2026-09-27 — China signals it may let ByteDance and Alibaba buy Nvidia’s RTX PRO 5500

• China’s industry ministry has signaled it intends to approve purchases of Nvidia’s RTX PRO 5500 chip by firms including ByteDance and Alibaba, The Information reported on 2026-09-27, citing two people familiar. The ministry asked the companies to report their purchase plans, and some executives expect the chip to fall outside US export restrictions. Reuters said it could not immediately verify the report, and an Nvidia spokesperson said US firms remain restricted by outdated US export controls and by China’s own import limits. (CNBC, Reuters)

2026-09-02 — Broadcom’s Q3 print puts the fourth name in this thread’s own watch line on the record

• Broadcom reported Q3 FY2026 results on 09-02 (quarter ended 08-02): AI semiconductor revenue of $16.7bn, up 221% YoY and 54% QoQ, on total revenue of $29.6bn (+86% YoY). CEO Hock Tan guided Q4 FY2026 AI semiconductor revenue to $21.7bn (+236% YoY) and, on the earnings call, put fiscal-2027 AI semiconductor revenue at roughly $115bn — the clearest fresh number this thread has for the “Broadcom” name in its own watch line ("$ flowing to Nvidia/AMD/TSMC/Broadcom"), which until now had only carried the generic custom-silicon framing from the 07-27 crawl. ⚠️ The FY2027 $115bn figure is earnings-call commentary, not a filed number — the press release itself gives only the Q4 guide. (Broadcom investor relations — primary, CNBC)

2026-08-28 — A venture firm raises a fund aimed at the supply-chain limits this thread tracks

• Andreessen Horowitz announced a $1.1bn “Machine Age Fund” on 08-28 to back AI hardware and infrastructure — chips, memory, networking, storage and full systems including data centres, robotics and home AI appliances — framed on its own site around every layer of the AI stack “hitting the limits of today’s supply chain capability.” New capital aimed squarely at the compute and memory scarcity Memory Squeeze tracks from the price side, and a venture-scale vote that the bottleneck is physical, not model-side. Source: Axios Pro Rata’s 08-28 fundraising lead. a16z is not a watchlist entity, so this carries no entity tag. (a16z, fund page, Bloomberg)

2026-08-26 — Nvidia’s own memory-supply commitments more than double to $279bn

• Nvidia’s Q2 FY2027 10-Q and CFO commentary (both dated to the 08-26 earnings release) disclose total purchase and supply/capacity commitments of $279bn, up from $119bn the prior quarter — split $92bn due in the rest of FY2027, $87bn in FY2028 and $88bn in FY2029 — with the company describing the jump as “primarily related to the procurement of memory.” This is the dollar figure behind this thread’s 08-22 finding that pricing power inside the chip flow moved upstream to the memory makers: Nvidia is now locking in HBM/DRAM supply years out at a scale roughly 7x SK hynix’s entire two-fab, $38.1bn buildout logged here 08-07. ⚠️ Dating note: the figure sat inside Nvidia’s own 08-26 filing/call alongside the guarantees and receivables-concentration facts already logged in Nvidia’s Order Book’s 08-26 entry, but this specific number wasn’t picked up in trade coverage until 08-27-29 — the underlying disclosure is 08-26, this map’s pickup is 08-29. (SEC EDGAR, NVDA 10-Q, TrendForce)

2026-08-27 — A neocloud shows the price of GPU debt, three points apart on the same collateral

• IREN guided to $25-30bn of fiscal-2027 capital expenditure and disclosed the stack that funds it. FY26 results, released after the close and filed on an 8-K the same day, show AI Cloud revenue of $70.5M for the year against $4bn of contracted annualised recurring revenue for 2026 capacity, $1bn already operating after Microsoft accepted its Childress, Texas “Horizon 1” site. The financing: a $3.6bn investment-grade GPU facility at 6.0% tied to the Microsoft contract, plus $2.8bn of new GPU debt including a $2.4bn tranche led by Blue Owl and PIMCO at 9.0% for its Mackenzie site, ending the year with $14bn liquidity. The 300-basis-point spread is the entry: same borrower, same collateral class, priced three points apart on whether a hyperscaler contract sits behind it. That is the market pricing counterparty quality, not GPUs. ⚠️ Forward guidance plus freshly closed debt — not cash already spent. (GlobeNewswire)

