Key takeaway
The other side of the $390 billion purchase commitment is six expansion agreements published on November 18. The money did not mainly go to wafers. It went to substrates and advanced packaging — the two tightest physical links in the Rubin cycle.
1. The event: six suppliers expand at once
On November 18, 2026, NVIDIA announced expanded cooperation with supply-chain partners TSMC, SPIL (Siliconware Precision Industries), Kinsus Interconnect Technology, Samsung Electro-Mechanics, Amkor Technology and Foxconn, covering substrates, advanced packaging, wafer fabrication and contract manufacturing.
The company had previewed the moves on its November 17 earnings call. Interim CEO Colette Kress said demand continues to exceed supply and that NVIDIA is working closely with partners to accelerate capacity in wafers, substrates and advanced packaging. She also noted that third-quarter purchase commitments rose to $390 billion, reflecting confidence in the long-term sustainability of demand for NVIDIA infrastructure.
2. Key facts
| Partner | Segment | Commitment | Capacity target | Timeline |
|---|---|---|---|---|
| Kinsus | Advanced ABF packaging substrates | Multi-year agreement (amount undisclosed) | NVIDIA-dedicated capacity to more than 3.5x current | Production starts H2 2027; capacity lands Q3 2028 |
| Amkor | Advanced packaging (Arizona, Korea) | Multi-year agreements totaling nearly $3 billion, including prepayments | Expand capacity for long-term production needs | Multi-year |
| SPIL | Advanced packaging (Taiwan) | Billions of dollars in funding support | Substantial increase in NVIDIA-dedicated advanced packaging capacity | Before mid-2029 |
| Samsung Electro-Mechanics | Substrates | Amendment to existing supply agreement (amount undisclosed) | Improve operations and output at existing fabs | Not specified |
| TSMC | Wafer fabrication and CoWoS packaging | Not disclosed | Supports Vera Rubin NVL72 and Vera CPU volume production | Already shipping in Q3 |
| Foxconn | Contract manufacturing | Not disclosed | Next-generation GPUs and networking platforms | Not specified |
| Purchase commitments | — | $390 billion in Q3 (prior quarter $279 billion) | — | — |
Additional context: Kinsus's expansion is based in Taoyuan and extends a partnership of more than a decade; both the Kinsus and Samsung Electro-Mechanics deals modify or extend existing arrangements; the Amkor agreements are separate from the Amkor, King Yuan Electronics and Unimicron cooperation announced in October.
3. Why the money went to packaging and substrates
NVIDIA's own language is that it is expanding "substrate, advanced packaging, wafer and contract manufacturing capacity." The dollar amounts tell a narrower story: only two partners received quantified funding — Amkor at nearly $3 billion and SPIL at "billions" — and both are advanced packaging. On the wafer side, the announcement cites TSMC's 5nm, 3nm and CoWoS technologies as enabling Rubin, without new funding.
This follows the bottleneck migration of the past two years. Compute chips are no longer constrained by lithography but by three physical links:
Advanced packaging (CoWoS). Rubin-class accelerators need CoWoS-type packaging to connect logic dies with HBM stacks. Packaging capacity directly caps how many accelerators can ship.
ABF substrates. Every high-performance chip needs large, high-layer-count ABF substrates, and such capacity has long lead times and slow yield ramps. Kinsus's 3.5x is the most concrete number in the announcement.
HBM supply. The earnings call attributed much of the purchase-commitment increase to multi-year memory agreements, pointing to HBM.
Of the three, NVIDIA is directly funding packaging and substrates.
There is a structural detail worth noting: Kinsus and Samsung Electro-Mechanics make substrates, while Amkor and SPIL do packaging. In traditional semiconductor division of labor these are separate steps with different equipment, materials and yield curves. On Rubin-class products, both must be ready at the same time for anything to ship. Locking both lines down on the same day suggests NVIDIA sees the constraint not as any single supplier's output but as the combined delivery cadence of substrate plus packaging. That is also why the announcement does not single out one partner's dollar figure — missing either link makes the rest irrelevant.
4. Timeline versus product cycle
Lay the four timelines side by side and a dividing line appears:
- Kinsus: production starts H2 2027, capacity lands Q3 2028;
- Amkor: multi-year, no single date;
- SPIL: expansion complete before mid-2029;
- TSMC: already supporting Rubin production today.
The Rubin systems shipping now use existing capacity. The newly signed capacity only begins arriving after the second half of 2027. This expansion is not for the current generation; it is for platforms in 2028-2029. That matches the time structure of the $390 billion purchase commitment, which covers future years rather than the current quarter.
5. Reading the three buckets of money
First, purchase commitments are obligations, not spending. The $390 billion is a total commitment for future purchases that flows into cost of goods as deliveries occur; it stood at $279 billion last quarter.
Second, funding support can take several forms. The Amkor agreements explicitly include prepayments; the SPIL arrangement is described as billions of dollars in funding support. Prepayments typically sit in prepaid assets or long-term purchase arrangements. The specific line item depends on upcoming quarterly disclosures, and should not be read as current-period expense.
Third, expansion costs pressure gross margin first. Fourth-quarter gross margin guidance is 71%-72%, which management attributes partly to the initial ramp of the Rubin platform. New line commissioning, yield improvement and parallel capacity all show up as cost in these quarters.
The timing gap deserves attention: cash begins flowing in 2026, but capacity only forms after the second half of 2027. In the interim these outlays do not convert into revenue; they appear as prepayments, depreciation provisions or gross-margin pressure. That window is the test of whether the spending pays off — if demand holds in 2028, the upfront cost is absorbed; if demand slows, the early commitments become idle capacity.
6. The immediate supply-chain reaction
In early trading on November 19, advanced packaging names on the A-share market jumped: Dagang Co. hit the daily limit, while Canxin, Lanjian Electronics and others rose more than 10%, with Yongsi, Xinqi Micro, Kangqiang and JCET following. The direct catalyst was the Kinsus 3.5x headline.
Note that A-share names do not map one-to-one onto NVIDIA's direct suppliers; most of the move is sentiment and expectation rather than booked orders.
7. Boundaries
- The six agreements, expansion targets, dollar figures and timelines come from NVIDIA's November 18, 2026 announcement, cross-checked against Cailianshe and Sina Finance reports;
- Kinsus and Samsung Electro-Mechanics amounts were not disclosed and are marked as such, with no estimates added;
- "Nearly $3 billion" and "billions of dollars" are the announcement's own phrasing, not precise contract values;
- Purchase commitments are total future obligations, not current-period spending;
- A-share moves reflect market reaction, not confirmed supply relationships with NVIDIA;
- This article makes no valuation judgment and no price forecast.
8. Sources
- NVIDIA official announcement: NVIDIA Blog
- Earnings call and agreement details (Cailianshe): Sina Finance
- Amounts and timelines: Sina Finance (Caiwen)
- A-share packaging reaction: Cailianshe
