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

Summary

Note: Analyst assigned ticker CSCO (Cisco) to Cerebras — this is incorrect. Cerebras trades as CBRS on NASDAQ.

Theses (episode spine)

• DSA-aligned socialist candidates endorsed by NYC Mayor Mandami swept all three targeted congressional Democratic primaries (NY-10, NY-13, NY-7), with Polymarket having put the full trifecta at only 26% odds; Sacks, Gavin, Chimath, and Travis all see this as dangerous and accelerating, while Gavin attributes it primarily to Mandami’s singular political talent.
• China’s GLM 5.2 (744B parameter, MIT-licensed open-weight model from Z.AI) scored highest ever on the Artificial Analysis Intelligence Index for any open-weight model, beating GPT-5.5 on SWE coding benchmarks and trailing Claude Opus 4.8 by under 1 percentage point at 85% lower API cost; Z.AI’s founder told Elon Musk open-weight fable-level capabilities will arrive before Q1 2027.
• Gavin Baker confirmed large-scale distillation from US frontier model APIs is definitively happening in China — tens of thousands of devices harvest reasoning traces through masked API accounts — and argues this, combined with GLM 5.2 now being good enough to run its own RL, may mean the cat is fully out of the bag.
• Micron reported blowout earnings with revenue up 4x YoY to $42B, beat expectations by 16%, and guided Q4 to ~$50B vs. $43B expected; Gavin calls DRAM the single most important AI bottleneck and notes new supply agreements with floor pricing above prior cycle peak gross margins could be transformational for industry structure; Gavin’s 2025 HBM call is up 14x.
• Gavin values Anthropic at roughly $3 trillion as a public company, citing expected revenue ‘well over $100B’ exiting 2026 and 85% gross margins on inference at scale; he dismissed market absorption concerns for the $4–6 trillion AI IPO wave as merely a shift from private to public capital markets.
• Cerebras broke IPO deal price within two days of its first public earnings report, triggering price-insensitive institutional selling; Gavin argues the key metric is megawatt capacity ramp — at 50 MW per month added in 2027, the company would exit that year at roughly a $9B cloud revenue run rate against a sub-$40B market cap.
• Sacks warns the US is on a ‘shot clock’ with only months before China reaches frontier parity: the US rolled back Fable and is delaying GPT-5.6 through new approval hoops; China is 6 months behind on models but 24 months behind on silicon, yet only a few months behind in total capability.
• DRAM will be 30–40% of all hyperscaler capex next year according to Gavin; a 1 GW terrestrial data center costs roughly $35B in semiconductors plus $25B in power and cooling; orbital compute via Starship could cost only $40B per GW all-in once launch costs fall to $5B, potentially making it cheaper than terrestrial within 3–4 years.
• Apple announced Mac price increases of 14–25% (MacBook Neo from $699 to $799; Mac Studio up 25%) due to AI data centers consuming DRAM supply; Xbox, Switch, and PlayStation price increases are also coming as consumer electronics compete with AI infrastructure for scarce memory.
• Social media bans for under-16s in Canada, UK, and Australia: Chimath believes these will reduce youth radicalization by cutting DSA-style movements off from their primary recruitment channel; Travis and Gavin warn the real agenda is forcing adult deanonymization to enable censorship of political dissent.

$MU (Micron Technology)

Convergence / divergence: Both speakers bullish. Gavin Baker is the primary analyst here.

Speaker calls:

• Gavin Baker (bullish, long-term): Micron’s blowout quarter (revenue up 4x YoY to $42B, Q4 guidance ~$50B vs $43B expected) combined with new supply agreements with floor pricing above prior cycle peak gross margins could be transformational for industry structure. DRAM is the single most important AI bottleneck. HBM DRAM will be 30–40% of all hyperscaler capex next year. Entire 2026 supply is sold out.
• Jason Calacanis (bullish): Credited Gavin’s 2025 prediction show call on HBM makers like Micron as the best-performing asset since that time, now up 14x (Gavin cited 14x on the show; current data shows approximately 10x from trough).

