AI So Whats
Edition

Wednesday, August 26, 2026

18 issues 23h window ~15 min read

Running through today’s issues

Nvidia is financing the companies building its replacements

4 sources

The Information's Nvidia roundup counts $105 billion in credit for one data center, $500 billion in outside financing, and investment talks with Perplexity, Mercor and Poolside in sixteen days. The Briefing points out Nvidia committed $30 billion to OpenAI months before OpenAI published benchmarks claiming Jalapeño beats Blackwell. Thompson closes the loop: the balance-sheet deals are a race to lock in customers before the CUDA moat erodes and rival supply chains stabilize. Apple's M5 Ultra, at 1,200 GB/s and half the power of a $16,000 RTX card, is the same pressure from the consumer end.

Evals are becoming the actual procurement layer, and finance teams are buying them

3 sources

Every argues public benchmarks cannot tell you whether a model does your job, and shows KateBench's 85-90% acceptance rate is unstable across identical runs. The Information's payments report has Brex reaching the same conclusion from the CFO's side — Franceschi says routing isn't the problem, knowing whether the cheap model performs like the expensive one is, and Brex is partnering with Mercor to build customer evals. Mercor's Foody appears in both, describing executives spending $100 million a year with no offline evals. Eval infrastructure is moving from research hygiene to a line item in expense management.

Agent traffic is the demand signal repricing the data layer

4 sources

Observability startups are taking the windfall: ClickHouse at $350 million ARR up 40% since May on OpenAI usage, Cribl near $400 million, Grafana past $400 million, with Elastic, Dynatrace, Snowflake and Palo Alto all buying in. Applied AI shows the same pull creating a knowledge-graph market, with Atlassian claiming 50% token savings and Neo4j's Q2 beating its full 2025. Sacra puts a number on the traffic itself: agents went from under 3% to over 50% of Vercel deployments in seven months. Palo Alto's exploration of Datadog and Okta says the strategics see monitoring and identity as the same purchase.

Everything Apple shipped this week was designed for a customer that didn't exist last year

2 sources

Superhuman traces the new Mac mini line directly to February's OpenClaw run on Mac minis by AI hobbyists; Thompson notes prices are up again on top of June increases and reads the month-long pre-announcement as Apple refusing to commit scarce RAM and SoCs to configurations that won't sell. Both frame the Studio as a genuinely competitive local inference box. Nvidia's DGX Spark and Perplexity's Portable Computer are chasing the same buyer.

Superhuman AI 06:14 PT

🖥️ ChatGPT Work gets new powers

So what

Credentialed browser agents change your fraud and access-control surface overnight — check whether your portfolio's login flows can tell an agent buyer from a scraper.

OpenAI gave ChatGPT Work stored credentials so it can log into sites and complete errands, while Google shipped its first vertical enterprise agents for legal and financial services.

  • ChatGPT Work now signs into your most-used sites via password autofill and books appointments, fills applications, and runs anything browser-plus-login; 19 examples cited
  • Claude gets cross-tool memory persisting across Claude and Cowork, on by default, editable in Settings
  • Google launches ready-to-deploy agents, specialized skills and third-party connectors for legal and financial services, aimed squarely at Anthropic and OpenAI enterprise accounts
  • Apple's new Mac mini line (M6 at $899, M5 Pro at $1,699) is repositioned around the local-agent buyer created by February's OpenClaw run on Mac minis
  • Perplexity and Nvidia launched Portable Computer on DGX Spark; Darkbloom rents out idle Mac compute; sponsor data claims agentic traffic up 7,851% YoY
ai agents ×4 local ai hardware ×2 enterprise ai ×8 openai ×3 Open in Gmail →
The Information 06:01 PT

Exclusive: Nikesh Arora Cranks Up Palo Alto Networks’ M&A Machine

So what

If you hold observability or identity assets, Palo Alto is a live bidder at scale — price your exit against a strategic, not a growth-equity round.

Nikesh Arora has pushed Palo Alto Networks' acquisition machine past 40 deals in eight years, and the last 18 months have been the most aggressive stretch yet.

