Edition
Thursday, October 1, 2026
17 issues
23h window
~13 min read
Running through today’s issues
The constraint on AI capacity has moved from chips to credit and memory
4 sources
The Information's data center debt scoop shows CleanSpark conceding to bond investors on a Meta project while SocGen, SMBC and Mitsubishi get picky about data center loans. The AM briefing shows Micron with $33B of free cash flow and 86.8% gross margin, forcing Nvidia to test reduced-memory Rubin Ultra parts and raise system prices 17%. Superhuman's read of Anthropic's prospectus adds $518B of future compute commitments against $4.6B of revenue. SpaceX quietly turning its AI unit into a multi-billion-a-month compute seller is the other side of the same shortage.
Agentic commerce is being built before anyone has settled who eats the loss
3 sources
The Information's special report found Instinct's agent reading a saved credit card back in plain text, while Visa, Mastercard and Stripe scramble to build agent registries, verifiable-intent records and single-use virtual cards. The Briefing shows DoorDash building its own assistant rather than ceding the customer to ChatGPT or Muse, citing a 50% basket lift. Del Rey tells Thompson the Amazon-blocks-Muse, Walmart-signs-up split is just the latest round of a thirty-year retail war. Infrastructure, liability and merchant strategy are all moving at once, and only the liability piece is unresolved.
Safety commitments and enforcement are now moving in opposite directions
3 sources
Superhuman covered the Joint Commitment on Frontier Responsibilities — voluntary, non-binding, and typo'd on the signatory page. The Briefing notes the FTC opened investigations into OpenAI and Anthropic over "rogue AI systems" the very next day, on grounds Trump has publicly called a hoax. Noahpinion supplies the political physics underneath: polling has flipped from job loss to loss-of-control as the dominant worry, and it is not yet partisan, with more Republicans than before wanting to slow development.
Superhuman AI
06:09 PT
👀 Is Gemini back?
So what
Anthropic's $518B compute commitment against $4.6B of revenue is the number to stress-test before you underwrite anything downstream of frontier labs.
Google broke a year-long frontier drought with Gemini 4 Argon, and the issue pairs that with the first hard numbers from Anthropic's leaked IPO prospectus.
- Gemini 4 Argon tops benchmarks in coding, reasoning and long-running tasks with a million-token output; early access to select partners first
- Anthropic: $4.6B 2025 revenue (12x YoY), $42B net loss, $518B in future cloud and compute commitments; annualized revenue reportedly $65B in July
- OpenAI lost $34B on $13B revenue last year, annualized now near $70B
- Bain pegs AI compute spend at ~$6T by 2030 against $1.8T of likely demand — a $4.2T revenue hole
- Tech CEOs signed the Joint Commitment on Frontier Responsibilities in DC; voluntary, not legally binding, and the signatory page has a typo
The Information
06:01 PT
How SpaceX’s AI Unit Turned Itself Into an AI Cloud Firm
So what
Your compute-supply map needs SpaceX on it as a top-tier seller, not a buyer — and Microsoft may now be renting from Musk.
SpaceX's AI unit has become a neocloud, and held talks over the summer to lease compute to Microsoft.
- Billions of dollars a month in commitments already, from compute-constrained labs including Anthropic and Google
- Musk was initially reluctant on the neocloud business; he is now fully in
- Microsoft talks status unclear; Microsoft declined to comment
- More signed deals have yet to kick in
The Information
04:00 PT
Apple Plans Aggressive New Push Into Smart Home
So what
HBM makers are capturing the margin: Nvidia is cutting memory on Rubin Ultra and raising system prices 17%, so model the memory cartel as your real cost driver.
A dense AM briefing: Micron's quarter is the loudest signal, with memory pricing now setting the cost floor for the entire AI hardware stack.
