Ep. 001 - Claude Code, Memory Mania, CPUs are Back | Jordan Nanos, Doug O'Laughlin, Myron Xie

Summary

• The memory cycle has whipsawed from the worst downturn since 1996 — which “bankrupted a third of the industry” that year — to a historic shortage with little near-term relief. Myron Xie’s mechanism: HBM yields several times fewer bits per wafer than conventional DRAM, every incremental GPU/XPU pulls wafers toward HBM, and post-COVID there were virtually no additional wafer starts or cleanroom space — “we’re adding virtually no additional supply while demand is booming.” Doug O’Laughlin: memory and NVMe prices rose ~20% in a single day and ~100% month-on-month, and it stays tight “for a few years at least.”
• The classic cycle-top signal — Samsung capitulating and adding capacity last — may not work this time. Doug: “I think they’re gonna build capacity, and it’s still not gonna be enough. The supply-demand gap is the largest it’s ever been, and it gets worse from the work that we see.” His “galaxy brain take”: the EUV bottleneck could pull 3D DRAM forward from the expected ~2035 timeframe toward ~2030, though not before 2030.
• CPUs are back, and even Intel is still — or probably will be — sold out, “which tells you things are tight.” The discussion frames Intel as the marginal supplier, while Arm is “such a bigger threat than it’s ever been” and AMD “continues to crush.” Phoenix CPUs are an interesting change, particularly for how Arm’s business model may adapt.
• Claude Code now authors ~4.7–4.8% of public GitHub commits, per Doug’s scraped daily series — and Anthropic linked the article containing the estimate in the press release for its $30B raise at a $380B valuation. The share rose from ~2% on January 20 to ~4% on February 5 and is still growing; Doug built the tracker after a tweet tipped him that Claude Code signs its own commits.
• Opus Fast and OpenAI’s GPT-5.3-Codex-Spark (on Cerebras) are price-elasticity experiments above the $2,400/year Max tier — the “caviar trade.” Opus Fast charges ~6x per token for roughly 2–2.5x speed; Jordan Nanos describes the Cerebras trade as “10x the money for 4x the performance, or really more like 20x the money for 10x the performance or something like that.” Jordan says the Fast mode is working on him: “Now that I’m addicted to Fast tokens… I don’t think I could go back.”
• China Model Week delivered real breakthroughs: MiniMax 2.5 “at Opus performance… 1/10 the price,” GLM-5, and above all ByteDance’s Seedance/CDance 2 video model — “animation done, cooked.” Doug calls it the first “total victory” over Google’s Veo/Genie lead, and floats a conspiracy: “the surveillance state gives everyone better training data” — 10% of all hard-drive demand is Chinese surveillance.
• Jordan’s storage thesis: SSD/HDD demand will boom on logs, synthetic data, and generated video — not KV-cache offload, whose retained fraction can be “less than 1%” of generated video. He also warns against attach-rate shortcuts like Jensen’s 16TB-per-GPU figure, which makes ICMS look like <0.7% of drive shipments — “the right conclusion for ICMS, but not the right conclusion for the growth of drive demand.”

Deep dive

1. Claude Code’s commit share hit ~5% — and its article landed in Anthropic’s press release

• Doug’s origin story: he read a tweet — “You idiots, it all says it’s committed by Claude Code, you could turn this off” — asked Claude Code whether GitHub could be scraped, and built a daily series of public commits authored by Claude Code. From ~2% in the January 20 core-research chart to ~4% in the February 5 public piece; “it’s like already almost 5%… 4.7 or 4.8,” still growing. Anthropic then linked the article containing the estimate in the press release for its $30B raise at $380B.
• Doug’s urgency, verbatim: “Anyone who’s paying attention can figure this one out… we need to be the first” — the paid data continues for institutional subscribers, likely with a public dashboard.
• On AI-assisted writing, both refuse full generation. Doug uses Claude Code for outlines and cuts (“what’s the least relevant portion of this paragraph?”) but writes “manual tokens, you know, with the fingers” — “it’s our last artisanal point… this hand-churned butter is all we got.” Myron only polishes phrases. Inside baseball worth keeping: Myron writes many of the bangers — “Do you think Dylan’s writing these articles? No, brother. It’s a team.”
• Live data point on agent use: Doug had seven agents on one task, seven windows open, and one of those windows running seven subagents.

2. CPUs are back: “even Intel is sold out”

• The CPU discussion’s key proof point is the marginal supplier: nobody would have guessed the largest, most mature compute install base would run short and Intel would catch a bid. The panel says Intel is still sold out, or probably will be sold out — “that tells you things are tight.”
• Doug’s framing is that Arm is “such a bigger threat than it’s ever been” — the Phoenix CPUs are an interesting change in how Arm’s business model may adapt — while AMD “continues to crush.” Even with more competition and fragmentation, Intel is still likely to be sold out.
• Jordan’s structural read: AI growth is compute consumed through an API via tokens, not Fortune 500 companies spinning up virtual machines in the cloud-era style — CPU data centers support cheap GPU data centers, per Microsoft’s Fairwater campus photos. VMware is “gutted” post-Hock Tan, Pat Gelsinger’s VMware arm tattoo notwithstanding.

3. Memory mania: the worst cycle since 1996 whipsawed into the best

• Myron’s supply mechanism: HBM yields several times fewer bits per wafer than conventional DDR DRAM — TSV keep-out zones eat die area, yield is “terrible” at HBM’s performance requirements, and 8- or 12-high stacking loses yield again at the package. Every incremental GPU/XPU pulls wafers to HBM and shrinks total industry bits, while CPU growth tightens conventional DRAM — into virtually no incremental wafer starts, because after the COVID hangover “there’s literally not much clean room space available to actually put equipment.”
• Doug’s history: the downturn was literally the worst since 1996, and the memory industry did something historic — shuttering machines rather than eating underutilization charges: “No, no, let’s try to convince everyone to turn it off.” Now there’s nowhere to turn the spigots back on. Memory and NVMe prices rose ~20% in a day and ~100% month-on-month, with “a little bit of panic buying, just a wee bit.”
• NAND supply has been worse than DRAM and is less profitable, so manufacturers will allocate incremental capacity to DRAM first — “it’s gonna be crazy for a few years at least.”