2026-08-22 — Nvidia raises AI server prices more than 15%, and blames memory

• Nvidia notified the contract manufacturers building AI servers for Microsoft, Google and Oracle that prices rise more than 15% in many cases on systems shipping in early 2027, including Vera Rubin and Grace Blackwell configurations, with the increase varying by chip generation and memory configuration. The stated cause is soaring DRAM and HBM prices, with Samsung, SK Hynix and Micron holding the leverage as demand outruns supply. This thread has tracked the dollars flowing to Nvidia, AMD, TSMC and Broadcom; the new fact is that the pricing power inside that flow has moved upstream of all four, to the memory makers. (Bloomberg, Fortune, CNBC)

2026-08-07 — SK hynix approves $38.1B for two new AI-memory fabs

• SK hynix’s board approved 54 trillion won (~$38.1B) for two new AI-memory fabs — 35.2 trillion won for “Y2” (DRAM, Yongin, cleanroom opening June 2029) and 19.1 trillion won for “M17” (NAND, Cheongju, cleanroom December 2028) — explicitly framed as securing mid-to-long-term production capacity for AI memory demand, directly on this thread’s own watch line ("$ flowing to Nvidia/AMD/TSMC/Broadcom"). (SK hynix Newsroom)

2026-08-04 — AMD’s Q2 print: data-center revenue more than doubles, stock falls anyway

• AMD reported Q2 2026 revenue of $11.536B (record, ~+50% YoY); data-center segment revenue $6.718B — up 107% YoY, now 58% of total revenue — the clearest fresh number this thread has on where AI silicon dollars land outside Nvidia, directly on this thread’s own watch line ("$ flowing to Nvidia/AMD/TSMC/Broadcom"). Q3 guidance ~$13B (±$300M, ~41% YoY, ~13% sequential). Stock fell ~7-9% after-hours despite the beat — a sell-the-guide reaction, not a demand-miss one. (AMD IR · CNBC/Stock Titan corroborating)
• CEO Lisa Su: “Instinct deployments scale and Helios begins to ramp” — the call highlighted MI400-series launches (MI455X, MI430X) and MI350P, and reiterated Anthropic’s commitment to deploy up to 2GW of MI450-series GPUs in AMD Helios racks. That Anthropic commitment itself reads as restated on the call, not newly announced this quarter — its original announcement date wasn’t independently re-confirmed this pass, flagged rather than dated as new. (AMD IR)
• 💡 Adjacent, not added to this thread’s record: multiple secondary outlets (TechPowerUp, Tech Times, biggo.com) reported 2026-08-06 that TSMC is sitting on roughly $1B / NT$32B+ of finished Apple 2nm iPhone-18 chips it cannot ship because of a DRAM shortage the same coverage frames as AI-demand-driven — i.e. the AI memory crunch spilling into a non-AI consumer supply chain. Could not confirm via a primary source or a reputable direct fetch (TrendForce’s own news page has no matching item, direct fetches to the secondary outlets 403’d) — noted here as a lead worth a primary-sourced check on a future pass, not logged as a timeline fact.

2026-07-30 — Samsung’s earnings put a number on the memory side of the silicon spend

• Samsung posted record Q2 chip-division profit (~250x YoY per multiple outlets), driven by AI-linked HBM/DRAM shortage pricing — the clearest fresh data point on where AI silicon dollars land outside the GPU line itself. ([CNBC] · [WSJ])
• Samsung locked ~70% of memory output into long-term deals with “5 major Big Tech firms” (unnamed) and sees the shortage extending to 2028 — direct evidence of hyperscalers contracting AI memory supply years out. ([finance.biggo.com])

2026-07-27 — The silicon split: undisclosed at all four

• The industry’s common non-disclosure: none of Google / Microsoft / Amazon / Meta publishes its own-chip vs Nvidia dollar or unit split — confirmed independently by all four capex crawls.
• Own-silicon state: Google TPU v7 “Ironwood” (+8t/8i next-gen) · Microsoft Maia 200 (dual-purpose: OpenAI inference + MAI training) · Amazon 1.4M Trainium deployed, T3 claims ~50% cost edge · Meta MTIA at “millions” of H100-equivalents (600K is stale), “Iris” to production Sept-26, Broadcom extended thru 2029.
• Clearest external volume signals: Anthropic on 1M+ Google TPUs (>1GW) AND 1M+ Trainium2 — the labs disclose what the hyperscalers won’t.

2026-07-24 — Opened

• Thread opened — meta-thread — groups the child threads below.

← Backstory