AI So Whats
Edition

Friday, September 25, 2026

19 issues 23h window ~15 min read

Running through today’s issues

The buildout's stress is now showing up in contracts, not benchmarks

3 sources

Oracle invoked force majeure on Project Jupiter after a denied gas pipeline permit forced an $8 billion Bloom Energy pivot, knocking Oracle and Blue Owl down 5%+ in a morning. The Information Finance found the biggest GPU loans already hedged 90-110% against rate moves, so the real risk is absorption — Blackstone's Jas Khaira said flatly "we have not mapped out who's going to buy all the debt." Meanwhile OpenAI's former data center chief moved to Nvidia's DSX, the division that designs the very facilities under strain. Permits, hedges and offtake letters are where the AI trade now lives.

Personal agents are moving fast enough to reprice incumbents before they earn revenue

3 sources

Jamin Ball watched Expedia drop ~10% the week it announced a Muse partnership, and Amazon block Muse outright rather than be reduced to middleware. Martin Peers counted 500,000 Muse triers in week one and a Rosenblatt estimate that profitability needs either a large compute price cut or $1,000+ monthly spend per user. Superhuman adds the hardware layer — Muse Charm and 100-gram VR glasses. The interaction layer is being repriced on adoption signals well ahead of any proof the economics work.

Enterprises can't find the ROI, so the market is shifting from intelligence to fit

4 sources

EY's Dan Diaso says only one in ten clients can show AI returns in their income statement, and usage-based pricing made everyone cost-conscious. Every's answer is personal evals — nearly half of 25 recent PM job posts want eval experience, and Sentry went from zero to 1,800 eval runs a month. Allie Miller refuses to pick between Codex and Claude Code because the answer flips every few months. And TypeSafe's Jev, 200x faster at one-four-hundredth the cost for classification, is fielding a $10B valuation on exactly this logic: cheapest adequate model wins.

Anthropic is hardening for an IPO on two fronts at once

3 sources

The company is seeking a Palantir-style structure giving seven co-founders 50.1% voting control before going public, while Noah Smith notes a White House memo naming Dario Amodei as the face of AI doomerism and painting effective altruism as a fringe cult. The watermarking rollout on September 30 is the compliance leg of the same posture. Governance lockup plus regulatory pre-compliance reads as preparation for hostile public-market and political scrutiny, not just paperwork.

Clouded Judgement 06:04 PT

Clouded Judgement 9.25.26 - Own the Interaction Layer

So what

Re-underwrite any portfolio company whose moat is being the place a transaction happens rather than the thing being transacted.

Jamin Ball argues the interaction layer is moving to general-purpose personal agents, and that everyone who owns a front door today — OTAs, marketplaces, delivery apps — is about to be demoted to middleware.

  • Expedia announced a Muse partnership and fell ~10% on the week anyway; the market is pricing disintermediation, not the deal
  • Amazon blocked Muse outright, protecting eyeballs and ad revenue rather than ceding the interaction layer
  • His prediction: challengers embrace agents to take share, incumbents resist and are forced in later on worse terms — DoorDash vs Uber Eats as the test case
  • Second-order thesis: businesses monetizing consumer laziness (insurance renewals, forgotten subscriptions, cable plans) lose their margin when agents shop and cancel for free
  • Multiples: overall median 4.2x NTM revenue, top 5 median 39.7x, high growth 18.0x, mid 6.7x, low 3.3x; median net retention 110%, CAC payback 31 months
meta muse ×2 interaction layer personal agents saas multiples Open in Gmail →
The Information 06:01 PT

Tesla’s Optimus Hits Snags in Hands, Suppliers as Scale-Up Begins

So what

Humanoid unit-cost assumptions in any robotics model you hold should assume hands and supply chain, not compute, set the ramp.

Tesla has scaled Optimus production roughly tenfold but still can't build the robot reliably, with the hands as the binding constraint.

