Claude Code for Finance + The Global Memory Shortage: Doug O'Laughlin, SemiAnalysis
Summary
Doug O’Laughlin’s core call is that Claude Code crossed from novelty to economically useful information worker around Opus 4.5. It can compress what looked like a “PhD project” into a day or two, yet “this crap makes mistakes all the time”; today it resembles a junior analyst, not an autonomous expert. Shawn compared the paid tool to a perfectly compliant junior analyst. The immediate payoff accrues to experienced reviewers who recognize the “artisanal last 5%,” while junior data-analysis roles look increasingly exposed.
Claude Code’s adoption curve suggests coding agents are becoming the interface for all information work, not merely programming. Doug described Claude-attributed commits reaching roughly 4% of public GitHub in “2 weeks or something”; Shawn said the updated chart was around 5%. Doug’s deliberately sandbagged year-end prediction was 25%, while Shawn said the current rate pointed closer to 50% and that 25% fit a 95% confidence interval. The broader claim is sharper: “Excel is the IDE for analysts,” and Excel, Bloomberg, PowerPoint, and other human-oriented interfaces are vulnerable once agents can retrieve trusted data and render the answer directly.
The most concrete hardware call is a memory shortage that Doug thinks cannot clear for roughly two years. One bit of HBM effectively consumes three to four bits of conventional DRAM capacity after process complexity and yield, just as suppliers emerge from a severe bust that froze clean-room investment. SemiAnalysis sees scope for DRAM prices to rise another 100%, forcing demand destruction, delaying data centers and devices, reviving old DDR4 through CXL, and making “context rationing” a plausible product reality.
Google’s TPU V7 has a temporary cost window, but Nvidia’s control of memory and the wider supply chain may close it with Rubin. Doug thinks nearly every major lab would consume more TPU V7 if supply were unconstrained, and the shared discussion put a possible TPU business near $1 trillion at roughly 30% share. Yet TPU V8’s HBM3 position versus Rubin’s HBM4, Nvidia’s aggressive supplier management, and Google’s limited available capacity make this an installed-base race rather than a clean architectural victory.
Microsoft has “the most to lose” because AI attacks the horizontal software where humans perform information work while Azure finances the attackers. Doug compared OpenAI to “barbarians at the gate”: Microsoft can remain a highly profitable compute supplier while its Office abstractions are disrupted, or redirect investment toward defending Office and building internal models. Oracle’s attempt to absorb the displaced buildout was not necessarily a bad underlying investment, in Doug’s view, but its huge, poorly staged debt issuance overwhelmed market liquidity and turned financing cadence into a bottleneck.
The AI buildout already resembles railroads more than the internet, implying multiple booms and busts rather than one clean cycle. Shawn cited railroad investment at about 4.8% of GDP and 25% of gross fixed capital investment, while also comparing Stargate with roughly 2% of U.S. GDP. Doug will not claim demand rises forever, but Claude Code changed his own elasticity: he would value access at $20,000–$30,000 a year or more because it behaves like many parallel junior analysts.
Doug’s semiconductor framework remains the anchor: Moore’s Law slowing while scaling-law demand accelerated transferred value to companies that could still deliver system-level performance. His 2020 conclusion that Nvidia was the primary beneficiary proved directionally right, though even he did not anticipate “the most valuable company in the world.” The durable investing lesson is not precision EPS maintenance; it is finding the one to three technological variables capable of moving billions in revenue, then using agents to investigate them without outsourcing judgment.
Deep dive
1. ASML turned a quality-focused investor into a semiconductor obsessive
Doug corrected the simplified origin story: no mentor “nerd-sniped” him into semiconductors. While looking for quality compounders in 2018, he found ASML, “fell in love with it,” then worked downstream through textbooks, manufacturing equipment, and the companies capable of producing each layer.
What held him was that semiconductor manufacturing felt like “all science fiction.” Even before intelligent conversational robots arrived, fabricating each new generation of chips required technologies that appeared impossible, yet investors treated those advances as routine features of a supposedly mature industry.
That fascination became a career decision. Doug had often spotted trends early, including becoming obsessed with TikTok in 2019, but ASML revealed “a really big wave” in which he had enough conviction to reorganize his life and go all in.
