Economists distinguish the short run, in which capacity is fixed and firms make do with what they have, from the long run, in which plants are built and structures adapt. Almost everything in the October 2026 material on Nvidia can be sorted along that line. In the short run, supply sets the pace. Management said supply was the limiting factor 123 and the bottleneck 118. At GTC Berlin Jensen Huang said "Capacity is the constraint" 82. In the long run, capacity is being added, and that is where equity investors have turned their attention.
The interesting question is not whether Nvidia's demand is strong. The evidence says it is. The question is why the constraint persists, who holds the scarce inputs, and what happens to prices and profits when supply loosens. The answer runs through three tiers: the foundry (TSMC), memory (SK hynix and Micron), and Nvidia's own attempts to stretch both the supply side and the customers' ability to pay. We take them in turn, then ask what the market is discounting.
One caution on evidence applies throughout. Much of the technical and economic material is vendor-originated, and many figures come from single sources or social posts. Where a claim is well corroborated, we say so. Where it is isolated or inconsistent, we say that as well.
The foundry tier: demand exceeds capability
TSMC's September-quarter report on October 15 is the cleanest reading of AI demand in the material. It matters for Nvidia because TSMC makes Nvidia's chips 56,64 and virtually all of Nvidia's and AMD's most advanced chips 94. Revenue was US$46.72 billion 33,111. Net income rose 68.7% year over year 33,79,111. Gross margin was 66.4% 101. High-performance computing, which includes AI accelerators 58, was 68% of revenue 101. It was 29% in the first quarter of 2019 58. Q4 revenue is guided to US$52.3–54.3 billion 33,64,111.
Revenue growth is less settled. One source gives 41.2% year over year in dollar terms 33,64, and a data summary gives 36.4% 30,67. The material does not state the basis for the difference, so the headline growth rate should be held loosely.
The management commentary is more informative than the figures. C.C. Wei said demand is "far above our capability today to supply" 100. TSMC is working to narrow that gap 111, and the CEO would not say whether it will widen or narrow 111. The history explains why the adjustment is slow. Annual reports have historically shown wafer capacity growth in the mid-to-high single digits, sometimes lower 111. Capacity grows by increments, and demand can move faster than any increment.
The tightest point is advanced packaging. CoWoS is described as the most critical bottleneck for GPUs, CPUs and custom ASICs 89, and in July Wei said packaging was so tight that it limited customers' growth 83. TSMC is leaning on outsourced partners including ASE, SPIL, Powertech and Amkor 93. It is also developing a panel-based CoPoS technology to cut future costs 111. A scarce input normally commands a rising price. Social posts say CoWoS prices may rise 10–20% 48,55. TSMC gave no hard figures 93 and did not say whether any increase would apply to existing contracts 111. That is a plausible expectation, but the material does not confirm it.
The long-run response is large and costly. The 2026 capex budget was raised to US$60–64 billion 9,79,111, and management expects 2026–2028 outlays to significantly exceed the prior three years 101,111. Annual N2-family capacity is expected to more than double in 2027 101,111. The second Arizona fab is scheduled for volume production in Q3 2027 100,101,111, and Texas is only under evaluation 111. TSMC says dilution from overseas fabs will be closer to 2% for full-year 2026 111.
There is one telling counterpoint. TSMC said it is planning prudently for 2027 111, and noted that economic uncertainty is challenging consumer and price-sensitive end markets 111. The US ADR fell as much as 5.3% the morning after the report 86. It was down 9.0% over the five trading days to October 16, against a 5.7% average decline for its closest peers 27,30. The strongest quarter on record was not enough to lift the stock, and the question of why belongs to the later sections.
The memory tier: a scarce input with three suppliers
Memory shows the same structure at a different point in the chain. HBM is the part Nvidia does not make and cannot readily substitute. The most corroborated supply claim is that AI demand has pulled fab capacity toward HBM 89. The physical cost is plain: HBM is said to use about 3x the silicon per GB of DDR5 103. DRAM supply sits in a three-firm oligopoly 32, which sustains extended constraints 49. Redirecting capacity to HBM then amplifies scarcity in commodity DRAM 32.
The supplier statements are stark. SK hynix's CEO said 2027 will be the worst supply year in industry history 103 and that demand will exceed supply past 2030 103. Micron's COO said the company has no line of sight to when supply and demand balance 52,87. IDC expects HBM4 output not to meet demand until mid-2027 89. Other horizons are longer but less corroborated. The New York Fed's John Williams put the earliest closing of the gap at late 2028 81. A report that all 2027 DRAM and HBM capacity is already sold has not been confirmed by Samsung, SK hynix or Micron 103.
Price evidence is consistent in direction, though not in magnitude or dating. TrendForce forecasts a 121% rise in blended HBM selling prices during 2027 26,103, and Bloomberg expects HBM contract prices to rise 70% to 80% in 2027 52. One headline places the 121% figure in 2025 29, which looks like a dating error. The figure is better treated as indicative.