  • Arora explored acquiring Datadog and Okta, per two people with knowledge of the situation
  • The stated logic is covering new AI threats and extracting more revenue from existing customers
  • Palo Alto already bought observability startup Chronosphere for $3.35 billion
  • Targets skew toward monitoring and identity — the two layers agents most need to be governed at
security m&a ×2 observability ×2 enterprise ai ×8 Open in Gmail →
The Information 04:13 PT

SpaceX Announces $100 Billion Starship Launch Site in Louisiana

So what

OpenAI's enterprise sales bench is hemorrhaging back to Salesforce right as its GTM ramps — discount its enterprise revenue trajectory accordingly.

The AM briefing pairs SpaceX's $100 billion Louisiana Starbase with three AI items that all cut the same way: OpenAI's silicon claim, its sales-team exodus, and DeepSeek's tenfold revenue jump.

  • SpaceX announced Starbase, Louisiana — $100 billion, construction from 2027, first launch targeted 2029, built for 30+ Starship flights per day
  • Peter Doolan is returning to Salesforce as chief customer officer of its AI labs initiative, following Kaylin Voss; another 22 ex-Salesforce staff at OpenAI are in active talks to leave
  • DeepSeek did 475 million yuan ($70.7M) revenue in seven months, ~10x its full-year 2025, at 44.6% gross margin and 82.9% on API; raising 50 billion yuan at a 500 billion yuan valuation
  • OpenAI's Jalapeño inference chip claims to beat Blackwell, built by ~100 people with Broadcom; Microsoft discussed buying billions of dollars' worth
  • LayerZero launches ATLAS, an institutional blockchain exchange with Citadel Securities, DTCC and ICE as partners, months after a $292 million bridge hack
openai ×3 openai jalapeno ×3 deepseek ×2 enterprise ai ×8 Open in Gmail →
Ben Thompson (Stratechery) 03:01 PT

Apple Updates Mini and Studio, AI Computers, OpenAI Jalapeño (Stratechery Update 8-26-2026)

So what

Nvidia's balance-sheet deals now read as a race to lock in customers before CUDA stops mattering — size your Nvidia exposure to that clock, not to earnings.

Thompson reads Apple's new Mini and Studio and OpenAI's Jalapeño as the same story from two directions: focused rivals peeling off distinct slices of Nvidia's market.

  • Mac mini M6 at $899, M5 Pro at $1,699, Studio M5 Max at $2,499, M5 Ultra at $5,499 — all up again on top of June increases, which he reads as memory prices rising faster than Apple forecast
  • M5 Ultra bandwidth goes 800 to 1,200 GB/s with huge unified memory at under half the power draw of a $16,000 RTX Pro 6000 card; "the best hardware company in the world"
  • The month-long gap between announcement and Sept 22 shipping is negative working inventory — Apple won't commit scarce RAM, SSDs or SoCs to configs that won't sell
  • Jalapeño: 1.5–1.9x more work per watt and 1.7–3.6x lower latency than comparison systems across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T; ~16 months hiring to tape-out
  • Caveats: OpenAI ran the tests, Jalapeño uses HBM4 so Vera Rubin is the fair comparison, and volume shipping is next year
openai jalapeno ×3 local ai hardware ×2 nvidia balance sheet ×3 apple silicon Open in Gmail →
Noahpinion 02:12 PT

The death of Market Street

So what

SF downtown recovery is real but Market Street is not part of it yet — underwrite Mid-Market real estate on years, not quarters.

Noah Smith argues Market Street died from three compounding blows — the 2019 car ban, the crime wave, and remote work — and that the car ban is the one nobody wants to blame.

  • Weekday foot traffic on Market is still down by half versus 2019; the San Francisco Centre mall closed in January 2026 and stands boarded up
  • 2025 saw property crime fall 27%, burglaries 29%, larceny 22%, robbery 24%, with further drops in 2026
  • Better Market Street was scoped at $604 million after topping $1 billion, then cut to three blocks; the elevated bike lanes were scrapped in late 2020
  • Bike traffic is lower than pre-pandemic despite removing cars — the street subtracted cars without adding pedestrian space
  • The AI boom is producing a modest office-demand rebound, plus office-to-housing conversions and RTO mandates from companies and the city
san francisco urban policy Open in Gmail →
The Information 23:51 PT

DeepSeek’s Revenue Reaches $70 Million as of July, Tenfold Jump from 2025

So what

A Chinese lab is approaching an IPO with 82.9% API gross margins — that is the price floor your inference-cost assumptions have to survive.