2. Moore’s Law’s slowdown overturned the mature-hardware playbook
Doug’s radicalizing belief was that Moore’s Law was ending. Industry primers still described semiconductors as consolidated, mature, and slow-growing, while the dominant investing playbook assumed CPUs would simply improve every generation and hardware would remain less valuable than software.
His causal chain was simple: scaling laws would drive compute demand sharply higher, while slowing transistor scaling would remove the automatic supply-side gains that previously made performance cheaper. When “all those free gains” disappear, expertise in chip, networking, packaging, and complete-system design gains pricing power.
Nvidia became his cleanest example because it mastered “every aspect of it from the chip to the networking to the design to the scale-up.” Parallel computing mattered, but the system-level competence mattered more once performance could no longer arrive automatically from the next CPU node.
3. The Nvidia thesis was right beyond even Doug’s conviction
In a 2020 piece about GPT-3 and “the writing on the wall,” Doug argued that scaling-law demand plus broken Moore’s Law should benefit semiconductors, with Nvidia “pretty much the only one” positioned to capture the upside. That remains his favorite long-range call.
His uncertainty is important: he believed the thesis deeply but did not foresee the magnitude. Being told Nvidia would become “the most valuable company in the world” would still have sounded implausible, even though he had written the premises himself.
Doug met Dylan Patel because Dylan seemed like “the only person who was as semiconductor-pilled in the entire world as me.” Their complementary perspectives—Dylan technology-first, Doug with more financial training—supported their early thesis and collaboration at SemiAnalysis.
4. SemiAnalysis focuses on inflections that can move billions
Doug’s critique of sell-side research is structural. The model descended from banks needing nominally independent analysts to help distribute securities, then evolved into “mechanical maintenance” around buy, hold, sell ratings and penny-level EPS estimates rather than differentiated technological research.
“No one’s saying, ‘My estimate is always one cent tighter than everyone else’s, and that’s why I’m good at stocks.’” A model that is 5% more accurate rarely determines the investment; identifying whether a product inflection happens at all can alter revenue by billions.
His example was AMD’s Helios rack: whether it is genuinely on time and ready to produce tokens at launch matters far more than another iteration of quarterly spreadsheet precision. The same applies to a networking bottleneck whose timing controls an entire deployment.
The analyst-PM conversation therefore reduces to finding the “one or like three things that actually matter.” The difficulty is spanning PhD-level depth across many supply-chain components while retaining enough width to see how one obscure component or optical technology changes the economics above it.
5. Opus 4.5 crossed Doug’s one-shot threshold
Doug had long used SemiAnalysis’s hiring case study as an informal benchmark: could an agent complete a multistep financial-analysis task that takes a human roughly 24 hours, and when would it beat the weakest applicants? Earlier systems showed promise but were not reliably there.
Opus 4 could build side projects, but it demanded heavy feedback and frequently broke. Doug also tried Codex earlier without getting the seamless agentic behavior he wanted; the qualitative break arrived in late December—roughly around December 20–27, with Doug saying it was after returning from Christmas—when Claude Code with Opus 4.5 began one-shotting complete tasks and small projects.
The important change was not perfect output. It could handle a dashboard, spreadsheet, or similar project, accept a request for improvement, and continue iteratively without forcing Doug to reconstruct everything after each turn.
That experience produced his current catchphrase: “You can just do things.” Once generalized projects could be completed from natural-language intent, every missing implementation became at least partly “a skill issue”—a question of framing, context, review, and tool use.
6. Portfolio analysis became an extensible judgment system
Doug began with a mundane request: ingest his positions and notes, organize them, calculate basic portfolio risks, and maintain the result. Once that worked, he asked Claude Code to encode his investment style, build a framework, grade holdings, and attack the underlying assumptions.
The value came from iterative extensibility. What started as copy-pasting notes became a reusable representation of how he thinks, capable of accepting new data, applying rubrics, comparing positions, and producing dashboards without a conventional development project.
Rubrics help manage stochastic behavior: specify the dimensions that matter, score each out of 10, and make the agent expose where its work is weak. Doug described both running the task with the rubric and evaluating it in a separate pass; the conversation suggested that a fresh evaluation context can reduce bias and sycophancy.
Opus 4.6’s tendency to agree makes that separation more useful. When creation and grading share one context, prior reasoning can contaminate the critique; a clean context reduces sycophancy and prevents the model from rationalizing decisions it already made.
