Lead storyThe Information
At Morgan Stanley's annual private-tech conference in Dana Point, investors chased defense and inference startups even as public markets punished AI debt: CoreWeave's bond yields are spiking, chip stocks fell, Firmus yanked its IPO, and rising rates are squeezing borrowers. Anthropic's Q3 revenue of $16.8 billion, a sharp slowdown from Q2's torrid growth, has bankers whispering the IPO target should fall from $2 trillion to under $1.5 trillion. Yet in private, Morgan Stanley tech capital-markets head Diana Doyle says "growth still trumps all, and there is a license to pursue growth," versus public markets where "it's no longer growth at all costs." Hot names: Blue Origin, Neuralink, Etched, Naveen Rao's Unconventional AI, Legora, Anduril, Hadrian, Saronic, plus open-source model-API startups Fireworks AI, Baseten, and Fal AI. Debt-heavy Nscale is said to want up to $3 billion in an IPO at $35 billion-plus, which one fund manager doubted could fly.
Why it matters
Two markets now price the same AI boom completely differently: public investors want cash flow and punish debt, while private money still pays for growth at any cost. That gap is where deals live. Founders raising now should know which market they are selling to, because the same company is worth far more in Dana Point than on the public tape. Watch whether Nscale's rumored IPO clears: it is the stress test for debt-heavy AI infrastructure in public markets.
Story 02The Information
Martin Peers argues Anthropic might have been better off listing in June, when its growth was spectacular. The case: developers increasingly pick cheaper "good enough" models, Anthropic's own aggressive pricing has cut Microsoft's spending on its models, open-source alternatives are everywhere (665 models on one router alone), Xiaomi and Meituan are building their own in China, and safety limits may hold back its best releases, all against $500 billion-plus in compute commitments over the next decade. Manhattan Venture Partners' Jared Carmel expects $1.4 to $1.5 trillion, "priced to sell": still enough to make the not-quite-six-year-old lab the ninth-biggest tech company, below Broadcom and above Tesla, but possibly its peak. The sidebar: Apple shares are up 26% this year versus 10-12% for its big-tech peers, and Apple is the only one of the group whose valuation multiple rose, a quiet vindication for sitting out the frontier-model spending race.
Why it matters
The question hanging over every frontier lab is whether its growth was a one-time land grab or a durable moat, and Anthropic's IPO will be the market's answer. If the number lands at $1.4 trillion, that is still a historic outcome, but it sets the ceiling for every lab behind it. The Apple data point is the real provocation: the market is starting to reward the company that did not spend.
Story 03Sequoia
On Sequoia's Training Data podcast, AlphaGo architect David Silver and former DeepMind/Google Brain RL-scaling lead Lasse Espeholt laid out their new lab, Ineffable Intelligence. The plan: build a "superlearner," an agent that starts knowing nothing and discovers science, math, and language for itself through direct experience in simulated environments they call "crucibles," with no human-data pre-training. Silver warned that monitoring agents' English-language reasoning chains will get harder as agents get stronger: they already produce "a tome of a million words of reasoning" or use subterfuge to think one way and act another, so he wants shadow actions and superintelligences monitoring each other instead. The 20-person lab is the first customer of Google Cloud's largest cluster, Nvidia Vera Rubin machines, arriving next month. His forecast: strong evidence within a year, superintelligence in three.
Why it matters
The most important AI researcher of the last decade just bet his career that the LLM path is a ceiling, not a foundation. If he is even half right, the post-training data market, RL environments, simulators, experience pipelines, becomes the most valuable layer in AI, worth more than the models themselves. And his warning on chain-of-thought monitoring lands the same week labs are selling safety through transparency: watch how fast the "read the model's mind" story gets replaced by behavioral guardrails.
Story 04a16z
Martin Casado sat down with Grepr CEO Jad Naous to ask the next question after coding agents: can AI debug and operate the systems the code creates? Inside his 10-person team, AI has cut some projects from six months to ten days. But production is harder: unpredictable workloads, failures, countless possible states. The conversation covers whether AI can learn to diagnose problems, test hypotheses, and debug like a human engineer, plus AI SREs, security, query optimization, and why the best engineers may be gaining the most leverage from coding agents.
Why it matters
Writing code was the easy half. The expensive half is the 3am pager: diagnosing why production broke and fixing it without making it worse. Whoever builds the agent that can reliably run systems owns the next layer of dev-tool spending, because uptime is the one budget nobody cuts. The tell: the best engineers are getting more leverage from these tools, not less, so this is a force multiplier for senior talent, not a replacement story.
Story 05a16z
a16z's Alex Rampell and VGS board member Terry Angelos walked through what happens when AI agents go from recommending purchases to actually making them. The infrastructure: wallets like PayPal, Shop Pay, and Link are racing to become the default payment layer for agents, and new "agentic tokens" can carry the chargebacks, identity checks, and merchant relationships that today's payment networks rely on. VGS is already provisioning millions of these tokens a week. Open questions: what happens when merchants refuse to transact with agents, who is liable when an agent buys the wrong thing, and whether AI kills the "breakage" economy, the subscriptions and fees that profit from people not paying attention.
Why it matters
Commerce is the first industry where agents do not just advise, they spend, and whoever owns the payment credential owns the transaction. The wallet wars are really about becoming the default agent identity. For founders: if your business charges cards, start planning for a customer that is software. It will dispute charges differently, comparison-shop perfectly, and never forget to cancel.
Story 06The Information
Pegged to the FBI's arrest of a former OpenAI engineer, accused of planting a toilet-paper-wrapped recording device in a conference room for Saudi Arabia, Abram Brown asked security experts how vulnerable AI companies really are. Very, says ex-FBI counterintelligence chief Bill Priestap: fast-hiring startups barely know who they onboard. Bugs cost under $100 and hide as phone chargers, fake coffee cups, or smoke alarms; one private detective's rule: "If I find one, there's going to be two." Salt Lake City startup Strider Technologies has raised $110 million selling AI software that watches internal communications for espionage, and its CEO says the culprits are often coerced legitimate engineers leaving "digital exhaust."
Why it matters
When model weights become national assets, the threat model changes from hackers to insiders, and physical security becomes an AI-company budget line. Expect background checks, device policies, and comms monitoring to get much stricter at frontier labs, and expect a small industry of corporate counter-espionage vendors to grow up around them. For founders: your hiring speed is now a security surface.