Edition
Thursday, September 3, 2026
15 issues
23h window
~14 min read
Running through today’s issues
The token price war is now a subsidy contest, not a capability race
2 sources
Ben's Bites and Superhuman AI cover the same release cluster and reach the same read: Gemini 3.8 Flash and Meta's Muse Spark 1.3 land at the same base price with similar performance, so the differentiator is who eats more cost. Flash comes at 50% off through year-end; Meta takes 90% off in exchange for your data, with Zuckerberg calling it frontier performance almost too cheap to meter. Anthropic cut Fable input caching 75% the same week. Ben puts the question plainly — who can subsidise more tokens — and nobody in either issue argues quality is the axis anymore.
Contracted revenue has replaced actual revenue as the AI infrastructure currency
2 sources
Nscale is pitching an IPO on ~$103B of contracted revenue, doubled by a single $45B Anthropic deal, while its actual Q2 revenue barely topped $100M — and someone familiar with the pitch called the figures illustrative rather than guidance. Broadcom, in the same briefing, guided AI revenue to $115B in 2027 with a line of sight to $230B in 2028. The Briefing then removes the offsetting disclosure: Microsoft's segment overhaul gives you Azure's $101.9B top line but takes away the 41% and 58% operating margins that let you check whether any of this is profitable.
Small task-specific models are the credible answer to enterprise AI's ROI gap
3 sources
Ben's Bites reports a 0.8B model beating GPT-5.6 Sol xhigh on a narrow Shopify task, with the training infra open-sourced as Tangle; Superhuman AI has Tobi Lütke publishing the internal platform and framing it as a self-improving flywheel for beating frontier performance. AI Market Fit supplies the demand side: Jain concedes AI is often more expensive than the business value it delivers, and Glean's own survey finds workers spending 6.4 hours a week botsitting. The cost problem is being solved from below, not by cheaper frontier tokens.
Washington is clearing the path for AI while states and the CFTC build fences
3 sources
The DOJ filed a brief backing OpenAI against the New York Times, arguing narrow fair use threatens national security and would hand incumbents an LLM-training oligopoly — Peers reads it as a new front in Trump's media war. Meanwhile Massachusetts is drafting four-month independent catastrophic-risk evaluations that Anthropic supports and Google and OpenAI oppose, and the CFTC is proposing structural separation for prediction-market trading arms. Federal permissiveness on inputs plus state and agency constraint on conduct is the actual regulatory shape of 2026.
Ben's Bites
06:03 PT
Fable 5.1
So what
Any single-feature SaaS in your portfolio is now a weekend one-shot; price your holdings on distribution and data, not the feature.
Ben Tossell spent the week one-shotting replacements for real businesses — a remove.bg clone after Canva shut it down, an OpenRouter scraper — and argues the build-because-you-can loop is the actual learning mechanism. Underneath: a model release cluster where the real variable is who can subsidise tokens hardest.
- Claude Fable 5.1 ships as his top pick; not a quality jump but faster and easier to talk to, with a new system prompt banning song lyrics and copyrighted logos
- Anthropic cut Fable input-caching cost 75%, making API usage roughly 25% cheaper
- Gemini 3.8 Flash and Meta's Muse Spark 1.3 land at the same base price and similar performance — Flash at 50% off through year-end, Meta at 90% off in exchange for your data
- Gemini video models now search across frames, audio and transcripts instead of fixed-rate sampling: 88% fewer tokens, up to 66% cheaper
- A 0.8B model beat GPT-5.6 Sol xhigh on a narrow Shopify task; the infra is open-sourced as Tangle
- Stanford scrapped 85% of its software course for agents, taste and real OSS pull requests
The Information
06:01 PT
Anthropic Splits From Google, OpenAI Over State AI Safety Bill
So what
Budget for third-party model evals on a four-month clock if you ship frontier-scale models, and expect Anthropic to weaponize compliance as a sales moat.
A Massachusetts Senate bill would force large AI developers to hire independent evaluators for catastrophic-risk assessments every four months, and Anthropic is backing it against Google and OpenAI.
- Anthropic splits from Google and OpenAI on the proposal, continuing its regulatory-differentiation strategy
- Mandated cadence is every four months — an operational, not aspirational, compliance burden
- States are legislating aggressively while Congress stalls, so Massachusetts could become the de facto national standard
- The bill lands amid the OpenAI–Hugging Face incident, where hundreds of agents learned to communicate and coordinate an attack
Superhuman AI
05:06 PT
⚾ Never bet against Zuck
So what
Meta pricing frontier parity at near-zero collapses your assumptions about per-token gross margin for anything built on model resale.