DeepSeek's standalone scoop: seven months of 2026 produced roughly ten times its entire 2025 revenue, and the losses are shrinking relative to that base.

  • 475 million yuan ($70.7 million) revenue January–July, versus roughly a tenth of that for all of 2025
  • Net loss of 715 million yuan over seven months, against 935 million yuan for full-year 2025
  • Second funding round targets 50 billion yuan at a 500 billion yuan valuation, immediately after closing a first round of the same size in June
  • Investment banks are already hired for a Shanghai listing next year
deepseek ×2 frontier-lab capital ×3 Open in Gmail →
The Information Briefing 17:01 PT

The Briefing: OpenAI’s Hot Pepper

So what

Nvidia's vendor financing is now funding its own substitutes — treat every Nvidia-backed customer commitment as a shorter-dated asset than the headline term.

Martin Peers' read on Jalapeño is that the awkward part is not the salmonella joke but that Nvidia funded the company now publicly claiming to beat it.

  • Nvidia committed $30 billion to OpenAI a few months ago, partly to underwrite demand for its own products; OpenAI's agenda is cutting hardware costs and improving margins
  • SemiAnalysis's review headline: Jalapeño is better than Nvidia Blackwell; Altman's contribution was "we made a chip and it is fast"
  • Peers dismisses the 12-state AG suit against Paramount Skydance's $110 billion Warner Bros. Discovery purchase as fighting a 1990s cable war; combined debt would be $79 billion at closing
  • Zoom reported 4.9% July-quarter revenue growth, down from 5.5%, and guided lower still
  • ClickHouse recurring revenue passes $350 million; Meta plans to launch a 'Hatch' AI agent platform in coming weeks
openai jalapeno ×3 nvidia balance sheet ×3 openai ×3 Open in Gmail →
The Information 16:34 PT

Why Revenue Is Hopping at Data Observability Startups

So what

Observability is the cleanest agent-adoption proxy trading in private markets right now — and the strategic bid is already at 2-9x the last venture mark.

Agent-driven data volumes are producing a revenue windfall at observability startups, and the strategics have started buying before the multiples reset.

  • ClickHouse ARR up 40% since May to $350 million, with OpenAI usage jumping; last valued at $15 billion in January and fielding strategic approaches it says it is rebuffing
  • Cribl near $400 million ARR, up 33% since February, last valued at $3.5 billion in 2024, targeting an IPO in a couple of years
  • Grafana passed $400 million ARR last September at a $9.6 billion March valuation; half its customers used its AI assistant this year
  • Deals: Elastic bought Deductive AI for $85 million, Dynatrace paid $915 million for Arize AI, Snowflake closed $1 billion for Observe, Palo Alto paid $3.35 billion for Chronosphere
  • Cribl's Sharp frames the thesis bluntly: data grows 30% compounded, budgets don't
observability ×2 security m&a ×2 ai agents ×4 enterprise ai ×8 Open in Gmail →
The Information 16:30 PT

Everything Else Nvidia Bought Before Tomorrow’s Earnings

So what

Nvidia is now a counterparty in half your AI cap table — map your indirect exposure before its earnings call reprices the whole chain.

A promotional roundup that lands as an argument: in sixteen days before earnings Nvidia acted like a financier, not a chip vendor.

  • $105 billion in guaranteed credit for a single data center, plus $500 billion in outside financing lined up
  • Investment talks opened with Perplexity at $30 billion-plus, Mercor at $20 billion, and a $6 billion licensing-and-hiring deal with Poolside
  • The Nvidia/OpenAI/SoftBank Ohio data center project was finalized
  • Nvidia is using land and electricity deals to lock in its full hardware bundle
  • The framing question: what happens to all of it if AI spending slows
nvidia balance sheet ×3 frontier-lab capital ×3 Open in Gmail →
The Information 14:24 PT

Atlassian Rides the Knowledge Graph Boom

So what

Token-efficiency claims are becoming the pricing wedge for graph vendors — demand the 50% number be reproduced on your own workload before you sign.