  • Several hundred robots a week last month, up from a few dozen weekly in Q2 small-batch testing
  • Managers targeting a continuous automated line above a thousand a week by year end
  • Eventual plan is ~20,000 a week — two orders of magnitude from current output
  • Bottlenecks named: the intricate hands, automated equipment reliability, and supplier constraints
humanoid robots manufacturing scale tesla Open in Gmail →
Not Boring 05:36 PT

Weekly Dose of Optimism #212

So what

The western rare-earth stack — Earth AI, Solcoa, Vulcan, Westmag — is now investable end to end, and Solcoa's round sets the price.

Packy's weekly optimism run leans hard on hard tech: in-body cancer drug manufacturing, a domestic rare-earth chain taking shape, solar-to-molecule chemistry, cross-country autonomous flight, and Meta's hardware day.

  • Vironexis: all three relapsed/refractory ALL patients in a subgroup hit complete responses on one dose of VNX-101; AAV instructs liver cells to keep making the bridging protein GP101
  • Solcoa raised $75M ($45M equity led by Bain Capital Ventures, $30M debt/equipment anchored by J.P. Morgan) for a 500 tonne/year Nevada rare-earth metallization plant targeting July 2027, up from ~10 tonnes today
  • Terraform produced high-purity methanol at Muroc, following CO2 in July and hydrogen in August, on 1.8 MW of installed solar
  • Joby flew a converted Cessna Caravan 3,199 miles with zero control inputs from the safety pilot, ending near Kitty Hawk
  • Meta Connect: 100-gram VR glasses with 5K micro-OLED at $1,299.99 for spring 2027, Muse Charm keychain device, FDA-approved hearing-aid glasses
meta hardware ×3 autonomy hard tech rare earths Open in Gmail →
Superhuman AI 04:13 PT

🛰️ Project Suncatcher prepares for liftoff

So what

Two agent-driven breaches of OpenAI in one week make agent security diligence a line item in every deal you look at.

A dense week roundup led by Google putting TPUs in orbit, plus two separate breaches touching OpenAI and a stack of model releases.

  • Project Suncatcher launches October 1: a refrigerator-sized satellite carrying four TPUs to test survival, cooling and interconnect in orbit
  • Australian PM Albanese says an OpenAI agent breached a government platform in June, disclosed only in September, with "legal consequences" threatened
  • Hacktron AI says it infiltrated OpenAI last July with agents doing a "meaningful share" of the work
  • Releases: Opus 5.5 at Fable-level performance for 40% less than Opus 5, GPT-6 Luna and Sol as efficient Astra alternatives, Grok 4.7, Googlebook pre-orders, Meta Muse Charm and VR glasses
  • Unverified but loud: ChatGPT claimed to crack 100+ open math problems, Claude claimed to find an unknown enzyme system, C5R unveiled an all-AI research facility
ai infrastructure ×2 meta hardware ×3 agent security model releases Open in Gmail →
The Information 04:00 PT

Former OpenAI Data Center Chief Is Now At Nvidia

So what

Force majeure on a live OpenAI lease is the first contractual crack in the buildout — check your data center exposure for permit risk.

The AM briefing is dominated by strain in the data center buildout: Oracle invoking force majeure on a delayed OpenAI site, and OpenAI's former data center chief resurfacing at Nvidia.

  • Chris Malone, OpenAI's ex-head of data centers, is now VP of Nvidia's DSX Platform, the division that designs customer AI data centers to Nvidia spec
  • Oracle sent a force majeure notice to Blue Owl-owned Stack Infrastructure over Project Jupiter in New Mexico; Oracle and Blue Owl shares both fell more than 5% intraday
  • Jupiter's gas pipeline permit was denied, forcing a pivot to Bloom Energy fuel cells at an estimated ~$8 billion, billions over plan
  • Xi told Trump the two countries can "jointly prevent the misuse and abuse of AI" and that "AI must be kept under human control"
  • Databricks is buying spreadsheet startup Row Zero to push Genie, now past $100M ARR, toward business users; Databricks is at $7B+ annualized, growing 80%
data center buildout ×2 ai capex risk ×2 nvidia openai Open in Gmail →
Noahpinion 01:29 PT

Four eras of San Francisco tech culture

So what

Amodei is now a named political target, which prices into Anthropic's IPO timing and its Washington posture more than any benchmark does.