The suppliers' results show where the scarcity rent lands. SK hynix's Q2 operating profit was about $41.6 billion, a figure with six sources 6,108, at a record operating margin of about 76% 6,108. Q3 revenue was 104.62 trillion won 91 against a 98.34 trillion won estimate 91. Micron reports an 86.8% GAAP gross margin 35,119, with more than 75% of fiscal 2027 shipments committed 52,87,103 at prices well above this year's 103. It also reports 26 multi-year strategic customer agreements 103 and customer financial commitments of about $32 billion, mostly cash 103. Micron supplies HBM for Nvidia GPUs 60,61,62 and announced volume shipments of HBM4 designed for Nvidia Vera Rubin 104.
Not all of the profit is a volume story. Korean media reported that SK hynix's Q3 HBM4 shipments may fall below original expectations for a second quarter running 108. That suggests, as interpretation, that pricing and the wider memory cycle are doing much of the work.
Nvidia's own response is to hedge its dependence, not remove it. It is working with six suppliers to expand substrates 43, and it signed a Unimicron deal reserving a significant portion of that supplier's capacity 53. It is paying a $1.1 billion advance to KYEC for new US testing facilities 66, and reporting describes advance payments as a way to secure production and reduce exposure to bottlenecks 115. The headline targets deserve care. The substrate agreements are expected to deliver a 2.5-fold capacity increase over three years 40, and that is a target, not existing capacity 66. Nvidia is also developing a custom memory branded NVHBM 51. The sources disagree about who the partners are, naming SK hynix 57, TSMC 51 or memory manufacturers generally 51. The best-supported reading is that Nvidia is customising the interface, not taking over memory production 102.
Where scarcity reaches the consumer
When one use of a scarce input expands, other uses are crowded out, and the consumer GPU line shows it. The GeForce RTX 5090 has largely vanished from normal retail channels. It is reported missing from Nvidia's direct US catalog 121, and Micro Center reportedly showed "Not Carried" 72. Street prices are far above the $1,999 launch price 71,73. Quotes range from about $3,000 74 to nearly $10,000 76. The spread probably reflects dates, regions and seller types, so any single figure is only indicative.
Discontinuation is the contested point. A leaker claimed production had stopped 114, and Nvidia has denied it 39,114. Partners say backorders are no longer possible 129 but they have not been told the card is end-of-life 129. Nvidia attributes the shortages to memory and component constraints 129. The proposed mechanism is that GB202 silicon, shared by the RTX 5090 and the RTX 6000 PRO 90, is being redirected to workstation cards. That is a leaker's claim 75. Reuters reportedly put orders for the workstation card at more than US$35 billion at sticker prices 78, and Nvidia says orders exceed planned production 47. Nvidia has not confirmed the redirection, and the evidence is circumstantial. A scarce die with a far more profitable alternative use is the textbook case for reallocation, however, and Nvidia's explanation is not necessarily incompatible with it. The effect is not confined to Nvidia: PC shipments reportedly fell 20% in Q3 2026 32, and memory shortages were constraining them 107.
The product cycle: Vera Rubin and the economics of the gigawatt
Current earnings are still carried by Blackwell. The Q2 FY27 margin improvement was attributed to Blackwell Ultra mix 36, and the Blackwell Ultra ramp drove Q2 Data Center results 36. Vera Rubin is the forward narrative. Full production is the best-corroborated claim, with seven sources 2,3,4,11,17,64, and Nvidia reportedly holds purchase orders from every major hyperscaler, AI cloud and system OEM 68,80,126. The shipment dates conflict. Accounts give July 2026 shipments 1,34, production shipments from early August 68, and a fall 2026 window for the first cloud cohort 44. These may describe a narrow early ramp followed by a wider one, but the sources do not say so.
The commercial case is expressed in revenue per unit of power. Management frames the opportunity as rising from about $25 billion per gigawatt for Blackwell to about $40 billion for Vera Rubin 18,59,68. The supporting performance claims are up to 30 times higher throughput per megawatt than GB300 NVL72 on agentic workloads 15,16,70,95,96, one-fourth the GPUs for training mixture-of-experts models 113, and one-tenth the cost per million tokens 113. All of these are vendor-originated, and the sources do not independently verify them. Single-source late-October reports also say Nvidia revised Vera Rubin Superchip specifications, with HBM bandwidth reportedly reduced 37,38. Bandwidth figures should therefore be treated as unsettled.
A related physical constraint is power. Sources conflict on rack density: one has it moving from 120–130 kW toward about 600 kW 116, and another has Rubin clusters exceeding 1 MW per rack 127. Delivered hardware is not energized hardware. AWS is reportedly taking about $5 billion of GB300 racks into warehouses 46 before data center power is available 46.
