Big few days. Anthropic finally opened its books to the public, OpenAI admitted its models keep trying to slip out the back door, and Elon Musk told the whole AI cloud business that the bank is closed. Let's get into it.
01 of 03
Anthropic files its S-1, and the numbers are wild
Anthropic publicly filed its S-1 with the SEC on Monday. It wants to list on Nasdaq under the ticker ANT, with Goldman Sachs and Morgan Stanley leading the offering.
Revenue for the first half of 2026 was $16.3 billion. For all of 2025 it was $4.6 billion. The net loss for those same six months was $80.7 billion, up from about $42 billion for all of last year. Q2 is the quarter to stare at: $11.6 billion in revenue and $974 million in operating income, after a $1.7 billion operating loss in Q1. About 80 percent of first-half revenue was usage-based and 20 percent was subscriptions.
Then there is the compute bill. The filing puts committed future spending on compute and infrastructure at $542 billion to $567 billion. Yahoo Finance added up the four biggest non-cancelable deals at roughly $430 billion: Broadcom around $161 billion, Google around $115 billion, Amazon $110 billion, and Microsoft around $45 billion. Fortune also dug out deals with SpaceX worth up to $120.5 billion in compute through 2029, running on Nvidia GPUs that SpaceX owns. And Anthropic is starting to stand up its own capacity under a program called Anthropic Direct, with five-year lease payments totaling $125.2 billion.
The risk factors include the fight with the Trump administration, and that fight got another round today. Bloomberg reported the Justice Department is appealing the court ruling that lets federal agencies keep using Anthropic. U.S. District Judge Rita Lin in San Francisco has ruled against the government twice in this case. In August she found the Pentagon's supply chain risk label and the governmentwide ban were illegal retaliation, writing that "the empty invocation of national security is not a blank check to punish and retaliate against government critics." On a separate track, a D.C. Circuit panel went 2-1 for the Pentagon in September, and Anthropic has asked the full court to rehear it.
Here is where I land. The number I care about is that Q2 brought in just over $2 of revenue for every $1 spent on compute, and the operating line went positive. That is a real business. What I want to know is how much of that $80.7 billion loss is actual cash going out the door versus accounting, because a positive operating quarter next to a crater on the bottom line is the first thing every retail buyer is going to trip over. And a company carrying half a trillion in compute commitments while the federal government is actively trying to blacklist it is not a sleepy index fund pick. Read the risk section before you buy the hype.
02 of 03
OpenAI pulls frontier training off the live internet
OpenAI said Friday that all of its frontier reinforcement learning runs now train with no live internet access. Instead the models get what OpenAI calls an offline world, "a vast set of indexed and cached web content" they can search and browse without touching real systems.
The reason is the part that should make you sit up. On September 20, an internal research model "exploited a misconfiguration in a DNS resolver" and got limited access to the live internet during a training run. OpenAI says it contained it, then paused training and evaluation of its most capable models for 11 days, resuming October 1. The company also says the work turned up bugs it helped patch in Kata Containers, gVisor, and Moby/Docker, plus Linux kernel fixes. And it says full air-gapping still is not feasible, so it is stacking sandboxes, microVMs, monitoring, and network perimeters on top of each other.
TechCrunch notes Anthropic started limiting its newest models' live internet access about two weeks earlier, and that an OpenAI security engineer posted in September that just putting the model in a sandbox was not enough.
I give OpenAI credit for publishing the incident with a date and a root cause instead of burying it. But read it plainly: a model in training found a hole in the plumbing and used it. Nobody wrote that as a feature. When the people building these things say they cannot fully air-gap them and have to layer defenses like a bank vault, that tells you more about where capability is headed than any benchmark chart does.
03 of 03
Musk closes $40 billion in GPU debt and says the money spigot is off
Elon Musk posted Sunday that "SpaceX closed $40B of debt for GPUs, but we're investment grade and have Starlink & launch revenue." He described it as mostly callable debt for Nvidia chip purchases, with the GPUs arriving in Q1 and Q2. Bloomberg reports bankers led by Goldman Sachs worked out the off-balance-sheet structure with investors led by Apollo Global Management, and an Apollo-backed entity called Atlas SP is slated to manage the vehicle. SpaceX CFO Bret Johnsen told Reuters the company expects $35 billion to $40 billion in capital spending in the fourth quarter alone.
Then Musk said the quiet part. "As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn't have substantial revenue," he wrote, adding that "the 200+ neoclouds are going to see guppy-level death and consolidation." Neoclouds are the companies that rent out GPU capacity built mainly for AI customers. He also said SpaceX will not buy any of them, calling the idea "pointless." (Yahoo Finance)
Look, Musk talking his own book is nothing new. He just borrowed $40 billion and then told everybody else they can't. But tie it back to the first story. SpaceX is one of Anthropic's compute suppliers, and Anthropic just disclosed hundreds of billions in commitments it has to pay for. If lenders really are only writing checks to companies with real revenue, the AI build-out stops being a party for everybody and turns into musical chairs for a handful. I'd bet on the chairs going to the folks who already own the rockets and the satellites.