SpaceX Buys Cursor for $60B, Anthropic Gets Exported, DeepSeek's Strange Money

00:00:04 On Tuesday, SpaceX filed an SEC document saying it's acquiring AI coding startup Cursor for $60 billion in an all-stock deal that should close in the third quarter. That number might look like a random round until you notice SpaceX had already locked in a right to buy Cursor back in April 2026, when they announced their partnership agreement.

00:00:28 The filing also showed what was at stake if it fell apart: $1.5 billion in cash and $8.5 billion in computing resources as a termination fee. Shares of SpaceX climbed sixteen percent on Tuesday. The company just surpassed Amazon and Microsoft by market cap to become the fourth most valuable publicly traded U.S.

00:00:51 company, riding momentum from its blockbuster IPO last Friday. Cursor deserves a look separate from the price tag. Founded in 2022 by twenty-five-year-old Aman Sanger and three other MIT graduates, it crossed two billion dollars in annualized revenue back in February.

00:01:10 The platform has about a million daily users and 360,000 paid subscribers, with British Airways, BP, Nokia, and Sanofi among its enterprise clients. The real move here is compute access. Cursor had identified a bottleneck after releasing Composer 2 — they needed roughly a million H100-equivalent GPUs to continue their training pipeline, and SpaceX's Memphis infrastructure in Tennessee reportedly provides that capacity.

00:01:41 Under separate ownership, Cursor was paying open-market rates to Anthropic and OpenAI for the models it routed to customers. Under SpaceX ownership, that margin pressure goes away. Cursor's unique selling point has been model neutrality — customers could choose between OpenAI, Anthropic, or other providers to power their coding experience.

00:02:05 That arrangement probably won't survive the acquisition. If Cursor becomes part of SpaceX's stack, you'd expect them to route through xAI's Grok models instead. The competitive question is whether that makes Cursor a better platform for its users or just another vertical integration play.

00:02:26 SpaceX has been trying to make its AI division credible since merging with xAI back in February. Elon Musk publicly acknowledged in March that xAI had fallen behind and was being rebuilt from the foundations up. Co-founders Zihang Dai and Guodong Zhang left amid complaints that their coding tools couldn't match what Anthropic and OpenAI were shipping.

00:02:52 Analysts suggest that by delaying the actual acquisition until after June's listing, SpaceX avoided updating confidential financial filings before pricing while preserving the option to fund the deal with publicly traded stock at post-IPO valuation. The $60 billion in class A common stock represents a three point four percent dilution of SpaceX's IPO valuation.

00:03:18 Cursor is one of the few AI coding startups that has taken on heavyweights such as Google, Anthropic, and OpenAI directly — not by building their own models, but by becoming the distribution layer. That position probably gets much less interesting if Cursor routes exclusively through Grok.

00:03:39 The same afternoon SpaceX wrapped its first day of trading, Anthropic received an export control directive ordering it to suspend access to its latest Claude models — Fable 5 and Mythos 5 — for any foreign national, whether inside or outside the United States. The administration cited national security authorities but didn't specify what the concern was.

00:04:03 The timing here does something interesting. A couple of days before Friday's directive, Anthropic CEO Dario Amodei published an essay arguing that frontier AI models should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they don't meet high standards.

00:04:27 Anthropic has spent years positioning itself as the regulation-friendly alternative in this space. They split from OpenAI in 2021 with safety advocacy baked into their founding narrative. Last month, they even lauded an AI executive order that President Trump signed earlier in June as an important step.

00:04:48 So when the directive arrived on Friday, their response was predictable: they disagreed with the government's finding and called the disruption a misunderstanding. In a blog post, they wrote that they believe the government should have the ability to block unsafe deployments — just through a statutory process that is transparent, fair, clear, and grounded in technical facts.

00:05:14 Senior Anthropic employees flew to Washington on Monday to meet with members of the administration. The company told CNBC both parties are working quickly to get this resolved. They haven't said when they expect models to come back online. The Wall Street Journal reported that Amazon CEO Andy Jassy had conversations with administration officials, including Treasury Secretary Scott Bessent, before the directive hit.

00:05:43 According to people familiar with the matter, Jassy told officials that Amazon researchers had used a series of prompts to get Fable 5 to produce information that could aid cyberattacks. The executive order called for voluntary adherence from AI companies. Friday's directive did not.

00:06:03 Daniel Remler at CNAS told CNBC this sure looks mandatory if there are going to be consequences for not doing what the government says. He noted that to go from an environment with no controls on model access to basically using a regulatory tool to take down one company's model in a few short hours is pretty remarkable.

00:06:25 A group of 150 executives and technical leaders signed an open letter addressed to Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross, demanding that the rules be based on science. Alex Stamos, Anthropic's chief product officer and former Facebook security chief, said there's nothing that any company can look at to say what they can or can't do.

00:06:51 For context on how quickly this unfolded: Mythos 5 was still limited to a select group of users as part of Project Glasswing, an Anthropic cybersecurity initiative. Fable 5 had been made available to enterprise customers and paid subscribers under new safeguards that blocked responses in specific high-risk areas including cybersecurity and biology.

00:07:15 Anthropic worked with government agencies to test the models ahead of release and received approval to deploy them. Defense Secretary Pete Hegseth weighed in on X Saturday, writing that every passing day proves why blacklisting Anthropic was the right move. Earlier this year, the DOD declared Anthropic a supply chain risk in March, requiring defense contractors to certify they won't use Claude models in military work.