4. Samsung capitulation is the top signal — but this time it won’t be enough

• Jordan’s setup: historically the cycle tops when Samsung, last after Hynix and Micron, finally adds capacity. Doug: last cycle Samsung held out roughly three quarters longer than usual against pleas to cut, and its eventual cut marked “the pico bottom.” His call now: “I think they’re gonna build capacity, and it’s still not gonna be enough” — the supply-demand gap is the largest ever and worsening in SemiAnalysis’s work.
• The binding constraint above fabs: ASML can only ship so many EUV tools a year, logic is also tight, and how tools get split between TSMC and the memory makers is a live internal debate — “that’s a question for ASML’s strategy team, and they should probably buy our models.” Myron’s view: since logic needs memory and memory needs logic, the system should optimize toward completed-system ratios, but “the trade ratio itself is honestly a very hard question… I don’t think many people know.”
• Doug’s on-the-fly math, hedges intact (“I’m vibing it out”): leading-edge logic CapEx per 100K wafer starts may run ~2–3x DRAM’s, but at a ~3x DRAM trade ratio HBM may be the most expensive wafer in the world — “I guess I’m defeated. I think HBM’s the hardest.”
• The galaxy-brain consequence: the EUV squeeze “might” pull 3D DRAM from its expected ~2035 timeframe toward ~2030 — “not gonna be before 2030, to be clear” — because 3D DRAM uses no EUV, a system-level throughput unlock.

5. The caviar trade: Opus Fast and GPT-5.3-Codex-Spark are pricing experiments

• Jordan’s segue — SRAM-only accelerators trading ~6x the money for ~2.5x performance — draws Myron’s line of the episode: “That’s like saying eggs are a bit more expensive, I’m gonna eat caviar for breakfast instead.” SRAM is far costlier than DRAM and logic is already tight, but “there’s an appetite to pay.”
• The tape: Opus Fast at ~6x per-token pricing and roughly 2–2.5x speed; OpenAI’s GPT-5.3-Codex-Spark running on Cerebras, which Jordan describes as “10x more money for 4x the performance, or really more like 20x the money for 10x… or something like that.” Jordan’s read: the labs are walking up the demand curve above the $2,400/year Max tier — “there has to be a higher tier of pricing… they’re finding out the price elasticity curve.” His own evidence: “Now that I’m addicted to Fast tokens, I don’t think I could go back, dude… You moved me from heroin to fentanyl.”
• Myron’s balance point: HBM still holds the best cost-bandwidth-density trade-off for now, and GDDR per-gigabyte pricing is closing in on HBM anyway — non-HBM architectures “always had trade-offs.”

6. China Model Week: “the most Chinese era of my life”

• Doug makes a speculative DeepSeek V4 President’s Day call, based on an SCMP read from November or December, and defends his release-timing theory against Jordan’s pushback that President’s Day isn’t well known in China: DeepSeek drops on American holidays — Thanksgiving, Christmas Eve — “to troll the shit out of America.” “If they really wanted to mog us, they would put it out on July 4th.”
• The releases: MiniMax 2.5 “at Opus performance… 1/10 the price” with ~10B active parameters; GLM-5 “an incredible release” from Zhipu, alongside discussion of a Chinese Indeed/LinkedIn-like company training foundation models at ~70% on Soy Bench; and ByteDance’s video model, referred to as Seedance and later CDance 2 in the discussion, generating minute-long scenes with consistent characters. Doug: “this isn’t gibberish anymore” — Opus 4.5-style breakthrough energy, “at least animation done, cooked,” and the first “total victory” over Gemini’s long Veo/Genie lead in video.
• Doug’s conspiracy theory on why: “the surveillance state gives everyone better training data” — 10% of all hard-drive demand is Chinese surveillance, meaningful enough that he says COVID shutdowns further hurt HDD companies because that demand was significant.

7. Storage demand is video and logs, not KV cache — and beware the attach rate

• Jordan’s running battle with questions from people in the industry: everything through a model gets stored to logs and “they don’t throw away data”; synthetic data output is stored; every camera writes to disk — sources are 10x to hundreds of times bigger than the models’ active parameters or inputs. The KV-cache share of generated video is “less than 1%.”
• Doug’s carrying analogy: open your iPhone storage — ~20% apps, ~70% photos and videos. “That’s literally the reason why” — mass video generation à la CDance, and above all sending and watching it, is the real storage driver.
• On heuristics: SemiAnalysis rebuilt the full BOM — every port, east-west, north-south — to test the 1.5x transceiver attach rate, and got 1.5x. “Some really powerful heuristics get you pretty far in life”; CapEx, or at least revenue monetization per gigawatt, has so far been the cleanest yardstick. But the popular attach-rate shortcut — Jensen’s 16TB-per-GPU from CES, implying ICMS is <0.7% of drive shipments — is “the right conclusion for ICMS, but not the right conclusion for the growth of drive demand.”
• Jordan says Jensen has a “good nose for saying the right thing at the right time,” redirecting memory panic to memory attached to his GPU. Doug’s read is that the $100B OpenAI deal “happened afterward to screw over AMD”; he also points to the Cerebras Spark announcement declaring “NVIDIA is still the heart of what we do.” Closing line: “It’s Jensen’s world we’re all living in.”