OPENAI GOING PUBLIC: CONFIDENTIAL IPO FILING IMMINENT, STOCK MARKET DEBUT TARGETING SEPTEMBER
Let’s get straight to it: Sam Altman’s little nonprofit-that-could is about to become one of the most expensive stocks you can buy. OpenAI is preparing a confidential S-1 filing with the SEC right now, working with Goldman Sachs and Morgan Stanley, and the plan is to ring the actual bell on an actual stock exchange sometime around September 2026.
Think about that for a second. Seven years ago this was a research lab that published papers about AI safety and played games with GPUs. Now it is eyeing a valuation somewhere north of $300 billion based on its last private round. That is more than PayPal, more than Netflix, more than most banks that have been around for a hundred years. And they have not made a dollar of profit yet.
The company has had an interesting few months leading up to this. It just won the Elon Musk lawsuit, technically on a statute of limitations technicality, which feels very appropriate for a company that has been in legal chaos for years. It restructured from a nonprofit into a for-profit capped-benefit company, which required rewriting its entire founding charter and probably some tense phone calls. It is now offering equity-for-tokens deals to entire startup accelerator cohorts. And it is heading toward a public offering while simultaneously spending billions on compute, fighting with Apple over their partnership deal, and watching Anthropic breathe down its neck from behind.
The Goldman and Morgan Stanley pairing tells you everything. These are the banks you bring in when you want the full prestige play: the roadshow, the institutional allocations, the Wall Street Journal front page, the whole thing. OpenAI’s CFO Sarah Friar said earlier this year they would also carve out shares for retail investors, which means Altman genuinely wants this to be a people’s IPO moment. Whether he can pull that off while navigating the complexity of a company that is simultaneously trying to race toward AGI and justify a $300 billion price tag to pension funds is a different question entirely.
What makes this interesting beyond the obvious biggest tech IPO in years angle is the timing. OpenAI is going public right in the middle of an AI gold rush where everyone, Anthropic, xAI, Mistral, Meta, is gunning for the same enterprise budgets and developer mindshare. Going public gives OpenAI a currency. Suddenly they can do acquisitions in stock. They can offer employees liquidity without doing secondary rounds. They can, critically, make the token deal they just offered to every YC startup look a lot more attractive when those tokens have a listed market price.
The downside? Every quarter, Sam Altman will have to get on a call and explain to analysts why they are burning money at a rate that would make a Vegas casino blush. The OpenAI model of spend everything on compute, win the AI race, figure out revenue later does not play well with quarterly earnings calls. At some point a very serious person in a very boring suit is going to ask about margins, and “we are racing toward artificial general intelligence” is not an acceptable EBITDA explanation.
Still, this is happening. Possibly by September. Mark your calendars.
ANTHROPIC GIVES AI THE ABILITY TO DREAM: AGENTS NOW REVIEW THEIR OWN MISTAKES OVERNIGHT AND SHOW UP SMARTER IN THE MORNING
Here is the most interesting sentence you will read today: Anthropic has given its AI agents the ability to dream.
Not dream like a screensaver runs while your computer sleeps. Dream like: the agent finishes its workday, and then while you are getting dinner, it sits there reviewing everything it did wrong, figures out what worked, writes itself a playbook for next time, and shows up sharper in the morning. This is not science fiction. They demoed it live on stage at the Code with Claude developer conference in San Francisco earlier this month.
The feature is called, plainly enough, Dreaming. The way it works is this: between sessions, a scheduled background process reviews the agent’s entire past history, all its runs, all its mistakes, all the moments where it zigzagged when it should have gone straight, and it extracts reusable patterns. It then writes these as structured playbooks in plain text that future sessions can reference. The agent does not change its own model weights. It is not self-modifying in the scary Terminator sense. It is more like an intern who keeps a very good notebook.
The results are genuinely absurd. Harvey, the legal AI company that has been quietly becoming one of the most interesting AI businesses in America, saw task completion rates go up roughly six times after turning on Dreaming. Six times. That is not an improvement, that is a replacement of whatever you were doing before. Wisedocs, which does medical document review, cut its review time in half. Netflix is running multi-agent systems simultaneously processing logs from hundreds of software builds.
Anthropic’s Alex Albert described it as the difference between doing a workflow once and just living through it, versus recording the path from A to B so the next person, or in this case the next AI session, does not have to rediscover it from scratch. That is actually a profound idea. Right now, every AI session starts with complete amnesia. You explain who you are, what you are doing, what you tried last time, every single conversation. Dreaming is the beginning of AI that actually accumulates institutional knowledge the way a good employee does.
What is worth paying attention to here is the enterprise angle. The knock on AI agents for serious business work has always been reliability. You can have the smartest model in the world, but if it makes the same mistake twice, your legal team shuts the experiment down. Dreaming is Anthropic’s answer to that concern. And based on early numbers, it is a compelling answer.
The competitive framing matters too. Anthropic spent a lot of this conference talking about closing the gap between what AI can do and what it is actually doing for people. That is a direct shot at OpenAI, which has been winning the consumer market while Anthropic has been quietly winning enterprise. The message is clear: they are not building a better chatbot. They are building AI that learns your company’s way of working, accumulates your institutional knowledge, and gets better at your specific job over time.
