Doug O'Laughlin Claude Code & Memory Shortage Investment Thesis

Source: Claude Code for Finance + The Global Memory Shortage: Doug O'Laughlin, SemiAnalysis, Latent Space (Swyx), February 24, 2026.

The Framework: Demand Outruns a Post–Moore's Law Supply Chain

Doug O'Laughlin's lens combines two converging forces: agentic AI collapsing the cost of information work (Claude Code and peers) and physical supply chains that cannot scale at software speed (memory, fabs, debt markets). The organizing equation is supply divided by demand — demand from scaling laws and agent adoption is exploding while semiconductor supply growth has structurally slowed.

Layer	Doug's View	Investable Read
Agentic software	Claude Code crossed a capability threshold; all information work is downstream	Anthropic ecosystem; horizontal software losers
Memory (HBM → DRAM)	3:1 to 4:1 wafer trade-off; worst shortage since the 1990s	HBM vendors, semicap, memory capacity builders
Merchant silicon	Google TPU at peak TCO gap now; window closes at Rubin/HBM4	Google TPU external sales; Nvidia supply-chain lock
CapEx funding	AI buildout = railroad-scale (~5% GDP); debt absorption is a bottleneck	Self-funded hyperscalers vs over-levered builders
Horizontal software	Excel/Bloomberg are human IDEs — structurally obsolete	Microsoft, Salesforce as disruption targets

Investment Thesis #1: Memory Is the Most Obvious Leg of the AI Trade Right Now

Doug argues the memory market just exited the worst NAND/DRAM shortage in decades — comparable to the 1990s — and immediately faces a new wave of demand that evaporates the middle layer of DRAM because HBM and KV-cache offload consume wafer capacity at a 3:1 to 4:1 ratio.

"We could see DRAM prices go up 100% again... The most obvious logical leg of the AI trade is effectively investing in memory capacity."

No one invested in clean rooms or capital equipment during the 2021–2024 downturn; lead times are ~two years. Hyperscalers may eventually play chicken and pull orders forward, but aggregate power-purchased demand still dwarfs supply. CXL — previously dead technology — is coming back as operators attach legacy DDR4 into racks to stretch capacity.

Trigger: DRAM spot price moves; SK Hynix/Micron/Samsung HBM ASP guidance; consumer device BOM inflation (iPhone memory costs); CXL adoption in cloud racks.

Names: Micron (MU), SK Hynix, Samsung — HBM and DRAM pricing power; Applied Materials (AMAT), Lam Research (LRCX) — semicap tied to memory fab capacity expansion.

Investment Thesis #2: Context Rationing Locks In the Memory Bottleneck for Years

Doug pushes back hard on the consensus that context windows will scale to 100 million or trillion tokens. Physical memory constraints — not algorithms — cap the curve.

"We've been effectively stuck at a million for 2 years now... This is it for like 5 years, 10 years pretty much."

Long-context models in papers "drop off" — they do not use full attention across all tokens. The practical implication: premium 1M-token context becomes a mansion; everything else gets rationed. Doug coined "context rationing" — users may soon face daily context budgets or vouchers as DRAM scarcity bites. That structurally sustains HBM demand even if algorithmic efficiency improves.

Trigger: Model providers introducing tiered context pricing; stagnation in advertised max context beyond ~1M; Apple or hyperscaler commentary on memory-driven COGS increases.

Names: Same memory stack (MU, SK Hynix, Samsung) — physical constraint beneficiaries regardless of which lab wins on algorithms.

Investment Thesis #3: Claude Code Is a Demand Shock for AI Compute and a Death Sentence for Human IDEs

Doug had his "awakening" around December 27, 2025, when Opus 4.5 started oneshotting projects that previously required iterative feedback. He now tracks Claude-signed GitHub commits and estimates ~4–5% of GitHub commits are Claude Code — growing at a rate that puts 25–50% within his 95% confidence interval (he publicly sandbagged at 25%).

"Coding is a big subset of all information work... They're telling me the hard one got automated. Why can't the easy one get automated?"

