06-reference

stratechery memory chips china microsoft deepseek

2026-06-23·reference·source: Stratechery·by Ben Thompson

Why this is in the vault

Ben Thompson's June 23 Daily Update covers two tightly linked stories: the structural causes of the memory chip shortage (HBM demand crowding out commodity DRAM/NAND, the DRAM oligopoly's self-inflicted wound, and Chinese entrants CXMT/YMTC), and Microsoft's incentive to route Copilot/Cowork through cheaper Chinese open-source models (DeepSeek V4) to protect margins under usage-based pricing. Both threads are directly relevant to RDCO's chip-fab/memory capital-cycle investing thesis and to the AI model economics that underpin the COO-agent infrastructure.

Summary

Section 1 — Memory Chips and China

Thompson opens with Apple raising prices due to the memory shortage — the last outcome the Trump administration wants but no quick fix exists. He ties this to a January 2026 prediction: AI "crowd-out dynamic" — Samsung, SK Hynix, and Micron shifted capacity to high-bandwidth memory (HBM) for Nvidia/AMD AI chips, starving commodity DRAM. The effect has cascaded from chips to grid power to turbines.

Structural root cause: The DRAM oligopoly (three players) learned discipline after decades of boom-bust shakeouts. They deliberately restrained capacity expansion to protect margins — Micron's gross margin hit record ~80%. Now that discipline is backfiring. HBM commands premium pricing, so all three prioritize Nvidia's HBM orders; everyone else takes the remainder. NAND faces the same dynamic (the big three plus Kioxia and Sandisk).

Chinese entrants:

Apple (Tim Cook: "everything needs to be on the table"), HP, and other consumer-tech makers are lobbying the Trump administration to loosen restrictions on Chinese memory. Apple had YMTC lined up as an iPhone NAND supplier in 2022 until congressional pushback killed it. HP is in talks to use CXMT chips in Asia-market products. DRAM fungibility means a company can mix-and-match: CXMT for Asia, big-three for Western markets. Volume plus state subsidies will close the quality gap over time.

Thompson's verdict: "The big three have probably already massively damaged themselves in the very long run: because they were comfortable enough with their oligopoly to not bring on more supply faster, they created the conditions for Chinese competition to arise, and I don't think they're going to like it very much."

Section 2 — Microsoft and Chinese Models

An Axios report revealed Microsoft's Copilot Cowork will shift to usage-based compute pricing and is exploring a fine-tuned DeepSeek V4 as a lower-cost alternative to Anthropic/OpenAI models currently powering the product. Thompson had a preview from his June 2026 Satya Nadella interview, where Nadella mentioned Cowork supporting multiple models — not yet public at the time.

The margin squeeze: Usage-based AI pricing destroys traditional SaaS bundling. Microsoft won Teams over Slack by making it "effectively free" in the M365 bundle. That model breaks when customers pay per-token — enterprises scrutinize marginal value and may prefer standalone Claude or GPT-4o. The new Microsoft E7 plan ($99/employee/month) bundles Copilot and Cowork, but heavy frontier-model usage within that bundle destroys margins.

Why DeepSeek solves it: Open-source DeepSeek V4 costs dramatically less to serve, allowing Microsoft to set a usage ceiling high enough that most employees never hit it — restoring the bundled-unlimited feel of traditional SaaS while protecting economics.

Broader pattern: "The biggest proponents and beneficiaries of Chinese companies offering a cheaper alternative will be American companies. This is a pattern that isn't going away."

Mapping against Ray Data Co

Memory cycle thesis — strong signal. Thompson's analysis validates the Phase 2 framing in RDCO's Markov capital-cycle tracker: HBM premium absorbs all disciplined capacity, commodity shortage ensues, Chinese entrants (CXMT/YMTC) are now credible Phase 2 disruptors. Key data points to feed into the thesis:

AI model economics — medium signal for RDCO infrastructure. Microsoft's DeepSeek routing logic is the enterprise mirror of RDCO's own model-cost decisions. The same margin math that pushes Microsoft toward cheaper open-source applies at RDCO scale — Claude as COO primary + open-source for high-volume, low-stakes inference. Thompson's framing of "usage-based AI destroys SaaS bundling" is worth tracking for the phData Deal Solutions Architect pipeline: enterprise clients are about to re-examine Copilot/M365 AI add-on pricing.

phData angle — weak-medium. Thompson's Cowork vs. Copilot distinction (different products, different model needs) is a useful framework for client conversations about AI stack rationalization during presales discovery.

Related