7. Claude-attributed commits revealed an exponential adoption curve
Wanting evidence beyond online “psychosis,” Doug noticed that Claude Code could sign public commits. He asked the agent how to scrape the signature systematically, query the available GitHub data, calculate daily totals, and express them as a percentage of overall activity.
Doug described the result as roughly 4% in “2 weeks or something.” Shawn said the updated chart was around 5%. Doug said he had never seen an exponential trend remotely like it.
The chart itself was agent-produced—Doug cited Opus 4.5 or 4.6—not hand-built. He had also asked an agent to read roughly 70 visualization books, compress their useful lessons into a tiny style skill, and apply SemiAnalysis’s colors, formatting, and watermark.
His point was economic rather than aesthetic: “The cost of doing this is nothing.” If synthesizing 70 books costs effectively the same as three, the agent can absorb the larger reference set, retain only the rules that matter, and repeatedly generate charts from new data.
8. Code is becoming the language beneath all information work
Shawn observed that Anthropic’s production traffic still showed software engineering near 50%, but asked whether apparently separate categories such as data analysis were themselves becoming software engineering. Doug’s answer was effectively yes: code is where machines and the rest of the world currently interoperate.
Finance already uses abstractions. An Excel model encodes relationships and expresses what an asset may be worth; coding is arguably harder, Doug said, so “you’re telling me the hard one got automated—why can’t the easy one get automated?”
This is why Doug finds Claude for Excel worse than Claude Code using Python and then depositing output into Excel if required. Fitting an agent into a legacy workbook is “a car engine” forced into “a horse carriage”; machine-oriented representations should become primary, with human formats generated only at the edge.
9. Context hygiene beats elaborate agent scaffolding
Doug’s current setup emphasizes a small set of strong skills, API access, and internal SemiAnalysis data exposed through controlled services. At the beginning of a session, he specifies a concrete goal that should finish inside one context window, then lets focused subagents gather bounded pieces.
The new one-million-token context is a large improvement because project instructions consume a smaller percentage of the window and subagents can work in their own contexts. By contrast, repeated compaction begins “compression of the noise” and steadily degrades fidelity.
He underuses hooks by his own admission. Earlier enthusiasm for Ralph loops and Gas Town-style orchestration gave way to “less is more”: current models still lack enough fidelity for extravagant multistage automation, while compact skills and explicit context more reliably reach completion.
Context rot remains visible. The agent can become garbled, lazy, or forget an API already documented in its
CLAUDE.md; Doug invoked the Of Mice and Men meme in which an exhausted session must finally be “put down.”
10. Kimi K2.5’s swarm worked where Claude’s agent teams did not
Shawn’s controversial assessment was that Claude’s experimental agent-team feature had not received the reinforcement learning needed for coordination. Splitting a large, multi-company KPI dashboard across the team made performance “meaningfully worse,” because prompt-level delegation lacked situational awareness.
Subagents worked better because each received a cleaner bounded task and returned a result. Doug reported that Kimi K2.5’s swarm was actually good; Shawn said it meaningfully improved model performance in his experiments and contrasted it with Claude’s weaker agent teams. Doug noted that Anthropic had described reinforcement learning and games in a post, so the training explanation remained contested.
That let Doug run a set of problems 20 times, measure performance across models, and then compare the qualitative differences—an internal benchmarking exercise unavailable to “a normal guy” three months earlier. Shawn said running the swarm required roughly 16 H100 nodes, highlighting that swarms are also a scale-out compute story.
11. OpenClaw’s promise arrived before its security model
Shawn’s first OpenClaw experience was “really euphoric”: it could read email, see his calendar, and act across personal information. He then recognized how prompt-injectable the setup was and withdrew from exposing sensitive accounts, deciding focused Claude Code use was enough for now.
Shawn mitigates the risk with separate email accounts, promoting an agent only after it proves useful. He said Clawdbot did not impress him; people reacting to Moltbook were overlooking how often a terminal agent still ignores explicit tools or loses task focus.
Doug nevertheless distinguished completion from novelty. Zapier could implement some of the same workflows more securely, but often demanded hours of rigid clicking; an agent that reaches the outcome in four and a half minutes represents a different mechanism. Shawn’s summary was memorable: “Your priors become your prison.”
