS, Google Cloud and Microsoft Foundry
opus 5.5 ×4 model pricing ×6 developer tooling ×2 Open in Gmail →
The Information Briefing 17:01 PT

The Briefing: AI, SI, UN and Xi

So what

The underrated position here is the Chinese labs' deteriorating relationship with Beijing, which prices into DeepSeek and Moonshot IPOs you may see marketed.

Martin Peers reads the UNGA-and-Xi week as theater, with the one substantive thread being the Anthropic-triggered Chinese regulatory probe.

  • Trump decreed AI be called super intelligence in U.S. government paperwork because artificial sounds fake; Peers notes it accidentally gifts Sutskever's SSI the perfect brand
  • Altman and Amodei both due to address the U.N. Security Council; OpenAI calling for international safety standards
  • CAC investigating DeepSeek and Moonshot over data security breaches after Anthropic's report; WSJ says Xi is not bringing corporate executives to Washington
  • Bessent says U.S. and China officials have discussed a joint AI safety framework
  • Alibaba unveiled a new training-and-inference chip, evidence of fast domestic semiconductor progress
china ai ×4 ai safety politics ×2 us-china tech ×4 Open in Gmail →
The Information 16:28 PT

Exclusive: AI Startup Avathon Sees Sales Boost From Barrick Deal

So what

Non-tech issuers are now buying AI partnerships as listing collateral, which makes vendors like Avathon a levered bet on the IPO calendar.

Avathon's Barrick Gold deal is a clean data point that industrial physical AI is converting to revenue, not just announcements.

  • Five-year partnership will generate $10M–$20M a year, roughly 10% of Avathon's expected next-year revenue
  • ARR jumped to $50M+ from about $15M a year ago; raising $100M+ by year-end against a $1.4B January 2022 valuation
  • Base platform fee plus outcome-linked fees; ~200 models scanning mine camera footage, predicting equipment failure, and negotiating supplier contracts
  • Customers include Aramco Digital (largest), Boeing (owns 9%), Airbus and the DoD; Barrick could be number two within three years
  • Barrick is floating a minority stake in its North American unit by year-end — the AI credential is pre-IPO positioning
enterprise ai spend ×4 industrial ai physical ai Open in Gmail →
The Information 15:30 PT

Inside the Scramble to Define AI Safety Before Someone Else Does

So what

Kill switches and pacing commitments are becoming procurement checkboxes; budget engineering time for them in any enterprise AI product you own.

A Tech & D.C. roundup arguing the pacing debate has escaped the labs and become a fight over who gets to define safe AI.

  • Microsoft backs pacing and pledges kill switches in its own AI products, raising the bar for every enterprise vendor
  • The White House publicly rejected the slowdown calls; Trump announced an AI Force and will appoint an AI czar
  • Zuckerberg reframes safety as competitive differentiator rather than constraint
  • New safety watchdog nonprofits are proliferating, several funded by the labs they evaluate
  • Bessent confirms U.S.-China talks on a joint safety system, making pacing a bilateral diplomatic item
ai safety politics ×2 enterprise ai spend ×4 us-china tech ×4 Open in Gmail →
Sacra 14:41 PT

openevidence vs. chatgpt for clinicians

So what

At 50x on ad revenue against a free OpenAI clone, OpenEvidence's therapeutics pivot is the only thing that justifies the entry price.

OpenEvidence has reached $300M annualized on a free, ad-funded, bottom-up motion — and that's exactly why OpenAI, UpToDate, Abridge and Doximity are all now aimed at it.

