The chips
Ten Chips, From AI Accelerators to a Phone SoC
Six AI accelerators from NVIDIA, AMD, Google, AWS and Microsoft, the Huawei Ascend 910C on China's own supply chain, Tesla's AI5, the SK hynix HBM4 stack inside most of them, and Apple's A19 Pro as the consumer counterpart.
The rack-scale rings on DatacentersX stop at the accelerator; each ring here opens it. Rows are banded by the most US status any supplier on that chip has.
| Chip | Designer | Logic node | Package | Memory | Rows | Best US status |
|---|---|---|---|---|---|---|
| NVIDIA Blackwell Ultra B300 | NVIDIA | TSMC 4NP | CoWoS-L | 288GB HBM3E 12-high | 11 | Some US volume |
| NVIDIA Rubin | NVIDIA | TSMC N3P | CoWoS-L | 288GB HBM4 | 12 | US planned |
| AMD Instinct MI455X | AMD | TSMC N2 (compute), N3P (base and I/O) | SoIC plus CoWoS-L | 432GB HBM4 | 11 | Made abroad |
| Google TPU Ironwood | Google and Broadcom | TSMC N3P | CoWoS | 192GB HBM3E | 8 | Made abroad |
| AWS Trainium3 | AWS (Annapurna Labs) | TSMC N3P | TSMC CoWoS-R, two assemblies | HBM3E 12-high, 144 GB | 8 | Made abroad |
| Microsoft Maia 200 | Microsoft | TSMC 3nm | CoWoS (inferred) | 216GB HBM3E | 6 | Made abroad |
| Huawei Ascend 910C | HiSilicon (Huawei) | SMIC N+2 (7nm-class); some TSMC 7nm dies | Two 910B dies, domestic advanced packaging | Stockpiled Samsung and SK hynix HBM; CXMT HBM3 next | 7 | Made abroad |
| Tesla AI5 | Tesla | Samsung SF2T (reported); TSMC node not disclosed | Not disclosed | None; LPDDR5X reported | 7 | US planned |
| SK hynix HBM4 12-high | SK hynix | 1b DRAM; TSMC 12nm-class base die | TSV stack, MR-MUF | 36 GB, 2,048-bit | 5 | US planned |
| Apple A19 Pro | Apple | TSMC N3P | InFO-PoP (inferred) | None; LPDDR5X | 5 | Made abroad |
US share
What Is Made, or Planned, in the US
- NVIDIA B300, logic die: TSMC Arizona, Fab 21 (some us volume). First Blackwell wafer at Fab 21 on Oct 17 2025; share of B300 volume not disclosed; wafers return to Taiwan for packaging
- NVIDIA Rubin, final test: KYEC (US site) (us planned). NVIDIA prepaid for KYEC US test capacity; high-volume test targeted April 2027
- Tesla AI5, logic die: Samsung Foundry, Taylor fab (us planned). Samsung SF2T (2nm) per SEDaily; tape-out done spring 2026; prototypes marked KR2613 came from a Korean line; volume around 2027
- Tesla AI5, logic die: TSMC, Fab 21 Arizona and Taiwan fabs (us planned). Musk (Nov 2025): slightly different AI5 versions at TSMC and Samsung; node not disclosed; Hankyung lists TSMC Taiwan and Arizona
- SK hynix HBM4, packaging: SK hynix, West Lafayette plant (us planned). Broke ground Aug 2026; HBM output 2H 2029
Everything else in the rings is made in Taiwan, South Korea, Japan or China, or its plant is not disclosed. The packaging that joins the dies stays in Taiwan until Amkor's Peoria, Arizona campus opens in 2028; see Foundry Captive Packaging.
Method
How the Rings Are Built
- Components: logic die, memory (HBM, or LPDDR for phone and car chips), HBM base die, interposer or bridge, package substrate, ABF film, advanced packaging and final test.
- Evidence: Confirmed (the company says so), Committed (an announced agreement), Claimed (credible press or analyst report) and Inferred (deduced from a supplier's known role; the reason is in the notes).
- US status: the same five levels as the DatacentersX rings: made in US, some US volume, US planned, made abroad and not disclosed.
Record
Dataset Record
- Dataset
- sx_chips (10 chips) and sx_chip_rings (80 supplier rows)
- Sources
- Company releases and filings, TrendForce, SemiAnalysis, DigiTimes, Tom's Hardware, Reuters; a source link on each row
- Last verified
- October 2026