Capacity Is the New Chip War's Real Battleground

Economists separate the short run, in which capacity is fixed and firms make do, from the long run, in which plants are built and structures adapt. Nvidia is in the short run, and the evidence says so. Demand is described as strong ⁶⁷. Nvidia is said to be selling chips faster than it can produce them ⁵⁴, and the source material notes that supply capacity may constrain shipments ⁵. Jensen Huang put it plainly: "Capacity is the constraint." ⁶³

Three questions follow from that. What is Nvidia doing to relax the constraint? Which dependencies does that effort leave untouched? And in the one large market where it cannot freely sell, China, what do enforcement and substitution show about how well the controls work? The evidence is uneven in quality, so I note the source count and the degree of confirmation as I go.

Buying Down the Back End

Nvidia's response is to pay in advance for the stages of production that bind. At GTC Berlin, Nvidia and six supply-chain partners announced capacity expansions ⁶³. The partners are Amkor Technology, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, SPIL and Unimicron ⁶⁶. Together they cover substrates, interposers, optical engines and testing. Nvidia describes the multi-year agreements as meant to strengthen the resiliency and redundancy of its global supply chain ⁷.

The best-corroborated commitments, each reported by four or five sources, are these:

• Interposers. Five sources report that Nvidia and Amkor are targeting an increase of more than 300% in interposer wafer capacity by the end of 2028 ^{7,15,34,35,66}. The work runs through advanced packaging expansion in South Korea ^15,34,41, and Amkor's October 21 announcement included a new $1.3B Nvidia agreement supporting it ^64,73.
• Optical engines. Five sources report that SPIL is targeting a 15-fold increase in optical-engine capacity for co-packaged optics by mid-2027 ^7,26,34. The partners plan another doubling by the end of 2028 ^15,41.
• Substrates. Agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron are expected to drive a combined 2.5x increase in capacity over three years ^34,63. Unimicron's new capacity begins in Q2 2027 ^15,35 and reaches full capacity by mid-2028 ^15,35.
• Testing. Four sources report a $1.1 billion capacity reservation advance to KYEC for new US testing facilities and equipment ^7,34,41,66. The expansions are meant to double total testing capacity in 2027 versus 2026 ^15,35. Expanded Singapore lines are expected to reach high-volume production in February 2027 ^34,41, and the US facilities are targeted for April 2027 ²⁷. That US date is a target, not existing capacity ²⁷.

The pattern is that Nvidia is not buying finished chips. It is reserving the stages around the chip, and Nvidia says the agreements will reserve production for it ²⁵. The scale of the advances suggests it expects demand to persist, though that is my inference and not a sourced finding.

Three qualifications matter. First, these are commitments to build, not capacity in hand. Nvidia has given no timeline for how much they will ease the constraints ²⁵. Second, the terms are patchy. Nvidia's announcement gives no starting capacity figure for substrates ²⁵, and the size of the Samsung Electro-Mechanics arrangement is undisclosed ⁶⁶. A single source's claim that Tesla and Nvidia are investing $2.8 billion in Samsung Electro-Mechanics ²⁰ should be treated as unconfirmed. Sources also disagree on when the KYEC payment was agreed: one says October 13 ¹⁵, another places the announcement around October 21 ⁷². Third, some of the broader scarcity narrative comes only from commenters. Examples are claims that Nvidia has absorbed much of wafer production for 3–5 years ⁵⁰ and that its chips are sold out for years ⁵¹. I treat those as unverified.

Why the Back End, and Why It Is Not Sufficient

The choice of target is explicable. Advanced packaging such as TSMC's CoWoS was a severe chokepoint through 2024 and 2025 ³³. Packaging problems at TSMC limited Blackwell supply ³⁶, and tight limits on FC-BGA substrates have held back full GPU cluster assembly ²⁶. IDC expects CoWoS capacity to grow about 80% in 2027 ⁴⁵ while demand grows close to 2x ⁴⁵. If both forecasts hold, the shortfall continues, which would explain why Nvidia is seeking capacity beyond TSMC. That is my inference.

We must be careful, though, to distinguish one bottleneck from the system's bottleneck. The sources disagree about where the binding constraint sits. In March 2026, SemiAnalysis judged front-end chipmaking capacity to have become the dominant bottleneck ³⁶. Another source argues that added packaging capacity is useless without front-end wafers ³⁶. The upstream evidence is consistent with that view. ASML reportedly may not fully meet customer demand in 2027 or 2028 ^42,58,60, even after planned capacity increases of about 30% in both years ⁵⁸. Nvidia places all its leading-edge production with one foundry, which it can plan with five years out ⁷⁶. Huang also noted that twice the transistors now cost twice as much ⁷⁶.