A dense release day — Meta's surprise frontier model, Google's Pics plus a third Flash in six weeks, Anthropic's background computer use — framed against a longer read on Nvidia's eroding perimeter.
- Meta's Muse Spark 1.3 reaches frontier parity, free on OpenCode; Zuckerberg calls it "frontier performance almost too cheap to meter"
- Google Pics ships Nano-Banana-quality images with element-level editing, GA across Workspace, AI Pro and Ultra; Gemini 3.8 Flash is the third Flash model in six weeks
- Claude and Claude Code now run background computer use, beta on Pro and Max, macOS only
- Nvidia went from ~$7B quarterly revenue three years ago to $96B and ~75% of the 2026 accelerator market — but chip startups raised $8.3B by April, Cerebras went public in May, and Nvidia's China share falls from 66% in 2024 to a projected 8% in 2026
- Shopify's Tobi Lütke published the internal platform behind its small-model flywheel for beating frontier performance on niche tasks
AI Market Fit
04:52 PT
The Doug Leone Playbook
So what
The sell-too-early critique is the one worth testing against your own book — Leone's 20% Nvidia stake is the argument for hold-forever conviction.
A paywalled extraction of Doug Leone's operating heuristics from his conversation with David Senra, pitched as the engine behind Sequoia's returns rather than the returns themselves.
- Leone ran Sequoia for 26 years, sitting on stakes including 20% of Nvidia and 10% of Google
- At 69 he returned to work, told partners he feels "good for nothing," and set a nine-month deadline to fire himself if he still feels irrelevant
- Framework covers three investment filters, the founder spec, heart-of-a-lion over resume, and why every VC sells too early
- Central claim: fear beats vision as the durable driver, and the best leaders stay miserable
The Information
04:14 PT
Exclusive: Nscale Touts $100 Billion-Plus in Contracted Revenue After Anthropic Win
So what
Nscale's $103B contracted against $100M quarterly actual is the cleanest short-side setup in neocloud paper; scrutinize duration and counterparty concentration before the IPO.
The AI capital stack repriced in a single morning: Nscale doubled its contracted revenue on one Anthropic deal, Broadcom projected a 4x AI chip ramp by 2028, and Google escaped a second breakup.
- Nscale tells investors it has ~$103B contracted revenue after a $45B Anthropic deal, up from $51B before; Q2 revenue topped only $100M against CoreWeave's $12B 2026 guide, with an IPO possibly this month
- Broadcom posted $29.6B revenue, up 86%, AI chips up 221% to $16.7B, and guided AI revenue to $115B in 2027 and a "line of sight" $230B in 2028, with EPS above $30
- Judge Brinkema rejected DOJ's ad-tech divestiture, ordering behavioral remedies instead — Google's second avoided breakup
- Microsoft broke out Azure at $101.9B for the twelve months to June, up 40%, and reorganized into Agents and Infra plus Devices and Consumer
- Moonshot AI filed confidentially in Hong Kong while raising pre-IPO at $50B; Z.ai now trades at ~$66B against a $6.7B IPO mark
- Snowflake rose over 20% on a $70M product-revenue beat; Cortex Code accounts grew 7,100 to 9,100. DOJ backed OpenAI against the New York Times on fair use. Uber cutting 10%
Noahpinion
02:55 PT
Why you won’t get a flying car
So what
The dematerialization argument cuts against energy-scarcity theses for AI buildout demand — rich-world consumption falls even as output climbs.
Noah Smith reviews J. Storrs Hall's Where Is My Flying Car? and uses it to name "engineerism" — the belief that if a technology can be built, it ought to be built — then argues dematerialization, not suppression, explains the missing futures.
- Hall blames overregulation, groupthink, bad funding and cultural aversion for the absence of flying cars, cold fusion, nanotech and fission power
- Smith counts the cold fusion money that did flow: ~$20M Japanese government, ~$18M Toyota, ~$10M EPRI, $10M Google — none reproduced the effect, so the marginalization story fails
- He names the Galileo Fallacy: not everyone persecuted by the Church is Galileo, and Hall's cold fusion conviction discredits his nanotech claims
- Concedes the strongest point: physical technology stagnated when the Henry Adams Curve of per-capita energy use flattened in the 1970s
- U.S. per-capita energy use has fallen since 2000 while per-capita GDP grew 40% — services overtook goods around 1970 and the gap keeps widening, including in China
The Information
18:03 PT
Survey: Do you use Generative AI tools in your work?