  • 65%+ monthly and 50%+ daily active among U.S. physicians, 90% gross margin, four-figure CPMs against only 5% of inventory sold
  • $15B valuation on a $250M September raise led by Byers Capital and a16z, ~50x; revenue up 1,800% YoY in 2025
  • UpToDate does ~$690M growing 8% with 80% institutional; Abridge $100M ARR at $5.3B; Doximity $645M at $4.7B
  • OpenAI launched ChatGPT for Clinicians in April 2026, free for verified clinicians, leaning on a $1B annualized ads run rate and 1B+ WAU
  • OpenEvidence is expanding from RAG into its own medical models and rare-cancer therapeutics — an AI-native Regeneron pitch
ai valuations ×3 healthcare ai vertical ai Open in Gmail →
Every 14:22 PT

Vibe Check: GPT-6 Sol vs. Opus 5.5

So what

Route by job, not vendor: Sol for volume reading and writing, Opus 5.5 for the ambitious builds where a failed run costs more than tokens.

Dan Shipper's head-to-head puts the tradeoff plainly: Sol 6 is the cheap daily driver, Opus 5.5 owns the top end of coding and visual work.

  • Sol at $2/$10 per MTok, half of GPT-5.6 Sol and half of Opus 5.5's $4/$20
  • Sol hit six of eight writing dimensions versus Astra's seven, and matched Astra on 17 of 18 computer-use attempts on the Hands benchmark
  • Opus built a working webcam heart-rate monitor; Sol couldn't finish, and missed instructions mid-run on other tasks
  • Opus clearly better at 3D worlds and short-film generation at equal low effort; Opus buries the lede, Sol leads with the interesting idea
  • Codex's stricter security checks halted approved long-running tasks, forcing babysitting
opus 5.5 ×4 gpt-6 family ×2 model pricing ×6 model evals ×2 Open in Gmail →
The Information 13:59 PT

Wall Street’s GPU Futures Push Stalls at the CFTC

So what

Compute hedging stays unavailable this year, so your data center and inference exposure remains unhedgeable except through equity proxies.

CME's GPU rental futures, the instrument that would make compute an investable asset class, just slipped past its October target.

  • CFTC is extending its review, digging into contract mechanics and soliciting industry comment, per a letter seen by The Information
  • CME had hoped to launch as soon as early October; approval won't arrive by then
  • The contracts reference rental prices on Nvidia GPUs
ai financialization compute markets Open in Gmail →
Every 11:46 PT

Vibe Check: Opus 5.5 Is Pulling Our Codex Converts Back to Claude

So what

Cap Opus 5.5 runs explicitly — the 5.9M-token burn is what your cost savings look like when the model does extra credit instead of the task.

Every's full Opus 5.5 review finds the Fable-at-60%-less claim credible, with a real failure mode: the model overshoots and loses the assignment.

  • Opus's Ruby code handled 427 requests per second and met 17 of 20 latency budgets, beating GPT-6 Astra on latency
  • A 251-line reviewable Rails patch passed four of five automated checks, while another app burned 5.9M tokens before core screens errored
  • A one-prompt golf game ran nearly two hours; a negotiation task settled after 12 rounds; a timed training task produced handouts but never the requested schedule
  • Kieran Klaassen replaced Fable 5.1 with Opus as daily driver; Mike Taylor and Tyler Nishida are reconsidering their Codex moves; Dan Shipper stays on Codex for writing
  • Guidance: keep Fable for problems too big to inspect, keep Astra or GPT-5.6 Sol for deadline work, and always give Opus a budget and stopping point
opus 5.5 ×4 model evals ×2 developer tooling ×2 Open in Gmail →
The Information 10:38 PT

New Data Show Anthropic, OpenAI and Other Upstarts Are Eating into Software Budgets

So what

Your portfolio's SaaS vendors are funding retention discounts with usage-based upside; model the 13x bill shock Pega ate before you sign.

Procurement data confirms the budget shift everyone has been asserting: AI natives are taking real share from legacy software line items.

  • OpenAI, Anthropic, Cursor and Sierra were 8% of Zip customers' software spend in the year to August, up from 1.4%
  • Those customers average 2,400 employees and spent $18B on software over four years; overall budgets rose a median 13%
  • Pegasystems' GitHub Copilot bill went from $20,000 to $260,000 a month after Microsoft moved to usage-based pricing
  • Microsoft offered Pega one million credits worth $10,000 for a month of Copilot Cowork after executives balked — small versus what other customers got
  • Workday, HubSpot, Microsoft and Amazon are all discounting or free-trialing AI while simultaneously introducing usage-based charges
enterprise ai spend ×4 model pricing ×6 saas disruption Open in Gmail →
The Information 10:33 PT

Opinion: Why America’s AI Dream Is Failing to Launch

So what

Site risk now dominates chip risk in your data center underwriting: $130B delayed in one quarter and Texas frozen.