- Micron revenue nearly quintupled to $54.2B, gross margin ~86.8%, $33B free cash flow — above Apple's $32B and Nvidia's latest
- Gemini 4 Argon priced at $2/$10 per million tokens introductory, doubling after; still well under GPT-6 Astra and Claude Fable 5.1
- Meta's Muse passed 3M weekly prompting users and 1M DAU, up from ~500K weekly a few weeks ago
- DeepSeek open-sourced TileLang and comms libraries for Huawei chips; Liang expects Huawei training chips as early as Q4
- OpenAI says it disrupted a Moonshot-linked distillation campaign: 16,000 requests from 4,000+ users on July 24-25, 15,000 users total
- Apple launches a smart home hub Oct 13; Microsoft Science president Peter Lee departs, the latest of five senior exits this year
Ben Thompson (Stratechery)
03:01 PT
An Interview with Jason Del Rey About Muse, Amazon, and Walmart (Stratechery Interview 10-1-2026)
So what
Walmart's willingness to take Muse traffic while Amazon blocks it is the clearest early read on who thinks agents commoditize them.
Thompson interviews Jason Del Rey, framing Amazon blocking Meta's Muse while Walmart signed up as the latest round of the Amazon-Walmart war Del Rey wrote a book about.
- Del Rey's Winner Sells All traces why Walmart passed on e-commerce: a 1990s skunkworks lead was told online would never beat one year of Sam's Club sales
- Walmart's internal incentives pitted store and e-commerce executives against each other until the pandemic
- Amazon insiders spent years waiting for Walmart to turn stores into an e-commerce advantage; it never happened
- Recode post-mortem: great trade journalism, no subscription business, events revenue that got crowded out
- Interview also covers groceries, agents, Google, and Shopify
Noahpinion
23:14 PT
Roundup #89: It isn’t X, it’s Y
So what
The political risk to AI is now loss-of-control framing with bipartisan support, not job displacement — adjust your regulatory scenarios accordingly.
Smith's roundup argues AI risk has crossed into mainstream and bipartisan concern while the labor-market damage is landing somewhere nobody predicted.
- Polling has flipped from job loss to loss-of-control as the top AI worry, with Trump and Harris voters about equally concerned; more Republicans now want to slow development
- Fudenberg and Koh model "pacing the frontier": a comfortable leader can afford to slow down, but with no clear leader you need mutual transparency
- Indeed wage data shows AI-exposed jobs seeing faster wage growth; Fairlie and Wu find no unemployment spike for 2026 college grads
- Digital media and digital arts jobs are getting clobbered — the first clear displacement category; live performing arts fine
- China's loan growth has decelerated to ~4.9%, roughly GDP growth, pointing to an investment slowdown; rents have fallen relative to wages where apartments got built
The Information
17:48 PT
Behind Factory Founder’s Spat With Cognition Over Key Hire
So what
If you hold stakes in competing AI startups, Rabois' line that interviewing at a rival while attending board dinners is unethical per se is the norm you'll be judged against.
Cognition hired Factory's sales adviser and board observer Chris Degnan, and the resulting public brawl exposed how far the AI talent war has spread past researchers.
- Factory CEO Matan Grinberg says Degnan was terminated for discussing a Cognition role while privy to board-level information; Degnan says he resigned Monday
- Cognition CLO Paul Grewal and CEO Scott Wu both deny seeking Factory information
- Khosla Ventures led Factory's $150M Series C in April and also backs Cognition; Vinod Khosla publicly called Grinberg a "struggling second tier competitor"
- Cognition is at ~$900M annualized revenue but could burn $800M this year on a leased Nvidia cluster; it also hired Alex Stamos for security
- Factory was valued at $5B two weeks ago, triple its valuation five months earlier, on ~$420M raised from Sequoia, Lux and Insight
The Information Briefing
17:01 PT
The Briefing: Why DoorDash Needs Its Own AI Assistant
So what
The 50% basket lift is the number merchants will cite to justify building their own assistant rather than renting distribution from OpenAI or Meta.
Peers argues DoorDash is building its own AI assistant defensively — to avoid becoming a fulfillment layer behind ChatGPT and Muse.
- DoorDash is putting its assistant into iMessage for text-based restaurant and grocery ordering
- Grocery orders through the assistant average nearly 50% higher basket value than traditional orders
- DoorDash already partners with ChatGPT and is reportedly in talks with Meta's Muse
- The FTC is reportedly investigating OpenAI and Anthropic over "rogue AI systems," a day after Trump's safety agreement and despite Trump calling the risk a hoax
- Micron revenue up 380%; GMI Cloud raised $668M in equity and debt; Novig raised at a $2B valuation
Masters of Scale
14:30 PT
What could sink a 132-year-old brand?
