brain/
← all contexts
project

stock-market

Stock-Market Research Dispatch — 2026-05-22

Auto-generated by daily research routine. Not a source; not ingest-pending material. Summary of what changed in the wiki today.

Top of mind

China is not buying US chips — by choice. The dominant China/Nvidia narrative has inverted completely: the US relaxed H200 export controls (BIS case-by-case review, Jan 15, 2026), cleared ~10 Chinese firms (Alibaba, Tencent, ByteDance) to buy H200s, and Jensen Huang flew to Beijing. But not a single H200 has shipped — Beijing blocked domestic firms from purchasing to force adoption of Huawei Ascend. Nvidia provisioned a $4.5B H20 inventory/obligation charge and guided zero China DC revenue. The question is no longer "will BIS close the loophole?" — it's "will China ever let Nvidia back in, and does Ascend 950PR (2.8x H20, SMIC N+2, DeepSeek V4 trained on it) make US chips strategically irrelevant to China's AI buildout?" SMIC is the emerging public equity in this chain.

Nvidia is simultaneously locking in preferred US data center operators via equity warrants. The Nvidia/IREN 5GW DSX partnership (May 7: $2.1B equity right, Sweetwater TX 2GW flagship) establishes a new pattern: Nvidia deploys capital as equity warrants to lock in preferred operators before chips ship. The Sweetwater 2GW campus alone is 10-15% of the current US AI DC pipeline — and that 2GW campus needs power, cooling, and electrical distribution (ETN, GEV, VRT, PWR) at Huang's "30K truckloads per 500MW" scale.

New wiki pages

PageTypeOne-line
smicEntitySMICY; fabricates Huawei Ascend 950PR (N+2 7nm-class); DeepSeek V4 trained on domestic silicon; new hypothesis as China's public AI chip beneficiary
lsccEntityLSCC NASDAQ; Q1 2026 $170.9M (+42% YoY), Compute +86% YoY; AMI pending Q3 close; ~60% to graduation
irenEntityIREN NASDAQ; Nvidia DSX preferred operator; 5GW pipeline; $2.1B equity warrant; Sweetwater TX 2GW flagship
iren-nvidia-dsx-ai-factory-operatorQuestionHypothesis: IREN as Nvidia's public DSX flagship — graduation: Sweetwater Phase 1 (500MW) + first chip orders
silicon-photonics-ai-cluster-networkingQuestionHypothesis: Amazon/STM Feb 9 opens AI cluster backend networking as undercovered picks-and-shovels (STM, MRVL, AVGO, COHR)
tsmc-kumamoto-physical-ai-hubQuestionHypothesis: Sony SSS/TSMC May 8 JV in Kumamoto → physical AI (automotive/robotics) co-located with TSMC logic fab

Significant wiki updates

China Blackwell thesis (china-blackwell-offshore-compute-routing): Completely reframed. The May 19 framing (Malaysia loophole durability) is superseded. Current reality: US cleared the path, China blocked the purchase. Nvidia $4.5B charge, zero China DC revenue guide. Huawei Ascend 950PR (1.56 PFLOPS FP4, 2.8x H20 performance, SMIC N+2, Q1 2026 available) + DeepSeek V4 (Reuters April 2026: first frontier model trained entirely on domestic Chinese silicon) = credible domestic alternative. Beijing's self-block is policy-driven, not capability-constrained. B30A (single-die Blackwell, ~$6.5-8K, June 2026 production target) is Nvidia's China-compliant product — commercial viability uncertain while Beijing's domestic-chip mandate holds. SMIC entity created as new hypothesis.

Samsung foundry (samsung-foundry-as-third-alternative, samsung): Qualcomm status clarified by Digitimes April 23 — "mulling" not decided; WCCFTech: full flagship switch "impossible," dual-source (lower/mid-tier Snapdragon to Samsung, flagship to TSMC) most likely. New: TSMC/LPU competitive angle — TSMC is targeting LPU market that overlaps Qualcomm's addressable TAM, giving Samsung structural leverage. AMD nearing deal (TrendForce Dec 2025, unconfirmed) would be a significant US AI/DC CPU anchor win for Samsung. Samsung Q1 2026 consolidated records (KRW 133.9T revenue, 57.2T op profit) confirmed — but HBM-driven, not foundry.

LSCC hypothesis (lattice-fpga-ai-cluster-management-layer): Advancing to ~60% graduation criteria. Q1 beat ($170.9M, +42% YoY; Compute $106.6M, +86% YoY = 62% of revenue). AMI close Q3 2026 on track. Combined run rate post-close $220-240M/quarter ($950M annualized). LSCC +50% YTD — thesis is being priced in. Missing: AMI close + combined guidance, and a named hyperscaler BOM confirmation. Q3 2026 AMI close is the next gate.