Cross-check:

• Price: $1,129.00 (P/E 23.59 TTM; forward P/E 9.76; mkt cap $1.27T). Q3 FY2026 earnings reported June 24, 2026. Next quarterly report approximately September 2026.
• Recent headlines worth knowing: Q3 FY2026 EPS $25.04 (vs $1.69 prior year), revenue $41.5B (+346% YoY). Beat by 16%. Q4 guidance $50B vs $43B expected. Supply chain agreements with floor/ceiling pricing covering ~50% of revenue with four customers. Chinese DRAM maker CXMT going public — potential future competition for consumer-grade DRAM. New York Micron fab construction halted due to environmental issues.
• Inconsistencies: Gavin said “up 14x” since the 2025 prediction show; current market data suggest approximately 10x from the 2025 low. Either the 2025 base was different or the stock has retreated somewhat from its high. Forward P/E of 9.76 suggests the market expects rapid earnings growth even at $1,129/share.

$NVDA (Nvidia)

Convergence / divergence: Chamath is the only speaker with an explicit NVDA view. No dissent expressed by others.

Speaker calls:

• Chamath Palihapitiya (bullish): Called Nvidia the American open-source AI champion and argued it could release a model competitive with GLM 5.2 or better whenever it chooses; noted this creates channel conflict incentive for Nvidia to start a model company if OpenAI (launching its own Jalapeno chip built by Broadcom) continues competing on hardware.

Cross-check:

• Price: $195.74 (P/E 29.84 TTM; down from 12-month avg 45.95; mkt cap $4.88T). Next earnings: 2026-08-25.
• Recent headlines worth knowing: Nvidia is world’s most valuable semiconductor company at $4.88T market cap and #1 company by market cap overall. OpenAI launched Jalapeno chip (made by Broadcom) — creating hardware channel conflict. Elon’s Terrafab (SpaceX AI chip fab) targeting DRAM production. Groq (Nvidia-acquired) and Cerebras being used for inference disaggregation with Nvidia H100/A100 chips.
• Inconsistencies: P/E compression to 29.84 from historical ~46 avg may reflect growth expectations moderating or earnings catching up with price. Chamath’s “American open source champion” framing is somewhat loose — Nvidia’s research division releases models, but Nvidia is not primarily an open-source company.

$CBRS (Cerebras Systems)

Convergence / divergence: Only Gavin explicitly discussed Cerebras. He is neutral-cautious pending execution on power capacity.

Speaker calls:

• Gavin Baker (neutral/cautious): Cerebras broke its IPO deal price ($185) within two days of its first public earnings report, triggering price-insensitive institutional selling from funds that exit any stock breaking deal price. Key investment metric is megawatt capacity ramp — at 50 MW per month added in 2027, company would exit that year at ~$9B cloud revenue run rate against a sub-$40B market cap. The quarter showed slower growth relative to rest of AI sector; that was the trigger. Warned companies going public to price so they don’t break deal price in the first nine months.

Cross-check:

• Price: $181.59 as of June 26 (P/E N/A — pre-profitability; IPO price was $185; mkt cap ~$55B at current price, peak ~$95B at IPO). No earnings date set for next report.
• Recent headlines worth knowing: Cerebras IPO priced May 13, 2026 at $185 (raised $5.55B — largest tech IPO since Uber 2019). Stock opened at $350, peaked near $385, closed first day +68% at $311. Subsequently declined back below $185 deal price. Q1 public earnings showed slower growth relative to rest of AI. OpenAI is a major customer ($20–25B contract signed December 2025; impact on revenue won’t show until ~Labor Day 2026).
• Inconsistencies: Analyst assigned ticker CSCO (Cisco Systems) to Cerebras — this is WRONG. Cerebras trades as CBRS. The company’s ticker symbol on NASDAQ is CBRS, not CSCO. All data above is for CBRS.

$AAPL (Apple)

Convergence / divergence: Jason bearish near-term on AAPL due to margin pressure/demand destruction from price hikes. Gavin more neutral — sees Chinese DRAM competition as a long-term relief valve.

Speaker calls:

• Jason Calacanis (bearish, near-term): Apple announced price increases across its Mac line — MacBook Neo from $699 to $799 (up ~14–15%) and Mac Studio up 25% — because AI data centers are consuming DRAM supply that previously went to consumer electronics. Called it “inflation has come to the desktop.” Notes similar increases coming for Xbox, Switch, and PlayStation.
• Gavin Baker (neutral): Noted Chinese DRAM maker CXMT going public may flood consumer-grade DRAM markets and could be “the cure for Apple’s ills,” but stressed AI-server-grade DRAM (HBM, LPDDR, SOCAM) is a completely different product that only three companies can make, so relief for consumer electronics doesn’t help the AI data center bottleneck.

