Chips
Designers, intermediaries, and memory — the silicon layer.
The metrics
- Posture
- expanding — 7 of 8 (expanding 7 · hedging 1) 8 members →
- Capital · available
- across 8 members — open each for the figure + sources 8 members →
- Capital · operating
- across 8 members — open each for the figure + sources 8 members →
- Capital · deployed
- across 8 members — open each for the figure + sources 8 members →
- Capital · in
- across 8 members — open each for the figure + sources 8 members →
- Capital · out
- across 8 members — open each for the figure + sources 8 members →
- Optionality
- across 8 members — open each for the figure + sources 8 members →
- Gravity
- across 8 members — open each for the figure + sources 8 members →
What's new this week
- Spot HBM is trading at four to five times long-term contract prices. Korean trade press puts spot HBM3E (36GB) near $2,100 against contract-equivalent pricing of roughly $350–490, and HBM4 (16-layer) spot near $3,500. The attributed causes are HBM4's lower yield consuming more DRAM wafer capacity per unit, and Samsung's already-recorded July commitment of 70% of capacity to long-term agreements starving the spot market. The report adds that Nvidia, Microsoft and Google reportedly cannot get their full contracted volumes. ⚠️ Figures are converted from Korean won — directional, not exact. (Seoul Economic Daily)
- 🕰 IBM built native Arm support into its next mainframe chip, confirmed from both companies' own newsrooms — alongside Arm shares moving on a reported $2bn AI-chip backlog. Dated 08-24, caught by the cold rotation. ⚠️ The backlog figure and the mainframe design win are separately sourced; only the latter is confirmed from primary channels.
- 🕓 CXMT has begun small-batch production of HBM3E — the high-bandwidth memory Nvidia H200/Blackwell-class accelerators are built around — the first time a Chinese maker has produced it at all. The Information, citing two insiders, reports Alibaba's T-Head chip-design unit and Cambricon are already testing the silicon for integration into commercial processors as early as 2027. The qualifiers matter as much as the fact: yields are around 25%; CXMT is assessed three to five years behind Samsung, SK Hynix and Micron, who are already in HBM4 mass production a full generation ahead; and CXMT's own ~$8.6bn Shanghai IPO prospectus reportedly earmarked no capital specifically for HBM. Read against this morning's two CXMT items, the picture changes shape. The morning had CXMT beating incumbent gross margins (87.59%) and first to mass-produce LPDDR6. This adds that the same company has now entered the one memory category export controls were aimed most squarely at keeping from it — at a yield that makes it a demonstration, not yet a supply source. (the-decoder, Investing.com)
- Nvidia is investing $3.5bn in MediaTek via convertible bonds — ~90% of MediaTek's entire $3.9bn overseas offering — its largest direct investment outside the US, deepening a partnership built around NVLink Fusion, which lets MediaTek's own custom silicon plug into Nvidia's rack-scale architecture. The bonds convert to shares later rather than a straight equity buy today, extending the stake-ladder pattern this thread tracks (Nebius, Naver, Intel, Groq) to a chip-design partner rather than a compute customer. MediaTek shares rose as much as 10%. (Bloomberg, TechCrunch)
- Anthropic signed a $35bn cloud deal with Lambda, a Nvidia-backed provider, for capacity at a Texas data center where Nvidia itself holds the lease — Nvidia supplies the chips, has invested in the cloud middleman, and now sits in the real-estate chain of a lab it has no direct equity stake in. Anthropic's second such deal this month, after $45bn with Nscale, as it works through a compute shortage. Reported via WSJ, corroborated by Reuters/Bloomberg; neither party has confirmed it officially. (Bloomberg)
- Nvidia opened higher while the broad market opened lower, which is the first real evidence on what Friday's selloff was actually about. NVDA opened around $219, up roughly 0.3–0.7% from Friday's $217.55 close, while the S&P 500 fell 0.45% to 7,677, the Dow 0.62% to 53,229 and the Nasdaq 0.43% to 26,290, on the Larak/Jordan/UAE exchange plus continued rate-hike repricing. This thread has carried two competing explanations for Friday's 4.74% Nvidia session — the vendor-financing pullback, or the hawkish Fed repricing. If financing were still the dominant driver, Nvidia would be expected to keep lagging; instead it is the outperformer against a macro-driven dip. ⚠️ One premarket and early session is not conclusive, and Micron, Intel and Broadcom are described in same-day coverage as choppy rather than clean movers either way. (CNBC, Yahoo Finance)
- Nvidia and the semiconductor index kept outperforming the tape into the afternoon, with Nvidia +0.62% and the SOX +0.06% against a −0.42% to −0.63% broad market. This strengthens the morning's own discriminator: if Friday's chip selloff had been principally about Nvidia's vendor-financing pullback, chips would be lagging on a risk-off day rather than leading it. They are leading it.
