E278
Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter
2026-06-26 spoken.md · speaker-labeled ▶ watch ← E277 all episodes E279 →
1
ideas born
6
ideas moved
7
captures · 2 voices
1
dissenting
+114.2
conviction added
-34.2
decay · 56 silent
Every number here is replayed from score_events — the same ledger the pool ranks on. Decay is what the 56 ideas nobody mentioned gave up this week; it applies only when an episode is processed.
Tier crossings
conviction thresholds crossed by this episode — 65 / 45 / 15 · ideas born here show where they landed. why 65 isn't always green
Kill dates that landed since E277
0 hit · 0 partial · 5 miss — windows that closed after 2026-06-19 and up to 2026-06-26, auto-scored against price data and never hand-set. verdict · R · α
| idea | verdict | R | α | closed |
|---|---|---|---|---|
| 🏛️ Taiwan's US protection expires as fabs onshore — TSMC carries the risk | MISS | -359.4% | -435.9 | 2026-06-24 |
| 📈 AI adoption disperses the S&P — short the index, own the category killers | MISS | -19.1% | -44.9 | 2026-06-21 |
| 📈 Apple has completed the transition from growth company to cash cow and loses the AI cycle | MISS | -48.5% | -74.3 | 2026-06-21 |
| 🤖 Consumption-based software pricing destroys the business — Snowflake is the case study | MISS | -8.7% | -34.5 | 2026-06-21 |
| 📈 Generative video (Veo 3) collapses Hollywood inside a year | MISS | -145.3% | -171.1 | 2026-06-21 |
Who moved the board
each voice's force on conviction this episode — supports and opposes, weighted exactly as the replay applied them · share = % of this episode's movement
+113.0 / -11.9
→ net +101.1
What got argued (6 ideas)
ordered by how hard each idea moved · quotes are verbatim from the transcript, timestamps deep-link into the episode
DRAM — specifically HBM/SOCAM/LPDDR for AI servers — is THE binding constraint of the AI buildout: only three makers (Micron, SK Hynix, Samsung), 2026 supply sold out, new supply-chain agreements with price FLOORS above prior-cycle-peak gross margins (~50% of Micron revenue), DRAM heading to 30-40% of all hyperscaler capex, and the stocks still trade cross-sectionally cheap vs the rest of AI. Risks: CXMT floods consumer-grade DRAM from China; Elon's TerraFab adds supply in years, not quarters.
plays
MU ·primary
SMH
evals 2027-12-26
memory capacity and bandwidth are foundational to the performance of every AI model. ... this bottleneck is going to be with us for a while. ... the floor pricing in these new contracts is ahead of prior cycle peaks from a gross margin perspective. ... DRAM is probably going to be 30 to 40 percent of all hyperscaler capex next year.
to stand up a one gigawatt data center, it's $35 billion in semiconductors and ... $25 billion of power and cooling equipment, and that is clearly inflationary ... When Starship is reusable, it's going to cost $5 billion to put a gigawatt of compute into space. ... you're $40 billion to put the gig into space. ... everybody who's a SpaceX investor or employee has had a chance to sell every six months for the last 10 years. ... Almost half the employees at SpaceX bought on the IPO.
There is very, very meager supply and there's effectively infinite demand. So that probably pulls forward the economic equation to want to go to space. But then again, that's going to prefer SpaceX and their compute stack and their compute decisions over the hyperscalers ... you're going to have a cost of an output token, I think terrestrially, particularly from the hyperscalers, be a little economically lopsided versus SpaceX once they get it to scale.
you can lift H100s, A100s out of some old data center, put ... a Grok or a Cerebras in front of it, and you can get a very competitive solution. ... we're going to be using GPUs for seven years, ten years, twelve years, and that's great because it lowers the cost of finance them, which makes this AI revolution more financeable.
▲ GOOGL
🤖 Anthropic IPO (~Oct 2026) is attractively priced at $1.5-2T
watch 54
CONTESTED
▲ +15.1
42.7 → 57.8
I think Anthropic is worth $3 trillion today. ... that is roughly where it would probably trade as a public company. ... they're going to end this year well over $100 billion. ... It will be very profitable at that scale because it'll be inference dominated and people are reporting they have 85 percent gross margins on inference.
▲ MSFT
🤖 Frontier AI labs keep durable premium pricing (not commoditized)
ember 27
CONTESTED
▼ -11.9
11.9 → 0.0
how good GLM 5.2 is, has challenged some of my beliefs. ... I profoundly believe the future is composable models. ... half the queries are going to go to the open source model, maybe 85%, ... shift economic value from the margins of the frontier labs to the infrastructure. ... To date, frontier tokens are capturing 90% of the economic value. And open source tokens are probably 80% plus of tokens processed.
here's a shipping container that you can just drop on a concrete pad somewhere, plug in the power and let it rip. ... if Tesla can make them available, folks like us for my data center project, we would be enormous buyers of these things ... you have like a 90-day build cycle, which is unheard of.
