Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Cerebras IPO, Warsh Confirmed Fed Chair, Musk-OpenAI Trial Nears End | Diet TBPN
TL;DR
Cerebras’ IPO blew past even bullish expectations — the AI chipmaker opened around $300-$350 a share, implying roughly a $64 billion market cap after hosts had considered a $50 billion IPO optimistic just days earlier.
The real bull case for Cerebras is that customers are proving they’ll pay up for speed — SemiAnalysis said it spent 80% of its AI budget on Anthropic Opus 4.6 fast mode despite a 6x price premium for roughly 2x-2.5x faster interactivity, and OpenAI already has a 750-megawatt deal with Cerebras.
Cerebras’ wafer-scale design has gone from ‘that’ll never work’ to clearly working, but scaling remains the question — redundancy solved early yield fears, yet its WSE-2 to WSE-3 memory only grew from 40GB to 44GB, raising concerns about larger models, 128k+ context windows, and networking versus Nvidia’s NVL72 systems.
Kevin Warsh was confirmed Fed chair in a narrow, political vote — the Senate confirmed him 54-45, and the hosts framed his big challenge as possibly managing stagflation, while contrasting that with Jerome Powell’s unusually broad bipartisan support of 80+ Senate votes in prior confirmations.
The Musk v. OpenAI trial appears to be ending faster than expected, and prediction markets have turned against Musk — what was once a 58% chance of victory on Polymarket-like markets fell to about 30% as closing arguments focused on whether Musk can prove an actual agreement OpenAI breached.
Figure’s humanoid robot demo impressed people, then triggered a fresh autonomy debate — a 24-hour stream running Helix 2 drew 3.4 million views, but a strange head-touching motion reignited questions about whether the robot was truly fully autonomous or had some kind of hidden teleoperation.
The Breakdown
Cerebras’ huge debut and the joke detour it inspired
The show opens with Cerebras stealing the day: an IPO so strong it essentially doubled the company’s valuation overnight, landing around a $64 billion market cap. The hosts immediately detour into a very TBPN riff about building an LA “museum of Instagrammable objects” — fake private jets, Lambos, confetti, hospital-kids photo ops — before snapping back to the point: this IPO looked nothing like a normal chip debut.
Why Cerebras suddenly looks real instead of theoretical
They explain Cerebras in the simplest memorable way possible: instead of cutting a wafer into many chips, it uses the whole wafer — “the biggest chip.” That idea carried a lot of FUD for years, especially around defect rates and yields, but the hosts note Andrew Feldman’s workaround of redundant cores solved the early critique, and now the chips are plainly in market and working.
Speed, not just intelligence, is becoming the product people buy
The liveliest discussion is around inference speed and why it matters more than many expected. They cite SemiAnalysis spending 80% of its AI budget on Opus 4.6 fast mode despite paying 6x for only about 2x faster responses, and compare it to landing on a fully loaded Wikipedia page versus waiting for text to dribble in like someone typing. Their Amazon-style analogy sticks: if every 100 milliseconds of latency really hurts conversion, then slow AI responses are probably killing usage and attention in exactly the same way.
The headwinds: memory, context windows, and the ‘boss model’ future
After the excitement, they hit the limits. Cerebras may struggle with bigger models and larger context windows because its on-wafer SRAM isn’t scaling fast enough — WSE-2 had 40GB, WSE-3 only 44GB — while Nvidia’s NVL72 systems are built to wire many chips together for giant workloads. Their framing is that future AI systems will likely be hybrid: a smartest “president” or “vice president” model delegates sub-tasks to faster Cerebras-style workers.
The people behind the company and a classic venture-capital ending
They shout out a lovely investor story from Ho Nam: Pierre Lamond backed Cerebras through Eclipse after earlier stints at Sequoia and Kleiner, and he did it in his 80s. The cap table history gives the arc some drama too — from a $100 million Series A in 2016 to $1.6 billion in 2018, $8 billion in early 2025, $23 billion later, and then a public debut near $48.8 billion before trading higher.
Kevin Warsh gets the Fed chair, and Powell gets a weird kind of praise
The hosts pivot to Kevin Warsh’s confirmation, first joking that he’s best known for interviewing Alex Karp while Karp spun a notebook on his finger and looked like he’d just popped a nicotine pouch. Then they turn serious: Warsh was confirmed 54-45 and may inherit a stagflation-style problem set where the usual rate-cutting playbook doesn’t work. They also spend a surprising amount of time defending Jerome Powell, arguing his tenure was strong but didn’t feature the same economy-saving crisis theater as 2008-era Fed legends.
Jensen in China, robot skepticism, and the orangutan loophole
There’s a quick China check-in with Jensen Huang and GPU sales politics, but the more entertaining segment is Figure’s humanoid robot stream. The hosts were genuinely impressed by the robot’s package-handling performance, then got pulled into internet skepticism after viewers noticed a weird motion where it touched its head. The bit escalates into a joke about an orangutan in a VR headset technically counting as “no humans in the loop,” which captures the mood exactly: impressed, but not fully convinced.
Musk vs. OpenAI heads toward the finish with courtroom absurdity intact
The final big story is the Musk-OpenAI trial entering closing arguments earlier than expected after only three weeks instead of four. The hosts lean on Mike Isaac’s colorful courtroom thread — bad AV, savage executive photos, anti-billionaire rhetoric, Microsoft fading into the bushes — while noting prediction markets now give Musk only about a 30% chance of winning, down from 58%. OpenAI’s legal position is blunt: unless Musk can prove a specific agreement about how his nonprofit donations had to be used, there’s no case.
A bizarre Tim Draper ice-bath finale
As a palate cleanser, they end on Tim Draper posting that he took 52 startup pitches in 52 minutes in sub-40-degree cold. Instead of admiring the grind, the hosts mostly marvel that the setup looks less like billionaire biohacking and more like a “prison ice bath,” complete with trash bags and decrepit walls — a perfectly odd image to close the show.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
