Ep. 022 - Market Drawdown, Historic Bubbles, Funding The Buildout, AI Politics (Doug is Back)
Summary
The memory trade is unwinding from euphoria, not necessarily from broken fundamentals. Doug says DRAM and NAND prices roughly tripled over the past year but may rise only 30–50% next year; that slowing rate of change, SK hynix’s earnings miss after shifting more volume into LTAs, and widespread leverage helped turn the KOSPI’s 40% fall into forced selling. “What goes up really, really fast often has a little bit of gravity.”
A 40% correction need not end the cycle. Doug compares Korea’s behavior—not its magnitude—with Taiwan’s late-1980s bubble, where banks reached 500 times earnings and the market endured two 40% retracements before continuing higher. His qualified call: “I’m not saying it’s over. I don’t think it’s over,” but a boom-bust this violent usually takes time to repair.
Chinese memory may compress margins before it eliminates scarcity. CXMT is incentivized to maximize provincial output rather than shareholder returns and could “ruin the party” by accepting roughly 10% gross margins, but Doug still sees more demand than supply. Apple turning to CXMT while accusing Micron of price gouging earns his verdict: “No crying in the casino, Apple.”
AI demand is the trillion-dollar unknown. Doug’s bear case is that chips, utilization, inference, and models improve faster than useful workloads expand—the fiber-boom pattern, where one strand eventually became 500,000 times more performant. The host counters that models grow, think longer, attract more users, and unlock new markets: SemiAnalysis went from roughly nine to 90 coding-agent users, then perhaps 10× usage per person, producing 100× AI spend.
H100 obsolescence produced the sharpest disagreement. Doug expects sufficiently large models to make old accelerators uneconomic and says a B200/B300 pricing divergence would confirm it; the host argues Hopper facilities cannot simply be swapped into B300, GB300, or Rubin because power, cooling, permits, and layouts differ. “We live in a world with friction,” and that friction can preserve older hardware value.
The buildout could outrun its financing even if AI ultimately works. Nate Silver, as labeled in the transcript, estimates roughly $150 billion of ecosystem ARR against $1 trillion of capex already committed; at a hypothetical 50% profit margin, that is only a 7.5% return. Revenue reaching $500 billion may support $2–3 trillion of investment, but $5 trillion against $500 billion of revenue is “a house that you cannot pay for.”
Capital, electricians, and politics may bind before chips do. Nate Silver cites a roughly 100,000-person US electrician gap and about $450 billion of hyperscaler debt raised year-to-date; doubling again could demand higher yields and crowd out other borrowers. Nate Silver and the host also discuss AI becoming a cost-of-living scapegoat in the midterms, creating regulatory risk even if voters do not treat AI itself as a top-three issue.
Deep dive
1. Korea’s memory boom hit leverage before it hit a fundamental wall
Doug dates the setup to the end of Q2, which he calls effectively the best semiconductor performance in history. The subsequent unwind may look “technical”—various factors, excess leverage, and forced selling—but his simpler explanation is that “what goes up really, really fast often has a little bit of gravity.”
The speculative behavior in Korea echoed earlier Asian bubbles: investors borrowed aggressively and in some cases took second mortgages to buy stocks. Doug’s deliberately harsh history lesson is that Korean investors bought banks in 2007, CMBS in 2007, and SaaS in 2021 before going “ultra mega yolo into itself.”
SK hynix missed consensus, partly because it shifted more volume into long-term agreements, making future price increases more conservative than the market’s euphoric expectations. DRAM and NAND may have tripled over the prior year, but another tripling was not expected; even 30–50% price growth looks disappointing when “finance brains are just absolutely broken. It’s all about rate of change.”
Doug says the KOSPI was down 40%; anyone above 2× leverage could therefore be wiped out, creating margin calls and reflexive selling. Yet he expects markets to overshoot in both directions: “Things are never as bad as feared, and they’re never as good as you think they will be.”
2. Taiwan’s historic bubble leaves room for another rally—and a warning
The comparison is about behavior, not equal magnitude. Doug says late-1980s Taiwan produced roughly a 100× per-capita bubble, including a bank at 500 times earnings; recreating it would require Korea’s market to become perhaps 15–20 times more valuable.
That Taiwanese bubble nevertheless contained two 40% retracements before moving higher. Doug therefore refuses to declare the current memory cycle finished, while warning that a leveraged boom-bust “usually takes a little bit of time to retrace out.”
The host presses the bullish fundamentals: LTAs remain in place, businesses are sold out for years, and new production takes time. Doug agrees there may be smart money waiting, but argues that lower future price increases also deserve a lower capitalization multiple—even if this cycle proves “bigger and longer and stronger” than prior ones.
3. Chinese memory threatens pricing, while the real demand curve remains invisible
Doug’s historical rule is that “everything they touch gets dumped” because Chinese producers can prioritize volume over returns. His framing is provincial competition for GDP: the government is effectively the shareholder, so CXMT can accept around 10% gross margins rather than protect shareholder returns.
CXMT is still, in Doug’s view, clearly number four, entering during a shortage and therefore likely to make substantial money first. Apple is reportedly using CXMT after accusing Micron of price gouging, but Doug dismisses the complaint: “The reality is you have to buy it at the market price.”
The supply curve is comparatively knowable; the demand curve is not. Coding agents and knowledge work clearly add consumption, but the market cannot yet tell whether the increase is 50%, 100×, or something in between or larger. Supply will keep “ramping blindly into this curve” until it finally crosses demand.
