AI So Whats
Edition

Sunday, August 23, 2026

17 issues 23h window ~14 min read

Running through today’s issues

AI ROI is being redefined as a headcount question, not a productivity question

3 sources

Uber's Andrew Macdonald says per-employee AI ROI is nearly impossible to measure because saved hours get reabsorbed, so capture it in OpEx by capping headcount. Robbie Allen arrives independently at the same place: track headcount plans, team scope and operating metrics, because there is no invoice for the person you didn't hire. Paul Baier's CFO use cases are the counterexample that proves the rule — the savings that are actually bookable are the consultant fee and the BI license, not the analyst hours.

The winning AI investment is shared infrastructure, not better tooling

3 sources

Sachin Rekhi argues elite teams get 2-3x gains because they standardized on one agentic platform and built shared skill libraries, citing Fin's CTO saying permissive tool choice produced gains only in isolation. Uber's 30-engineer pod embedded with process owners is the same idea in enterprise form, cutting a 15-hour pricing process to two. Every's Headway story extends it: when the shared layer can't be bought, 650 of 900 employees end up on something built in-house on the Claude Code SDK.

Nvidia's pricing power is now the largest single variable in AI capex

2 sources

Server makers are telling customers Grace Blackwell 300 and Vera Rubin 200 systems will cost about 17% more next year, adding at least $5B per gigawatt of data center. Martin Peers puts that in context: Nvidia is expected to post 97% revenue growth and roughly $213B of free cash flow this year, more than Apple, funded by capex flowing out of Alphabet, Meta and Amazon. The price hike is not margin defense against memory costs; it is Nvidia taking more of a pool it already dominates.

China's lead is now showing up in physical manufacturing, not models

3 sources

A DOE-commissioned report writes off U.S. semisolid-state battery startups as negligible while five Chinese firms build plants and SAIC already ships 440-mile EVs. LandSpace landed an orbital booster, the third private company ever to do it. Unitree popped 460% and Chery's humanoid unit filed to follow — though The Information's own Sunday Insights cautions that mainland listing mechanics, not demand, explain most of that pop.

20VC 15:18 PT

20VC Newsletter - 23rd August 2026

So what

Stop asking portfolio companies for per-seat AI ROI and start writing capped headcount into their operating plans instead.

Two episodes: Uber's president on extracting AI efficiency through OpEx rather than per-employee ROI, and Lemkin/O'Driscoll splitting on whether Stripe's OpenRouter buy is a mistake.

  • Uber runs a pod of 30 elite AI engineers embedded with G&A owners; weekly pricing went 15 hours to 2, marketing QA two weeks to two days
  • Macdonald's rule: measure AI at the OpEx line by capping headcount, because saved hours get reabsorbed and vanish
  • At ~$250B gross bookings, new Uber products must show a path to billions in transaction volume or get starved
  • The bear case on OpenRouter: high-reasoning B2B workflows standardize on one model to avoid drift, so routing is a developer niche
  • The bull case: a premium for a deployable product beats building it, and it buys immediate access to inference flows as a second growth engine
  • Uber burned $52M/week on subsidies at the DiDi war's peak; capital deployment bought a negotiated exit
ai roi measurement ×3 enterprise ai deployment ×5 m&a stripe / openrouter Open in Gmail →
The Information Briefing 15:02 PT

The Briefing: Nvidia, Salesforce on Deck

So what

Wednesday gives you both sides of the trade in one day: dollars leaving Alphabet, Meta and Amazon capex and landing in Nvidia's cash pile.

Peers frames Nvidia's Wednesday earnings as the whole AI trade in one print: growth is reaccelerating and free cash flow now dwarfs Apple's.

  • Analysts expect 97% July-quarter revenue growth, up from 85% in Q1; full-year growth of 83% after last year's 65%
  • Q1 free cash flow rose 85.7% to $48.6B; projected $55B in Q3 and $61B in Q4, roughly $213B for the year versus Apple's expected $144B
  • Nvidia is raising chip prices even as memory costs climb, funding both buybacks and stakes across data centers, labs and data-labeling
  • Salesforce reports the same day guiding to 10-11% growth; USDA is already cutting Salesforce usage in favor of AI-native suppliers
  • Alibaba raising HK$80B (~$10.2B) for AI capex; OpenAI's Americas sales VP Kaylin Voss resigned a week after CRO Denise Dresser left
nvidia earnings ×2 ai capex ×2 enterprise software disruption ×2 china tech ×4 Open in Gmail →
The Information 14:01 PT

Geely and Other Chinese Companies Steal a March on U.S. Solid-State Battery Makers

So what

Your solid-state battery exposure is likely a hedge, not a position — the defensible U.S. bet is silicon anodes and non-automotive niches.