The customer tier: capital as the new bottleneck
When supply is fixed, the remaining question is who can pay for what is built. Huang called capital the new bottleneck 41 and said neoclouds "just need working capital" 63. Nvidia's mechanisms include financing support to neoclouds, repeated across three sources from July through October 5,132, a minimum revenue guarantee program 112, and residual-value support to cloud operators 56. Its $3.5 billion investment in MediaTek convertible bonds is the most corroborated financing claim here, with 11 sources 12,13,14,19,20,21,22,23,24,25,84. Management's line on the earnings call was "We're not making loans" 112. Whether guarantees differ from loans in substance is contested, and the terms are not disclosed in detail.
The strain shows at the lenders. Banking partners reportedly objected to the financing plans over the residual value of Nvidia's GPUs 124, and lenders questioned whether the chips are reliable long-term collateral 54. Collateral value depends on useful life, and the sources do not agree on it. They cite roughly three years for frontier use 50, three to five years 65,92, or eight or more 65. It also depends on utilization. Rising GPU rental prices, including for older chips, appear in seven sources 85,88,97,98,120. Yet a VentureBeat survey found 79% of 218 respondents reporting GPU utilization at 50% or below 125. The most defensible reading is that scarcity at the frontier coexists with idle capacity elsewhere, and the evidence does not settle which dominates. The pieces are interdependent: financing works only if GPUs hold their value, and supply expansion is sized against demand that partly depends on the financing. This reading is interpretation, not a claim in the sources.
Substitution at the margin: the software layer
How elastic is demand for Nvidia's products with respect to rivals' offerings? CUDA, built over more than fifteen years, creates switching costs 56, and NVLink and InfiniBand extend the lock-in to the systems level 56. The competitive effort is now aimed at that switching cost directly. AWS's TorchNeuron is meant to let developers deploy PyTorch-based models by changing one configuration setting 42,69. Google's open-sourced TorchTPU aims at minimal code changes 106. Oracle plans to offer Google TPUs, with deployments beginning in 2027 77.
The frictions are visible in the same sources. Trainium's nightly CI sweeps more than 100,000 models but with an 80–90% pass rate 105, and TPUs and Inferentia2 are compiled per model and shape 113. On one agentic benchmark, every Blackwell system and the MI355X reach roughly 600–670 tokens/s for a single agent at batch size one, and differences appear only under load, power or cost constraints 99. The portability and economics claims are mostly vendor-stated. They show intent and direction, not measured switching. The credible reading is that custom silicon takes share at the margin, among large buyers with procurement leverage, and does not displace Nvidia wholesale. A source also notes that pressure from alternatives is not equivalent across training and inference workloads 110.
What the market is discounting
The equity data matter here. The Philadelphia Semiconductor Index was about 16% below its early-October high 31 and about 25% below its June 2026 peak 10,117. The two figures use different reference points. Nvidia was down 6.61% for the week ending October 16 122 and drew a Hold rating after the Berlin keynote 45. Micron trades at about 5x fiscal 2027 estimates 87,109, and SK hynix at about 5x forward P/E, a figure with four sources 7,8,133. Micron fell 14.59% in the week ending October 16 28,31, even as 29 analysts raised its estimates in the preceding 90 days 131.
The market is therefore not disputing current strength. It is questioning durability. The scarcity story and the de-rating are compatible, because a physical bottleneck that produces extraordinary margins also invites the capacity that ends it. Micron's capex was $27.37 billion in fiscal 2026 and plans imply more than roughly $50 billion in fiscal 2027 52. TSMC's capex is rising 101,111, and IDC says capacity additions will matter competitively as they enter production in 2027–2028 100. Morgan Stanley's downside case has spending stabilizing before supply adjusts, leaving volumes flat while prices fall 130, and it cites the late-2022 memory downturn, when prices fell 20%–30% 128. A survey finds most companies expect slower AI spending growth 107.
Multi-year contracts and prepayments, such as Micron's commitments and Nvidia's advances, are designed to dampen the usual boom-bust pattern. They may blunt the cycle without removing it. Nothing in the material settles how long peak conditions last, and the next datapoint on Micron is its December 16 earnings report 131.
Conditional conclusions
Under current conditions the evidence supports four statements.
- Supply sets the short-run pace. Management at Nvidia and TSMC, and suppliers at SK hynix and Micron, all describe demand above capacity 100,103,123. The constraint is packaging, memory and substrates, and the fixes are future-dated.
- Nvidia is hedging dependence, not escaping it. Prepayments, capacity reservations and interface customisation reduce exposure to bottlenecks 51,53,66. They leave Nvidia reliant on a small group of Taiwanese and Korean suppliers.
- The weakest link may be financial, not technical. Customer financing, collateral value and utilization 54,112,125 are the least settled elements of the demand case, and the disclosure on them is thin.
- The market is pricing the transition, not the present. The sell-off 10,31,117 against reported strength suggests investors expect margins to normalize as capacity enters production in 2027–2028 100.
What would change this view is observable. Confirmed HBM and CoWoS price increases would strengthen the scarcity case. A TSMC capex plan that outruns demand, a visible fall in rental prices or utilization, or financing terms that tighten would strengthen the transition case. Until then, the evidence describes a system with capacity fixed and prices doing the adjusting.