AI Chip Shortage 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: August 21, 2026 | License: CC BY 4.0
NVIDIA has pre-committed $119 billion to manufacturing, supply and capacity as of April 26, 2026 — four times the $29.8 billion it had committed a year earlier, and more than a full quarter of its own revenue paid forward to hold a place in line. That single line item in a 10-Q describes the 2026 AI chip shortage better than any demand forecast.
Quick Answer
The AI chip shortage in 2026 is real but misdiagnosed. According to Axis Intelligence Research, the Axis Chip Scarcity Index (ACSI) reads 177.9 in its inaugural reading of August 21, 2026, against a base of 100 set at the year-earlier comparable period. The scarcity sits in forward capacity commitments and memory pricing, not in idle fab floor space: US semiconductor plants ran at 75.4% of capacity in July 2026, 3.7 points below their 1972–2025 average.
Key Findings
- According to Axis Intelligence Research, the Axis Chip Scarcity Index reads 177.9 as of August 21, 2026, with memory pricing and forward capacity lock-in contributing 143 of its 178 points.
- NVIDIA’s manufacturing, supply and capacity commitments reached $119 billion as of April 26, 2026, up 299.3% from $29.8 billion a year earlier, per its Form 10-Q filed with the SEC.
- Axis Intelligence Research finds that NVIDIA’s Capacity Lock-in Ratio — forward commitments divided by trailing-quarter revenue — rose from 0.68 in April 2025 to 1.46 in April 2026.
- Micron’s revenue grew 345.7% year over year in the quarter ended May 28, 2026 while its cost of goods sold grew 10.5%; Axis Intelligence Research calculates that each additional dollar of revenue carried just 1.9 cents of additional cost.
- US semiconductor capacity utilization was 75.4% in July 2026, below the 79.1% long-run average for the series, according to the Federal Reserve’s G.17 release of August 18, 2026.
Is There Really an AI Chip Shortage in 2026?
Yes — but the word “shortage” is doing unusual work here.
A classic shortage looks like full factories and empty warehouses. That is not what the 2026 data shows. The Federal Reserve’s G.17 industrial production release published on August 18, 2026 puts capacity utilization for semiconductors and related electronic components (NAICS 3344) at 75.4% in July, up from 73.2% in the second quarter and 72.7% in the first. The published 1972–2025 average for that same series is 79.1%. American chip plants, in the middle of the largest AI infrastructure buildout on record, are running below their historical norm.
Output is not the problem either. The semiconductor production index stood at 191.9 (2017 = 100) in July 2026, up 11.9% year over year — the strongest contributor inside a high-technology group that grew 11.8%. The reason utilization stays flat while production surges is in Table 8 of the same release: capacity for selected high-technology industries is expanding at a 13.2% annual rate in 2026. Capacity is outrunning output. Utilization is a ratio, and the denominator is winning.
Meanwhile the money says scarcity. The Semiconductor Industry Association reported on August 6, 2026 that global semiconductor sales hit $403.3 billion in the second quarter, a 35.1% sequential increase, with June alone at $134.5 billion — 123.6% above June 2025. Americas sales rose 160.9% year over year. For how that spending divides between vendors, see Axis Intelligence Research’s AI chip market share analysis.
Sarah Mitchell: The 2021 automotive chip shortage was a capacity story — 28-nanometer lines were physically full and Ford idled plants. This one is not. Two things are tight: the specific step that stacks logic silicon against high-bandwidth memory, and the memory itself. Everything else in the stack has slack. Treating the two as one shortage is why most 2026 forecasts have been wrong in both directions. Our semiconductor market statistics page tracks the aggregate revenue picture that sits above this supply question.
AI chip shortage indicators at a glance
| Indicator | Reading | Period | Source |
|---|---|---|---|
| US semiconductor capacity utilization | 75.4% | July 2026 | Federal Reserve G.17 |
| Long-run average, same series | 79.1% | 1972–2025 | Federal Reserve G.17 |
| Semiconductor production index (2017=100) | 191.9 | July 2026 | Federal Reserve G.17 |
| High-technology capacity growth, annual rate | 13.2% | 2026 | Federal Reserve G.17 |
| Global semiconductor sales | $403.3B | Q2 2026 | Semiconductor Industry Association |
| Global sales growth, month over year-earlier month | 123.6% | June 2026 | Semiconductor Industry Association |
How Big Is the AI Chip Supply Gap? The Axis Chip Scarcity Index
Compiling other people’s forecasts does not settle an argument about whether supply is tight. Measuring the behaviour of the firms inside the constraint does.
The Axis Chip Scarcity Index (ACSI) — an original Axis Intelligence Research metric — reads 177.9 as of August 21, 2026. This is the inaugural reading and the base of 100 is defined as the year-earlier comparable period for each component; no earlier readings have been computed and none are implied.
