Podwise Weekly Picks, Week 44: Inference, Anthropic's S‑1, and the Power Bottleneck
Ten episodes from the week of October 26: why inference is the real business, what Anthropic's IPO filing reveals, Ethan Mollick on productivity, and the grid as the limit.
Weekly picks6 min read
The through-line this week is arithmetic. Inference costs more than training, Anthropic's filing puts real numbers on the compute bill, and Brookfield says the thing that actually slows the buildout is not money or demand — it is land, grid connections, and power.
These are the ten episodes from the Podwise Weekly Picks shelf for the week of October 26. Each entry says what the episode is about and who it is for; the show name links to the episode in Podwise if you want the summary and transcript.
The ten
Why Inference Matters 12x More Than Training
Y Combinator · 24 Oct
YC's Paper Club walks the full life of an inference request, from the API call down to the kernel on the GPU. The number that sticks: for every FLOP spent training, roughly six go to serving, and getting that part right has produced over $220 billion of market cap in a decade. If you build on top of models and have never looked below the API, this is the hour that explains your bill.
An 86-minute guide to finding PMF
Lenny's Podcast · 25 Oct · 86 min
Rob Snyder spent two years pushing a product nobody wanted before finding one that worked, then advised 50 startups on the climb to $1M ARR. This is his step-by-step framework for product-market fit, told through the failures rather than the highlight reel. Long, but the most useful kind of long.
Satya Nadella on open vs closed models
The Twenty Minute VC · 26 Oct · 26 min
Microsoft's CEO on whether every frontier model should be assumed compromised, why companies may need to own their own intelligence, and how vertical integration from chips to models changes the competitive map. Short, and dense with positions rather than predictions.
When AI agents use computers faster than you
The a16z Show · 26 Oct · 45 min
OpenAI's Ari Weinstein and James Sun argue that computer use is entering its scaling moment: better models, faster execution, and safety controls good enough for real work. The interesting claim is that the graphical interface, not an API, becomes the universal way for agents to touch software. Worth it if you are deciding what to expose to an agent first.
Separating the brain from the hands at Anthropic
Latent Space · 27 Oct · 107 min
Katelyn Lesse and Angela Jiang on the engineering behind the Claude platform: launching Opus 5.5, splitting the model's "brain" from the agent's "hands," and what it takes to run managed agents reliably. The architecture discussion is the part to listen for; the launch details age quickly.
Inside Anthropic's S-1
SemiAnalysis · 27 Oct
Revenue per dollar of compute is up 5.2x since early 2025, a customer who spent $1 now spends $9, and incremental gross margins are near 90% — against half a trillion in compute commitments. The team works through profitability, customer concentration, and the 2027 capacity question. Read this one with the Odd Lots episode below; they are describing the same week from different seats.
Ethan Mollick on the productivity gap
Big Technology Podcast · 28 Oct · 66 min
Why the models keep getting better while the economy-wide productivity numbers barely move, and what companies would have to change about how they operate for that to shift. Mollick is good at naming the organisational excuse. Useful if your team has tools but no measurable gain.
What Wall Street really thinks about the AI trade
Odd Lots · 29 Oct · 41 min
A week of wobbles, an S-1, and hyperscaler earnings, explained by someone who talks to institutional investors all day. The takeaway is less "bubble" than "we do not have a valuation model for this yet." A sober counterweight to the more excitable episodes above.
Coding agents for the rest of the enterprise
No Priors · 29 Oct · 38 min
Serval's Jake Stauch on putting code-generation agents in the hands of IT teams rather than engineers, and what it takes to challenge ServiceNow at the system-of-record layer. A concrete look at where agents land inside a company that is not a tech company.
The real bottleneck in the AI buildout
Goldman Sachs · 30 Oct
Brookfield's Connor Teskey says the constraint is neither capital nor demand but the operational capacity to develop land, secure a grid connection, and build power. He also pushes back on the idea that data-centre demand is fragile. If you only listen to one episode about the buildout, make it this one — it is the physical version of the argument.
How these are picked
The shelf is curated by Podwise from the week's episodes, and the order here follows the collection. Summaries and transcripts for all ten are already processed, so opening any of them costs no credits.
Browse the full shelf, including past weeks and themed collections, on Explore.