He spends $20,000–$30,000+ annually on API access and argues GDPval results — models consistently outperforming industry experts on white-collar tasks — are his definition of AGI for information work. Excel, Bloomberg, and PowerPoint are "human IDEs" that agents with Python and verifiable data APIs will replace. SemiAnalysis is actively trying to "rip Bloomberg out" for API + Claude Code workflows.

Trigger: Anthropic revenue/ARR disclosures; Claude Code commit-share continuing exponential growth; financial/data terminal churn among research firms.

Names: Anthropic (private) — direct beneficiary of Claude Code adoption; Nvidia (NVDA) — every agentic workflow still runs on GPUs at scale.

Investment Thesis #4: Microsoft Has the Most to Lose — Horizontal Software Meets the Barbarians

Doug's spiciest software call: Microsoft faces the largest disruption surface of any company because Excel, PowerPoint, and email are the general-purpose IDEs for information work — and Claude Code is building the replacement in real time.

"I cannot paint a bigger target... They're the horizontal software company that humans use their software to do information work."

The Azure–OpenAI partnership is "renting barbarians at the gate" — each year the barbarians grow stronger while Microsoft's moat dilutes. Satya Nadella is managing for cash flow and ROIC, not betting the company on AGI. On earnings calls, Microsoft signaled it could grow Azure faster but is reinvesting in internal capabilities instead — pulling back GPU rentals to defend Office. Ex-Azure, Doug thinks Microsoft would trade at 8–10x earnings. Microsoft should have built "Claude for Excel" itself; Anthropic is doing it instead.

Trigger: Claude Cowork / Claude for Excel adoption metrics; Microsoft Azure growth deceleration vs Google/Meta capex acceleration; Office 365 seat growth stalling.

Names (bearish — Key Risks only): Microsoft (MSFT), Salesforce (CRM) — same horizontal-software exposure.

Investment Thesis #5: Google's TPU Window Is Open Now — and Nvidia's Supply Chain Lock Is the Counterweight

Doug agrees with SemiAnalysis's collective view that Ironwood TPU v7 is at peak TCO advantage vs Nvidia — the widest gap before Rubin closes it with HBM4 vs HBM3. Google is selling TPUs externally (Anthropic) to gain merchant-silicon market share while the aperture is open; internal TPU value could approach ~$1 trillion at ~30% merchant share.

"If you want market share, now's the time... TSMC is the biggest blocker."

Nvidia's counter-moat: Jensen Huang personally secures HBM and fab allocation (Doug cites Jensen drinking with Samsung's chairman in Asia). Google/Sergey are not playing the same supply-chain relationship game. V7 silicon was designed 3–4 years ago; talent dispersion from the original TPU team is a long-term risk. The TPU story is good for one to two years, then Rubin narrows the gap.

Trigger: Anthropic TPU deployment scale; Google external TPU revenue; Nvidia Rubin HBM4 ramp timing; TSMC allocation shifts between Google and Nvidia.

Names: Alphabet (GOOG) — TPU merchant window + consumer AI distribution; Nvidia (NVDA) — supply-chain relationships and Rubin roadmap.

Investment Thesis #6: AI CapEx Has Reached Railroad Scale — Funding Mechanics Matter

Doug draws a direct parallel to the Gilded Age railroad buildout: ~4.8% of GDP and ~25% of gross fixed capital formation — levels the current AI capex wave is already matching (Stargate alone ~2% of US GDP). Railroad cycles had three boom-bust waves over 45 years; AI cycles will compress faster but will not be one instantaneous up-down.

Oracle's aggressive GPU buildout and debt issuance is Doug's cautionary tale: hyperscalers historically self-funded, but Oracle issued so much high-yield debt it moved the entire TMT debt index — a bottleneck he never expected ("supply of debt into the market"). Microsoft could have internally funded equivalent buildout at near-sovereign borrowing costs but chose not to.

Trigger: Oracle CDS spreads and debt rollover; hyperscaler free-cash-flow approaching zero (Google, Meta); high-yield TMT index supply gluts.

Names: Oracle (ORCL) — execution and funding risk on aggressive buildout (watch for distress, not endorsement).

