OpenAI Misses Targets, Codex vs Claude, Elon vs Sam Trial, Big Hyperscaler Beats, Peptide Craze

The episode opens with the hosts playing a clip from a podcast called Miss Thing, which runs a recurring bit called "gay name straight name." In the clip, the comedians riff on the hosts' names, judging David as a "straight" name and joking about Chamath's sweaters, wine talk and gym selfies. The hosts react with laughter. One host says the comedians recorded it as a favor rather than for pay, and gives them a shout-out.

The formal introduction follows. The host calls All-In "the number one podcast in the world" and lists the panel as David Friedberg, Chamath Palihapitiya and David Sacks, with the host introducing himself as Jason Calacanis. (The transcript garbles several of these names.)

The first topic is a Wall Street Journal investigative report published on a Tuesday. According to the host's summary, OpenAI had expected ChatGPT to reach 1 billion weekly active users during 2025. It missed that goal and, four months into 2026, still had not reached it. The report also says OpenAI missed its 2025 revenue target for ChatGPT. The host notes that the exact figure was not specified and that industry accounting for revenue run rates has some unresolved nuance. The run-rate number in the transcript is garbled.

The host lays out why the misses matter. OpenAI reportedly has $600 billion in compute spending commitments, roughly what the company trades for on secondary markets. In the host's words, "the entire value of the OpenAI enterprise equals their spend commitments in the coming year." CFO Sarah Friar is reportedly worried that revenue is not growing fast enough to keep up with expenses. OpenAI wants to IPO later this year. The WSJ reports that Friar does not think the company is ready for public reporting standards, while Sam Altman wants to move faster. The two then released a joint statement. The host waves the statement away ("This is ridiculous. Yada yada yada.") and asks whether the misses are real headwinds or just expectation management by the leader of the race toward superintelligence.

Prediction markets have moved too. Polymarket puts the chance that OpenAI goes public by the end of 2026 at 32%, down from 60% in December. The host recalls an earlier warning on the show that IPO capital is limited. The SpaceX IPO is going first, and if Anthropic lists this year while OpenAI does not, that would create "an interesting dynamic."

One panelist offers what they call "a little bit of a contrarian take." They agree OpenAI had a bad week in the press. Beyond the WSJ story, which raised doubts about whether it can afford its data center commitments, the company was also in the middle of its lawsuit with Elon Musk. At the product level, though, the panelist argues the past week or two went well. ChatGPT 5.5 got strong reviews from developers and coders this panelist talks to in Silicon Valley.

At the same time, Opus 4.7, which is the latest Anthropic release, appears to be a bust.

The panelist reports user complaints that Opus 4.7 rations compute and cuts thinking time, that it is "not as good," and that Claude had some bugs. In many cases, they say, users are rolling back to 4.6. GPT 5.5, by contrast, is built on a new base model called SPUD. The panelist describes it as OpenAI's first base model upgrade in over a year, though they hedge on the timing ("I don't know"). A new base model, they argue, will make future improvements possible. They also say they are seeing posts on X from developers who are moving their coding work from Opus to GPT 5.5.

I think that SAM may end up being right, but for the wrong reason.

The panelist explains this "right for the wrong reason" thesis. Altman's large compute commitments were based on forecasts of a billion consumer users and the matching revenue. The consumer business turned out weaker than forecast. Meanwhile coding became the most important part of AI, and OpenAI now has more compute than Anthropic. The panelist describes Anthropic as token constrained, which limits its ability to serve Mythos and led it to engage in "compute gating" with Opus 4.7. They call that a prudent business decision by Dario Amodei and say explicitly that they are not criticizing it. Their conclusion is that OpenAI missed on consumer, but its enterprise business is "going gangbusters" and gives it room to catch up in coding. The host adds that Grok and Cursor teamed up the previous week.

Another panelist expects both OpenAI and Anthropic to be multitrillion-dollar companies. They argue the coverage misreads what is happening in AI, because one specific choke point limits everything: access to the power needed to generate tokens. In this view, any miss by OpenAI points to too little compute capacity today, and the problem is getting worse. As evidence, they cite Anthropic. Anthropic found a way to "economically induce" Amazon to give it enough capacity that customers no longer have to go through Bedrock to reach Anthropic models. It has also signed new deals with economic participation, including with Google, to gain more capacity.

Everything in this market is power constrained. The reason that these folks may miss a number or a forecast have nothing to do with demand.

The panelist says backlogs are now growing for physical components as well as for power access. These include not just reciprocating engines (rendered "RESIPs" in the transcript) and natural gas turbines, but also transformers and other grid infrastructure. They cite a gap between announced and actual construction: "less than half of it is actually being built." Most of the rest is stuck in red tape and supply chain delays. They also claim that 40% of announced projects will be canceled, because 40% of all projects over the last four years were canceled. They attribute this partly to the AI industry's failure to build a positive public "halo." The host adds that bad feelings about data centers and AI jobs are creating headwinds, and that some people are committing violent acts and blaming data centers and AI. The host chooses not to dwell on it.

Winners and losers under the power-constraint thesis, as the panelist lays them out: - Most hurt: Anthropic and OpenAI, which need outside capacity. - Beneficiaries: the hyperscalers, specifically Oracle, Amazon, Meta, Microsoft and Google. - The trade-off model labs face: how much equity and control to give up for compute, against how badly they miss their growth forecasts without it. - An opening for Grok and SpaceX, which the panelist says have "a ton of excess capacity." They call the Cursor deal "the appetizer" and suggest Musk could do a large deal with Anthropic, or perhaps not OpenAI given the lawsuit. They say he and Dario "should do a deal tomorrow."

The host first restates this as a compute constraint and is corrected: "No. The limiting resource is power." The host then gives the chain as power, which runs compute, which produces tokens, which serve developer and co-work demand from consumers and enterprises.