My reading is that the back-end deals relax one constraint while leaving two dependencies intact: TSMC wafers and, further upstream, ASML lithography tools. The packaging commitments do not address either. Output may therefore rise by less than the headline multiples imply.

Memory is a third dependency, and a more ambiguous one. Samsung's preliminary Q3 operating profit was about 107.4 trillion won ⁶¹, up 782.5% year over year ⁶¹. DigiTimes reported that Samsung, SK hynix and Micron had sold all of their 2027 DRAM and HBM capacity ⁶¹. A Micron HBM architect said HBM's roughly 3x silicon-per-GB penalty versus DDR5 is not shrinking ⁶¹. Every gain in HBM supply therefore costs more conventional DRAM, which keeps pricing tight and gives memory makers leverage over Nvidia. Seoul Economic Daily reported, citing unnamed sources, that Nvidia asked the three suppliers to build the NVHBM base die on TSMC 3nm and that Samsung agreed ⁶¹. None of the companies has confirmed this ^30,61.

The cyclical counterweight is that TrendForce reports cloud providers weighing less HBM per chip on their next in-house accelerators ⁶¹. Goldman Sachs' Korea team says semiconductor valuations imply a 65% decline in 2027 earnings ⁷⁹. The Bank of Korea warned that South Korean chip exports could fall US$50–100 billion below projections in 2027 if AI investment slows sharply ⁴⁴. A supplier with sold-out books and a market pricing a downturn is a reminder that quasi-rents in a capacity-constrained industry last only as long as the constraint does.

China: A Closed Market With a Leaky Channel

The policy history is well established. Export controls began in October 2022 ^53,57. Trump announced in December 2025 that H200 exports would be allowed ^2,3,31,39. On January 15, 2026 the Bureau of Industry and Security moved H200 license review to case-by-case, a change reported by three sources ^14,39,49. The change carried conditions: a 25% tariff, a 50% volume cap and mandatory third-party end-use testing ³⁹. Nvidia then reportedly halted H200 production aimed specifically at China ³⁹. It also took a $4.5B charge on H20 chips for China ⁸. The net result is that, for data center compute, China remains largely closed to Nvidia ⁶. Guidance excludes China data center compute, as several sources state ^{5,32,40,55,70}.

An order book that cannot become revenue

Against that backdrop, Reuters reported, citing anonymous sources, that ByteDance, Alibaba and Tencent placed orders worth about $35 billion ^13,55. The sources disagree on the product. One version says more than 3 million Blackwell chips ⁵⁵, another more than 3 million RTX PRO 5500 workstation cards ^67,68. The latter is the more widely repeated framing, and it matches reports that the card was designed to meet U.S. export restrictions ²⁸. That points to an export-compliant product, not restricted accelerators. Other posts carrying these figures are unconfirmed ^10,22.

The orders look unlikely to be filled. Nvidia reportedly told customers they exceed planned RTX Pro 5500 production capacity ²⁸. One Reuters source said Nvidia might fill only 5% to 10% ⁵⁵. Barron's reported that even rumors of a large Chinese purchase failed to lift the stock ^19,78. This is consistent with guidance that already excludes China. The sensible reading is that the China book is a signal of demand and not a source of revenue.

Enforcement is moving into the channel

The strongest evidence on leakage is the indictment of three individuals tied to Super Micro Computer, alleging $2.5 billion in AI-training server sales ⁴⁷. It also alleges $510 million in servers diverted to China in a single three-week window between late April and mid-May 2025 ⁵⁷. The reported mechanics are specific. Ninety-two servers went from San Francisco International Airport to Kuala Lumpur, and court documents show the onward Hong Kong consignee was a Chinese customer ⁵⁶. Real hardware was reportedly en route to Chinese buyers while dummy servers were staged ⁵⁷, and the operation was reportedly orchestrated in significant part from Taiwan ⁵⁷. Super Micro has disclosed a grand jury subpoena from the SDNY U.S. Attorney's Office ³⁸ and faces an active BIS probe ³⁸.

The jurisdictional gap is instructive. Exporting advanced chips to mainland China is not currently a criminal offense under Taiwanese law, according to two sources ⁴⁷, and assembled servers are not among the categories restricted under the Strategic High-Tech Commodities list ⁵⁷. That gap is now being addressed. Taiwan has confirmed plans to criminalize unauthorized exports of advanced AI chips to China ⁵⁷, fulfilling a February 2026 U.S.–Taiwan trade agreement obligation ⁵⁷. No implementation date has been set ⁵⁷. My interpretation is that the legal perimeter is being redrawn around the server channel, which is where the alleged leakage occurred. BIS red-flag guidance on product capability points the same way, with three sources ⁴⁸. Whether the proposals become law is not established.