Formula:
ACSI = (0.35 × C1) + (0.30 × C2) + (0.20 × C3) + (0.15 × C4)
Each component is rescaled so that its own comparison base equals 100.
| Component | What it measures | Current | Base | Sub-index | Weight | Contribution |
|---|---|---|---|---|---|---|
| C1 Forward capacity lock-in | NVIDIA commitments ÷ trailing-quarter revenue | 1.4581 | 0.6763 | 215.6 | 0.35 | 75.46 |
| C2 Memory scarcity pricing | Micron consolidated GAAP gross margin | 84.6% | 37.7% | 224.4 | 0.30 | 67.32 |
| C3 Leading-edge concentration | TSMC advanced-technology share of wafer revenue | 77% | 74% | 104.1 | 0.20 | 20.82 |
| C4 Fab capacity response | US semiconductor capacity utilization vs long-run average | 75.4% | 79.1% | 95.3 | 0.15 | 14.29 |
| ACSI | 100 | 177.9 |
Sources: NVIDIA Form 10-Q (SEC EDGAR) · Micron fiscal Q3 2026 results · TSMC Form 6-K, Q2 2026 and Q2 2025 · Federal Reserve G.17. Computed by Axis Intelligence Research, August 21, 2026.
Why these weights. Forward commitments (C1) carry the heaviest weight because a purchase commitment is the least ambiguous scarcity signal a public company emits: nobody pays cash forward for capacity they expect to be able to buy on demand. Memory pricing (C2) is next because DRAM and HBM are the components where 2026 price movement has been most extreme. Leading-edge mix (C3) is weighted lower because node migration reflects product roadmaps as well as scarcity. Fab utilization (C4) is the deliberate counterweight — the one component that pushes the index down, so that ACSI cannot mistake a demand boom for a capacity ceiling.
How to read it. ACSI above 100 means the supply chain is behaving more scarcely than it did a year earlier. What moves it: a fall in Micron’s gross margin or a flattening of NVIDIA’s commitment balance would pull it sharply toward 100. Rising US fab utilization would push it up — which is the counterintuitive but correct signal, because it would mean capacity had finally stopped outrunning output.
Sarah Mitchell: The index is 178 and 143 of those points come from two companies’ balance-sheet behaviour. That is the honest shape of this shortage. It is not distributed across the industry; it is concentrated in the accelerator-plus-memory pairing, and it shows up in cash commitments long before it shows up in a utilization statistic.
How Much Are AI Chip Buyers Paying to Lock In Supply?
The most underused number in the AI chip supply chain sits in Note 10 of NVIDIA’s Form 10-Q for the quarter ended April 26, 2026. Manufacturing, supply and capacity commitments: $119 billion, of which $95 billion falls due in the remainder of fiscal year 2027.
The trajectory is the story:
| As of | Commitments | Trailing-quarter revenue | Capacity Lock-in Ratio |
|---|---|---|---|
| April 28, 2024 | $18.8B | — | — |
| April 27, 2025 | $29.8B | $44.06B | 0.68 |
| April 26, 2026 | $119B | $81.62B | 1.46 |
Source: NVIDIA Forms 10-Q, SEC EDGAR. Ratio computed by Axis Intelligence Research. Revenue figures for 2025 and 2026 are drawn from the same filing.
Axis Intelligence Research computes the Axis Capacity Lock-in Ratio (ACLR) as forward manufacturing, supply and capacity commitments divided by trailing-quarter revenue — both taken from the same filing, so no cross-period estimation is involved:
ACLR = 119,000 ÷ 81,615 = 1.46 (April 2026)
ACLR = 29,800 ÷ 44,062 = 0.68 (April 2025)
Below 1.0, a vendor is buying components roughly in step with what it ships. Above 1.0, it has bought more forward capacity than it sold in the entire preceding quarter. NVIDIA crossed that line during fiscal 2027.
For scale — and Axis Intelligence Research labels this as a scale comparison, not a market share, because a multi-year commitment balance and a single quarter of sales are not the same kind of quantity — $119 billion equals 29.5% of the $403.3 billion the whole world spent on semiconductors in the second quarter of 2026.
The inventory line tells the same story from a different angle. NVIDIA’s raw materials rose from $3.81 billion at the January 25, 2026 balance sheet date to $6.65 billion at April 26, 2026 — a 74.6% increase in one quarter, against a 20.5% rise in total inventories. Companies that expect components to be available do not stockpile substrate and memory at that rate. The vendor-level context for those commitments is covered in our Nvidia GPU market share breakdown.
Is the Memory Shortage Worse Than the AI Chip Shortage?
On the evidence, yes. And the cleanest proof is not a price index — it is an income statement.
In its fiscal third quarter ended May 28, 2026, Micron reported revenue of $41.456 billion against $9.301 billion a year earlier — growth of 345.7%. Over the same twelve months, cost of goods sold moved from $5.793 billion to $6.400 billion: growth of 10.5%.
