Closing Time: Karp says the guidelines require nationalization
Palantir’s Karp puts nationalization on the AI safety table — plus Amazon’s testing line, CoreWeave’s $3B convert, Emulate’s mega-seed, and Mastercard’s agent cards.
CNBC reports Palantir CEO Alex Karp told Squawk on the Street that AI needs enforced “reasonable guidelines,” with civil and criminal liability as the first line of defense — and argued that frontier labs may ultimately need nationalization to cap lawsuit risk over stolen IP and proprietary data.
Reuters reports Amazon entered the pacing debate, saying models should ship only when they are ready and safe after rigorous testing and strong safeguards, while stopping short of endorsing an industry-wide slowdown.
Reuters reports Nvidia-backed CoreWeave launched a proposed $3 billion convertible senior notes offering due 2033, with a $500 million greenshoe and an ATM program for up to 35 million shares to fund operations and capped-call hedges.
Bloomberg reports Emulate, a UK startup founded by former Google DeepMind world-model researchers, is in talks to raise about $700 million in a seed round at a roughly $3 billion pre-money valuation to simulate and predict the physical world.
The Wall Street Journal reports Mastercard is rolling out an agentic payment option with Alchemy this week so users can issue restricted virtual cards to AI shopping bots, with American Express also readying its own agent rails.
Today’s Biggest Moves
Historic market snapshot — September 17, 2026 (day % change as of that edition; not live)
Biggest winners
| Ticker | Name | Price | Day |
|---|---|---|---|
| ALAB | Astera Labs | 293.44 | +16.2% |
| SMCI | Super Micro Computer | 40.35 | +13.2% |
| INTC | Intel | 109.14 | +12.4% |
| CIFR | Cipher Mining | 16.93 | +12.2% |
| CRDO | Credo | 168.16 | +11.8% |
Biggest losers
| Ticker | Name | Price | Day |
|---|---|---|---|
| TTD | The Trade Desk | 14.23 | −5.10% |
| INTU | Intuit | 313.44 | −4.81% |
| CRM | Salesforce | 243.43 | −4.78% |
| IBM | IBM | 238.10 | −4.14% |
| PATH | UiPath | 13.79 | −3.13% |
Karp puts nationalization on the safety table
CNBC reports that Palantir co-founder and CEO Alex Karp told Squawk on the Street that the industry has to set “reasonable guidelines” that are actually enforced, and that the first line of defense is liability for your own actions. He pushed for civil and criminal exposure for builders of technology that can upend the world, and he argued the philosophical frame at some frontier labs — that a small group deserves all the IP — would require severe liability protection that only the U.S. government can provide. “These businesses have to be nationalized because if you don’t nationalize them, every single one of my clients is going to sue,” he said of AI labs. The comments land in the middle of the week’s pacing fight, with Amodei, Altman, and Musk on one side of a slowdown argument and Trump calling AI risk a “hoax.”
That matters because Karp is reframing the safety debate as a lawsuit and IP problem, not only an alignment seminar. If enterprise buyers believe model training vacuums up their alpha, the pressure shifts from voluntary pacing essays to indemnity, nationalization, and who pays when the data leaks into a competitor’s model.
The one move is watching whether Congress’s kill-switch and pre-release review talk starts citing liability caps the way Karp does — because that is where guidelines become a balance-sheet story. Read CNBC
Amazon wants testing without the slowdown slogan
Reuters reports that Amazon told the wire models should be released when they are ready and safe to use, which comes from rigorous testing and strong safeguards, and that the company does not see a choice between progress and safety. The Seattle firm stopped short of backing the industry slowdown Amodei’s essay set off, while saying risks rise if progress does not happen collectively with government. Amazon builds models for Alexa and AWS Bedrock and is also a major infrastructure landlord for rival labs.
That matters because Amazon had stayed quiet while OpenAI, Anthropic, DeepMind, and Microsoft lined up behind pacing language. A hyperscaler that both sells and hosts frontier systems is now drawing a line: test hard, ship when ready, skip the slowdown branding.
