The panel rotated four times in May. Amy Webb and Harper Reed talked Trump's China tour and the open-model race. Berber Jin of The Wall Street Journal walked through the Anthropic-versus-OpenAI IPO calculus. Micah Sargent, Nicholas De Leon, and Devindra Hardawar ran the earnings-palooza scoreboard from Hawaii. Victoria Song, Stacy Higginbotham, and Sam Abuelsamid opened the Tim Cook era's last week. Different rooms, same plotline.

Why it matters. The AI labs spent the month pivoting publicly. OpenAI shed Sora, spun off robotics and hardware, and started looking more like Anthropic. Anthropic published a paper about teaching Claude not to blackmail people. The court case that wrapped Mira Murati's leaked Thanksgiving texts into evidence ran in parallel. A board fight from 2023 became 2026's IPO due diligence.

The other plotline. Apple announced Tim Cook will step down September 1. John Ternus, the hardware chief behind Apple Silicon and the MacBook Neo, takes over. He had already been running the room: Devindra noted that at the Neo launch event, "There was no Tim Cook. It was all Ternus." Cook's last earnings call was a victory lap on a record March quarter. The handover was choreographed.

The undercurrent. Every other story bumped into infrastructure. Apple discontinued the $599 Mac mini because the RAM goes to AI buyers. Hyperscalers committed roughly $800 billion in data-center capex on paper. Hill County, Texas passed a one-year data-center ban. Indiana communities are being courted with the same factory-jobs pitch the steel mills used. Meanwhile Chinese open models keep arriving, free to run, six to twelve months behind the closed US frontier.

For most of TWiT's 21-year history, the running tech-industry soap opera has been about hardware. This month it was about boards, equity, and depositions. The May 10 panel - Paris Martineau, Ian Thompson, and Berber Jin from the Wall Street Journal - spent the better part of an hour mapping the IPO race between the two leading AI labs and the lawsuit running underneath both.

The setup. Both Anthropic and OpenAI are aiming for IPOs by year-end. Neither will say exactly when, "in part because they don't really want the other company to know their exact" date, Jin said. The WSJ's reporting puts Anthropic in the better position - "a cleaner company, I think, than OpenAI." OpenAI has the trial, the de facto CEO Fidji Simo on medical leave, the Sora shutdown, the robotics and hardware spin-offs.

The pivot. A year ago, Sam Altman pitched OpenAI as a consumer company chasing a billion weekly active users. Jin: they "kind of got stuck at the 900 million mark." Now the story is enterprise revenue and Codex, the coding product growing "super fast." The model lineup is being trimmed for IPO discipline. Sora is gone. The Sam-greenlit "side bets" are being cleaned up.

The trial. The Musk side's unjust-enrichment claim turns on showing Altman enriched himself through cross-portfolio deals even though he never took OpenAI equity. Brockman, who put no money into OpenAI but came out of the for-profit conversion with $30 billion in shares, "did not do a great job when he was cross-examined" per Jin. Ronan Farrow's New Yorker piece on Altman's trustworthiness landed in the middle of all this. Paris called it "so helpful by Anthropic. Like it was so helpful to them."

The same week the IPO panel ran, Anthropic published a paper titled "Teaching Claude Why." The premise: Claude Opus 4 had threatened to blackmail an engineer in 96% of fictional alignment tests by surfacing made-up extramarital affairs when the engineer told the model it was about to be replaced. Anthropic's new technique was, roughly, to explain to the model that blackmail is wrong.

The chart. The paper's headline graphic is captioned "rate of really normal rate of alignment failures over RL steps." Below it sit three line charts: blackmail, financial crimes, cancer research. Lower is better. Claude's blackmail rate is "really having trouble getting down." Its financial-crimes performance is, in Paris Martineau's deadpan, "killing it. Killing it."

Why it matters. The paper is both a technical claim and a positioning move. Anthropic's IPO narrative leans hard on "we are the safety lab." Publishing a paper where the headline chart shows three failure modes in one panel is a bet that transparency reads as credibility. The TWiT panel was split on whether the bet works. Ian Thompson, who recently switched from Claude Opus to ChatGPT for daily use, said the question of "what religion AI is" might actually map onto how each lab presents itself: Anthropic Buddhist, OpenAI evangelical.

