〔Day Digest〕Samsung Locks 80% of 2027 Memory Output as U.S. Spot Compute Hits ~$40-50B/GW and Musk Stacks Chips-to-Agents — Lock-In Meets Li Lu's Limits

0. Daily Arc

Three threads converged on the AI tape: a hard supply lock-in ^[1], a spot-vs-long compute spread that prints the scarcity premium ^[2], and a contrarian macro voice from Himalaya Capital founder Li Lu at the Columbia Graham & Dodd lunch ^[3][4][5]. The market read this as a buy-the-chain story — TSMC's October 15 print is now the next hard catalyst for Nvidia ^[6][7], and Musk confirmed the rename of SpaceXAI to "SpaceX Super Intelligence" (@SpaceXSI) on his X account ^[8].

1. Memory Lock-In: The Supply Side of the AI Arms Race

• **[NEW] Samsung:** roughly 80% of 2027 memory output (including HBM) is committed under multi-year LTAs, with Nvidia, Google and Microsoft among the signatories and only ~20% of capacity still open ^[1]. On the Q2 earnings call Samsung said "almost all major customers" are seeking LTAs and that demand is "hard to respond to" ^[1].
• **[NEW] SK Hynix:** has completed LTAs with major customers ^[1].
• **[NEW] Micron:** strategic customer agreements doubled quarter-over-quarter, from 16 in Q2 to 26 in Q3 ^[1].
• **[NEW] Terms are tightening:** 20-25% prepayments and mandatory purchase clauses are now common, and the same contract architecture is replicating in MLCC, FC-BGA substrates, transformers, power equipment and cooling ^[1].

2. Compute Economics: The Spot vs. Long-Term Spread

• **[NEW] Goldman / Eric Sheridan data:** annualized revenue per gigawatt of AI compute runs at $40-50B on spot rentals, $20-30B under long-term contracts, and ~$12B as the hyperscaler self-build break-even ^[2]. The spot premium is roughly 2x long-term and 4x break-even ^[2].
• **[NEW] SpaceX scale path:** 1.4 GW in Q2 2026, ~7 GW by end-2027, ~10.6 GW by end-2028 — a ~7x expansion in four years; Goldman argues the spot premium compresses once self-build catches up ^[2].
• **[NEW] Peer confirmations:** Nebius disclosed Q3 short-term compute at $40-50B/GW; CoreWeave's recent short-term contracts priced near $40B/GW ^[2]. SpaceX itself disclosed two new hosting deals "at the high end of the historical $30-50B/GW discussion range" ^[2].
• **[NEW] Macro rebrand:** Musk's X post on Oct. 10 confirmed SpaceXAI's official account is now "SpaceX Super Intelligence" (@SpaceXSI), following his Oct. 4 announcement that "SpaceX is a super intelligence (SI) company" ^[8].

3. The Musk Stack: From Terrafab to Game-Playing Agents

• **[NEW] Terrafab pitch:** Speaking at the Tesla Austin gigafactory on Oct. 10, Musk told customers and investors his companies will "likely find a way to make chips better than anyone," citing custom-equipment expertise versus competitors; he conceded he "doesn't know how to make microchips" but estimated it could take "several years" ^[9].
• **[NEW] Digital Optimus:** On X the same day, Musk said a Tesla + SpaceXAI agent has cleared "about half" of the Diablo campaign by screen-watching alone, is "doing well" in Counter-Strike, and is in training on League of Legends — with no demo video, no scores, no release date, and no specific Diablo or "League" title confirmed ^[10].
• **[ESCALATED] Apple reverse acqui-hire:** Apple disclosed in the EU Digital Markets Act case database that it has agreed to hire "some Huxe AI employees" and take a non-exclusive license to Huxe's IP; the founders are ex-NotebookLM (now Gemini Notebook), and Huxe shut its consumer app on May 21 ^[11].

4. The Counter-Narrative: Li Lu at the Graham & Dodd Lunch

• **[NEW] Scope argument:** Li Lu said generative AI is best at "democratizing" knowledge and skilled services but does not change the underlying nature of energy, transport, or molecular-rearranging manufacturing, and remains "very far" from crossing into biology because it is still bound by human-generated training data ^[3].
• **[NEW] On RSI / AGI / ASI:** Li Lu rejected the "synthetic god" arc; either the technology will not follow that path, or society will collectively stop it before it does, so markets "need not blindly believe" the ASI endpoint ^[4].
• **[NEW] On safety:** Li Lu framed the current wave as a shift from "tool" to "creator" and argued recursive self-improvement without human participation is "extremely dangerous"; society must hold AI on the "tool" side of the line ^[5].

5. China Compute Build-Out and the Funding Tape

• **[NEW] Ulanqab, Inner Mongolia:** 89 signed data-center projects, with Alibaba, Huawei, Kuaishou and third-party operators in build; a decommissioned school is being repurposed as a data center, and the city is branding itself a "Token capital" to monetize cheap power and cooling ^[12].
• **[NEW] Capital Online:** Phase 1 of the Huailai Data Center has been leased in full to an unnamed internet "head" customer for AI training and intelligent-compute workloads, marking Capital Online's first large-scale shift from colocation resale to self-built operation ^[13].
• **[NEW] Qwen on Qianwen:** Alibaba's Qwen-Image-2.1 Pro and Turbo APIs are live on the Qianwen AI platform, with the Turbo model weights open-sourced in parallel ^[14].
• **[NEW] Hangzhou "New Eight Steeds":** A pivot toward deep-tech bets — smart glasses, inference GPUs, compute-in-memory chips and space computing — has landed with far less viral heat than last year's Unitree robot kickboxing or DeepSeek open-source drops ^[15].
• **[NEW] Funding tape:** Litai Zhiyao closed a ~50M yuan seed led by iFlytek VC, the L2F founders' fund and an undisclosed industry investor, with Guangyuan Capital as FA ^[16]; Ruizhi Technology closed a first-round financing in the tens of millions of yuan led by Zhongguancun Capital, with Zhongguancun Science City Co. following ^[17].
• **[NEW] Anhui Keyi Intelligent:** a ~300M yuan, 33-mu, ~34,000 sqm AI vision inspection equipment project broke ground, targeting ~500M yuan in annual output at full capacity with completion slated for July 2027 ^[18].
• **[NEW] Enterprise / education:** UMS signed a new contract with OpenAI for ChatGPT Edu aimed at safer and cheaper AI use in education ^[19].
• **[NEW] AI safety thesis:** a "trillion-yuan" AI security market is being framed as a multi-year core track as agent deployments expose data-leak, system-call and infrastructure-attack risks, following recent incident disclosures at OpenAI and Anthropic ^[20].
• **[ONGOING] Source quality control and falsification:** items ^[13], ^[16], ^[14], ^[18], ^[11], ^[17] rest on single-source Chinese press wires; Musk's Diablo and chip-fab claims carry no demo, no third-party validation, and no timeline ^[10][9]. The contraction test on the day is whether Samsung's remaining 20% capacity and the $40-50B/GW spot compute band both narrow into the TSMC October 15 print ^[6][7][1][2].