A congressionally mandated DOE report says China has already won semisolid-state batteries, and Steve LeVine asks whether that is a race worth entering.

  • SAIC ships two EV models on semisolid cells with up to 440 miles range; the MG4 Urban hits the U.K. this year, Geely pilots next year
  • The Avicenne report projects solid state at 15% of China's EV market in four years and a fifth of global EVs by 2035, ~15M vehicles a year
  • U.S. players SES AI, QuantumScape and Factorial were excluded outright as "negligible"; none has a plant, a supply chain or a design win
  • Five Chinese firms are building large semisolid plants; 106 Chinese companies sit in the segment across the supply chain
  • The counter-argument: silicon anodes with liquid electrolyte deliver most of the benefit, and the Pentagon just put $1.4B conditionally behind Sila Nanotechnologies
  • Noteworthy: Pony.ai plans 4,000 robotaxis outside China; Chery's AiMoga humanoid unit files to IPO after Unitree
china tech ×4 china robotics ×2 batteries industrial policy Open in Gmail →
Masters of Scale 10:03 PT

How Barnes & Noble made a comeback

So what

Before you install a data-driven operator at a struggling retail asset, check whether the metric being optimized actually describes the customer.

James Daunt's Barnes & Noble turnaround is a case against the standard distressed-retail playbook: he removed data, hierarchy and marketing rather than adding them.

  • B&N had closed 150+ stores over a decade before Elliott bought it in 2019 and installed Daunt, fresh from Waterstones
  • Board books sell brilliantly per square foot, but toddlers reread one title; following that data left too little space for 10-plus readers and made both groups unhappy
  • He flattens store org charts deliberately and refuses to assign blame, so junior staff will say "Are you sure, guys?"
  • He stocked the Kindle at Waterstones mainly as a management move, to kill the fear and remind booksellers discovery was their moat
management ×2 retail turnarounds Open in Gmail →
The Information 10:00 PT

Why Scientists See AI Leaders as False Prophets

So what

The credibility risk in AI-for-bio pitches now sits with the founders' timelines, not their models — discount decade-scale cure claims in diligence.

Amy Dockser Marcus documents a growing scientific backlash against Hassabis and Amodei's disease-cure timelines, and it is coming from people who are otherwise AI enthusiasts.

  • Cardiologist Eric Topol, a self-described Hassabis superfan, publicly criticized the 60 Minutes claim that AI will cure all disease within a decade
  • Topol's line: no cures exist for diabetes, Alzheimer's or heart disease, and there is no precedent for drugs changing that in five to ten years
  • Stanford's Daphne Koller likened the AI-for-science narrative to believing magic wands are real
  • The criticism has moved from conference side conversations to blogs, podcasts and X — a reputational shift, not a technical one
ai hype cycle ×3 biotech ×2 ai in science Open in Gmail →
McKinsey & Company 09:42 PT

The books media executives can’t put down

So what

Nothing actionable; skim the list only if you want to know which AI framings media CEOs are absorbing this year.

McKinsey's annual book list, media-executive edition, with nearly 100 picks. Two of the four highlighted titles are about living with AI.

  • Nicholas Thompson of The Atlantic picks Carissa Véliz's Prophecy, on what happens when AI predictions harden into prophecies
  • Rebecca Blumenstein of NBC News picks Joanna Stern's I Am Not a Robot, a year of outsourcing her life to AI
  • Adam Grant flags Danny Meyer's What Could Possibly Go Right? on scaling an enduring culture
  • Stephanie Mehta picks Daniel Mendelsohn's An Odyssey, timed to Nolan's film
books ×2 ai hype cycle ×3 media Open in Gmail →
Superhuman AI 09:07 PT

💉 New vaccine stops cancer from returning

So what

LandSpace's leg landing resets launch-cost assumptions for any space or satellite position you hold that priced in SpaceX's monopoly.

The Sunday science edition, no AI news. The two items that matter are a Moderna cancer-vaccine readout and China's third-ever private orbital booster landing.

  • Moderna and Merck's personalized mRNA vaccine plus Keytruda kept high-risk melanoma patients cancer-free longer in a large human trial
  • LandSpace recovered the first stage of its Zhuque-3, joining only SpaceX and Blue Origin; the rocket is rated for 20 reuses
  • Chinese researchers wind-tunnel-tested an 85-meter-wingspan flying-wing concept carrying 800 passengers
  • Scientists used AI to redesign antibodies into charge-tuned intrabodies that survive inside cells, aimed at Alzheimer's and Parkinson's proteins
china tech ×4 biotech ×2 space Open in Gmail →
Sachin Rekhi 08:07 PT

How to Build a Compounding OS

So what

Mandate one agentic platform across your companies this quarter; the tool-choice debate is costing you more than picking the wrong tool would.

