Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter
All-In #278 (Freeberg out; Travis Kalanick and Gavin Baker guesting) spans politics, China and — most usefully for a builder — AI-infrastructure economics. **Politics:** the DSA swept NYC's Democratic primaries (Mamdani's slate 3-for-3), which the table reads as a generational, downwardly-mobile-progressive + migrant coalition; Sacks frames the future as **'communism... of the Democrat Party, or nationalism in the Republican Party,'** while Chamath reframes the whole backlash through AI — **'AI is the greatest economic leveler [we] will ever find in our lifetime'** that Silicon Valley has branded so badly it created a vacuum socialism filled. **China:** Z.AI's open-weight GLM-5.2 lands a tick below Opus 4.8 on coding at ~85% lower cost, reportedly trained on Huawei chips; Gavin concedes heavy **distillation** ('a cheat sheet for other models to catch up') but argues the future is **composable models — Karpathy's 'council of LLMs'** routing most queries to a cheap open model and only the hardest to frontier checkers, and that open source 'shifts economic value from the frontier labs to the infrastructure.' Sacks (now running PCAST) hammers the race: GLM is as good as the *currently available* US models while Fable/GPT-5.6 sit in regulatory purgatory — **'we are going to lose if we keep doing this stuff to ourselves.'** **The infra core:** Micron's blowout (HBM sold out, stock ~14×) headlines a **DRAM bottleneck** Gavin calls the one that matters — **'memory capacity and bandwidth are foundational to the performance of every AI model'** — DRAM heading toward 30-40% of hyperscaler capex and pushing 'AI-flation' onto Apple/consumer kit. And the first-principles bet: data centres are getting *more* expensive (~$35bn silicon + ~$25bn power/cooling per gigawatt terrestrially), so **reusable Starship at '~$5 billion to put a gigawatt of compute into space'** makes orbital compute pencil — 'racks in space linked with lasers' — while modular 'Megapods' and distributed/disaggregated inference (Groq/Cerebras decode in front of old GPUs) reshape where compute lives. **Markets:** Gavin pegs Anthropic at **'$3 trillion today'**, says the IPO wave is just private-to-public reshuffling capital markets can absorb, and dissects Cerebras breaking deal price (forced 'sell-no-matter-what' PMs → auction, don't over-price).
Key points
- NYC's DSA sweep (Mamdani's slate 3-for-3, incumbents toppled) is read as a coalition of downwardly-mobile, elite-educated progressives plus a migrant base; Sacks lays out the radical platform and frames the macro choice as 'communism... of the Democrat Party, or nationalism in the Republican Party.'
- Chamath's reframe: 'AI is the greatest economic leveler [we] will ever find' — it turns the world's knowledge into per-person expertise (everyone gets 'an equivalent Travis Kalanick as co-founder') — but Silicon Valley branded it so poorly (doomerism, jobs/water FUD) that it left a vacuum socialism is filling.
- China's open-weight catch-up: Z.AI's GLM-5.2 (744B params, MIT license) scores just below Opus 4.8 on SWE coding at ~85% lower cost and reportedly trained on Huawei Ascend chips — China is ~6 months behind on models, ~24 on silicon, yet only a few months behind in total.
- Distillation is the mechanism (Gavin): harvesting frontier 'reasoning traces' via masked API accounts is 'a cheat sheet for other models to catch up' — but once a model is good enough to do its own RL, 'the cat may be out of the bag.'
- The future is composable models — Karpathy's 'council of LLMs': route most queries to a cheap RL'd open-weight model and only the hardest to frontier checkers; open source isn't bad for AI, it 'shifts economic value from the frontier labs to the infrastructure' (Nvidia is 'the American open-source champion' that simply chooses not to ship, to avoid channel conflict).
- Sacks on the race (now running PCAST): GLM is as good as the *currently available* US models while Fable and GPT-5.6 sit in regulatory purgatory; clamping down doesn't slow China (out of US jurisdiction), so the answer is white-hat vulnerability-finding + fast upgrade cycles + pro-export — 'we are going to lose if we keep doing this stuff to ourselves.'
- The DRAM bottleneck is THE bottleneck (Gavin): Micron's blowout (HBM sold out for 2026, revenue up ~4×, stock ~14×) reflects that 'memory capacity and bandwidth are foundational to the performance of every AI model'; only 3 firms make HBM, Elon is pointing TeraFab at memory, and DRAM is heading toward 30-40% of hyperscaler capex.
- 'AI-flation' has reached consumers: DRAM scarcity is forcing Apple price hikes (MacBook Neo +14%, Mac Studio +25%) and pressuring Xbox/Switch/PlayStation — AI demand is price-insensitive, so consumer electronics get squeezed; meanwhile CXMT going public may flood cheap consumer-grade DRAM.
- Orbital-compute economics from first principles (Gavin): a 1GW data centre is ~$35bn silicon + ~$25bn power/cooling (the latter inflationary, labour-heavy); reusable Starship could put a gigawatt up for '~$5 billion' of launch, so 'racks in space linked with lasers' start to pencil while terrestrial sites become scarce 'diamonds' — favouring SpaceX's stack on token cost.
- Markets: Gavin pegs Anthropic at '$3 trillion today' (ending the year >$100bn revenue, ~85% inference gross margins) and argues the IPO wave is just private-to-public reshuffling global capital markets can absorb; the Cerebras lesson is structural — 'sell-no-matter-what' PMs dump anything that breaks deal price, so price via auction and don't price to perfection. Plus modular 'Megapods' and distributed/disaggregated inference (prefill vs decode; Groq/Cerebras in front of old GPUs) reshape where compute lives.
Notable quotes
I think AI is the greatest economic leveler will ever find in our lifetime.
the choices of the future are gonna be communism, or if you want to call it socialism, of the Democrat Party, or nationalism in the Republican Party.
you're gonna have what Andrej Karpathy called the council of LLMs.
This is the most important bottleneck simply because memory capacity and bandwidth are foundational to the performance of every AI model.
When Starship is reusable, it's going to cost $5 billion to put a gigawatt of compute into space.
I think Anthropic is worth $3 trillion today, and it's very important.
We are going to lose if we keep doing this stuff to ourselves.
Themes
- AI infrastructure economics (memory bottleneck + orbital compute)
- China's open-source catch-up
- the IPO froth and market absorption
- the political backlash (DSA socialism)
- AI as an economic leveler
Mentioned
People
Companies
Ideas
- AI as economic leveler
- DSA socialism vs nationalism
- China open-source catch-up (GLM 5.2)
- distillation as a cheat sheet
- composable models / council of LLMs
- DRAM/HBM as the key bottleneck
- AI-flation
- orbital compute economics
- modular data centres (Megapod)
- distributed / disaggregated inference (prefill vs decode)
- IPO absorption / breaking deal price / auction-vs-underwrite
- Anthropic ~$3tn