An opinion piece arguing America's AI lead has eroded on both the model and the buildout side in four years.

  • Local opposition blocked or delayed roughly $130B of data center projects in Q1 2026
  • 71% of Americans surveyed this spring opposed a data center near them
  • Texas, the fastest-growing data center state, has frozen new grid-connected projects pending impact studies
  • Chinese models have taken major global token share; U.S. giants are now eyeing Chinese logic and memory chips
  • Data centers are framed as a bipartisan sleeper issue for the election cycle
us-china tech ×4 china ai ×4 data center buildout Open in Gmail →
McKinsey & Company 09:24 PT

Author Talks: Read the room before you lead it

So what

Skip this one unless you want a framing device for your next board-meeting read of the room.

An Author Talks with Calibrate CEO Pamela Meyer on reading social signals before leading a room, plus notes on cognitive overload and brain-powered org design.

  • Meyer's claim: the cues that shape power, trust and influence are visible and learnable, and missing them causes leadership missteps
  • Companion pieces cover how cognitive overload multiplies every cognitive bias
  • Also five principles for designing brain-powered organizations
  • No AI, market or deal content in this issue
leadership management research Open in Gmail →
Forge Global 09:10 PT

Anthropic debuts advisor-focused Claude

So what

Two AI chip startups raised $1.2B combined in a week — secondary pricing on SambaNova and Lightmatter is the read on whether that supply clears.

Forge's weekly private-market digest shows AI infrastructure and vertical AI absorbing the largest late-stage rounds.

  • Positron raised $875M Series C; Harvey $550M Series H; Temporal Technologies $550M Series E; EnCharge AI $310M Series C
  • Anthropic launched Claude for advisors with Schwab and BlackRock
  • Cohere in talks for up to $3B at a $20B valuation; Shield AI seeking a round at $20B
  • Most-viewed names: OpenAI, Databricks, Stripe, Kalshi; actively traded: SambaNova, Lightmatter, Dataminr, ConsenSys
  • Whoop fund offering closes September 25
ai valuations ×3 ai chips private markets Open in Gmail →
The Information 09:01 PT

Exclusive: Luckey’s Erebor Bank Plots Syria Expansion

So what

Erebor is building a geopolitical-arbitrage deposit base, not a tech bank — underwrite it as sanctions-timing risk, not fintech risk.

Palmer Luckey's Erebor Bank, having built itself into a U.S.–Venezuela financial bridge, is now targeting Syria.

  • Erebor has told staff and some investors it could reconnect Syria to the global financial system
  • The pitch is billions in deposits from Syrian businesses and consumers
  • Washington only recently reopened Syria to U.S. firms and investment, making Erebor an unusually early mover
fintech geopolitical arbitrage Open in Gmail →
AI Market Fit 07:28 PT

How to Follow General Catalyst’s New Strategy as an Investor

So what

GC and a16z are both converting owned assets into proprietary dealflow; if your fund's only asset is capital, your access is depreciating.

General Catalyst's premise is that value concentration has made the billion-dollar outcome not worth a founder's career, so the firm restructured around owning assets rather than writing checks.

  • Framing data: Nvidia adding a trillion of market cap in 100 days, Anthropic adding billions of revenue a month
  • New capital vehicles are described as the product, not the funding instrument
  • GC owns a hospital as a living lab so it sees clinical problems before the market does
  • The Janus flywheel: owning a half-trillion-dollar asset manager to lower portfolio companies' cost of capital
  • Taneja concedes indexing might have beaten GC's surgical AI bets
venture capital ×3 a16z ×2 ai valuations ×3 Open in Gmail →