That is a fundamentally different product pitch. And it might be the one that actually matters when the dust settles.
ALTMAN PLAYS SANTA WITH A CATCH: OPENAI OFFERS $2 MILLION IN FREE TOKENS TO ALL 169 Y COMBINATOR STARTUPS IN EXCHANGE FOR EQUITY
At a Y Combinator event last Tuesday, Sam Altman walked in, waited for the right moment, and then offered two million dollars worth of OpenAI tokens to every single startup in the current batch. All 169 of them. In exchange for equity. No cash changes hands. Just tokens for shares.
The room, reportedly, went quiet. Then it went loud.
This is what Silicon Valley people call a mic drop moment. It is what the rest of us would call a move so audacious it deserves its own Harvard Business School case study. Altman, who helped found Y Combinator and still shows up to mentor their batches, decided to just buy a stake in every company in the room. Not with cash, which he has plenty of, but with credits to use OpenAI’s AI models to build their products.
Think about the mechanics of this. One hundred and sixty-nine startups, each getting $2 million worth of OpenAI tokens. The deal converts to equity when the startup raises a proper priced round, typically a Series A. It is structured as an uncapped SAFE, which in plain English means the higher the startup’s valuation at conversion, the less of the company OpenAI receives. Founders’ lawyers will note this is actually a reasonably founder-friendly structure. Everyone else will note it is still equity for compute credits, which have been dropping in cost every six months as inference gets cheaper.
The genius of the play is that it works on two levels at once. Level one: OpenAI gets a tiny slice of 169 potentially high-growth companies for almost nothing in real cost. AI token costs keep falling as inference improves, so what costs $2 million to give away today might cost half that to actually deliver over the life of the agreement. Level two: every one of those 169 startups now has a strong incentive to build on OpenAI rather than Anthropic or anyone else. It is the classic platform lock-in play, executed before the startups are even big enough to be worth locking in.
Jason Calacanis, who runs his own competing accelerator and is therefore not exactly a neutral observer, immediately warned founders on X that OpenAI could study their ideas and copy them. This concern is not crazy. But it also ignores that Altman already has complete access to every YC batch as a frequent guest speaker. The argument of do not give them equity because they might steal your idea does not quite land when they can already see your idea for free.
The more honest risk for founders is simpler: what if you burn through $2 million in tokens faster than expected, you have given up equity, and you still do not have a working business? Tokens are infrastructure, not runway. But infrastructure is expensive, and for a twenty-something founder building an AI product, not having to pay that bill for a year or two is genuinely life-changing money.
Altman knows this. That is exactly why he made the offer.
ANTHROPIC’S COMPUTE CRISIS GOES TO MICROSOFT: AI FIRM PAYING SPACEX $1.25 BILLION A MONTH AND STILL RUNNING OUT OF SERVERS
Anthropic is paying SpaceX, yes Elon Musk’s SpaceX, one billion two hundred and fifty million dollars per month for computing power. That is $15 billion a year. That is enough to fund a mid-size country’s government. And even that is apparently not enough.
Because on top of the SpaceX deal, on top of the $100 billion-plus arrangement with Amazon Web Services, on top of the Google TPU arrangement, Anthropic is now reportedly in talks with Microsoft to use its custom Maia 200 AI chips. Microsoft, which also owns a massive chunk of OpenAI. Anthropic’s main competitor. Yes, really.
This is the compute crisis in its most raw, unfiltered form. Dario Amodei said at the Code with Claude conference earlier this month that Anthropic planned for ten times annual growth and got eighty times instead. That sounds like a great problem to have until you realize what eighty times means for your server bills. You cannot just call your cloud provider and say add more GPUs please. There are not enough GPUs. There are not enough data centers. There is not enough power on the electrical grid. So Anthropic has been throwing money at every compute provider they can find, Amazon, Google, SpaceX’s Colossus supercomputer, and now potentially Microsoft.
The Microsoft angle is particularly interesting from a competitive dynamics standpoint. Microsoft already owns a significant chunk of OpenAI. They are the ones keeping OpenAI running on Azure compute. The idea that they would also supply chips to Anthropic, OpenAI’s main challenger, would have seemed like a joke twelve months ago. Now it is CNBC reporting it as a confirmed negotiation in progress. That is how desperate the compute situation across the entire AI industry has become.
Microsoft’s Maia 200 chip offers over 30% better tokens per dollar than standard Nvidia GPUs, according to CEO Satya Nadella. Anthropic, which is burning through compute at a rate that would make a central bank nervous, would presumably find that 30% savings very attractive. So would their investors, who are presumably doing the math on how long this can continue before someone asks a difficult question at a board meeting.
The $1.25 billion monthly SpaceX number came out this week from a SpaceX IPO filing disclosure. That figure deserves a moment of silence. Anthropic, which has been operational for about three years, is spending $15 billion annually just on compute at one facility. The company raised roughly $12 billion in total venture funding through last year. It spent more than that on SpaceX servers alone in under a year.
The AI boom is real. The compute bill is realer. And somewhere, a very tired Anthropic CFO is on the phone with a fourth cloud provider, asking about availability.