Rekhi argues the AI productivity gap between elite and average product teams is widening because elite teams built shared infrastructure — a "compounding OS" — instead of collecting individual hacks.

  • Elite teams see 2-3x gains in merged PRs; average teams get 20-30% and drown in slop
  • Three steps: standardize on one agentic platform, build a shared skill library, make company context machine-legible
  • His own skills include /conduct-nps-analysis (Notejoy moved from quarterly to automated weekly NPS), /update-competitor-pricing, /generate-competitor-teardown, /answer-data-curiosity
  • Fin CTO Darragh Curran: maximal tool permissiveness produced gains "mostly in isolation" until they standardized on Claude Code
  • Lock-in fear is overstated — MCP is cross-platform, skills are converging on a standard, context lives in markdown, and porting configs takes ten minutes
enterprise ai deployment ×5 agent skills ×3 ai productivity ×3 Open in Gmail →
The Information 08:01 PT

Why Unitree’s 460% IPO Stock Pop Wasn’t Unusual in China

So what

Discount mainland first-day pops heavily when marking your China robotics or memory comps — the pricing mechanism, not demand, is doing the work.

Unitree's 460% Shanghai debut looks like proof of Beijing's homegrown-champion strategy, but Yang and Osawa argue the mechanics of the mainland IPO market explain most of it.

  • Unitree popped 460% on Wednesday; CXMT popped 472% on its first day last month
  • The pattern is now a series, not a one-off, across mainland tech listings
  • Read as euphoria plus a structurally supply-constrained listing process, not purely as a valuation signal
china robotics ×2 china tech ×4 ipos Open in Gmail →
Robbie Allen (ACG on AI) 07:55 PT

The overage bill is the good news

So what

Write down your portfolio's 18-month headcount plan by function today — that list is the only defensible AI ROI baseline you will have in 2027.

Allen predicts 2027 is when CFOs demand AI returns, and warns that the value is structurally harder to measure than the invoice — so start keeping score now.

  • Sequence: 2025 was "what is AI?", 2026 was "get it deployed," 2027 will be "what is it doing for me?" with an invoice attached
  • Mid-market firms spend five to six figures a year, large enterprises seven; the overage bill is proof adoption worked, not proof of return
  • His answer to CEOs asking about AI ROI: what financial return did you get from PowerPoint?
  • Two separate scoreboards: capacity value (tracked via headcount plans, team scope, operating metrics) and project value (baseline, business owner, operating cost, review date)
  • A project saving $200K that costs $80K to run is a good project, not a $200K return; nontechnical employees' builds eventually need managed infrastructure and an owner
ai roi measurement ×3 enterprise ai deployment ×5 ai hype cycle ×3 Open in Gmail →
Peter Yang 07:03 PT

How to Build Better AI Evals with Claude Code in 5 Steps | Shreya & Hamel

So what

If your teams' eval suites are all rules written before launch, they are measuring compliance and missing every failure mode that matters.

Shreya Shankar and Hamel Husain audit Peter Yang's evals live and find the common failure: strong top-down rules, almost no bottom-up error discovery.

  • Top-down evals are rules defined upfront ("is every takeaway 240-330 characters?"); bottom-up evals come from diffing AI output against your own final edits
  • Their fix: run the skill on 10-20 past examples before promoting any failure into an eval, to avoid overfitting to a handful of cases
  • They released a free Error Discovery skill that runs in Claude Code or Codex and walks through the five steps
  • The episode also covers spinning up agents to grade each criterion and where human judgment still can't be removed
agent skills ×3 ai productivity ×3 ai evals Open in Gmail →
Eric Ries 07:02 PT

The force that kills companies

So what

Useful frame for board seats where the exit conversation has started to reshape what the founding team is willing to defend.

Ries introduces "financial gravity" — the pull of future transactions on behavior and values — as the explanation for what he calls unusual corporate failure, with Polaroid as the case.

  • Polaroid's board forced out Edwin Land in 1982; Jobs called it "the dumbest thing I've ever heard" and Land "a national treasure"
  • R&D atrophied, scientists left, spending collapsed, bankruptcy followed in 2001 — not competition, not creative destruction
  • Three symptoms: investors killing golden geese, drift into bureaucracy and hollow culture, and advisors loyal to the financial system over the client
  • His claim: the more valuable a company becomes, the more vulnerable it is; success causes the failure
  • Free workshops Aug 26 on mission/purpose/vision/ethos and Aug 27 on regenerative growth in the age of AI
management ×2 books ×2 governance Open in Gmail →
Bain Inside Software 07:00 PT

Inside Software: Are You Pricing the Old Workflow or the New One?