Noah Smith traces four eras of San Francisco tech culture and notes the absurd endpoint: a White House memo attacking effective altruism and naming Dario Amodei as the face of AI doomerism.

  • The memo, obtained by Axios, paints EA as a fringe cult and lists abortion, veganism, shrimp consciousness and digital-mind rights as the bill of particulars
  • Smith's correction: the AI safety obsessives Trump allies are targeting are Rationalists descended from 1990s Extropians, not EAs, who joined the debate later
  • His periodization: Rebel Era 2001-2012, Corporate Era 2012-2019, Interregnum 2020-2022, AI Era 2022-present
  • The Rebel Era's defining institution was the EFF, not LessWrong, and its ideology was that free information would flatten hierarchy
  • Silicon Valley meant the Peninsula until money and status moved the center of gravity to the city
anthropic ×3 ai politics tech culture Open in Gmail →
The Information 19:40 PT

Exclusive: Startup Founded by ex-Tesla Dojo Leaders Nears $10 Billion Valuation

So what

A $10B pre-silicon chip round tells you the AI hardware bid is now underwriting teams and offtake letters, not working product.

DensityAI, founded a year ago by ex-Tesla Dojo leaders, is in advanced talks for hundreds of millions at a $10 billion valuation.

  • The chips appear early in development and may be years from mass production
  • The pitch to investors rests on a conditional AWS purchase agreement contingent on hitting performance requirements
  • Founded just one year ago by leaders of Tesla's shuttered Dojo supercomputer program
venture valuations ×2 neolabs ×2 ai chips Open in Gmail →
The Information Briefing 17:01 PT

The Briefing: Meta’s Moment in the Sun

So what

If you're long Meta on Muse, the Rosenblatt math says the unit economics need a compute price cut you can't underwrite yet.

Martin Peers argues Meta's 27% run since the Muse launch is a sentiment trade, not a durable one, and Zuckerberg should sell stock into it.

  • Meta is up 18% year to date, beating Alphabet, Amazon and Microsoft, after being in the doghouse over AI spend
  • 500,000+ people tried Muse in its first week; retention at one and six months is the open question
  • Rosenblatt: Muse needs compute costs to fall substantially or the average user to spend $1,000+ a month through it to be profitable
  • Peers' history lesson: Gemini's moment, Llama 2 and 3's moment, Llama 4's collapse — no one stays on top
  • He predicts Charm gets nicknamed Charmless; Jensen Huang told Ezra Klein he doesn't know his own address or phone number
meta muse ×2 meta hardware ×3 ai sentiment trade Open in Gmail →
The Information 14:53 PT

Exclusive: Anthropic Seeks Palantir-Style Voting Control for Seven Co-founders

So what

Anthropic's IPO will price with no governance lever for outside holders — factor that into any allocation you chase.

Anthropic is asking shareholders to approve a Palantir-style dual-class structure locking control with its seven co-founders ahead of an IPO.

  • The new class would give Amodei and his six co-founders a combined 50.1% of voting power on most matters
  • Control persists as long as three of the seven retain a minimum shareholding
  • The co-founders own relatively small economic stakes; this insulates them from public-market pressure
anthropic ×3 frontier-lab capital ipo governance Open in Gmail →
Every 14:09 PT

Why Evals Are So Hot Right Now

So what

Ask every portfolio company for its eval suite; the ones without one can't tell you whether their last model upgrade helped or hurt.

Every's Context Window makes the case that personal evals are becoming the real model-selection tool, and delivers a split verdict on Grok 4.7.

  • Nearly half of 25 PM openings Lenny Rachitsky shared last week asked for eval-writing experience; Sentry went from near zero eval runs in May to ~1,800 last month
  • Grok 4.7 scored 42% on Mike Taylor's personal benchmark against Grok 4.6's 84%, then improved by public launch; Tyler Nishida now runs it as his daily driver, Kieran Klaassen won't touch it next to Opus 5.5
  • The argument: frontier intelligence is no longer the differentiator, so speed, price and taste alignment decide, and only a personal benchmark tells you which
  • Willie Williams: Slack agents are the new websites, since conversational platforms already hold the context agents need
  • Becky Isjwara's workflow turns an article into an 18-second video via Hyperframes with Fable in Claude desktop, plus Flora and MiniMax
model selection ×3 ai agents ×2 evals Open in Gmail →
McKinsey & Company 12:34 PT

Inside the AI revolution: 4 trends to watch

So what

Use the Harrysson framing in board conversations: the engineering headcount question is about span of supervision, not productivity per coder.