12. A failed memory-price model still compressed a PhD project into days
Doug asked agents to gather historical NAND and DRAM prices, choose and fine-tune Chronos-2, incorporate covariates, identify the GPU being rented, evaluate results, and eventually expose the work through an internal Vercel dashboard.
The forecasting thesis failed for a recognizably financial reason: memory markets move through regimes whose rules can invert, while treating each cycle as distinct destroys sample size. Shawn pushed back on LLM stock games for exactly this reason—past relationships work “until something fundamental changes.”
Shawn summarized the failed forecasting approach as ending in heuristics—“good luck, have fun”—while Doug’s conclusion was that the time-series model probably would not work. Yet the project left him with every historical series he could locate, paid API data, macro covariates, and a framework for describing the beginning, middle, and end of each regime.
He had previously built that history manually from old annual reports, GDP data, and narrative reconstruction across the 1980s, 1990s, 2000s, and 2010s. What once looked like “a lifetime of work” or a PhD project took one or two days, even though expert judgment remained necessary at the end.
13. Experts capture the upside because they can see the slop
Shawn’s pushback was reputational: SemiAnalysis cannot publish fluent but faulty work to paying clients, and an analyst who did not gather the evidence personally may lack the knowledge needed to challenge it. Doug agreed without softening the issue: “This crap makes mistakes all the time. All the time.”
His working metaphor is a junior analyst who gathers painful information for a senior decision-maker. What is missing is the historical transition in which the junior internalizes repeated cases, discovers where they have a reliable hit rate, and develops “meta-level thinking” into genuine expertise.
Doug hopes future systems acquire that learning—“everyone who’s spending one quadrillion dollars in the world thinks it will”—but does not believe it is present today. At SemiAnalysis, agents massively amplify existing experts because those experts already carry patterns, exceptions, and implicit overwrites in their heads.
The human contribution is the “artisanal last 5%”: spotting fabricated assumptions, knowing when a premise conflicts with the analyst’s own valuation framework, and correcting the premise rather than polishing the output. That makes agent use “a game of hygiene,” not permission to accept whatever appears complete.
14. Automation threatens apprenticeship as much as entry-level work
Doug worries that junior employees who skip evidence-gathering may never build the internal model needed for review. Checking an answer for visible mistakes is not the same as wrestling with the original problem until its exceptions and hierarchy become intuitive.
SemiAnalysis may therefore be less permissive with AI among new hires than among established researchers. “You have to still do some of that”; otherwise, there is cognition-free output whose sloppiness is obvious to someone who already paid the tuition.
Shawn reframed automation as more turns at the wheel: a human might attempt one analysis, while agents can generate several parallel versions and shift attention toward review. Doug accepted the leverage but returned to the same risk—without having once done the work, the reviewer may not know what deserves attention.
They discussed always-on heartbeat designs that could review sessions, extract lessons, and carry them into the next task. A specialized customer-service agent that can retrieve every prior interaction may possess more case context than any human, provided verification prevents accumulated errors from becoming permanent memory.
15. Economic AGI arrived before the “machine god”
Doug became “AGI-pilled” under a practical definition: can the system automate, transform, or eliminate a meaningful range of information jobs? After Opus 4.5 began completing longer projects, his answer became “yes, 100%,” even though he rejected claims that present agents are flawless or superintelligent.
Entry-level data analysis is his clearest case. Given a well-designed agentic system that scans quarterly data for interesting changes, he cannot imagine the average 22-year-old “murdering the hell out of” its performance consistently.
Shawn pointed to GDPval, where 50% represents parity between a model and an industry expert. He said newer systems including GPT-5.2 and Opus 4.5 had moved into the 70-something range, which he interpreted as models outperforming experts more often on the sampled professional tasks.
Their disagreement was mostly over vocabulary. Shawn called that an AGI definition; Doug agreed for white-collar work but separated it from ASI and “the machine god.” Moving the goalpost toward superintelligence obscures how much ordinary economic work has already crossed a meaningful threshold.
16. AI may increase output while making GDP harder to interpret
Doug described a progression in which economies move from agriculture to manufacturing, then white-collar work and a mature financial sector. Humans will invent new work and adapt, but the five-to-ten-year transition can still be socially abrupt even if a new economic layer ultimately appears.