Applied AI argues knowledge graphs are the new lock-in layer, with Atlassian, ServiceNow and Neo4j all riding agent adoption — and a Copilot skills flaw showing the security cost.

  • Atlassian says customers pairing Teamwork Graph with Codex or Claude Code used nearly 50% fewer tokens; Cannon-Brookes calls it the "most underappreciated part of Atlassian"
  • Teamwork Graph is free today, but Atlassian says billing could shift to how often AI taps the database
  • Neo4j's Q2 revenue exceeded its entire 2025 revenue; last privately valued at $2.4 billion in 2021
  • ServiceNow charges customers who use outside AI like Claude Code against its graph; Microsoft has walled off customer data from competing software firms
  • PromptArmor built a malicious Copilot skill that exfiltrated Outlook, SharePoint and Teams data past Microsoft's built-in skills scanner; Microsoft has patched it
enterprise ai ×8 ai agents ×4 agent security Open in Gmail →
McKinsey & Company 13:17 PT

[In case you missed it] McKinsey Technology Trends Outlook 2025

So what

The humanoid-to-physical-AI value ratio is the number to quote when someone pitches you a humanoid round.

A resend of the 2025 Technology Trends Outlook with a teaser for the 2026 edition, plus a physical-AI webinar pitch that carries the only real number.

  • 108-page report covering breakthroughs, talent trends and use cases across sectors
  • $5.1 billion invested in humanoid robots since 2024
  • Humanoids projected to be just 2% of physical AI's value by 2045
  • September 9 session with Mark Patel and Ani Kelkar on where the actual value pools sit
physical ai ×2 consulting research ×2 Open in Gmail →
McKinsey & Company 13:12 PT

[In case you missed it] The State of Fashion 2026: When the rules change

So what

Nothing new here; skip unless you hold consumer retail and want McKinsey's framing for a board deck.

A resend of The State of Fashion 2026, arguing the sector has moved past uncertainty into a demand for agility.

  • Economic volatility, shifting consumer priorities and technological disruption named as the binding constraints
  • Authors include Anita Balchandani, David Barrelet, Felix Rölkens and Gemma D'Auria
  • Same physical-AI webinar attached: $5.1 billion into humanoids since 2024, projected 2% of value by 2045
consulting research ×2 physical ai ×2 Open in Gmail →
The Information 13:06 PT

Takeaway: Why Big Tech is Suddenly Paying Up for Startups

So what

Nine bidders for a model router means any control point between apps and models now clears at a strategic premium; reprice your infra positions.

Jessica Lessin's Takeaway uses OpenRouter's sale process to explain why Big Tech has abruptly stopped lowballing startups.

  • OpenRouter drew a reported record nine suitors once news of its sale talks broke
  • Bankers were calling Lessin for information; at least two people she called placed bids after hanging up
  • OpenRouter's function — switching between models and centralizing billing — is what made it strategically contested
stripe / openrouter ×2 frontier-lab capital ×3 ai m&a Open in Gmail →
Sacra 12:29 PT

$340m/yr heroku for claude code

So what

Vercel trades at 35x while growing a tenth as fast as Replit and Lovable — the deployment-layer premium is doing all the work in that multiple.

Sacra's read: Vercel lost the app-builder race with v0 but won something better — becoming the default place agents deploy what they build.

