Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter

2026-06-27 - 101 min - source - Read full transcript

Jason Calacanis (host)David Sacks (host)Chamath Palihapitiya (host)Travis KalanickGavin Baker

Key insights

DSA candidates swept multiple safe Democratic seats in NYC's June primary on a low-turnout, highly organized base, and the DSA's own leadership frames this as a deliberate takeover of the Democratic Party rather than loyalty to it.

Chamath, Sacks and Gavin Baker walk through New York's 7th and 13th districts (Claire Valdez and an unnamed PhD-candidate winner) plus a similar upset in LA, noting DSA has roughly half of LA's city council seats and attributing the wins to low turnout (around 17%) rewarding organized, ideologically motivated volunteers. Sacks quotes DSA co-chair Josh Block describing the party as merely a 'ballot access vehicle': the group builds its own organization, caucuses with Democrats 'when it's useful,' and treats the establishment wing as an obstacle rather than a home.

nyc-dsa-socialist-sweep

Gavin Baker argues DSA's actual voting base is downwardly-mobile, wealthy white progressives, not the working-class or poor voters the traditional Democratic Party once represented.

He notes DSA is losing votes with working-class, poor, Black, and Hispanic voters even as it gains among the college-educated and high-income. He ties this to a class of elite-school graduates who went into the NGO/nonprofit sector instead of productive industry and became a well-funded activist vanguard.

nyc-dsa-socialist-sweep

Chamath's cross-country theory: Canada, the UK, and Australia veered into socialism-adjacent policy well before the US and show the resulting instability, and each has since banned social media for under-16s.

He frames the under-16 social media bans in those three countries as a circuit-breaker on youth radicalization via algorithmic feeds, predicting a less radicalized cohort will age into voting rolls there. Travis Kalanick counters that age-gating is really a pretext for de-anonymizing adults to enable broader censorship regimes, particularly in the UK.

nyc-dsa-socialist-sweep

China's Z.ai released GLM 5.2, an MIT-licensed, 744-billion-parameter open-weight model that scores 51 on the Artificial Analysis Intelligence Index - the highest of any open-weight model - and beats GPT-5.5 on frontier coding benchmarks while trailing Claude Opus 4.8 by under one point.

It runs roughly 85% cheaper than GPT-5.5 for comparable performance and is inference-optimized for Huawei Ascend 910B chips, which Z.ai claims trained the entire model, reflecting a deliberate Chinese push toward chip self-sufficiency (indigenousization).

open-source-vs-closed-models

Gavin Baker believes distillation via masked API farms is real and significant, but argues the frontier gap will persist because composable, multi-model architectures - not a single winner-take-all frontier model - are the actual future.

He describes 'iPhone farms' of masked accounts harvesting reasoning traces from Claude/GPT APIs to cheaply approach frontier capability via reinforcement learning, but says most enterprise queries (he estimates up to 85%) will route to cheap in-house open-weight models, with only the hardest queries escalated to frontier models - shifting economic value toward infrastructure rather than eliminating frontier labs.

open-source-vs-closed-models

Anthropic's Claude/Fable models were pulled back after a jailbreak report from a trusted partner triggered a government approval process Dario Amodei had publicly lobbied for, and Sacks argues this created a regulatory moat that backfired.

Sacks says Amodei has advocated for an 'FAA for AI' and a federal approval regime; when the government acted, it rolled back Fable itself, which Sacks calls Amodei getting 'hoisted on his own petard.' He argues the US cannot afford further self-imposed slowdowns because China is roughly 6 months behind on models (per Z.ai's own claim of hitting Fable-level capability by Q1) despite being ~24 months behind on silicon.

open-source-vs-closed-models

DRAM, not GPUs or power, is the single most important AI infrastructure bottleneck right now, according to Gavin Baker, and it is expected to consume 30-40% of all hyperscaler capex next year.

He argues memory capacity and bandwidth are foundational to every model's performance, citing Elon Musk's decision to have TerraFab focus specifically on memory rather than lasers, capacitors, or power supply chains. Only three companies (Micron, SK Hynix, Samsung) can make the specialized HBM/LPDDR DRAM needed for AI servers, versus many more suppliers of consumer-grade DRAM.

ai-memory-and-dram-bottleneck

Micron's new supply chain agreements set price floors and ceilings covering roughly 50% of revenue with four large customers, with floor pricing already above prior-cycle gross margin peaks - a structural, not just cyclical, re-rating.

