NVIDIA

Blackwell and Rubin · TSMC 4NP and N3P · CoWoS-L · HBM3E and HBM4

NVIDIA earned $89.0 billion from data center chips in the quarter ended July 2026, and every one of them passed through TSMC packaging in Taiwan. Blackwell wafers have been made in Arizona since October 2025, but each die still travels about 11,500 km to Taiwan for CoWoS packaging; no US packaging line opens before 2028.

The six steps that make an NVIDIA Blackwell Ultra B300 GPU and their US status. The logic die has some US volume at TSMC Fab 21 in Phoenix; HBM3E from SK hynix, Micron and Samsung, CoWoS-L packaging at TSMC in Taiwan, Ibiden substrates with Ajinomoto film from Japan, and final test at KYEC and ASE in Taiwan are made abroad; rack assembly has some US volume reported in Texas. An Arizona-made die travels about 11,500 km to Taiwan for packaging. ONE B300 GPU: SIX STEPS, ONE WITH US WAFERS 1 Logic die TSMC FAB 21 PHOENIX; FAB 18 TAINAN 4NP, TWO RETICLE-SIZE DIES 2 HBM3E, eight stacks SK HYNIX, MICRON, SAMSUNG KOREA, JAPAN, TAIWAN 3 CoWoS-L packaging TSMC, TAIWAN ONLY NO SECOND SOURCE 4 Substrate and film IBIDEN SUBSTRATE, AJINOMOTO ABF JAPAN 5 Final test KYEC AND ASE, TAIWAN KYEC US SITES PLANNED 6 Rack assembly FOXCONN, WISTRON, QUANTA TEXAS FROM 2026 (REPORTED) PHOENIX WAFERS TAIWAN PACKAGE AND TEST TEXAS, MEXICO RACKS About 11,500 km DIE FLOWN OUT FOR COWOS FINISHED GPU OR BOARD SOME US VOLUME US PLANNED MADE ABROAD STEPS FROM THE SEMICONDUCTORX B300 SUPPLIER RING. DISTANCE PHOENIX TO HSINCHU, GREAT CIRCLE, COMPUTED. SCHEMATIC, NOT TO SCALE.
Revenue, Q2 FY2027
$96.2B
Data center
$89.0B
CoWoS share, 2026
~60% (reported)
HBM4 suppliers
3
US-made steps
Wafers only
Hub
Spotlights

Route

An Arizona Blackwell Die Still Crosses the Pacific to Be Packaged

  • Some US volume
  • Made abroad
NVIDIA Blackwell Ultra B300 supplier ring
ComponentSupplierRoleWhereUS status
Logic dieTSMCPrimaryFab 18, Tainan, TaiwanMade abroad
Logic dieTSMC ArizonaSecond sourceFab 21, Phoenix, USASome US volume
HBMSK hynixPrimaryM16, Icheon, South KoreaMade abroad
HBMMicronSecond sourceHiroshima and Taichung, Japan and TaiwanMade abroad
HBMSamsungSecond sourcePyeongtaek, Pyeongtaek, South KoreaMade abroad
Interposer or bridgeTSMCSoleCoWoS-L lines, TaiwanMade abroad
Advanced packagingTSMCSoleCoWoS-L lines, TaiwanMade abroad
Package substrateIbidenPrimaryOgaki area, JapanMade abroad
ABF filmAjinomotoSoleKawasaki area, JapanMade abroad
Final testKYECPrimaryHsinchu and Miaoli, TaiwanMade abroad
Final testASESecond sourceKaohsiung, TaiwanMade abroad
Placeholder: rows from sx_chip_rings; full ring on B300 Supplier Ring

TSMC began Blackwell wafer production at Fab 21 in Phoenix in October 2025, the first NVIDIA flagship silicon made in the United States. The wafers still return to Taiwan, because CoWoS-L, the packaging that joins the two logic dies to eight HBM stacks on a silicon bridge interposer, runs only at TSMC plants there. Of the 11 supplier rows in the site's B300 ring, the Arizona logic die is the only one with US volume. Every flagship NVIDIA data center GPU since the P100 in 2016 has been packaged on CoWoS in Taiwan.

The ring for the next generation, Rubin, has no row with US volume yet, only a planned KYEC test site in the US: its N3P logic dies come from Taiwan until Fab 21's N3 phase runs.

Chokepoints

Three HBM Suppliers, One Packager

Number of qualified suppliers for each input to an NVIDIA Rubin GPU, from the SemiconductorX Rubin supplier ring: logic die one (TSMC), CoWoS-L packaging one (TSMC), ABF film one (Ajinomoto), package substrate one (Ibiden, inferred), final test two (KYEC, ASE), HBM4 three (SK hynix, Samsung, Micron). QUALIFIED SUPPLIERS PER INPUT, RUBIN GPU Logic die TSMC CoWoS-L packaging TSMC ABF film AJINOMOTO Package substrate IBIDEN (RUBIN AWARD INFERRED) Final test KYEC, ASE; KYEC US PLANNED HBM4 SK HYNIX, SAMSUNG, MICRON DARK: ONE SOURCE. ALL SIX INPUTS ARE MADE OUTSIDE THE US TODAY.