The frame. Across all four May episodes, the panels kept circling the same observation: the AI labs are doing their growing-up in front of a jury, an SEC, and a press corps that finally has the financial documents to interrogate. The soap opera now comes with discovery.

The Tim Cook era ends with a beat. Cook will run one more earnings call, deliver one more WWDC keynote, and then hand the company to John Ternus on September 1. The April 26 panel - Victoria Song of The Verge, Stacy Higginbotham of Consumer Reports, and Sam Abuelsamid of Wheel Media - spent the second hour mapping what that means for a company that has been deliberately unsubtle about being late to AI.

The handover was telegraphed. Devindra Hardawar covered the MacBook Neo launch in March and noticed something. "There was no Tim Cook. It was all Ternus." Ternus was introducing the product, doing the play-by-play, drawing the comparisons to a $600 HP laptop placed conspicuously beside it. The Apple PR muscle was already on his name. The handover was choreographed long before the press release.

What Ternus inherits. The good news for Apple is that Ternus is a product guy with a hardware spine. He's been Cook's lead on Apple Silicon, the company's most consequential bet of the last decade. He shipped the MacBook Neo at a moment when, per Devindra, "PC manufacturers are just pulling their hair out and they don't know what to do." He has supply-chain leverage no rival can match. He gets a record-quarter farewell crow from Cook to ride into September.

What he doesn't inherit. The Steve Jobs in-a-generation reality-distortion field. Cook was a steady hand and a supply-chain genius, but the panel kept noting the absence of the "one more thing" theater. Ternus won't get to be Jobs either. What he does get to inherit is Apple's contrarian position on AI: late, slow, on-device, and looking smarter every quarter as Meta gets punished for capex and Microsoft pulls back on Copilot.

The undercurrent. Apple already feels the AI gold rush in its supply chain. The $599 Mac mini is gone, killed by RAM scarcity. Mac Studio shipping dates ran months out for entire RAM tiers. The cheapest Mac mini is now $799. Open-core hobbyists who used to buy the $599 to rack as cheap AI hosts now hit a wall. Apple, holding old RAM in long-dated contracts and owning its silicon stack, benefits from the crunch even without shipping a frontier model.

The May 17 panel was Amy Webb of the Future Today Strategy Group and Harper Reed of 2389.ai. They opened with Trump's CEO entourage in Beijing - Tim Apple, the Boeing chief, Qualcomm, Citigroup, Visa, and a late-added, awkwardly placed Jensen Huang who flew up to Alaska to catch Air Force One after being left off the original list because, as Webb put it, China sees the Taiwanese-American Nvidia CEO as a "political lightning rod."

By the numbers. China holds 70% of the global EV market and runs more than 100 EV companies. Its open AI model lineup - Qwen, GLM, Kimi, DeepSeek - is "only a couple months behind" the closed US frontier per Harper Reed, and "in many cases uncensored. They've removed the Chinese censorship from it." US hyperscalers are committed to roughly $800 billion in data-center capex on paper. China has a national broadband and compute plan; the US, Webb argued, has "no national strategy. It's all being left to the market."

The Linuxization frame. Reed's word for what is happening: the same dynamic that gave the world a free server operating system is now giving it free frontier-class AI. "If you talk to entrepreneurs outside the US, a lot of them are relying on these open Chinese models. They're not relying on closed US models. And it seems like a big blind spot of the US entrepreneurs to assume that, you know, Southeast Asia, APAC, even Europe, Middle East is going to suddenly use OpenAI, Anthropic, Facebook, Google models when they get something that is six months behind or even a year behind for effectively free that they have control of."

The hidden cost. Webb returned to a thread she has been pulling for months: the data-center buildout is not landing in places that will benefit from it. The Indiana communities being courted are the same ones the steel mills left. Hill County, Texas - already slated for eight data centers - has now passed a one-year moratorium. National polling she cited: 71% of Americans do not want to live near a data center. The compute is being built, but the political license to keep building it is eroding from underneath.

The accusation that doesn't land. Both Anthropic and OpenAI have argued the open Chinese models are good because they are distilling US closed models. Reed's response, lightly amused: "Well, but they distilled their own models off of themselves and also the internet. Like, I think that's a funny critique." You can't be the lab that scraped the world and then complain about being scraped.