The Ecosystem Map

Doug's personal AI stack:

Primary tool: Claude Code (Opus 4.6) for coding-shaped information work
Also tracks: Codex 5.3 (coding-pilled, weaker for general research), Kimi 2.5 agent swarms (actually improves performance)
Skeptical of: Claude agent teams (no RL, "very bad"), OpenClaw/Clawdbot (security), most merchant AI accelerators pre-Groq/Cerebras

SemiAnalysis research focus:

Memory mania post (HBM:DRAM trade ratio, CXL revival)
Claude Code GitHub commit tracking
TPU v7 external sales / Anthropic partnership
CPU shortage from deferred cloud refresh + RL environments

Companies Doug cites as supply-chain winners:

Nvidia (Jensen's fab/HBM relationships)
Memory vendors + semicap for capacity investment
Google (TPU window)

Companies Doug cites as structurally threatened:

Microsoft, Salesforce (horizontal software)
Apple (memory BOM inflation; "buy your iPhone now")
Oracle (funding execution)

Key Risks

Memory demand destruction: Hyperscalers could delay orders if power constraints bind first — historically the mechanism that crashes memory prices.
Microsoft reverses course: A full AGI-pilled capex pivot (like peers) could re-rate Azure and internal AI investment.
Claude Code adoption stalls: If commit-share plateaus below 25%, the information-work disruption thesis slows.
TPU window closes faster: Rubin/HBM4 ramp or TSMC reallocation could narrow Google's TCO lead sooner than 1–2 years.
Oracle funding crisis: Debt-market indigestion could spill into broader AI infrastructure financing sentiment.
Agent slop risk: Doug stresses Claude Code "makes mistakes all the time" — expert hygiene required; firms that skip human review will produce garbage research.
Meta-level learning absent: Doug does not believe models yet develop expert judgment; if trillion-dollar AI spend fails to produce it, enterprise ROI narratives break.
Consumer electronics demand collapse: Smartphone volumes could fall from ~1.4B toward 500–600M as memory is rationed to AI — hurts Apple and Android OEMs.

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	Micron (MU) / SK Hynix / Samsung	DRAM + HBM shortage; 4:1 wafer trade-off; prices could rise 100% again	HBM ASP guidance; DRAM spot; iPhone BOM reports
1	Nvidia (NVDA)	Supply-chain lock on HBM/fab; agentic demand multiplies GPU need	Allocation priority vs Google; Rubin HBM4 timeline
2	Applied Materials (AMAT) / Lam Research (LRCX)	Memory capacity investment is the bottleneck; 2-year clean-room lead times	Memory fab capex orders; semi cap equipment backlog
2	Alphabet (GOOG)	TPU v7 at peak TCO gap; external sales to Anthropic; ~$1T merchant value potential	Anthropic TPU deployment; Google Cloud TPU revenue
3	Anthropic (private)	Claude Code commit-share growing exponentially; GDPval-level white-collar capability	GitHub commit attribution; enterprise ARR reports
4	CXL ecosystem	Dead tech revived to pool legacy DDR4 as memory crisis deepens	Cloud operator CXL rack deployments

Monitoring Checklist

Bottom Line

Memory is the trade everyone can see but hasn't fully sized: Doug thinks DRAM could rip another 100% because HBM's 4:1 wafer trade-off meets zero new clean-room capacity — buy the memory stack and semicap, not the narrative.
Claude Code is not a coding tool — it is an information-work platform: Doug's December 2025 awakening and GitHub commit data imply 25–50% of code generation migrates to agents; Bloomberg and Excel are the next casualties.
Microsoft is the biggest single-name loser in the agent era: Horizontal software plus Azure barbarians-at-the-gate is an innovator's dilemma Satya is managing for cash flow, not survival.
Google's TPU merchant window is open for ~1–2 years: Sell TPUs externally now at peak TCO advantage before Rubin and HBM4 close the gap — but TSMC allocation and Nvidia's relationship moat cap volume.
Context rationing, not context expansion, is the next UX paradigm: Physical memory limits cap context at ~1M tokens for 5–10 years — sustaining HBM demand regardless of algorithmic hype.