His “crackpot theory,” explicitly hedged, is that AI may be massively deflationary. Information work can expand in units while its market value collapses under abundant supply, making the service-hours logic embedded in GDP less representative of actual productive output.
That creates the possibility of an “AI Great Depression” during the adjustment—not because less work is performed, but because society has not yet priced or absorbed a flood of inexpensive cognition. Doug stressed uncertainty; his observed near-term response is simply that people use productivity gains to work harder.
17. Railroads imply several AI capex cycles, not one
Doug prefers the railroad buildout to the internet as an analogy. He said the internet’s real-dollar buildout was roughly $1 trillion and that AI had already passed it in absolute size, then described railroads as a much longer infrastructure cycle lasting about 45 years and containing three booms and busts.
Railroads did more than move agricultural output: their financing needs helped create modern banking, with rail debt at one point dominating the paper market. The infrastructure was so large and slow to deploy that new capital institutions had to emerge around it.
Shawn cited railroad capex near 4.8% of GDP and 25% of gross fixed capital investment; he compared Stargate with roughly 2% of U.S. GDP. Doug expects AI ultimately to exceed railroads, though faster information flows should compress the cycles.
His base case is not one uninterrupted ascent. Supply and demand curves will cross eventually, and every capital boom reaches “this must be built; it doesn’t matter the price” before investors discover that “that was a steep-ass price.”
18. Claude Code finally made token demand legible to Doug
Claude Code changed Doug’s confidence in the demand curve because he became a heavy consumer himself. A standard Max plan was “not even anywhere near enough”; he said he was on Fast with “$1 million on the API,” though the transcript does not clarify the unit.
Asked what the tool was worth annually, he estimated $20,000–$30,000 “easily, if not more.” The relevant comparison was a roughly $90,000 junior analyst who is perfectly compliant and can be instantiated in parallel many times.
That willingness to pay does not prove infinite demand, and Doug refused the intellectually dishonest “number go up forever” claim. It does show that a new capability can reveal elasticity far above consumer subscription prices once it reliably completes valuable professional work.
19. The analyst’s IDE is headed for the coder’s fate
After hearing the claim that the traditional coding IDE was dead, Doug concluded the same logic applied to finance. “Excel is the IDE for analysts. Bloomberg is the IDE for analysts,” and both preserve interfaces built around what human operators can manually navigate.
Shawn said categorically, “I will never make a chart in Excel again.” Doug agreed with the broader conclusion: an agent can query trusted sources, analyze relationships, and return a Matplotlib image or purpose-built dashboard faster than a human can manipulate cells, even if the output is slightly inconsistent with old templates.
Shawn’s earlier startup tried to challenge Bloomberg and taught him that its defensibility lay in messaging, journalism, and data feeds more than the interface. Shawn said SemiAnalysis was moving toward a FactSet API plus Claude Code, while conceding that traders and regulated deal work still depend on information networks outside the basic analyst workflow.
Shawn estimated that switching down could save $10,000–$20,000 of annual terminal cost for some analysts. More importantly, the workflow replaces tacit memorization of folders, keys, and functions with direct intent.
20. Claude leads general work while GPT-5.3-Codex is “coding-pilled”
Doug’s deliberately conservative year-end prediction for Claude-attributed GitHub share was 25%, which he said he had sandbagged. Shawn said 25% fit a 95%-confidence interval while the current trajectory looked closer to 50%; both acknowledged that public-signature tracking misses some activity.
Doug revised an earlier view that Anthropic’s advantage came mainly from token efficiency after GPT-5.3-Codex arrived: “They’re so back.” Its reinforcement-learning stack looked excellent for coding, and Doug expects a stronger future pre-training run paired with that stack could flip the race.
The limitation is specialization. GPT-5.3-Codex repeatedly tries to build scraping software when Doug wants it simply to ingest and reason over webpages; Opus 4.6 more naturally switches among research, rubrics, analysis, and coding because it is less “coding-pilled.”
Shawn raised a multi-model interface such as Conductor, where Claude and Codex can review each other. Doug had not used Conductor personally but agreed that cross-review could be useful, while distrusting middleware trapped between fast-moving, well-funded first-party platforms: the layer that appears to be a clean superset often gets “eaten by one or the other.”