Shortages corrupt the signal because a gigawatt-cluster builder may double-order equipment and triple-order memory, knowing excess parts can be resold. Factories interpret those orders as durable demand, add capacity over several years, then face a demand slowdown; utilization can fall from 100% to 50%, forcing price cuts—the semiconductor bullwhip in miniature.
4. AI usage could outrun efficiency—or repeat the fiber glut
Doug’s strongest bear case is useful-market saturation: perhaps Kimi K3 becomes good enough for data entry while chips, software, and models keep improving. The internet carried similar claims that demand doubled every 90 days, yet one fiber strand eventually became 500,000 times more performant and capacity could no longer be filled.
The host’s rebuttal is that four inflationary forces offset deflation: models get bigger, think more, reach more people, and are used more per person. Kimi K3 reportedly tripled in size and fits on a single B300 or MI355X node, but cannot run that way on Hopper.
The host also sees 100–1,000 times more potential users plus workloads extending beyond chat: coding, research, video, images, drug discovery, and material science. His best example is that few startups are pursuing weather prediction because everyone is currently distracted by coding—not because the opportunity set ends there.
Dylan says the breakthrough in Claude 4.5 changed his mind by crossing a capability threshold: work impossible the day before suddenly became feasible. SemiAnalysis then moved from fewer than 10 coding-agent users to around 90. Doug confirms average usage is higher but is unsure whether tokens rose 10× because his initial usage included intense 14-hour sessions; Dylan argues spend rose 10× per person through sub-agents and related tooling, taking company AI spend up 100× or more.
5. Model growth does not settle what old GPUs are worth
Doug speculates that old chips eventually become uneconomic: if inference requires 100 H100s, operators may prefer one newer system and “just let the old girl go.” He would look for a pricing divergence between B200 and B300 as confirmation.
The host “completely” disagrees because almost nobody can rip H100s out and replace them directly with B300, GB300, or Rubin. Hopper data centers have different power and physical designs; retiring them requires demand to fall below the operating cost of running them, or such extreme scarcity of permitted power and land that owners demolish functioning facilities.
Doug concedes the friction, while the host reduces the broad trade to two cases: stalled model progress pressures GPU prices, while continued progress supports them. Government restriction is the X factor—limiting access to frontier systems could reduce aggregate demand and particularly hurt older GPUs—though Doug declines the scenario as containing “too many what ifs.”
6. AI politics will likely arrive disguised as cost-of-living politics
Nate Silver, as labeled in the transcript, guesses AI will be a top-five but not top-three issue for most midterm candidates: widely mentioned, rarely a platform’s centerpiece. When an issue lacks that priority, he expects corporate lobbying to dominate.
The host points to the ROSA bill, which he tentatively describes as the Remote Access Security Act. It reportedly passed the House roughly 300–20 but remained stuck in the Senate amid lobbying—an example of how a fourth- or fifth-ranked public concern can lose institutionally.
Their shared political concern is scapegoating. Nate Silver expects cost of living to be voters’ number-one issue, with AI and “tech bros” blamed for inflation, housing, employment, healthcare, or climate pressures: not a referendum on whether America should sponsor AI, but AI as a “whipping boy” for existing grievances.
7. Cash-flow timing, debt capacity, and labor narrow the buildout path
Nate Silver’s central financing warning is that the future can arrive too late. A technology boom may ultimately justify its vision, yet investors can spend $1 trillion to reach $100 billion of revenue and wait five more years for that figure to become $1 trillion: “You built a house that you cannot pay for.”
His rough ledger is $150 billion of ecosystem ARR against $1 trillion of capex. At 50% profit, the implied return is around 7.5%; growing ARR to $500 billion could probably carry $2–3 trillion of investment, but a later doubling toward $5 trillion makes the path much narrower and demands adoption “like yesterday.”
The host sees enterprise, banks, telcos, defense, intelligence, and government agencies as more important than getting every grandmother to subscribe. Nate Silver’s concern is operational: chips alone do not create deployments, and the US already has a roughly 100,000-electrician gap; training takes perhaps 18 months, while journeyman-level electricians can make $250,000 and potentially $400,000–$500,000 on extreme hours.
Capital has its own supply curve. Nate Silver puts hyperscaler debt issuance near $450 billion year-to-date; more borrowing may require 7–8% yields and push up mortgage rates compared with hypothetical mortgage borrowing near 5%. Life insurers and annuities are important funding sources, but pensions are secular decliners and “everyone needs twice as much insurance” is not a scalable financing thesis.
8. Data centers spread the boom locally, but permitting can reverse it
Taiwan illustrates the physical limit: Dylan says TSMC and its supporting ecosystem may represent roughly 20% of the economy, while Taiwan’s GDP rose about 25% as chip output surged. The host also notes that TSMC directly employs fewer than 100,000 people in a population above 20 million and can expand abroad; Nate Silver answers, “No electricians in Arizona, bro.”
Their oil-boom analogy cuts both ways. Data centers can create construction and service economies in remote places because energy availability and permits—not natural deposits—determine location; a permitting or political shift can therefore strand a community as quickly as a commodity bust.
Nate Silver cites a poll suggesting that people who dislike data centers often do not live near one, while communities—especially younger people—can become more favorable after jobs arrive. Doug adds that construction work is accessible and honest employment, and uses a tentative estimate of 10,000 jobs per gigawatt: 70 gigawatts could mean roughly 700,000 jobs. Unlike an app built in San Francisco, a $10 billion information factory in a rural area creates “broad participation” and more sustainable investment.