So what

Ask each software company you hold which of Bain's seven triggers has fired; if none has, seats plus entitlements is a defensible answer.

Bain's rule for AI pricing: not every software company needs to change its model, but every one needs to know why it isn't. The trigger is whether the product now does the work rather than helping a human do it.

  • "Seats are dead" is the symptom; the real question is whether AI changes how work gets done, inside your product or elsewhere in the customer's stack
  • Informatica's cloud transition, under Permira and CPP, went from ~$150M to ~$800M cloud consumption ARR in about four years
  • Informatica Processing Units were a prepaid drawdown, not pay-as-you-go — more flexibility than an ELA, more predictability than post-paid
  • The real win was the second sale: committed currency removed the new SKU, new business case, new procurement cycle from every expansion
  • Seven diagnostic questions, including whether AI-heavy usage decouples value from cost-to-serve and whether you are creating more value than the model can capture
enterprise software disruption ×2 saas ×2 ai pricing models Open in Gmail →
Every 06:38 PT

Life After Automation

So what

Headway's wait-buy-build outcome is the template: compliance requirements, not capability gaps, are what push regulated buyers to build in-house.

Every's weekly roundup is about the operating cost of living at the frontier: a 230% AI bill jump, one engineer running a roster of agents, and a healthcare company building what it couldn't buy.

  • GPT-5.6 Sol pushed Every's daily usage from 11,520 to 26,685 credits; ops head Arielle Shipper chose parameters over hard token budgets, interrogating big runs on cost, purchase and learning
  • Naveen Naidu runs Monologue as specialized Codex agents handing work to each other like coworkers
  • Headway built Eddy on the Claude Code SDK when off-the-shelf tools failed compliance; every conversation runs in a sealed disposable container, and 650 of 900 employees use it daily
  • Thesis Statements collects 100 predictions on work after automation; first 25 live, including Linear's Karri Saarinen arguing design becomes AI's biggest problem
  • Every named a seven-person frontier team with explicit permission to prioritize weird experiments over urgent work
enterprise ai deployment ×5 ai productivity ×3 agent skills ×3 ai costs Open in Gmail →
Lenny's Newsletter 05:31 PT

How to close $100K+ enterprise deals, step by step | Jen Abel

So what

Hand this to any portfolio founder whose pipeline reviews keep surfacing deals stuck between "demo" and "proposal" with nothing in between.

Jen Abel's third appearance maps the full enterprise sales cycle as roughly 15 steps, not the five stages your CRM shows.

  • The standard five-stage pipeline is a forecasting artifact, not a sales process
  • The "pincer model": open at the executive and N-minus-one level simultaneously
  • Outreach is two to three sentences built around giving the buyer alpha, not describing a feature
  • Run the intro call to extract intelligence before any demo; structure pilots at two to three days with jointly defined success criteria
  • Guidance on when to charge for a longer pilot versus giving it away, and on surviving procurement and redlines without losing momentum
saas ×2 enterprise sales gtm Open in Gmail →
The Information 16:45 PT

Nvidia AI Chip Prices to Rise About 17%, Server Makers Tell Customers

So what

Reprice every gigawatt-scale data center model in your book; the chip line just moved by $5B per GW before any memory-cost pass-through.

Server makers are telling customers Nvidia flagship chip system prices go up about 17%, on systems shipping next year.

  • Applies to Grace Blackwell 300 and Vera Rubin 200 systems due for delivery in 2027
  • Adds at least $5B to the cost of a 1 gigawatt data center at current system pricing
  • Two sources whose firms were notified directly by server makers
nvidia earnings ×2 ai capex ×2 datacenters Open in Gmail →
GAI Insights (Paul Baier) 15:38 PT

26 AI Use Cases for the CFO: What Finance Teams Are Actually Building

So what

The BI license line is the cleanest AI savings you can book this year — audit Tableau and Power BI spend across the portfolio before renewals.

Baier catalogs 26 finance-function AI use cases pulled from hands-on CFO workshops, all built on Claude, with the savings attached.

  • One team insourced its 13-week cash flow forecast from a consultant, saving roughly $70,000 a year
  • Flux analysis and variance commentary save one controller 3 to 5 hours of analyst time per close cycle; one team targets cutting the close from 10 days to 2
  • Teams are cutting Power BI and Tableau licenses as Claude handles data cleaning, modeling and report generation directly
  • Long tail includes a 70-slide PE reporting deck, ASC 606 and 842 lookups against specific facts, PDF purchase-order-to-sales-order write-ups, and a 21-point source tie-out check
  • Risk control is procedural: humans review before anything enters the record, and the analyst signs off on anything auditors see
enterprise ai deployment ×5 ai roi measurement ×3 finance Open in Gmail →