21. Microsoft is renting compute to the barbarians at its gate
Doug rejected the clickbait claim that Microsoft is “out of AI,” but maintained it has “the most to lose of everyone.” Office, PowerPoint, Excel, and email are horizontal interfaces through which humans perform information work—the exact abstraction agents threaten.
Azure complicates the defense because Microsoft earns money renting infrastructure to OpenAI, a potential disruptor. Doug’s analogy was Rome hiring barbarians: each year the mercenaries become stronger while the walls they may eventually scale become more dilapidated.
The strategic fork is painful. Microsoft can emphasize Azure and risk becoming a premium “dumb pipe,” or redirect capital into proprietary models and Office defenses, sacrificing some cloud growth.
Claude for Excel and Claude for PowerPoint sharpen the indictment because “Microsoft should have built it.” Doug read management’s comments about internal capability investment as evidence of pulling resources back toward the walls, but remained skeptical after uneven execution and said Microsoft must choose a direction.
22. Oracle’s financing cadence turned capital supply into a bottleneck
Shawn asked whether Oracle was irresponsible for absorbing capacity Microsoft declined. Doug’s answer was that the setup was an own goal, principally because Oracle promised enormous scale—he recalled about $400 billion of RPO—then raised aggressively before establishing a smoother, partially self-funding glide path from deployed GPU revenue.
Shawn’s rough, explicitly imprecise comparison put investment-grade TMT debt near $500 billion and Oracle around $135 billion. Issuance that large must offer better terms to attract buyers, repricing competing debt and making the whole index sell off through simple supply pressure.
The surprise bottleneck is therefore “supply of debt into the market.” Hyperscalers historically funded themselves; suddenly asking credit markets for orders of magnitude more capital creates liquidity problems even when the underlying projects may ultimately pay.
Microsoft could have financed the same capacity internally or borrowed near U.S.-government rates, giving it a roughly two-percentage-point capital advantage over Oracle. Doug called declining that advantage a blunder while Oracle’s abrupt issuance helped trigger CDS anxiety and forced capitalism to “pump the brakes.”
23. TPU V7 has a narrow window to convert TCO into installed base
Doug interpreted Google’s willingness to sell current-generation TPUs externally as a market-share decision. Before Gemini 3, hoarding hardware mattered less if Google’s own products were losing; Ironwood TPU V7 now offers its widest expected TCO advantage before Nvidia Rubin arrives.
The shared discussion put TPU near $1 trillion at roughly 30% share. The key is establishing an installed base: customers who own current TPUs have a reason to upgrade, whereas AMD must repeatedly win users who are not replacing earlier AMD accelerators.
TPU hardware, networking, and software are mature enough, and Anthropic is an unusually capable external customer. Doug thinks Anthropic, OpenAI, and other labs would consume far more TPU V7 in an unconstrained world because its current price-performance is “the hottest kid on the block.”
Supply prevents that theoretical demand from becoming share. TSMC is probably the biggest blocker, and some of the engineers behind the original TPU program have dispersed across the industry; the window may last only a year or two.
24. Nvidia’s supply-chain control could close Google’s hardware gap
SemiAnalysis expects TPU V8 to compare less favorably with Rubin, particularly because of HBM3 versus HBM4 and Nvidia’s stronger memory-scale-up position. Larger, faster memory directly supports bigger contexts and more capable systems.
Doug argued GB200 would have “completely mogged” TPU V7 had it arrived on time and stable. Its delay and reliability issues created the current opening, but Nvidia has multiple ways to correct course.
Nvidia behaves like an F1 program, pushing every component to its limit, while Google designs stable, replicable pods consistent with its infrastructure culture. The trade-off favors Google when Nvidia stumbles; it favors Nvidia when the maximum-performance system arrives cleanly.
Jensen Huang’s supplier relationships are part of the product. Doug pointed to his meetings and “love shots” with Samsung, SK hynix, and other Asian partners: Nvidia secures priority for HBM, packaging, connectors, density, and future road maps in a way Google’s leadership has not visibly matched.
25. HBM turns every AI accelerator into a multiplier on DRAM scarcity
The memory mechanism begins with a roughly 3:1 to 4:1 trade ratio: producing one bit of HBM effectively removes several bits of ordinary DRAM capacity after added process steps and imperfect yield. Doug’s analogy was refining a newly essential jet fuel by consuming much more conventional fuel.