Cross-check:

• Price: $275.15 (P/E 37.16 TTM; forward P/E 34.24; mkt cap $4.47T). Next earnings: ~August 2026 (Q3 FY2026).
• Recent headlines worth knowing: Apple shares fell ~6% after announcing 15–25% price increases across Mac, iPad, and accessories lines. All-time high was $315.20 on June 2, 2026. Consensus 12-month target ~$313 (~14% upside from current). Leadership change approaching September 2026. CXMT (Chinese DRAM maker) going public as potential future consumer DRAM competitor.
• Inconsistencies: Apple’s stated rationale (DRAM scarcity from AI data centers) matches Gavin Baker’s bottleneck analysis precisely, lending credibility. The 6% stock drop on announcement suggests market viewed these as demand headwinds, not margin expansion.

Topics discussed

DSA Socialist Sweep of NYC Democratic Primaries

Summary: DSA-backed candidates endorsed by Mayor Mandami won all three targeted NYC congressional primaries: Brad Lander defeated two-term incumbent Dan Goldman in NY-10 primarily over the Israel issue; a 32-year-old unemployed PhD candidate beat a five-term Hakeem Jeffries ally in NY-13; and a DSA candidate won the open NY-7 seat. Polymarket had the full trifecta at only 26% pre-election.

Speaker views:

• David Sacks: Sees this as a genuine national DSA takeover of the Democratic Party enabled by open-border policies the Democratic establishment created; the DSA co-chair explicitly stated they use the Democratic Party as a “ballot access vehicle” while building their own organisation.
• Gavin Baker: Attributes DSA ascendancy primarily to Mandami’s singular political talent — the most talented politician Gavin has seen in his lifetime — not to the appeal of DSA ideas, which lead to measurably bad outcomes wherever tried.
• Chimath Palihapitiya: Silicon Valley has done a poor job representing AI as an economic leveller, letting personal conflicts spill public and ceding credibility; the vacuum has allowed DSA to gain momentum partly as a referendum on AI; proposed social media bans under 16 in Canada/UK/Australia may help reduce radicalization.
• Travis Kalanick: Frames the problem through two aphorisms: truth and justice are society’s immune system; communism is in all of us (laziness and wanting something for nothing) and ecosystems that allow those tendencies without consequence reach critical mass.
• Jason Calacanis: Noted Mandami has taken the Trump playbook — building a big tent, communicating with exceptional oratory — and will destroy the Democratic establishment from the inside out.

Potential impact: DSA primary wins will force mainstream Democrats to bend leftward to avoid primary challenges; Sacks expects DSA takeover of the Democratic Party to continue, with the Israel issue as a primary driver of generational realignment.

China’s GLM 5.2 Reaches Frontier-Level AI Performance

Summary: Z.AI released GLM 5.2, a 744B-parameter MIT-licensed open-weight model scoring highest ever on the Artificial Analysis Intelligence Index, beating GPT-5.5 on SWE coding benchmarks and trailing Claude Opus 4.8 by under 1 percentage point at 85% lower API cost. Z.AI founder told Elon Musk open-weight Fable-level capabilities will arrive before Q1 2027.

Speaker views:

• Gavin Baker: Confirmed large-scale distillation from US frontier model APIs is happening — tens of thousands of devices harvesting reasoning traces through masked accounts — and GLM 5.2 is now good enough to do its own RL, meaning the cat may be fully out of the bag; training claimed to be on Huawei Ascend 910b chips; predicts composable models (frontier + enterprise fine-tuned open-source) as the dominant future architecture.
• David Sacks: Argued the US has handed China months of advantage by rolling back Fable and subjecting GPT-5.6 to new approval hoops; China plans to package Huawei-chip-optimized models as “AI in a box” and sell them globally at a fraction of US prices within one to two years.
• Chimath Palihapitiya: The gap is insane — China is 6 months behind on models and 24 months behind on silicon yet only a few months behind in total capability; argued Nvidia is the American open-source AI champion and could release a competitive open-weight model whenever it chooses.