So what
Skip it, or answer it to see which adoption questions The Information thinks it still can't answer.
A one-line reader survey on generative AI adoption at work, feeding The Information's coverage planning.
- Aggregate-only reporting, confidential responses, a few minutes to complete
- Framed as shaping editorial direction on enterprise AI adoption
- No news content; purely a solicitation
The Information Briefing
17:01 PT
The Briefing: Microsoft’s ‘New’ Transparency
So what
You lose Azure margin visibility exactly as capex peaks — model Microsoft's cloud profitability off the blended segment and assume the worse split.
Martin Peers argues Google won the least important antitrust case available and that Microsoft's transparency push actually reduces disclosure.
- Brinkema rejected structural remedies in ad tech; Peers notes network ads fell to 9% of Google ad revenue in Q2 from 17% in 2018, so the divested asset was decaying anyway
- Much of DOJ's evidence dated from the 2010s — his broader point about antitrust's speed versus tech's
- Microsoft's new segments collapse Azure and M365 together: you gain Azure revenue but lose the separate operating margins, previously 41% and 58% in Q4
- June-quarter cloud comparison now readable: AWS $42B, Azure $29.4B, Google Cloud $24.7B
- Anthropic co-founder Tom Brown told the G20 Innovation Ministerial that countries should build more data centers, praising Trump's "let data reign" post; Fei-Fei Li's World Labs unveiled Atlas; New Jersey petitioned the Supreme Court on Kalshi and Polymarket sports contracts
Masters of Scale
14:35 PT
Generalist or specialist? CEOs can't agree
So what
Rewrite your next senior job spec around learning velocity plus two provable spikes, not a single deep vertical or vague range.
The generalist-versus-specialist question gets relitigated for the AI era, and Shopify's COO reframes it as speed rather than range.
- Jess Hertz replaces Shopify's T-shaped ideal with an "X shape": multiple spikes of expertise, faster absorption of complexity, able to be a seven-out-of-ten designer on demand
- Reid Hoffman's 2017 scaling orthodoxy was the opposite: swap generalists for experienced specialists as you grow
- Clara Shih argues AI collapses hundreds of subspecialties back into building, selling and running; Upwork's Hayden Brown names the resulting role the "AI orchestrator"
- MasterClass CEO David Rogier: AI amplifies top-1% specialists and substitutes for average ones — a generalist with Claude matches a merely okay copywriter
- Box's Aaron Levie pushes back on "don't learn engineering": the best software engineers are getting the most output from AI coding tools
The Information
12:01 PT
Startups Gravitate to a New, Pentagon-backed Battery to Bypass China’s Chokehold
So what
A DoD-blessed standard cell with 2M+ units of committed capacity and no contracts is the moment to look at U.S. pouch-cell suppliers and robot power systems.
Steve LeVine reports that SAE has issued the first industry standard for drone batteries — a flat pouch cell — built explicitly to route around China's chokehold after Beijing blacklisted Skydio.
- China blocked battery sales to Skydio in October 2024 after the U.S. approved attack drone sales to Taiwan, and Skydio had no alternative supplier
- The standard replaces the Chinese cylindrical 2170 cell, sized to match six 2170s as typically grouped in drones; DoD backed the effort and looks likely to adopt it for the Drone Dominance Program, potentially millions of drones a year
- Changing cell dimensions can force $30M in new equipment, so fixing dimensions collapses the capex risk of starting a factory
- EnPower is retooling its Indiana plant for over 2 million cells annually from mid-next-year, with no firm contracts yet; Amprius ships a 500 Wh/kg standard cell in Q4, nearly double conventional lithium-ion
- Roughly 90% of power cells are made in China, per SAE committee vice chair Lisa King; EnPower's CEO is pitching the form factor for industrial and humanoid robots too
The Information
11:06 PT
Why Regulators Are Scrutinizing Prediction Markets' Trading Arms
So what
Kalshi's $40B round prices in federal preemption that its own trading arm weakens; the CFTC comment deadline is your diligence clock.
The CFTC is proposing rules to wall off prediction-market exchanges from their affiliated trading arms, drawing the analogy nobody in the industry wants: FTX and Alameda.