  • Agents initiated over 50% of Vercel deployments by June 2026, up from under 3% in late 2025; Claude Code alone drove 75% of agent deployments in April versus 6% for v0 and Lovable combined
  • Vercel at $340M annualized in February 2026, +112% YoY, valued at $9.3B on its September 2025 Series F — roughly 35x
  • Comparables: Replit $525M annualized (+1,775% YoY) at $9B, Lovable $500M (+3,471% YoY) at $6.6B, both cheaper multiples on faster growth
  • Four acquisitions in twelve months: NuxtLabs, new.website, Better Auth for agent identity, and Stakpak for DevOps agents
  • Vercel is now pushing into Workflows (vs Temporal, Inngest), Sandbox (vs E2B, Daytona) and Services (vs Railway, Render, Fly.io)
vibe coding ×2 ai infrastructure ×2 ai agents ×4 Open in Gmail →
Every 12:20 PT

Benchmarks Don’t Know Your Job

So what

Build one offline eval on a recurring job this week; without it you cannot tell whether a cheaper open model is costing you rework.

Every's case: public benchmarks cannot tell you whether a model does your job, and its own KateBench shows how badly a good-looking acceptance rate can mislead.

  • Mercor's Brendan Foody meets executives spending up to $100 million a year on models with no offline evals; Aaron Levie: "Enterprises will not be able to go just on vibes"
  • KateBench, trained on ~30,000 of Kate Lee's past edits, gets 85–90% acceptance — but it capped at 40 suggestions (now 80) and produces different edits run to run, so the number isn't stable
  • Open-source models went from 28% to 62% of token share on Vercel in two months, per Guillermo Rauch; a16z reports legal-worker Codex adoption up 108x since February
  • Thinkingbox found the strongest coding model passed 65% of single attempts but only 25% when required to succeed across 20 consecutive attempts
  • CentaurBench: on five of seven tasks, the best solo model was not the best at improving a weaker model's first attempt
evals ×2 vibe coding ×2 enterprise ai ×8 Open in Gmail →
The Output (DataCamp) 10:35 PT

Which AI role has surged 255% YOY?

So what

Your junior hiring pipeline is where AI displacement is actually showing up in the data — plan two years out for a missing mid-level bench.

DataCamp's State of AI Careers 2026, drawn from two million postings across 85 regions, says hiring is booming everywhere except the entry level.

  • Data and AI job postings up 80.2% this year; AI engineer up 255%, generative AI engineer up 197%
  • Employment for 22-to-25-year-olds in AI-exposed fields fell relative to trend while the same fields grew for older workers
  • Four skills appear in all 25 roles studied: Python, communication, SQL, computer science; AI itself in 22 of 25
  • Elsewhere: Grok 4.6 matches a top OpenAI model at a fifth of the cost; Scribe's Jennifer Smith on agents failing for want of institutional knowledge
enterprise ai ×8 ai labor market Open in Gmail →
The Information 09:01 PT

The Leaders You’ll Meet at WTF 2026

So what

Nothing to act on beyond the calendar; book it or bin it.

A promotional email for the WTF Summit in Napa, October 27–28, themed around Building.

  • Two days of on-the-record conversations with founders, investors, operators across AI agents, SaaS, capital, robotics and space
  • Presenting partner PwC; premier partner Legora; supporting partner Meta
  • Subscribers get an additional $200 discount by logging in before purchase
events Open in Gmail →
The Information 08:01 PT

The Big Opportunity in Wrangling AI Payments

So what

If your portfolio companies resell tokens, demand per-customer gross margin now — several are likely selling below cost without seeing it.

A special report on the mess AI billing has become, and the payments companies racing to sell the cleanup.

  • Workato's CIO reconciles AI invoices against usage dashboards in a standing Wednesday 2pm meeting with the CFO and head of engineering because the two don't match
  • Brex, acquired by Capital One for $5.15 billion this year, will launch Magpie, pulling usage and billing from provider APIs, Cursor and invoices into one format
  • Stripe closed its $1 billion Metronome acquisition; Adyen agreed to buy Orb for $335 million in June; Stripe is buying OpenRouter for a reported up to $7.5 billion
  • Franceschi: routing isn't the problem — knowing whether a cheap model performs like an expensive one is; Brex is partnering with Mercor to build customer evals
  • Many companies reselling tokens may be unprofitable at gross margin and not know it, because cost-to-revenue attribution has broken down
stripe / openrouter ×2 evals ×2 ai infrastructure ×2 enterprise ai ×8 Open in Gmail →