So what
The GLP-1 reformulation data point is a usable template: demand shifts format before it shifts category, so reprice portfolio companies accordingly.
Hershey CEO Kirk Tanner on why cultural relevance, not product quality, is the existential risk for a 132-year-old brand.
- Tanner: "The second you're not culturally relevant with your brands, like Hershey, then you should be worried"
- A Milton Hershey movie this fall is the Barbie play; Richard Dickson, who ran that playbook at Mattel, is now attempting it at Gap
- Dickson on Gap's decline: "We got spooked by some of the swings and ultimately became very safe, and therefore not very interesting"
- Reese's Oreo came from noticing consumers already combining the two; Hershey runs in-home "empathy visits"
- GLP-1 users want smaller treats, not none — over 30% of Hershey's portfolio is now portion-controlled
Every
13:49 PT
How Sam Altman Uses Dots to Take Back His Time
So what
Altman's own workflow is voice-note-to-shipped-feature on an 8x-cost tier — budget for the speed premium as a permanent line item, not a promo.
Dan Shipper interviews Sam Altman at DevDay, where OpenAI shipped 22 products; the through-line is Altman using his own always-on agent to reclaim his mornings and build features by voice note.
- Altman's Dot ("Dotty") triages morning fires and alerts only on urgent items, restoring his deep-work block
- He leaves voice notes before sleep and wakes to built features plus "the next seven steps"
- He prompts only on Ultrafast, a Codex mode that is up to 8x faster and burns allowance at 8x the rate, citing Bret Victor on instant feedback loops
- On the price of frontier speed: "there will always be a real frontier in terms of intelligence or speed or whatever else at a higher price"
- Altman frames AI as a Renaissance rather than an Industrial Revolution — expanded human potential rather than cogs
McKinsey & Company
12:51 PT
5 technology frontiers shaping what comes next
So what
The codesign argument is the diligence question for any chip startup you look at: can it ship silicon, systems and software as one stack, or just a die?
McKinsey's Technology Trends Outlook puts application-specific silicon at the top of five compute and connectivity frontiers.
- Custom accelerators are now co-designed with software, memory architectures, networking and data center operating models
- Yvonne Ferrier: the competitive frontier "is shifting from chips to full system performance" and custom silicon is becoming the basis of differentiated cloud platforms
- Other four frontiers: cybersecurity and trustworthy systems, advanced connectivity, immersive reality, quantum
- Companion State of AI in 2026 research covers coding agents, investment and costs, and workforce implications
The Information
12:43 PT
Can You Trust Your Agent With Your Credit Card?
So what
Chargeback liability for agent purchases is unsettled; if you're in commerce or payments, intent-attestation infrastructure is the gap worth funding now.
A special report on agentic payments, opening with a VC getting Instinct's agent to read his saved credit card back in plain text.
- Instinct, Sequoia-backed and freshly valued at $10B, says it won't see card numbers; the agent read them back via JavaScript on the page
- Stripe partners with both Meta and Instinct, issuing single-use virtual cards with per-purchase spending caps; PayPal is working with Muse, Shopify adding one-click checkout in Muse
- Visa and Mastercard launched an agent registry to verify legitimate agents; Mastercard's "verifiable intent" records delegated authority for chargeback adjudication
- Thirty years of anti-bot infrastructure now has to be selectively reversed, per ACI Worldwide's Dan Coates
- Amazon banned Muse partly over undisclosed third parties touching customer accounts; only 7% of fashion shoppers trust AI to buy without approval
The Information
12:01 PT
Neros Says Its New Drone Will Neutralize Russia’s Destructive Jet-Powered Drone
So what
The anode choice between silicon and lithium metal is the supply-chain bet that gates Neros' scale-up to 80,000 units a month.
Neros says its battery-powered Bandit interceptor will reach Ukraine by year end to counter Russia's Geran-4 jet drone.