Gavin Baker frames this as transformational for the DRAM industry's business model, similar to how wafer-fab-equipment suppliers re-rated to permanently higher multiples in prior cycles. He contrasts this with Apple's price hikes (MacBook Air up 15%, Mac Studio up 25%) as an early sign of 'AI-flation' bleeding into consumer electronics, including Xbox and expected PlayStation/Switch price increases.

ai-memory-and-dram-bottleneck

The DRAM/power bottleneck means it may be getting harder and more expensive to build new data centers over time, not easier, inverting the usual tech-cost-curve assumption.

Gavin Baker puts a terrestrial one-gigawatt data center at roughly $35 billion of semiconductors plus $25 billion of power/cooling equipment, with the power/cooling portion inflationary due to labor costs and worsening siting/political constraints. He contrasts this with orbital compute economics: once Starship is fully reusable, launch cost could fall to about $5 billion per gigawatt, making a roughly $40 billion orbital gigawatt competitive with the $60 billion terrestrial equivalent within a few years.

orbital-and-distributed-compute

Inference is increasingly disaggregated into a memory-capacity-bound 'pre-fill' stage and a memory-bandwidth-bound 'decode' stage, letting older GPUs be repurposed rather than retired, but distributed/community compute pools only work for this inference case, not for training.

Sacks explains that decode-optimized chips like Groq (Nvidia-acquired) or Cerebras can be placed in front of older H100s/A100s pulled from legacy data centers, extending useful GPU life to 7-12 years. Chamath and Sacks note that physical distance between GPUs kills training efficiency even over short distances (facilities two kilometers apart on the same fiber connection suffer meaningful loss), while distributed inference is far more latency-tolerant, making community compute-sharing schemes (BitTensor, Venice, Pluralis, Targon, and a rumored Tesla Powerwall-plus-GPU bundle) plausible for inference but not training.

orbital-and-distributed-compute

Gavin Baker values Anthropic at roughly $3 trillion as a public company, projecting it will end the year with over $100 billion in revenue and roughly 85% gross margins on inference.

He argues the ~$4 trillion combined new equity supply from SpaceX, OpenAI/Anthropic-scale offerings, and Cerebras is not actually hard for public markets to absorb, since typical secondary offerings are a small percentage of a company's value and global capital markets dwarf even these headline numbers - the money is simply shifting from private to public markets, not appearing from nowhere.

trillion-dollar-ai-ipos

Cerebras broke its IPO deal price within two days of reporting its first public-company quarter, triggering mechanical, price-insensitive selling from portfolio managers who treat a broken deal price as a promise violated.

Gavin Baker says this creates a self-reinforcing pile-on: shorts target stocks approaching deal price to trigger the mechanical selling and profit from the resulting drop. He argues Cerebras also under-told its growth story (its OpenAI contract's revenue impact won't show up until roughly Labor Day due to the ~7-month chip-to-token production pipeline) and advocates for Dutch-auction IPO pricing to avoid the deal-price-break dynamic entirely.

trillion-dollar-ai-ipos

Books referenced

The Fountainhead - Ayn Rand - Sacks needles Travis Kalanick that this was his Twitter avatar for years, tying it to Kalanick's individualist streak versus DSA collectivism.

Media referenced

Bill Belichick 'onto Cincinnati' - other - Gavin Baker quotes the line after the SpaceX IPO to signal moving on to the next challenge rather than celebrating.
Andrej Karpathy's 'Council of LLMs' - other - Gavin Baker cites Karpathy's framing to describe the coming composable-model future where enterprises route queries across multiple frontier and open-source models.
Megyn Kelly interview (Chamath appearance) - other - Chamath references his own appearance on Megyn Kelly's show where he laid out his socialism-in-Anglophone-countries theory and the Jews/Israelis/Netanyahu distinction.