- Nvidia is investing $3.5bn in MediaTek via convertible bonds — ~90% of MediaTek's entire $3.9bn overseas offering, and Nvidia's largest direct investment outside the US. Extends the stake-ladder pattern this thread tracks (Nebius, Naver, Intel, Groq) to a chip-design partner rather than a compute customer; MediaTek shares rose as much as 10%. (Bloomberg)
- Anthropic signed a $35bn cloud deal with Lambda, a Nvidia-backed provider, for capacity at a Texas data center where Nvidia itself holds the lease — a fourth named entity (after AMD, Nvidia, OpenAI/Oracle) in this thread's circularity loop: Nvidia supplies chips, invests in the cloud middleman, and now holds real estate serving a lab it has no direct stake in. Second such deal this month after $45bn with Nscale. (Bloomberg)
- AMD, Cisco and HUMAIN (Saudi Arabia's PIF-backed state AI vehicle) brought their jointly-built AI infrastructure online in Saudi Arabia, per AMD's own investor-relations release. Published 7:15am ET 08-31, before this digest's cutoff, and missed until this catch. (AMD IR, primary)
- SB Energy, the SoftBank/OpenAI vehicle behind Nvidia's up-to-$105bn Ohio lease guarantee, filed for a US IPO — the first public numbers on its own book: a $439bn backlog against a $3.21bn six-month net loss. Nvidia is separately investing $1.5bn in the IPO itself. Full write-up in the global-capital digest.
- Bloomberg moved the Nvidia–Hugging Face deal from "agreed" to "could close this week," and raised the total to about $14bn. The base price is unchanged from the $12.9bn this map logged on 08-27; the new figure adds a reported $1bn employee-retention package on top. Both companies again declined to comment and no final agreement is confirmed by either side. What changed is the timeline, from open-ended to a specific week. (Bloomberg via Investing.com, timestamped 09-01 8:56pm ET)
- Reading SB Energy's actual S-1 rather than the first-day wire coverage doubled Nvidia's committed investment and reversed a correction this map made itself. The filing discloses $3.0bn total from Nvidia, in two instruments dated the same day (08-17): a $1.5bn Prepaid Forward Contract already funded and already spent, plus a $1.5bn Class N private placement closing with the IPO. This map's own 08-17 entry, sourced to Nvidia's press release, read the Prepaid Forward Contract as the whole story and explicitly "corrected" weekend reports of "up to $3B" as pre-announcement talk. The weekend reports were closer to right and the correction was the error — Nvidia's release described only one of two simultaneous instruments. Both affected timeline entries now carry the correction. (SEC EDGAR S-1, accession 0001628280-26-059639)
- SB Energy Corp — the SoftBank- and OpenAI-backed developer of the Ohio "PORTS" campus that Nvidia's up-to-$105bn lease guarantee underwrites — filed for a US IPO, disclosing a $439bn contracted backlog against a $3.21bn net loss on just $138.7m of revenue for the six months ended June 30, 2026 (versus a $215.5m loss on $83.3m revenue a year earlier). Nvidia has separately committed to invest $1.5bn in the IPO itself via private placement at the offer price. First time SB Energy's own numbers — not just the guarantee figure — have been public: a company with no data center yet online carrying a backlog more than 100x its trailing revenue. JPMorgan, Goldman Sachs, Morgan Stanley, Citigroup and Mizuho lead the offering; plans to list on Nasdaq as "SBE." (Bloomberg)