This demand arrived after what Doug called the worst NAND and DRAM shortage ever, with the last comparable analog perhaps in 1996. Suppliers went deeply free-cash-flow negative, stopped spending, and postponed clean rooms and equipment whose lead times can reach two or three years.
HBM absorbs the high end, while KV-cache offload and broader AI systems consume the middle layers. With no spare capacity added during the bust, “more demand than God” now cascades through every memory grade.
SemiAnalysis’s conclusion was that DRAM prices could rise another 100%, with supply unlikely to catch up for about two years. At that point, Doug expects real demand destruction rather than a painless transfer of higher component costs.
26. Scarcity will ration devices, context, and even old memory
Near-term demand destruction could mean hyperscalers delay marginal purchases, particularly when power-delayed data centers slip from 12 to 18 months. Customers first pull forward and double-order everything, then pause once inventories and power constraints catch up—the familiar trigger for a memory-price collapse.
Consumers will feel the allocation. Shawn advised buying an iPhone sooner because handset makers eventually enter the spot market and must pass through higher costs; low-end phones, gaming GPUs, and other products may simply be priced out while AI infrastructure wins scarce supply.
CXL may receive “a shot on goal” after previously losing relevance to HBM. Operators can collect old DDR4, place it into expansion racks, and attach it through CXL—resurrecting a near-dead architecture because every available memory bit now has value.
The one-million-token context window may therefore resemble “a mansion” rather than a universal default. Doug thinks physical memory could keep full contexts near today’s scale for five or even ten years, encouraging “context rationing,” while Shawn raised the possibility of differentiated pricing for scarce context and recursive language models.
27. Fixed-weight chips and CPUs reveal secondary shortage paths
Shawn asked about Taalas burning model weights directly into silicon, eliminating repeated memory transfer. Doug found the logic compelling: “The way to speed things up is to never transfer anything,” especially as production models are often smaller or distilled below frontier-training size.
The uncertainty is market breadth. A fixed efficient model could scale inference dramatically, but changing weights and the performance trade-off leave a difficult design space; Shawn remained skeptical of most accelerator startups because so many predecessors failed to reach production.
CPUs face a quieter squeeze. Clouds bought roughly $100 billion of CPUs and related equipment in 2020–2021 and now approach a five-to-six-year refresh after diverting two years of capex toward GPUs, just as agent-generated software, production agents, and reinforcement-learning simulations raise CPU utilization.
Doug floated—explicitly as “schizophrenic tinfoil-hat-brain” speculation—that recent web-service instability might combine vibe-coded production errors with aging cloud fleets. The firmer thesis is simpler: modest new demand hitting a severely underinvested category can create another shortage.
28. Doug writes by accumulating privately, then taking one clean shot
Before LLMs, Doug considered high-throughput reading and synthesis his defining information skill. He could push through textbooks or a friend’s PhD paper at adjustable depth, and writing completed the loop by forcing him to express what all that reading had produced.
Publishing weekly from October 2021 built the habit. He dislikes LLM-generated prose but uses models for ideation, outlines, and editing against lessons from On Writing Well; the words themselves still need to come from his own thinking.
His best technique is to gather evidence, outline, think hard, and then sleep. The next morning provides a “fresh context window”: he opens a new tab and writes in one shot, usually reaching 60%–75% before filling the remaining gaps.
29. Six months on the Continental Divide Trail clarified the tool user
In 2021 Doug chose the Continental Divide Trail because, among the three major U.S. through-hikes, it scared him most. He covered roughly 2,800–2,850 miles over six months, mostly alone, accepting that he might never again have an equivalent opening.
The opportunity cost felt real—he thought he missed an important Substack growth year—but the trip delivered something professional acceleration could not. It was a genuine “adventure,” with boredom, fear, hunger, isolation, and the “lowest lows and highest highs” of a compressed life.
Returning to the bottom of Maslow’s hierarchy made abstract information work feel “totally fake” beside staying warm, fed, and alive. Doug came back with a clearer understanding of his own limits and motivations, making Shawn’s closing synthesis apt: “Self-mastery is your most important tool of all.”