NVIDIA can buy HBM from three memory makers, and since 2026 SK hynix, Samsung and Micron all ship HBM4 for Rubin. It has one source for everything that turns those parts into a GPU: TSMC for the logic die and the CoWoS-L package, Ajinomoto for the build-up film in the substrate. Reported allocations give NVIDIA about 60% of TSMC's CoWoS output in 2026, from capacity estimated at 120,000 to 130,000 wafers a month by the fourth quarter; both figures are industry estimates, not TSMC disclosures.

  • CoWoS: the step NVIDIA cannot route around. No other company runs CoWoS-L at volume, and no CoWoS-L line outside Taiwan opens before 2028. See CoWoS.
  • HBM: three suppliers, but all three are capacity-constrained and none makes HBM in the US yet. See HBM.
  • Substrates: Ibiden is NVIDIA's main AI substrate supplier; its award for Rubin is inferred, not confirmed. See Substrates and Interposers.

Beyond TSMC

Samsung Makes NVIDIA's First Non-TSMC Platform Chip

NVIDIA silicon made outside TSMC
ProductMakerStatusSupply-chain point
Groq 3 LPX inference chipSamsung Foundry, KoreaFull mass production announced August 2026First NVIDIA platform chip of the Blackwell and Rubin era built outside TSMC; came with the Groq technology deal of December 2025 (reported at about $20 billion)
HBM4 for RubinSK hynix, Samsung, MicronShipping in 2026Three-way memory sourcing; SK hynix reported as the largest share
x86 CPUs and RTX SoCs with IntelIntelCo-development after NVIDIA's $5 billion Intel stake (2025)Gives NVIDIA a path to Intel Foundry and Intel packaging; no NVIDIA wafers at Intel announced

None of these move the core GPU. Every Blackwell and Rubin data center GPU is a TSMC logic die in a TSMC package, and the next generations on NVIDIA's roadmap, Rubin Ultra and Feynman, are also planned on TSMC nodes.

US build-out

US Packaging Arrives in 2028, Three Years After US Wafers

NVIDIA US manufacturing build-out by step. Wafers: Blackwell wafers at TSMC Fab 21 in Phoenix from October 2025, N3 phase targeted around 2027. Packaging: Amkor Peoria from early 2028 and TSMC Arizona CoWoS before 2029. Final test: KYEC US sites planned, prepaid by NVIDIA. Racks: Foxconn in Houston and Wistron in Texas from 2026, reported. NVIDIA IN THE US, STEP BY STEP 2025 2026 2027 2028 2029 Wafers TSMC FAB 21: BLACKWELL, OCT 2025 FAB 21 N3 PHASE (TSMC TARGET) Packaging AMKOR PEORIA, EARLY 2028 TSMC AZ COWOS BY 2029 Final test KYEC US SITES (NVIDIA PREPAID) Racks FOXCONN HOUSTON, WISTRON TEXAS SOME US VOLUME US PLANNED
  • Some US volume
  • US planned
NVIDIA US manufacturing sites
StepPartner and siteStatusUS status
Logic wafersTSMC Fab 21, Phoenix, ArizonaBlackwell wafers since October 2025Some US volume
CoWoS packagingAmkor, Peoria, ArizonaWorking with NVIDIA and Apple; production targeted early 2028US planned
CoWoS and 3D-IC packagingTSMC Arizona packaging plantConstruction begun; TSMC goal is before 2029US planned
Final testKYEC US sitesNVIDIA advancing about $1.1 billion; sites not disclosedUS planned
Server and rack assemblyFoxconn, Houston; Wistron, TexasPlants announced April 2025 for 2026 production; volume not disclosedSome US volume
Placeholder: HBM in the US: SK hynix West Lafayette, Indiana packaging plant (planned)

In April 2025 NVIDIA said it would produce up to $500 billion of AI infrastructure in the United States over four years. The plan covers every step except two that have no US date: HBM stacking and substrate making. Until the Arizona packaging lines open in 2028, a US-made NVIDIA GPU means a US wafer packaged in Taiwan. See US Reshoring.

China

NVIDIA Plans on No China Data Center Revenue

China, including Hong Kong, was about 13% of NVIDIA revenue in fiscal 2025, before the April 2025 license requirement on the H20, its China-only data center GPU, which cost NVIDIA a $4.5 billion charge in the first quarter of fiscal 2026. NVIDIA's outlook for the quarter ending October 2026 assumes no data center compute revenue from China at all. The gap is being filled by Huawei's Ascend line, made at SMIC; see Huawei and HiSilicon and China Bifurcation.

Record

Page Record

Covers
Where NVIDIA's Blackwell and Rubin GPUs are made, single-source inputs, silicon made outside TSMC, US manufacturing sites and China exposure
Sources
NVIDIA second-quarter fiscal 2027 results (August 2026); SemiconductorX B300 and Rubin supplier rings (sx_chip_rings, October 2026); The Register and iConnect007 on Arizona Blackwell wafers (October 2025); Reuters via WHBL on TSMC Arizona packaging (April 22, 2026); Silicon Analysts on CoWoS capacity and allocation (September 2026, estimates); Design & Reuse on Groq 3 LPX at Samsung (August 31, 2026); NVIDIA US manufacturing announcement via National CIO Review (April 15, 2025); FreightWaves on Foxconn Houston (June 2025); Phoenix to Hsinchu distance computed
Last verified
October 2026