SoftBank Roze (softbank-roze-robotics-datacenter-construction): Open question resolved — ABB deal confirmed cash-only ($5.375B EV, ~$5.3B net to ABB). ABB shareholders have zero equity exposure to Roze. The ABB indirect-play thesis is eliminated. S-1 not filed, ABB acquisition not yet closed, Roze pre-revenue. Low priority confirmed; revisit only at S-1 filing.

Section 232 (section-232-phase-2-us-fab-premium): No change. April 14 USTR/Commerce report filed but non-public. Phase 2 not triggered. July 1, 2026 Commerce DC chip market report (~39 days) is the next gate. Phase 1 exemptions (data center supply chain, R&D, startups, government) confirmed comprehensive.

Datacenter construction (datacenter-construction-electrical-picks-shovels): Nvidia/IREN 5GW DSX is the most concrete named US AI DC campus outside hyperscaler direct builds. Sweetwater 2GW Texas = ~120K truckloads at Huang's scale → electrical/cooling/construction suppliers (ETN, GEV, VRT, HUBB, PWR) are the first-order supply chain.

Watchlist

TickerThesisConvictionSnapshotUndervalued vs base?Next catalyst
MPus-critical-mineral-independenceMediumStale (403)Unknown — June 15 + Nov 10 H2 binary catalystsJune 15 mining controls scope announcement
INTCus-fab-capacity-bottleneckMedium-highStale (403)Yes vs base case (Apple A21 revenue ~2028, DoD 14A)VLSI June 14-18; Intel Q2 earnings (July)
AMATpicks-and-shovels-leading-edge-fab-buildoutHighStale (403)No — Citi PT $520 at 33x fwd P/E requires continued beat-and-raiseAMAT Q3 FY2026 earnings (August)
MUhbm-supply-bottleneckHighStale (403)UnknownQ3 FY2026 actuals vs $33.5B guide
LSCClattice-fpga-ai-cluster-management-layerMedium (hypothesis ~60%)Stale (403); +50% YTD knownMay be priced in at +50% YTDAMI Q3 2026 close
IRENiren-nvidia-dsx-ai-factory-operatorLow-medium (new hypothesis)Stale (403); "climbed following Nvidia partnership"UnknownSweetwater Phase 1 construction start

New chains (hypothesis-stage)

Step 4b — Valuation snapshots

WebFetch blocked (HTTP 403) for seventh consecutive run. All active concept pages carry stale snapshots. No price/P/E updates possible.

What I looked at

  • Step 1 topics: China Blackwell/export controls (major reframe), Samsung foundry (Qualcomm/AMD status), Section 232 Phase 2 (no change), LSCC/AMI Q1 2026, SoftBank Roze (ABB cash-only resolved), M&A/exec capex (IREN/Nvidia, Sony SSS/TSMC, Amazon/STM), Regulatory (quiet week)
  • Step 2 macro buckets: M&A partnerships + exec capex; regulatory antitrust
  • Podcast episodes transcribed: 0 (all 6 RSS feeds HTTP 403; 7th consecutive failure)
  • Sources promoted: 7; sources ingested: 7
  • New hypothesis chains opened: 3 (IREN DSX operator, silicon photonics networking, TSMC Kumamoto)
  • New entities created: 3 (SMIC, LSCC, IREN)
  • Calibration logged: 0 — China export control narrative is a framing reframe, not a verified-wiki-claim contradiction

Open questions for human review

  1. China H200 deal — will it ever close? Jensen Huang in Beijing; Trump involved; ~10 Chinese firms cleared. But Beijing's domestic-chip mandate appears structural. Does the B30A change the calculus? Nvidia's next earnings guidance will be the clearest signal.
  2. SMIC hypothesis worth pursuing? SMICY is a Chinese state-affiliated company with significant geopolitical/sanction risk. The structural thesis (China domestic AI chip = SMIC fab volume) is sound, but investability is constrained. Paul should decide whether SMICY makes sense to track.
  3. IREN valuation post-Nvidia announcement: IREN stock "climbed" but no price data available. Is the 5GW DSX thesis priced in? Need a price snapshot before forming a view.
  4. Qualcomm/AMD Samsung 2nm — will either deal close before Samsung's Q2 earnings (July)? This is the key evidence point for Samsung foundry becoming a genuine third option, not just a bilateral Intel/TSMC competition.
  5. Podcast feeds still blocked (7th run): The feed 403 pattern is consistent across all 6 feeds and all remote execution environments. Consider whether this is a network policy issue with the remote container that needs to be resolved at the infrastructure level.