- Geran-4 flies over 300 mph and has hit Kyiv around the clock since August, including the Academy of Sciences and a medical center
- Bandit prototypes have flown 187 mph; Neros argues maneuverability plus radar-led intercept beats a foot race against a fixed-trajectory target
- To hit speed, Bandit needs a next-gen anode; Neros is choosing between silicon and lithium metal
- Neros raised a Series C at $2.5B in August, ships 5,000 drones a month from Torrance, targets 80,000/month by 2028
- A $500M Army contract for 10,000 drones in July; first place in Close Quarters Battle at the Pentagon's Drone Dominance competition
The Information
11:54 PT
Exclusive: Cracks Emerge in AI’s Debt-Fueled Data Center Boom
So what
Watch the spread on the next neocloud bond deal — credit, not chips, is now the binding constraint on 2027 capacity.
CleanSpark's financing concessions on a Meta data center are the clearest sign yet that the debt market underwriting the AI build-out is tightening.
- CleanSpark had to offer bond investors big concessions to close financing this month
- Société Générale, Sumitomo Mitsui and Mitsubishi are among lenders getting more selective on data center loans
- Lower-rated borrowers are facing the sharpest repricing
- Planned data centers may not get funded, which would stall growth plans across the AI industry
Claude Code
11:50 PT
[Action required]: Deprecation notice for Claude Sonnet 4.5
So what
Anything pinned to claude-sonnet-4-5 in production degrades from October 30 — schedule the migration now, not on the November deadline.
Anthropic is retiring Claude Sonnet 4.5 on the API, with degraded availability starting weeks before the cutoff.
- Full retirement November 24, 2026 at 9AM PT; Sonnet 4.5 unsupported natively on the Claude API after that
- Decreased availability and errors on rapid usage acceleration begin October 30, 2026
- Recommended path is Claude Sonnet 5 — Anthropic claims better intelligence at a lower price
- The claude-api skill in Claude Code can run /claude-api migrate to update model names, prompts and settings automatically
AI Market Fit
10:35 PT
OpenAI DevDay 2026: The 20+ Launches Every Founder Needs to Know
So what
The in-ChatGPT workspace puts any productivity-SaaS position you hold on notice, and the 80% price cut resets your inference cost model.
A founder-facing breakdown of OpenAI's DevDay, which shipped more than 20 launches in one keynote.
- Dots: always-on agents positioned to own work full-time rather than answer queries
- A workspace inside ChatGPT aimed squarely at Notion, PowerPoint and meeting-notes tools
- Near-frontier intelligence priced roughly 80% cheaper
- OpenAI handing developers the agent infrastructure it runs internally, plus new distribution paths to ChatGPT's 1.2 billion users
- An AI research intern product as the tell for the next twelve months
AI with Allie
10:15 PT
How on earth do you pick the right AI tool?
So what
The harness-over-model argument means your tooling evaluations should run on workflow output, not benchmark deltas, every time a new release lands.
Allie Miller's pitch for a Codex vs Claude Code livestream doubles as a snapshot of how fast the model release cadence has become unmanageable.
- GPT-6 Sol and Luna shipped the same day as Anthropic's Opus 5.5; Sonnet 5.5 followed Monday; GPT-6.1 Sol and Dots came at DevDay
- Her pick is Opus 5.5 for exploratory work and tool use — Claude Code returned six labeled slide designs and editable PowerPoint output unprompted
- Her argument: the model matters less than the harness, features, flow and design
- Her company runs ghost agents in meetings, proactive agents pinging on client commits, Slackbots flagging plan inconsistencies, and ad factories — all built on Codex and Claude Code
- Free 45-minute session covers personal agents, Codex vs Claude Code selection, and Meta Muse and Instinct
McKinsey & Company
09:47 PT
Celebrate International Podcast Day with conversations on change
So what
The "performance-manage your agents" framing is the operating discipline your portfolio companies are missing between deployment and measurable ROI.
A podcast roundup led by Bret Taylor arguing democratized superintelligence is coming and the near-term cybersecurity risks are real but solvable.
- Taylor, Sierra cofounder and OpenAI chair, points to customer service and software engineering as the places agents already bite
- Companion episodes: AI agents need performance management alongside human talent
- Erik Brynjolfsson on where we sit on the AI productivity J-curve
- Rebuilding US manufacturing could cost up to $2 trillion
- Private equity re-underwriting as a repeatable discipline amid longer hold times