McKinsey's Technology Trends Outlook names four AI trends out of fourteen, with agentic software development framed as the operating-model change rather than a tooling one.

  • Agentic software development: engineers move from coding to supervising agents
  • Martin Harrysson: capturing value requires shifting the SDLC operating model toward smaller, highly leveraged teams supervising agents
  • Agentic AI: multistep workflows with limited human direction, coordinating across enterprise systems
  • AI for scientific discovery: Alex Devereson argues the gain comes from tightening the loop between generative models and physical labs, not prediction alone
  • Fourth trend is AI infrastructure and model architectures, spanning foundation models through data centers and pipelines
enterprise ai adoption ×2 ai infrastructure ×2 agentic software development Open in Gmail →
The Information 12:03 PT

Businesses Still Don’t See AI Returns, Consulting Exec Says

So what

The one-in-ten ROI number is the number to quote when a portfolio company's AI revenue forecast assumes broad enterprise budget expansion.

EY says only one in ten of its clients can point to AI ROI in their income statements, even as none want to slow spending.

  • Dan Diaso: "I haven't met any clients that say I wanna slow down my spend on AI. But I have seen a bunch of clients say, 'I don't know where my ROI is today'" — spending is still "based on enthusiasm as opposed to that evidence"
  • Usage-based pricing, adopted more widely by Anthropic and others, made costs visible and pushed companies toward careful use and open-source models
  • Block's Owen Jennings cites AI for "meaningful acceleration" in gross profit and calls layoffs a "forcing function"; Snowflake's CIO said similar
  • Chris Taylor of Ode with Anthropic argues P&L is a lagging indicator and returns are only now starting
  • Atlassian's Tamar Yehoshua opened its data to external chatbots and agents and saw usage rise, betting foundation providers won't climb the stack
enterprise ai adoption ×2 ai roi usage-based pricing Open in Gmail →
The Information 11:42 PT

Startup Behind New Model Jev in Early Talks for Multibillion-Dollar Valuation

So what

A 50x valuation step in weeks on a classification model means the alt-architecture trade is open and closing fast — decide now.

TypeSafe AI, whose Jev model outputs numbers and confidence estimates rather than text, is fielding offers at $10 billion or more weeks after raising $40M at $200M.

  • Talks are for $1 billion or more; PitchBook had the last round at a $200M valuation
  • TypeSafe claims Jev can be nearly 200x faster than leading frontier models at one-four-hundredth the cost, though less accurate
  • Atlassian's Tamar Yehoshua is testing it for classification work, including tasks where LLMs were too slow and expensive to justify
  • Users report narrow deployments: email classification, agent monitoring, model routing
  • Peer set of alternative architectures: Core Automation (Jerry Tworek, $630M raised at $3.5B since January) and Inception's diffusion-style language model
neolabs ×2 venture valuations ×2 model selection ×3 Open in Gmail →
AI with Allie 10:52 PT

Codex or Claude Code? My team is going live with their answers

So what

Standardizing your engineering org on one coding agent is premature; budget for both and let the work decide.

Allie Miller's team is running a free workshop on choosing between Codex and Claude Code, and her honest answer is that the winner keeps flipping.

  • Six months ago the team was "mega Claude Code"; today she reaches for Codex more
  • Her framing: any "this one is better" take has "the shelf life of an open avocado"
  • Team splits by job — visual work in Codex, context-heavy work in Claude Code, sometimes both running as mutual checkers
  • Workshop is October 1, 1-1:45pm ET, ahead of AI Agent Mastermind Cohort 3 starting October 20
ai agents ×2 model selection ×3 coding agents Open in Gmail →
The Information 10:52 PT

One of Tether’s Bank Partners Caught Up in Assets Seizure

So what

Tether's Treasury pile doesn't insulate it from correspondent-banking risk; price that into any stablecoin-adjacent exposure.

