SpaceX-Cursor Deal, SaaS Debt Bomb, New Apple CEO, SPLC Indictment, Colon Cancer Spike

Episode 270 of the All-In podcast opens with banter about the "SPLC files" and a joke about one host being "adjacent" to them. The hosts then explain that the recording was pushed back an hour because David Sacks was at the White House. In his account, he asked whether the president had time, the president made time, and they held a short meeting.

The speaker describing the meeting contrasts his own experience with media portrayals of a president who is "yelling at people" or "moody". He says the president has always been pleasant and genial with him, asks questions, and takes an interest in the subject matter. He adds one caveat: things might be different for someone who had "double crossed him". He says he has seen no evidence for the media's description.

The discussion moves to AI policy. Prompted to imagine an alternative history under a different administration, one participant claims there would be no data centers, AI would be used for censorship, and AI would promote DEI values, which he says appeared in the Biden executive order. He says the current president lacks "doomer neuroses" about AI. He also says this does not mean opposing all regulation.

we should have specific solutions for specific problems as opposed to being cowering in fear over this and just trying to halt all progress.

His main example is the president's position, taken more than a year ago, that AI companies should be allowed to build their own power generation "behind the meter". He presents this as better than what he calls the Bernie Sanders approach of "just shutting everything down." He also praises how the president gathers views from industry leaders and synthesizes them.

On Anthropic, the group recalls the president calling the company's founders brilliant and saying a deal was being worked on. The speaker who responds calls that assessment balanced. As he relays it, the president said Anthropic has a great product and is very left wing, but that this could be worked through and need not kill a deal. The president also said the company had tried to tell the Pentagon what to do, which the Pentagon did not like. The segment closes with the view that one of the president's core convictions is that he likes being around smart people.

The first main topic is a deal between SpaceX and Cursor, which the host introducing it calls the AI coding startup that defined its category. He says xAI and Cursor will build a new coding model meant to be "the world's best coding and knowledge work AI." As he explains it, SpaceX will either buy Cursor for 60,000,000,000, which is 10,000,000,000 more than Cursor was rumored to be raising at, or pay Cursor $10,000,000,000 for the collaboration. He cites Bloomberg's view that the $10,000,000,000 is essentially a breakup fee.

Figures cited on air: Cursor's run rate was 2,000,000,000 as of February. Cursor expects to end 2026 with a $6,000,000,000 run rate. SpaceX's projected 2026 revenue is 22 to 24,000,000,000. The SpaceX IPO is reportedly targeting a valuation of 2,000,000,000,000, described as roughly 80 times top-line revenue. Cursor's implied valuation would be 30x.

The same host gives some product background. Cursor was initially built on Anthropic's LLM and let users pick any model. In March it released Composer two, the second version of its proprietary model, which he places between GPT 5.4 and Opus 4.6 on a leaderboard shown on screen. His main argument is about infrastructure. Elon has 550,000 GPUs in Colossus and is scaling to 1,000,000, with plans to take compute to space. He says this matters a great deal for Cursor, which has been compute constrained.

So this is peanut butter and chocolate.

He predicts that within twelve months the combination will put SpaceX/xAI and Cursor at the front of the coding leaderboard. He presents this as his own prediction.

The next speaker treats the deal as effectively done. In his view, the deal was structured so that the S-1 does not go stale, which would otherwise force SpaceX to rewrite parts of the filing, its disclosures and its risk factors. He then works through the valuation. Suppose SpaceX is worth about a trillion today and, for the sake of argument, 2,000,000,000,000 later. A stock-for-stock deal worth 60,000,000,000 in "tomorrow dollars" would then amount to roughly a 50% discount, which he describes as getting Cursor effectively for 30,000,000,000.

He lists what SpaceX gets: a strong coding model and service in the area where near-term AI revenue is concentrated; hard-won usage patterns that are valuable for reinforcement learning; and what he calls a "very cracked team." He ties the deal to spare capacity at Colossus. Grok once offered a steep discount on output tokens, which drew a lot of experimental usage that later faded. In his account, this left low utilization that was redirected into buying what he calls "the most interesting and valuable third party wrapper service in AI right now."

Another participant agrees that the two companies complement each other. Cursor brings coding strength, training data, enterprise clients and coding experience. xAI brings compute and a foundation model. He describes Cursor's strategic problem this way: it began competing with generalists such as OpenAI and Anthropic, but those companies then moved into coding with Claude Code and Codex. Cursor depended on model providers that had become its competitors, "which was just not a good place to be."

The background includes a brief political point. One host argues that three years ago the Biden administration excluded Elon from an EV summit and that agencies were engaged in "lawfare" against him. A participant recalls a DOJ lawsuit accusing SpaceX of not hiring enough asylees, which the group says SpaceX could not do under ITAR. The participant then calls this "ancient history."

The group disagrees about whether Cursor will keep using third-party models. One host notes that Composer two reportedly uses Moonshot's Kimi K2.6 and doubts Elon would pay $60,000,000,000 without moving it onto Grok. Another pushes back: developers want a toggle to choose which model writes their code. Cursor's strength, he argues, is a well-built IDE whose user experience sits above Codex, Claude and the alternatives.

The same speaker makes a broader argument. Over the past roughly hundred and twenty days, people have realized how much of AI's value comes from writing software. Agents, he says, are basically quickly spun-up applications. Enterprises end up with too many of them, and they run inefficiently. He cites a story from that week about Amazon having "a million agents" running internally, with redundant data creation, redundant data stores and redundant API calls wasting money. His conclusion is that strong software engineering, delivered through an IDE like Cursor's, is still needed to centralize this work and fix it.

Another participant adds a cost argument. Enterprise token bills are rising month over month as employees build more software, and he is not sure anyone has yet had a reason to be efficient about it. He expects optimization to become a priority over the next several months, with middleware deciding which model to call and how much to spend.

it really only makes sense to go to a frontier model for a frontier task.

He also points to cyber as possibly "the white hot center" within coding. He says Mythos has shown that frontier models can be used for both cyber offense and cyber defense. He describes Mythos as roughly a 10,000,000,000,000-parameter model and notes reports that Anthropic lacks the compute to serve it. He adds that he is unsure it was ever meant to be a commercial model. He expects dedicated, cheaper cyber models comparable to Mythos and thinks this could be the hottest part of the market over the next three to six months, because IT departments and CISOs are worried about AI-powered hacking. Separately, one host predicts that xAI's products will be "dramatically" improved within six to twelve months.