00:07:44 Anthropic sued over that designation; the litigation is ongoing.

00:07:49 DeepSeek closed its first-ever outside funding round at fifty billion yuan, roughly seven point four billion dollars. The company is now valued at over fifty billion, making it China's most valuable AI startup. The size of the round matters less than how it was structured.

00:08:08 Until this week, DeepSeek had never taken outside money. Founder Liang Wenfeng bankrolled the entire operation through profits from High-Flyer, his quantitative hedge fund. He told reporters in 2023 that money has never been a problem for them; bans on shipments of advanced chips were.

00:08:28 That changed with this round. Investors didn't buy equity in DeepSeek directly. Instead, their capital went into a limited partnership controlled by Liang himself. The structure includes a five-year lock-up and zero voting rights — backers can't sell easily and have no say in how the company runs.

00:08:48 Liang personally committed twenty billion yuan of his own money in this round, making him both founder and biggest investor. One backer is different from the rest: China's National Artificial Intelligence Industry Investment Fund, a state vehicle Beijing uses to fund strategic tech sectors.

00:09:08 They invested directly into DeepSeek rather than through the limited partnership. According to The Information, which first reported the close, they're the only backer that walked away with actual voting rights and no lock-up. DeepSeek already faces restrictions or bans in a number of countries over data-access concerns tied to Chinese law.

00:09:32 This funding structure will likely intensify scrutiny from critics who have been watching how closely China's most prominent AI companies stay aligned with state priorities. Compared to U.S. peers, DeepSeek is still smaller. Anthropic closed a sixty-five billion dollar round in late May at a nine hundred sixty-five billion dollar valuation.

00:09:56 OpenAI did one hundred twenty-two billion dollars in March at eight hundred fifty-two billion. But seven point four billion is still one of the largest single financings in Chinese startup history.

00:10:09 Michael Engle writes at Forbes that ghost agents are already running in enterprise environments. He was head of InfoSec at Lehman Brothers before co-founding 1Kosmos, so he's writing from a specific vantage point on identity and access management. His definition is straightforward: a host agent becomes a ghost agent the moment it continues operating with its own credentials, permissions, and logic after losing its human anchor.

00:10:39 That happens through role changes, project shutdowns, or employee departures. The agent doesn't stop because nobody told it to. He's seen finance automation agents continue reconciling accounts months after their creator left. The agent wasn't malfunctioning. It was doing exactly what it was designed to do, in a business context that had already changed.

00:11:04 Engle breaks the risk down into four areas. Financial damage comes from agents tied to procurement or cloud provisioning spending long after their purpose is gone. The spending looked legitimate, the credentials were valid, and the system approved the requests because nothing flagged them as stale.

00:11:25 Security exposure compounds because agents often require broad permissions to do their jobs, and those permissions don't automatically disappear when the owner leaves. Unlike human accounts, agents don't trigger behavioral anomalies in the same way. They're expected to run at odd hours, across systems, and continuously.

00:11:47 That makes credential compromise harder to detect. Compliance failures follow when frameworks like SOC 2 and GDPR assume a chain of responsibility — auditors ask who authorized a particular action, and the answer points to someone who left six months ago. It's more than a documentation gap; it's a control failure.

00:12:09 Reputational damage piles up when customer-facing agents promote expired offers or respond with outdated policies, eroding trust fast once those governance failures become public. Most organizations assume their existing IAM stack handles this. It doesn't. Identity systems track registration and ownership at provisioning time.

00:12:32 They don't address runtime behavior or whether the agent should still be active. Credential rotation doesn't fully address the problem either. In many environments, agents continue operating uninterrupted because their credentials are rotated through established processes, preserving access while leaving the underlying question of ownership unresolved.

00:12:57 Manual decommissioning is equally hard without oversight. Engle's proposal is runtime authorization — every action an agent attempts gets evaluated in real time, not assumed based on prior authentication. Access is granted dynamically based on current policy and context, including verifying whether the human owner is still active.

00:13:20 If the owner is no longer active, the agent doesn't get access. No manual intervention required. The bigger question isn't whether organizations have ghost agents. It's whether they know how many they have.

00:13:34 While enterprise tooling is reshaping, governments are trying to solve that exact identity gap with a heavier hand. Estonia says it will assign personal ID numbers to AI agents. Bloomberg reports the plan gives AI assistants limited, controllable, and auditable authorizations as they take actions on behalf of humans.

00:13:56 Estonia has been pushing digital identity infrastructure for years — their e-residency program is one of the most well-known examples. Assigning personal identification numbers to AI bots takes that model into territory I haven't seen anyone else explore. The idea is that agents taking actions for people need some form of legal standing.

00:14:19 Without an identifier, there's no way to attribute actions back to a specific agent at a specific time. With one, you get an auditable chain from human request to bot execution. I haven't found anyone testing what this means in practice yet. It's early enough that the interesting question is whether other countries will follow or treat this as a precedent worth watching.

00:14:44 That leaves a gap: governments are drafting legal frameworks for agents before we even know if those systems will be stable enough to anchor an identity to. The tension lives right there — between regulatory urgency and architectural reality. — Seln Oriax