Companies

Adam (Travis Kalanick's new startup) - Kalanick's new venture, teased but not detailed; he jokes it's 'not super secret, it's super awesome' and references a trademark filing tied to modular AI data center hardware ('Megapod').
Z.ai (Zhipu AI) - Released GLM 5.2, a 744B-parameter, MIT-licensed open-weight frontier model that scores 51 on Artificial Analysis and claims to be trained entirely on Huawei Ascend 910B chips.
Micron - Reported blowout earnings: revenue up 4x year-over-year to $9-42 billion range with guidance beating estimates by 16%; stock up roughly 10x, 14x since Gavin Baker's 2025 HBM call.
Nvidia - Discussed as the incumbent GPU/compute stack; Sacks warns Nvidia may build an OpenAI-competing model business if OpenAI's in-house Broadcom-built chip effort threatens Nvidia's position.
OpenAI - Launched its own 'jalapeno' chip this week, reportedly built with Broadcom, seen by Sacks as a competitive shot at Nvidia.
Anthropic - Discussed as being 'rolled back' regulatorily after a jailbreak report triggered government scrutiny of Claude/Fable; Gavin Baker estimates its public-market value at roughly $3 trillion.
SpaceX - Recently completed IPO, discussed alongside Cerebras and Anthropic as historic new equity supply hitting public markets; stock has retreated from its post-IPO high.
Cerebras - IPO'd and then broke deal price within two days, triggering price-insensitive selling from portfolio managers who mechanically sell any stock that breaks its IPO price.
Apple - Raising prices across MacBook Air/Mac Studio lines (up 15-25%) because DRAM costs, driven by AI data center demand, are being passed on to consumers for the first time.
Huawei - China's national semiconductor champion; Z.ai claims GLM 5.2's full training run used Huawei Ascend 910B chips as part of a broader Chinese chip-indigenization push.
SK Hynix - One of only three companies (with Micron and Samsung) that makes high-bandwidth memory for AI GPUs; based in South Korea.
Samsung - Third HBM maker alongside Micron and SK Hynix; based in South Korea.
CXMT - Chinese memory maker going public, expected to flood the market with cheap consumer-grade DRAM even as AI-grade HBM stays a three-company oligopoly.
Tesla / SpaceX AI ('TerraFab') - Elon Musk's TerraFab initiative to build a chip/memory fab, discussed alongside a rumored Tesla-SpaceX merger and a trademark filing for modular 'Megapod' data center hardware possibly sited at Supercharger stations.
Groq - Nvidia-acquired chip company; cited alongside Cerebras as inference-optimized (decode-focused) alternatives that can sit in front of older Nvidia GPUs like H100s/A100s in disaggregated inference setups.
Crusoe - Cited as a company modularly assembling data centers, similar to the disaggregated shipping-container approach discussed for megapods.
Perplexity - Cited as an example of an American company that forked a censored Chinese open-source model and restored uncensored answers on topics like Tiananmen Square.

Techniques and frameworks

Distillation via masked API farms - Gavin Baker explains Chinese labs allegedly run 'iPhone farms' of masked accounts querying frontier APIs to harvest reasoning traces, then use them to fine-tune/RL their own models toward frontier performance at a fraction of the cost.
Composable models / Council of LLMs - Gavin Baker's thesis that enterprises will route most queries to a cheap in-house open-weight model and escalate only the hardest queries to frontier models, shifting economic value from frontier labs to infrastructure.
Disaggregated inference (pre-fill vs. decode) - David Sacks/Chamath describe splitting inference into a memory-capacity-bound 'pre-fill' step and a memory-bandwidth-bound 'decode' step, letting older GPUs (H100s/A100s) be paired with decode-optimized chips like Groq or Cerebras for competitive economics.
The Curley effect - Gavin Baker cites this political-economy concept (named for a Boston mayor) to describe pursuing policies known to hurt constituents because they drive out political rivals, entrenching incumbent power.
Dutch auction IPO pricing - Gavin Baker recounts using a Dutch auction to price an early Uber funding round (participants submit price/quantity sheets, clear at the price that fills the raise) and argues auctions avoid the deal-price-break problem seen with Cerebras.