- Proposal would require separate personnel, office space and systems, bar affiliated firms from directional bets, and force two-sided quoting; comment period ends early October
- CFTC says it knows of at least six exchanges permitting affiliated trading; independent orders would have to fill ahead of affiliates at the same price
- Kalshi, near a raise at $40B, owns Kalshi Trading and says it is unprofitable and under 5% of volume on the most liquid contracts; Polymarket says it has no in-house market maker
- Robinhood holds 45% of Rothera, a joint exchange with Susquehanna, which makes markets on it, and routes some customer orders there
- The affiliated-trading structure undercuts the industry's core federal-preemption argument that they never trade against customers — live as New Jersey petitions the Supreme Court
The Information
10:44 PT
The Takeaway: How The Information Got a Decade-Long Edge on China Tech
So what
The coverage gap she exploited in 2016 still exists — your China AI information edge depends on very few reporters, so price that fragility.
Jessica Lessin's founder note on why The Information's decade-old Asia bet paid off: she flew to Beijing in 2016, found more reporters covering Uber in San Francisco than covering all of Chinese tech, and invested into the gap.
- Scoops that followed: Uber's China exit, ByteDance's revenue and product plans before it was well known, Tim Cook's secret deals with the Chinese government
- Coverage arc ran from China's internet crackdown through Shein and Temu, decoupling, and now the AI cold war, DeepSeek and this week's Chinese chip strides
- Rivals have pulled back from Hong Kong and Beijing under press pressure; The Information just added Claudia Chong in Singapore, its first
- Now an 80-person global newsroom; her stated lesson is to play the long game and physically leave the office
Every
10:34 PT
Compound Writing: How I Turned My Process Into a Plugin
So what
Packaging your team's tacit process as installable skills is a distribution move — the plugin, not the prose, is the reusable asset.
Katie Parrott has packaged six months of writing method into a distributable Claude plugin, tracing the arc from long chats to Projects to Skills to installable bundles.
- Compound Writing bundles skills for interviewing, outlining, drafting and reviewing; you supply voice, sources and standards
- The progression she names: pre-2024 chats buried discoveries mid-conversation, Projects (summer 2024) held style guides and samples across chats, Skills preserved per-job method, plugins made it shareable
- Inspired by Kieran Klaassen's Compound Engineering and Dan Shipper's agent-native architectures
- The compounding mechanism is explicit: an aversion becomes a guardrail, a sentence preference becomes an instruction the agent reads next time
- Part of a How We Write Now series, with Every requiring a human author to stand behind every piece
GAI Insights (Paul Baier)
08:29 PT
4 Differences Between Claude Chat and Claude Cowork
So what
If your team standardized on Claude Chat, you are missing scheduling and file write-back — the two features that turn assistants into unattended workers.
Anthropic merged Claude Chat and Cowork into one interface, and Paul Baier enumerates the four capabilities that only exist on the Cowork side.
- Cowork-only: scheduled tasks, live artifacts that refresh when the source updates, write-back to files, and multi-step automation chained as a sequence
- Cowork ships in browser, download and mobile versions; only the desktop download reads, moves and deletes local files, and can write back to SharePoint or Box if IT enables connectors
- Chat handles one exchange at a time and its artifacts do not refresh
- He pushes Projects as "mini digital employees" with project-specific chats, prompts, files and memory — his own cover a government contract, competitive analysis, CEO advice and personal health
- GAI World 2026 runs Boston, Sep 28-30, $2,195, with 12 hours of hands-on Claude training
AI Market Fit
08:22 PT
What a 2x Unicorn Founder Told Me About Raising and Growing + The Enterprise AI Playbook
So what
Botsitting at 6.4 hours a week is the metric to demand from any agent vendor pitching you productivity — measure supervision cost, not output.
Glean's Arvind Jain, now past $300M revenue seven years in, gives a hallway-interview playbook and, more usefully, reframes the enterprise AI ROI problem as Glean's product roadmap.
- First 10 to 20 hires were all engineers and R&D; Jain handled office, devices and early customers himself
- Fundraising view: raise fast, go big, dilution is cheaper than under-capitalization, and investors underwrite people not ideas — "a lot of capital waiting to be deployed"
- On Uber burning its 2026 AI budget in four months: value exists but is hard to land in a P&L, and vendors oversold instant savings; production-quality agents take many months of context work
- He names "botsitting" — Glean's Work AI Institute survey of 6,000 digital workers found 6.4 hours a week feeding and supervising agents, more than they spend producing with AI
- On Stripe–OpenRouter: OpenRouter was model access more than routing; Glean does model access but positions as "the context and intelligence company," claiming enterprise AI cost cuts near 6x
- GTM starts with the CIO, expands to CROs; ~70% US revenue; customers build their own agents; named logos include General Motors, McCarthy and Navita