The one move is watching whether Amazon’s Bedrock and AGI unit policies change in public, because a procurement-facing test bar would move the market faster than another CEO essay. Read Reuters
CoreWeave borrows another $3 billion for the buildout
Reuters reports that Nvidia-backed CoreWeave plans to raise $3 billion through convertible senior notes due April 2033, with buyers able to take another $500 million, and launched an at-the-market equity program for up to 35 million Class A shares. Part of the note proceeds would fund capped calls against dilution; the rest would go to operations. The neocloud also said Q3 short-term capacity was pricing near $40 million per megawatt annualized and contracted power rose to about 4.2 GW.
That matters because the AI factory story is still a balance-sheet story. Convertible debt plus an ATM is how a public neocloud keeps buying GPUs when backlog grows faster than free cash.
The one move is tracking whether the convert clears and at what coupon, because that price is the market’s read on CoreWeave’s leverage into the next rack cycle. Read Reuters
DeepMind alumni shop a $700 million seed
Bloomberg reports that Emulate, founded by Jack Parker-Holder, Matthew McGill, and Philip Ball from DeepMind’s world-model team, is closing in on a $700 million seed at about a $3 billion pre-money valuation. The company builds systems meant to simulate and predict how the physical world behaves — another lab-alumni check written at frontier-lab scale.
That matters because seed rounds at unicorn marks are the new normal for world-model bets. If Emulate clears, it resets what “early” means for physical AI startups spun out of Google.
The one move is watching for a named lead and any compute offtake attached to the round, because a $700 million seed without a factory partner is still mostly a talent premium. Read Bloomberg
Mastercard hands the bots a spending card
The Wall Street Journal reports that Mastercard plans to roll out an agentic payment option this week through a partnership with Alchemy, letting users issue virtual cards to AI bots with caps on how much they can spend and on what. American Express is also preparing rails for shopping agents that act on behalf of cardholders, as Visa has already pushed into the same lane.
That matters because agent commerce only works if the payment network can constrain a bot the way it constrains a teenager with a debit card. Virtual cards with merchant and amount limits are the control plane the networks can sell.
The one move is testing whether Alchemy-style spend limits survive a real checkout path end to end, because a demo card that cannot clear fraud rules is still a press release. Read WSJ
Aigentic Exclusive
Seven AI Doomsday Scenarios and the Case Against Each
Warnings about runaway AI are making headlines again. Seven recurring scenarios explain what people fear, what evidence supports those concerns, and where the arguments become less certain. The Deep Dive separates the claims that track real capability trends from the ones that stretch past the evidence.
Watch
CNBC — Karp on reasonable guidelines
Alex Karp joins Squawk on the Street to argue for enforceable AI guidelines, liability for builders, and why he thinks frontier labs may chase nationalization to cap lawsuit risk over IP and proprietary data.
The one move is watching whether his nationalization frame shows up in the next Hill hearing the same week as the kill-switch bills.
Nate B Jones — agents start to buy
Nate B Jones walks through why AI agents are starting to initiate purchases and how Stripe is building the payment layer those agents will need — spend controls, identity, and rails that treat a bot like a constrained buyer.
The one move is mapping one checkout flow in your stack that an agent could run with a virtual card limit this quarter.
Dwarkesh Patel — OpenAI researcher on agent swarms and RSI
Dwarkesh sits with an OpenAI researcher on agent swarms and recursive self-improvement — what improves when you scale agents versus models, and where the research bets actually are.
The one move is separating swarm orchestration claims from raw model capability when you read the next RSI paper.
Y Combinator — startups move from bits to atoms
YC walks through why more founders are leaving pure software for hardware, robotics, and physical-world AI — and what changes when the product has to ship in atoms, not only tokens.
The one move is asking which part of your roadmap still assumes software margins on a physical bottleneck.