ALL-IN · EXTRACTED
All-In Podcast ft. Travis Kalanick & Gavin Baker
China's AI catch-up, orbital compute economics, and why the DSA is taking over the Democratic Party the same way Trump took over the Republican one.
Preview · 3 of 5 tactics
"Communism is in all of us. Have you ever in your life been lazy? Have you ever wanted something for nothing? The difference is, do you make that a way of life?" — Travis Kalanick
This is All-In episode 278 with guest hosts Travis Kalanick of Adams and Gavin Baker of Atreides Management, filling in for Friedberg. The three big threads are the DSA's sweep of New York's congressional primaries, China's GLM 5.2 model closing the gap with US frontier labs, and the emerging economics of orbital versus terrestrial compute. The political analysis is sharper than usual because Kalanick and Baker bring operating-company instincts to a conversation that normally stays in the realm of punditry. What ties all three threads together is a single underlying question: when incumbents refuse to defend their position, who fills the vacuum?
Read The DSA Sweep As A Takeover, Not A Protest Vote
The three New York congressional primary results on Tuesday were not a protest. They were a demonstration of a working acquisition strategy. DSA co-chair Josh Block said it plainly: "We're using the Democratic Party as a ballot access vehicle. Not because we share its goals. We build our own organization, get elected under the Democratic label, caucus with Democrats when it's useful, and push our own agenda from the inside. We see the Democratic establishment as an obstacle, not a home." Sax made the structural point clearly. Even the Democratic members who do not lose their seats now have to recalculate. Three major upsets in one night means every congressman in a deep-blue district now has to ask whether they might be next. The response is to tilt their voting and their rhetoric toward the DSA before a primary challenge forces them to. The establishment bends the knee without a single additional election being necessary. Gavin Baker added the demographic precision that changes how you read the polling. The DSA is not winning with working-class voters, poor voters, black voters, or Hispanic voters. It is winning with relatively wealthy, downwardly mobile white liberals. These are, in his framing, people who went to elite schools, grew up in comfortable circumstances, and moved into the NGO and nonprofit sector rather than into industry. Their economic outcomes diverged from peers who did productive things, and that divergence is now expressing itself politically.
THE PLAY
Map the congressional districts in your state or city that fit the profile: heavily blue, low primary turnout, high share of college-educated voters under 40. Those are the districts where a DSA primary challenge is viable in the next cycle. If you are a founder, operator, or investor with any interest in local policy, that is where your attention and resources need to go before the primary, not after. Low-turnout primaries are where the DSA excels because they are more organized and more motivated than anyone else on the field.
Distillation Is The Mechanism Behind China's Model Catch-Up
GLM 5.2 from Z.AI is a 744-billion-parameter open-weight model under an MIT license, meaning it can be downloaded, forked, and built on freely anywhere in the world with no regional restrictions. It scored 51 points on the Artificial Analysis Intelligence Index, the highest score any open-weight model has achieved. It trades within one percentage point of Claude Opus 4.8 on the Frontier SWE coding benchmark and beats GPT 5.5 on software engineering. API usage costs roughly 85% less than GPT 5.5 for comparable performance. Gavin Baker explained the mechanism that made this possible. Picture the Chinese iPhone farms you have seen videos of, but scaled to tens of thousands of phones, tablets, and computers all querying the cloud APIs of US frontier models through masked accounts with very specific questions. Every token of the response is visible on the API. Those reasoning traces get harvested and fed back into the model during reinforcement learning and pre-training. You do not need to invent the capability from scratch. You train against the output of the models that already have it. Baker said this has been going on for a long time and there is no question it happened here. The further implication Baker raised is that GLM 5.2 is now good enough to run its own reinforcement learning, which means the distillation pipeline is no longer strictly necessary. The cat may be out of the bag. The Z.AI founder told Elon Musk that open-weight frontier-level capability will arrive before Q1 2027. China is currently 9 months behind US models, plus or minus 3 months depending on the specific capability. On silicon, they claim GLM 5.2's inference is optimized entirely for Huawei Ascend 910b chips, with no Nvidia hardware in the training cluster.
THE PLAY
Treat China's model timeline as 9 months behind, not years. Sax's framing is: we are on a shot clock and we have a few months, not a few years. If you are building on US frontier APIs, identify which workloads can run on open-weight models today and test GLM 5.2 or its successors on those tasks now. The cost difference is 85% cheaper for comparable coding and long-context work. The strategic question is not whether to use open-weight models but which queries need to go to a frontier model at all.
Newsletter
Get each new playbook the day it drops
One email per drop. No spam. Unsubscribe anytime.
The Economics Of Orbital Compute Start With The $25 Billion Problem
Standing up a 1-gigawatt terrestrial data center costs $35 billion in Nvidia semiconductors and $25 billion in power and cooling equipment. The $25 billion is heavily labor-intensive and clearly inflationary. Entitlements are harder. Political opposition is growing. Gavin Baker estimated that since 2021 roughly 40% of all data centers get contested, and that number is going up. The effective supply of new terrestrial compute is constrained in ways that have nothing to do with semiconductor availability. Gavin Baker laid out the orbital math from first principles. The $35 billion in silicon is roughly constant whether the chips are on the ground or in orbit. When Starship is fully and rapidly reusable, the launch cost to put a gigawatt of compute into orbit drops to approximately $5 billion. Total cost in space: $40 billion. Total cost terrestrially: $60 billion, and rising. The $25 billion terrestrial power and cooling figure is inflationary. The $5 billion Starship figure is deflationary as reuse improves. In three or four years the spread may be $70 billion terrestrial versus $40 billion orbital. The ongoing operating cost difference matters too. Running and cooling a gigawatt of terrestrial compute runs roughly $1 billion per year in power. Orbital compute changes that calculation entirely. Baker's framing was that this is not a speculative thesis about space. It is basic arithmetic about where the cost curves are heading on both sides of the comparison, and the terrestrial side is moving in the wrong direction while the orbital side benefits from every incremental Starship improvement.
THE PLAY
Use the $60 billion terrestrial versus $40 billion orbital figure as your baseline when evaluating any long-duration compute investment. The specific number to track is the cost per Starship launch as full reusability matures. Baker's framing is that 40% of new terrestrial data center projects get contested. If you are allocating capital to AI infrastructure over a 5-year horizon, the orbital cost curve is a real input to that decision, not a science fiction assumption. Watch the Terrafab timeline and Starship reuse cadence as the two variables that will move the orbital number fastest.
Subscribers Only
Unlock the Full Playbook
2 more tactics + Action Plan
TACTIC 04
Composable Models Are The Enterprise Architecture, Not A Single Provider
TACTIC 05
IPO Pricing: Auction The Deal, Stop Trying To Manage The Number
Already subscribed? Log in
ALL-IN · EXTRACTED BY PODEX