Episode digest
written during extraction and stored in data/extractions/ep278.json — the auditable source of truth, including everything market-adjacent that did not earn a capture
Friedberg out ('mental health day after the socialist sweep'); Travis Kalanick + Gavin Baker guest (both now multi-episode regulars at full weight). Note: Sacks referred to as FORMERLY AI czar, now running PCAST. LABEL DEFECT (new pattern for the catalog — Chamath→Jason merge): from ~1:13:22 through ~1:21:40 multiple Chamath turns are labeled 'Jason Calacanis' — receipts: 'hundreds of megawatts to gigawatts... those are just lock them down, WHICH I OWN' (his Arizona sites), 'for MY data center project, we would be enormous buyers', basis-of-design/BOD expertise, liquidated-damages/SLA talk, and one turn addressing 'Jason, I think you've talked about BitTensor' (speaker can't be Jason). Both Chamath captures from this span rest on those receipts. One residual ambiguous turn at 1:25:22 ('I don't have any investments in the distributed compute space') conflicts with Jason's known TAO position — left uncaptured. SOCIALISTS SWEEP NYC: Mamdani-endorsed candidates go 3-for-3 in Dem primaries (Lander takes NY-10 beating pro-Israel Goldman — Sacks reads it as a straight Israel referendum, 80% of Dems now disapprove; Chevalier, 'end Western civilization' quotes, takes NY-13 from a Jeffries ally; Valdez wins the 'Commie Corridor'); Polymarket had the trifecta at 26%; DSA platform (abolish Senate/ICE/carceral state, public ownership of major corporations) + 'ballot access vehicle' quote; Chamath's social-media-ban-under-16 thesis (Canada/UK/Australia as socialism casualties; dates the radicalization pipeline) vs Travis' counter (age-gating = de-anonymization = censorship infrastructure — 'criminalizing disagreeing with them'); Gavin: DSA voter base is downwardly-mobile educated whites, Mamdani is 'one of the most talented politicians I've ever seen', NGO complex as organized corruption, 'if they get total control of these cities and drive out everyone productive, I don't know how you come back' (soft state-muni-stress color, no capture — politics); Chamath: key congressional races were AI referendums — Anthropic-funded anti-AI groups vs OpenAI-funded pro-AI groups, pro-AI won 'just barely'. CHINA CATCHES UP: z.ai's GLM 5.2 — open-weight (MIT license), 744B params, 1M context, 51 on Artificial Analysis (best open-weight ever), within 1pp of Opus 4.8 on SWE-bench at 85% lower cost, founder promises Fable-level open weights before Q1 2027; Gavin explains industrial-scale distillation (masked-account API farms harvesting reasoning traces) but concedes 'the cap may be out of the bag' — his composable-models thesis captured as frontier-premium oppose (router → 85% open-source queries, frontier as conductor); claimed trained entirely on Huawei Ascend 910B (Sacks: indigenization push, 'AI in a box' exported at a fraction of the cost, pro-export urgency, 'we're on a shot clock', cyber white-hat upgrade cycle now); Sacks on Gavin's is-this-what-Dario-wanted: 'on a certain level it is what they wanted... hoisted on his own petard or a FAFO situation' — wants Fable back on market once the jailbreak is fixed (softening, no capture); Fable still in purgatory, GPT-5.6 navigating new approval hoops; Gavin's aside 'NVIDIA is the American open source champion — they can release GLM 5.2 or better whenever they want' + OpenAI's Broadcom-built 'jalapeno' chip → Jason predicts NVIDIA could fund an OpenAI competitor (digest; ASIC-vs-NVIDIA incentive warning from Gavin). MICRON BLOWOUT births ai-memory-supercycle (captured, Gavin str-3): rev 4x y/y $9B→$42B, +16% beat, Q4 guide $50B vs $43B, 2026 supply sold out, stock 10x (Gavin's 2025 prediction-show HBM call now +14x on MU — scoreboard-relevant receipt for his tech cell); SCAs with price floors above prior-cycle peak gross margins covering ~50% of revenue; DRAM→30-40% of hyperscaler capex; Elon pointing TerraFab at memory (Intel partnership; 2-3.5yr fab timelines he may compress); CXMT IPO = Chinese consumer-DRAM flood risk (and 'the cure for Apple's ills'); NY Micron plant killed same-day by environmental issue → 'build fabs in Texas'. AI INFLATION hits consumer hardware: Apple raises prices (MacBook Neo +14%, Mac Studio +25%), Xbox up, Switch/PlayStation next — demand destruction in consumer while AI demand is price-insensitive (apple-ai-dark-horse cost headwind noted, no clean capture — Jason absent from the exchange). Gavin's gigawatt math captured on spacex (space $40B vs terrestrial $60B and diverging; racks-in-space-linked-with-lasers; contested-datacenter share >40% since 2021 per Chamath) + his anti-lockup-short case; SpaceX retreated from ~$200 highs. Tesla MEGAPOD trademark (filed 6-18): modular AI-compute-in-a-container, rumored for supercharger sites w/ existing power — Chamath's enormous-buyer receipt captured on tesla-second-act; Powerwall-GPU/Starlink distributed-inference riff + Elon-buys-T-Mobile rumor (digest); Travis exploring GPUs in his 500 CloudKitchens properties (natgas on site; the 2km fiber latency kill for distributed TRAINING vs Gavin's distributed-INFERENCE counter — captured on decentralized-ai-compute). Cerebras broke IPO deal price post-first-quarter → Gavin's price-insensitive-selling mechanics + how he'd tell the story (OpenAI $20-25B Dec-2025 contract shows up ~Labor Day; watch megawatts-brought-online, only CoreWeave/Crusoe/SpaceX-AI have done >1GW outside hyperscalers); Gavin's ANTHROPIC $3T call captured (str-3: >$100B revenue exiting 2026, 85% inference GM, 'the market's already absorbed it — it's just shifting from private to public'); Uber $17B Series-whatever Dutch-auction nostalgia w/ Travis (digest). Conflicts noted: Gavin long SpaceX (Atreides), Chamath owns gigawatt sites + Grok(-into-NVIDIA) exposure, Travis building Adams on physical AI.