Leakage claims run wider but are weaker. Lava reportedly found 32 restricted Blackwell Ultra B300 chips in China Telecom's network ²³. Reports describe technicians in China dismantling RTX 5090 cards to harvest chips ⁹. A CNAS estimate put smuggled chips at between 1% and 30% of China's inference computing capacity at one point ⁵⁷. The width of that range is the point: no one measures leakage well.

Substitution is real but capacity-limited

The other adjustment is domestic substitution. Huang has acknowledged that U.S. curbs spurred China to expand domestic chip manufacturing ⁵². Huawei was reportedly within six percentage points of Nvidia in China's AI chip market in the first half of 2026 ¹⁶. China has instructed domestic tech giants to accelerate adoption of homegrown alternatives ⁵⁷. Yet Huawei reportedly still cannot make enough chips to meet demand ²⁴, and the Ascend 950DT price reportedly rose 60% ⁵⁹. Substitution is therefore proceeding, but a supply-limited substitute leaves room for continued grey-channel demand for Nvidia hardware. That last step is interpretation, not a sourced finding.

Customers as Financiers and as Competitors

The demand side has its own frictions, and two matter. The first is that Nvidia increasingly supports its buyers' balance sheets. It agreed to buy CoreWeave's eligible unsold capacity through April 2032 ^37,69,75. As part of a $500 billion financing initiative, Nvidia guarantees up to $125 billion of residual value on financed GPUs ⁷⁷. One source notes that advance purchasing could sustain near-term revenue while increasing customers' inventory exposure ⁷⁴. These arrangements tie Nvidia's results to its customers' credit conditions, which is my interpretation. Demand quality then matters as much as demand volume.

The second is that customers are building alternatives, and the evidence is partial. Broadcom management says it has secured enough supply for about $115 billion of FY2027 AI semiconductor revenue ^4,5, and Broadcom is on track to deploy nearly 10 GW of compute in 2027 ^1,46. AMD has an agreement with Anthropic for up to 2 GW of Instinct GPUs ⁶². Custom chip adoption could shrink the addressable market for Nvidia and AMD by 15 to 25% ⁷¹. There are also software-level challenges. Oracle will reportedly offer Google TPUs on its cloud ¹⁷, and Google released a developer tool intended to let PyTorch users migrate applications off GPUs, carried by two sources ^11,12. Nvidia's moat is software and ecosystem as much as silicon, and these items attack that point directly.

The displacement is partial for now. Meta is working with Broadcom to reduce its reliance on Nvidia GPUs ⁴³. It also plans to expand its use of Nvidia's Spectrum-X networking and BlueField-4 ⁶⁵. One commenter describes the effect of custom silicon as a possible mix shift rather than a demand collapse ²¹. Mixed signals persist elsewhere. AWS reportedly plans to deploy two million more Nvidia GPUs in 2027 and 2028 ¹⁸, while Microsoft cancelled a lease for 23,000 Nvidia chips in Loughton ²⁹. Nvidia's locked-up packaging capacity could become a strategic asset if rivals cannot obtain the same substrates and interposers. That is an interpretation, since the sources do not say what rivals can obtain.

What the Evidence Implies

Under current conditions, the evidence suggests that Nvidia is strongest where it controls scarce inputs and weakest where it depends on policy or on customers' finances. Four conclusions follow, in order of confidence.

• Supply is the short-run constraint, and the response is well corroborated. Four- and five-source reports on interposers, optics, substrates and testing show capacity arriving between February 2027 and the end of 2028. Nvidia has offered no timeline for relief ²⁵.
• The long-run limits lie upstream. Single-foundry dependence, ASML's 2027–2028 supply warning and sold-out HBM mean packaging gains may not convert fully into output.
• China is upside that guidance does not count. The $35 billion order book may be only 5–10% fillable ⁵⁵, and the near-term risk lies in channel compliance rather than in the market itself.
• Demand quality now matters as much as volume. Residual-value guarantees and capacity buy-backs link Nvidia to customer credit ^37,69,77, while custom silicon is growing on large disclosed figures.

Several observations would change this picture: evidence that the packaging commitments are lifting shipped volume, Taiwan's legislation being enacted with a date attached, or accelerators shipping with less HBM and not merely planning to. Until then, the right posture is to treat Nvidia's position as a strong short-run one, with long-run adjustments under way on several fronts.