Members 9
capital ~$200-300B commandable/12mo (liquid $80.6B Apr-2026 + ~$119B/yr FCF + near-zero net debt); market cap ~$5T is a …
- opt
- free — no fiduciary mandate, near-zero leverage; ~$30B …
- grav
- ~$1.0T/yr GROSS orbiting AI-infra (Nvidia DC run-rate + the …
New 09-04: the loop got a filed number. Nvidia's own 10-Q puts its TOTAL EQUITY-INVESTMENT PORTFOLIO at ~$99bn, up from ~$7bn a year ago — a …
capital ~$50-60B commandable/12mo ($19.6B cash + ~$33B/yr FCF) but net debt ~-$45B (VMware) pre-commits much of it; …
- opt
- mixed-to-constrained — free cash earmarked to deleveraging …
- grav
- ~$63.9B/yr GROSS FY25 revenue — hyperscaler custom-AI ASICs …
New 09-02 (recorded 09-03): FISCAL Q3 PRINT — revenue $29.6bn (+86%), AI semiconductor revenue $16.7bn (+221%), both beats; Q4 guide $34.8bn …
capital ~$70B commandable/12mo ($39B cash+STI + ~$30B/yr FCF); market cap ~$930B is a separate label
- opt
- mixed — capex-cyclical; SK chaebol control
- grav
- ~$70-80B/yr GROSS direct (the HBM the whole buildout needs; …
New 09-04: deferred again, on its own clock. A Form 6-K filed with the SEC says "no matters have been determined" on a rumored 5-10 trillion …
capital ~$50-55B commandable/12mo ($26B cash+STI + ~$26B/yr FCF, net cash $24B); market cap $1.04T is a separate label
- opt
- mixed — FCF earmarked to US megafabs + CHIPS
- grav
- ~22% of DRAM revenue GROSS across DC+AI+PC+auto (DRAM/NAND); …
New 08-23: Micron put a number on the memory wall at Hot Chips 2026 (Stanford, 08-23 to 08-25, day one). Raghu Sreeramaneni's "Evolving …
capital state-backed IPO ~$8.5-9.8B fresh + Big Fund; market cap ~$524B is the mark (separate label) — corrected …
- opt
- locked — state-directed (36.29% state pre-IPO, …
- grav
- ~7.6% of DRAM revenue GROSS — China's domestic memory …
New 08-31 PM: began SMALL-BATCH PRODUCTION OF HBM3E, the high-bandwidth memory Nvidia H200/Blackwell-class accelerators are built around, …
capital ~$25-35B commandable/12mo ($12.35B cash + ~$8.6B/yr FCF, net cash $8.5B); market cap $851B is a separate label
- opt
- mostly free — net cash; self-earmarked to AI-GPU R&D + …
- grav
- GROSS — x86 server + PC + DC-GPU on AMD silicon (AMD-claimed …
New 08-29 (cold rotation, caught late): two developments the map missed while the thread was quiet. At Hot Chips on 08-24 AMD disclosed hard …
capital ~$25-30B commandable/12mo ($9.8B cash + ~$12.5B/yr FCF); market cap $176B is a separate label
- opt
- mostly free
- grav
- GROSS — premium-Android ecosystem (Snapdragon; ~39% …
New 08-16: no further movement all week (08-10→08-16) — no Qualcomm-specific development since 08-05 — the only mention across this week's …
capital ~$60B total assets (€54.7B, Jun-2026); market cap is a separate label
- opt
- free — strong balance sheet, little encumbrance
- grav
- GROSS — sole EUV supplier; no leading-edge chip is made …
The EUV monopoly — the tightest chokepoint on the board; rentier with an emerging AI-stakes tilt (Mistral 11%).
capital ~$5B/12mo ($3.6B cash + ~$1B/yr FCF) — TINY vs its gravity; market cap $278B is a separate label
- opt
- constrained — 87.1% SoftBank-controlled
- grav
- GROSS — ISA under ~99% of smartphones + DC (Grace/Graviton) …
New 08-24: IBM built native Arm support into its next mainframe chip, confirmed from both companies' own newsrooms — a design win in the one …