Summary

The episode opens with an extended, heated segment on the DSA's June primary sweep in New York City, where two guest hosts (Travis Kalanick and Gavin Baker) sit in for an absent Friedberg. Chamath, Sacks and the guests dissect several specific upset races - including a 32-year-old unemployed PhD candidate who has called to "eradicate Western civilization" defeating a longtime Hakeem Jeffries ally - and argue the DSA's real base is downwardly-mobile, elite-educated white progressives plus recent migrants, not the working-class voters Democrats traditionally claimed to represent. Sacks reads DSA co-chair Josh Block's quote describing the party as merely a "ballot access vehicle," and the group frames Zohran Mamdani as a singularly talented, Trump-style populist communicator whose charisma (illustrated by a viral Knicks victory speech) is doing more to drive DSA's ascendance than its actual platform. Chamath pivots to a broader theory that Canada, the UK, and Australia's earlier drift toward socialism-adjacent policy produced visible instability, and credits their subsequent under-16 social media bans as a possible circuit-breaker on youth radicalization - a claim Travis Kalanick pushes back on, arguing age-gating is really about de-anonymizing adults to enable censorship. The Israel/Gaza divide within the Democratic primary electorate (80% of Democrats now disapprove of Israel) and within the Republican coalition (generational split, with under-50 Republicans at 57% disapproval) gets a substantial side discussion, including Chamath's point that people increasingly conflate Jews, Israelis, and Netanyahu into one undifferentiated target.

The conversation shifts to AI competition with China's Z.ai releasing GLM 5.2, an MIT-licensed 744-billion-parameter open-weight model that scores highest-ever on the Artificial Analysis Intelligence Index, beats GPT-5.5 on coding benchmarks, and trails Claude Opus 4.8 by under a point - all allegedly trained entirely on Huawei's domestic Ascend 910B chips. Gavin Baker acknowledges heavy distillation from frontier APIs (via "iPhone farms" of masked accounts harvesting reasoning traces) but argues the real future is composable, multi-model architectures where most enterprise queries route to cheap open-weight models and only the hardest escalate to frontier labs - a shift that benefits infrastructure providers more than it threatens frontier labs outright. Sacks connects this to his own regulatory history, arguing Anthropic's Dario Amodei got the government-approval regime he'd publicly lobbied for, only to have his own Fable models rolled back by it after a jailbreak report - a self-inflicted setback Sacks says the US can't afford given China is roughly 6-9 months behind on models but closing fast.

A major middle section covers Micron's blowout earnings (revenue up 4x year-over-year, stock up roughly 10x since the report and 14x since Gavin Baker's 2025 call), which the group uses to argue DRAM - not GPUs, power, or any other component - is the defining AI infrastructure bottleneck, set to consume 30-40% of hyperscaler capex next year. Only three companies worldwide (Micron, SK Hynix, Samsung) can make server-grade HBM, and Micron's new long-term supply contracts lock in floor pricing above prior-cycle peaks, which Gavin Baker frames as a structural industry re-rating. The DRAM crunch is already visible in consumer prices: Apple raised MacBook Air and Mac Studio prices 15-25%, and the group expects similar "AI-flation" to hit gaming consoles. This bottleneck discussion extends into a long riff on the economics of terrestrial versus orbital data centers (Chamath argues a fully reusable Starship could make orbital gigawatt compute cost-competitive within a few years), disaggregated inference architectures that extend the useful life of older GPUs, and the practical and security limits of distributed/community compute pools, including a discussion of a rumored Tesla "Megapod" trademark filing.

The episode closes on public markets absorbing an unprecedented wave of AI-related equity: Gavin Baker pegs Anthropic's public-market value at roughly $3 trillion, and the group discusses SpaceX's post-IPO retreat, Cerebras breaking its IPO deal price within two days (triggering mechanical, price-insensitive selling from funds that treat a broken deal price as inviolable), and the case for Dutch-auction IPO pricing to avoid that dynamic - illustrated by Gavin Baker's own war story pricing an early Uber round via Dutch auction.

Notable Quotes

"I think AI is the greatest economic leveler will ever find in our lifetime... but we've done such a poor job in representing it, in bringing it to market, in talking about it." - Chamath Palihapitiya

"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." - David Sacks, quoting DSA co-chair Josh Block

"The bottleneck that matters is DRAM... memory capacity and bandwidth are foundational to the performance of every AI model. So this is the most important bottleneck." - Gavin Baker

"If you're going public and you're listening to this, tell your bankers: price this in such a way that we're not going to break deal price in our first nine months as a public company." - Gavin Baker

"We are going to lose if we keep doing this stuff to ourselves." - David Sacks, on self-imposed AI regulatory slowdowns versus China