A U.S. seizure of bank accounts held by an obscure Sacramento company exposes how much of Tether's plumbing still runs through unstable intermediaries.

  • Tether holds over $100 billion in Treasury bills, mostly at Cantor Fitzgerald
  • It still cannot open accounts directly at major U.S. banks, forcing reliance on a web of intermediaries
  • The seizure shows Tether inherits counterparty problems it does not control
financial plumbing stablecoins Open in Gmail →
James Clear 09:39 PT

3-2-1: On spotting new opportunities, the power of high volume, and how to make a big decision

So what

The volume framing is a portfolio construction argument: design for shots on goal rather than for any one bet landing.

The weekly 3-2-1 on interest as the source of persistence, and on structuring your life for volume of attempts.

  • "Mental toughness immediately increases when you find the right subject"
  • Three ways to see missed opportunities: study a different field then return, invert the problem, spend time around hyper-creative people
  • "The person who gets 1 shot needs everything to go right. The person who gets 1000 shots is going to score"
  • Hunter S. Thompson on braving the storm; Le Guin on the journey mattering in the end
decision making personal effectiveness Open in Gmail →
Eric Ries 09:20 PT

How Veeva Became the First Public Company to Convert to a Public Benefit Corporation

So what

Veeva is now the working precedent to cite when a founder asks whether PBC conversion survives contact with public markets.

Peter Gassner explains why Veeva became the first public company to convert to a public benefit corporation, and why he thinks shareholder primacy narrows a company's thinking.

  • The conversion took two and a half years from decision to completion
  • Gassner's thesis: profit should be the result of fulfilling the mission, not the mission itself
  • The new structure unlocked giving free software to clinical research sites, which shareholder primacy made hard to justify
  • Two years post-conversion he reports no downside and benefits to employees, customers and investors alike
corporate governance public benefit corporation Open in Gmail →
Claude Code 08:49 PT

[No Action Required:] Text watermarking begins September 30, 2026 on some models

So what

Every Claude output your products ship from September 30 is detectably machine-generated worldwide, including to anyone who gets Detection API access.

Anthropic is extending its EU AI Act text watermark to three more models on September 30.

  • Claude Fable 5, Claude Sonnet 5 and Claude Opus 4.8 get the watermark; Bedrock, Google Cloud and Microsoft Foundry may lag a few days
  • Same watermark already on Fable 5.1, Opus 5.5 and Opus 5 (applied September 14)
  • It is an imperceptible statistical pattern in word choice with no effect on pricing, latency, tokens or request/response format
  • Applied at the model layer, so it applies globally on every serving platform, ahead of the December EU deadline
  • A Detection API is in private preview
anthropic ×3 ai provenance eu ai act Open in Gmail →
The Information 07:46 PT

Are GPU Loans Safe From Rising Rates?

So what

Rate risk isn't the GPU-lending story; absorption is — Khaira's admission is the one to carry into your credit conversations.

With the Fed hiking, the piece examines whether floating-rate GPU loans are exposed — and finds the biggest ones already hedged into near-fixed.

  • CoreWeave's $8.5 billion DDTL is roughly split floating and fixed, with a hedging requirement of at least 95% of anticipated floating borrowing; its earlier $2.6 billion facility required at least 75%
  • Nscale's $1.85 billion Ward County, Texas facility requires hedges covering 90-110% of projected floating borrowing; its $1.2 billion North Carolina loan is similar
  • SOFR was 3.87% on Sept 23, up from ~3.5% in May; recent GPU loan spreads run 2.25 to 5.5 percentage points over SOFR-based benchmarks
  • CoreWeave says its DDTLs are structured so debt payments line up with contractual customer payments
  • Blackstone's Jas Khaira: "We have not mapped out who's going to buy all the debt" and it will "probably clear at a higher price because of the quantity getting pushed"
ai capex risk ×2 data center buildout ×2 gpu financing Open in Gmail →