Potential impact: China capturing global open-source AI market with Huawei-native models threatens US export-control strategy; US companies already adopting Chinese open-source models for workloads restricted by frontier lab safety policies.

AI Memory Crunch: DRAM as the Defining Infrastructure Bottleneck

Summary: DRAM scarcity — driven by AI data centers consuming HBM supply — is cascading into consumer electronics price increases and constraining AI infrastructure buildout. Only three companies globally (Micron, SK Hynix, Samsung) can produce AI-server-grade DRAM. DRAM expected to be 30–40% of all hyperscaler capex next year.

Speaker views:

• Gavin Baker: DRAM is the single most important AI bottleneck — more important than lasers, capacitors, NAND, or HDDs — and Elon is focusing Terrafab on memory for this reason; new Micron supply agreements with floor pricing above prior cycle peak gross margins are “potentially very transformational” for industry structure; Micron’s new supply deals covering ~50% of revenue with just four customers lock in above-peak pricing.
• Travis Kalanick: Distributed AI training is severely limited by physical proximity requirements — even GPU clusters 2km apart on private fiber suffer major efficiency drops; compute co-location is essential; distributed inference is more viable but latency still matters.
• Chimath Palihapitiya: Contested data center permitting — affecting ~40% of projects since 2021 and rising — combined with inflationary DRAM and power costs ($35B silicon + $25B power/cooling per GW) makes terrestrial buildout increasingly expensive and pulls forward the economics of orbital compute.

Potential impact: DRAM scarcity driving consumer electronics inflation (Apple, Xbox, PlayStation); long-term, only relief is new capacity from Micron, SK Hynix, Samsung (2–3 year ramp), or disruptive Chinese DRAM from CXMT for consumer grade.

AI IPO Wave: Anthropic Valuation, SpaceX Liquidity, Cerebras Stumble

Summary: Gavin valued Anthropic at approximately $3 trillion as a public company, projecting revenue well above $100B exiting 2026 and 85% gross margins on inference at scale. The hosts discussed how global capital markets absorb up to $6 trillion in combined new offerings (SpaceX, Anthropic, OpenAI, Cerebras).

Speaker views:

• Gavin Baker: Anthropic would trade at roughly $3 trillion as a public company today; market absorption is not a concern because these offerings simply shift capital from private to public markets; SpaceX employee base has already had liquidity via tender offers every 6 months for 10 years, so the lockup-driven supply overhang may be smaller than some shorts expect; Cerebras broke deal price due to execution of a simple playbook by shorts, not fundamentals — key metric is MW ramp.
• Jason Calacanis: Questioned whether retail or crypto-rotation capital is the marginal buyer for $4–6 trillion of new offerings; praised SpaceX’s IPO pricing and auction mechanics.

Potential impact: Gavin’s $3T Anthropic valuation would make it the second-largest company in the world if public today. Cerebras deal-price break creates a cautionary tale for future AI IPO pricing.

SpaceX Megapod, Orbital Compute, and Modular Data Centers

Summary: A trademark for “Megapod” filed June 18, 2026 describes self-contained modular AI computing hardware; speculation focuses on Tesla deploying GPU-laden pods at Supercharger sites. Panel discussed the economics of orbital compute (Starship-based) vs. terrestrial data centers.

Speaker views:

• Gavin Baker: When Starship is rapidly reusable, orbital compute will cost roughly $40B all-in per GW ($35B silicon plus $5B launch) versus $60B terrestrially today and rising; if the $25B terrestrial cooling/power cost is structurally inflationary while Starship launch costs are deflationary, the orbital economics crossover may arrive within 3–4 years.
• Chimath Palihapitiya: Modular deployable compute (shipping-container system on a concrete pad with prefab racks) can achieve 90-day build cycles; air-cooled legacy chips could be repurposed in these pods; looking at putting compute into his own restaurant kitchens for Adams.
• Travis Kalanick: Distributed training requires near-physical co-location to avoid dramatic efficiency drops; distributed inference is viable but still latency-constrained; true distributed training pools (like Bittensor) work for inference not training; modular single-operator clusters are viable.

Potential impact: Tesla Megapod product at Supercharger sites could accelerate distributed AI infrastructure deployment. Orbital compute remains speculative but Gavin Baker sees credible economics by 2029–2030.