Trump confirms Washington wants equity in the big AI labs, Google rents $920 million a month of SpaceX compute — and the "biggest ChatGPT overhaul since launch" isn't really an IPO story: it's about closing the widening gap between the people running agents in loops and everyone still just chatting.
The AI advantage gap is widening — and the labs are redesigning their apps to close it.
When agents became viable — roughly November to January — the gap between power users and casual users shifted into overdrive: people running agents see compounding value while chat users see linear gains, and the vanguard has already moved past prompting agents to writing loops that prompt them. Read the planned ChatGPT overhaul through that lens, not just the IPO one. The labs' best tool for getting everyone else into agent workflows is the interface itself — and that's why the way we use AI is changing.
Per the Financial Times, OpenAI plans to transform the chatbot into a "super app" combining coding tools and AI agents, rolling out in the coming weeks as website and mobile changes that steer users toward coding, image generation, and partner apps. Executives increasingly view ChatGPT as a gateway to higher-value products.
— Jenny Shao, Leonis Capital partner, in the FT. Both labs are aiming for IPOs, "and investors care more about money than dreams." The FT points to the shutdown of Sora as evidence of the new business focus.
— Hedgy Markets, on the ChatGPT overhaul. Their numbers: 900 million people use ChatGPT weekly, 50 million pay, and OpenAI still loses $14 billion a year — while enterprise is already 40% of revenue with a target of 50% by December. A chatbot, they argue, is hard to put a multiple on.
It's worth pausing to note how frequently the investor class cannot imagine that anything a company does isn't primarily about impressing them. What's actually going on is the embodiment of a much bigger trend: the most valuable categories of AI use cases, we've discovered, are simply not about chat.
OpenAI CFO Sarah Friar: free users average about seven turns a day, the first paid tier doubles that, Plus is about 3x, and Pro is about 11x. But the power users aren't just using AI more — they're using it differently.
OpenAI views Codex as its most successful product — at least the kind of success it wants — and anyone on AI Twitter can attest to the vibe shift of the last few months. In Ben Holmes' poll of nearly 2,100 coding-agent users, 51.1% use the Codex app, with CLIs in the terminal next at 30.9%.
The inflection hit between November and January, when people figured out coding tools weren't just for software engineers but for any knowledge worker who could use code and bespoke applications to solve problems. Since then, agent users are seeing compounding value while regular chat users see only linear gains.
The gap between seat-based and usage-based pricing is the gap between the $3 billion run rate Anthropic had last year and the $47 billion it's on now. The dominant theme of the past few weeks — the shift from the token-subsidy era to the token-scarcity era — is the same story.
— Peter Steinberger, OpenClaw creator, now at OpenAI. Even among power users there's a gap between the power users and the power users — a sign of just how quickly user-experience patterns are evolving.
— Boris Cherny, Claude co-creator. Six months ago Cherny stopped writing code by hand; then he was running five to ten Claudes in parallel. Now it's leveled up to the next abstraction — and he thinks this transition plays out through the rest of the year.
From the Ralph Wiggum loop to Karpathy's auto-research in March, the idea isn't new — but now both OpenAI and Anthropic have embedded loops into Claude Code and Codex via the /goal primitive: less human intervention, longer runs, agents that fix their own mistakes and take on more complex tasks.
— Just Jake of Railway, on agent loops. Shanu Matthew's "non-technical idiot guy" plea for resources drew over 300,000 views, and Cloudflare CTO Dane Ketch got about 250 responses asking how people use loops. The knowledge is evolving fast and dense — "like pulling a neutron star out of a magic hat."
Learning materials are one answer — OpenAI just published a list of Codex use cases. But the other approach is to lead people to agent workflows by redesigning the apps they already use. That, not the S-1, is the key point of the ChatGPT overhaul — and why the way we use AI is changing.
The president confirmed reports that the government is looking to take an equity stake in major AI labs, saying he's spoken to all of them and may meet "all the big ones" at the White House this week. On Bernie Sanders' 50% equity tax, he added: "we have certain things that aren't that far apart."
— President Trump, confirming the equity-stake concept to reporters Friday. Trump framed the plan as AI dividends for the public: "the American people can benefit from the success of AI, and by doing that, they're going to like it better."
CNBC reports Sam Altman met with Bernie Sanders on Wednesday, with sources saying OpenAI is proposing to donate equity to the US government to seed a public wealth fund — possibly paying dividends or allocated to individuals through the new Trump accounts for children.
The cynical read — Daryl Bostonjo's "the groundwork is already being laid for a government bailout of OpenAI" — shouldn't be dismissed out of hand. But the even more cynical read is that the government already views the leading AI labs as too big to fail, stake or no stake.
— David Sacks, former AI czar, on the Sanders proposal. Sacks said he can see why the 50% stake resonates — even on the right — but warned of "central government AI, a system with even more totalistic power over information, decision-making, and human behavior" than the digital currencies conservatives fear.
Brad Gerstner took both sides: government-owned shares are socialism and a "terrible slippery slope" toward crony capitalism, but he's encouraging founders to donate shares for the direct benefit of citizens — held in a pooled trust or in their Trump accounts, not by future politicians.
When Bernie Sanders and Donald Trump are floating not-dissimilar proposals for the government's relationship with AI companies, the debate has left its old boundaries — and it's going to get even weirder before it resolves. The optimistic case: this is how you make the AI revolution something the whole country can support.
The three-year deal — October through June 2029 — grants access to at least 110,000 Nvidia GPUs, structured like last month's landmark Anthropic deal for the Colossus-1 supercluster. Google calls it "bridge capacity" for surging Gemini Enterprise demand, and can terminate in October if SpaceX fails to deliver.
— Jim Chanos, short seller, on the Google–SpaceX deal. The skeptic read: with the SpaceX IPO later this week — and Google's 6% stake potentially worth $100 billion at the target valuation — both parties have a strong incentive to puff up the company before the listing.
Yu Chen Jin's tally: 550k GPUs, more than double CoreWeave, making GPU rentals bigger than Starlink's $15 billion ARR. Boring Business does the bull math: xAI spent $40 billion on data centers and the Anthropic plus Google deals pay $26 billion a year — an 18-month payback from just two customers.
The multi-year deal keeps SK Hynix as Nvidia's largest memory supplier heading deeper into the shortage, adds Nvidia as a design partner on new memory chips for physical, personal, and infrastructure AI, and secures high-bandwidth memory for the Vera Rubin ramp. It was reportedly sealed over chicken and beer with SK Group's chairman.
— Jensen Huang, to press in Seoul. Jensen's face-to-face supply-chain diplomacy — soju in Seoul, beers in Taipei — is part marketing, but it's also very clearly a critical part of Nvidia's strategy in a shortage year.