Axis Intelligence Research calculates the marginal cost of that incremental revenue:
($6,400M − $5,793M) ÷ ($41,456M − $9,301M)
= $607M ÷ $32,155M
= $0.019 per additional dollar of revenue
Micron added $32.2 billion of annual quarterly revenue and 1.9 cents of cost per dollar of it. Consolidated GAAP gross margin went from 37.7% to 84.6%. That is not a production story. It is a price story, and price is what scarcity looks like when it reaches an income statement.
Two supporting figures make the shortage reading harder to argue with. Inventories were $8.567 billion at quarter end, up 2.5% from $8.355 billion at the close of fiscal 2025 — essentially flat while revenue quadrupled. And the Cloud Memory business unit alone booked $13.769 billion, running an 83% gross margin. Micron guided fiscal Q4 to $50.0 billion at roughly 86% gross margin, and spent $19.6 billion on property, plant and equipment in the first nine months of the fiscal year.
Sarah Mitchell: Memory bit supply is set by wafer starts and process yield, and both take fiscal years to move. Micron is spending nearly $20 billion in nine months precisely because it cannot buy its way out of this quarter. When a supplier books 84.6% gross margin and its inventory does not build, the constraint is not commercial — the product is leaving as fast as it is made.
Micron fiscal Q3 2026 versus fiscal Q3 2025
| Line | FQ3-26 | FQ3-25 | Change | Source |
|---|---|---|---|---|
| Revenue | $41,456M | $9,301M | +345.7% | Micron FQ3-26 results |
| Cost of goods sold | $6,400M | $5,793M | +10.5% | Micron FQ3-26 results |
| GAAP gross margin | 84.6% | 37.7% | +46.9 pts | Micron FQ3-26 results |
| Cloud Memory revenue | $13,769M | $3,386M | +306.6% | Micron FQ3-26 results |
Which Manufacturing Step Is the Real Bottleneck?
TSMC’s own disclosure narrows the field. In its second quarter 2026 results, filed with the SEC on Form 6-K, revenue reached $40.20 billion — up 33.7% year over year — with gross margin at 67.7% against 58.6% in the second quarter of 2025.
The node mix is where the concentration shows:
| Node | Share of Q2 2026 wafer revenue | Source |
|---|---|---|
| 2 nanometer | 3% | TSMC Form 6-K, Q2 2026 |
| 3 nanometer | 30% | TSMC Form 6-K, Q2 2026 |
| 5 nanometer | 33% | TSMC Form 6-K, Q2 2026 |
| 7 nanometer | 11% | TSMC Form 6-K, Q2 2026 |
| Advanced technologies (7nm and below) | 77% | TSMC Form 6-K, Q2 2026 |
A year earlier the same disclosure put advanced technologies at 74%. Three points of mix shift in twelve months, on a revenue base that grew a third — the leading edge is absorbing an increasing share of an increasing total. TSMC guided third-quarter revenue to between $44.6 billion and $45.8 billion. Unit-level flows through this same leading edge are tracked in our GPU shipment statistics.
Here Axis Intelligence Research declines to publish a number that every other outlet is happy to estimate. Advanced packaging allocation — how much CoWoS capacity any single customer holds — is widely quoted as a percentage. No primary-source disclosure supports those figures, and neither TSMC’s earnings release nor any customer filing states an allocation share. A methodological construction built on an unsourced input is a guess wearing a formula, so ACSI uses TSMC’s published node mix instead. The packaging constraint is real; the circulating percentages are not verifiable, and Axis Intelligence Research does not restate them.
How Do US Export Controls Shape AI Chip Supply?
This is where the 2026 shortage acquired a legal definition, and it is the detail most coverage has missed.
On January 15, 2026 the Bureau of Industry and Security published a final rule at 91 FR 1684 moving certain advanced computing exports to China and Macau from presumption of denial to case-by-case review. The covered band is defined by two hardware thresholds: total processing performance below 21,000 and total DRAM bandwidth below 6,500 GB/s — the second of which is, in the agency’s own note, a memory-bandwidth test that explicitly counts co-packaged high-bandwidth memory.
The conditions attached are, functionally, a scarcity test written into export law. Before a licence can be reviewed case-by-case, the applicant must certify that domestic supply is sufficient such that the export would not delay any existing or new US order taking normal lead times into account, and that global foundry capacity otherwise destined for US end users will not be diverted. Aggregate exported processing performance to China and Macau is capped at 50% of what has shipped to US customers for US end use. BIS estimated the change would generate roughly 100 additional licence applications a year.
The commercial imprint is already visible in filings. NVIDIA’s China and Hong Kong revenue was $4.550 billion in the quarter ended April 26, 2026, against $9.659 billion a year earlier. The demand side of that equation is covered in our China AI statistics research.
Sarah Mitchell: Read that certification again. A US agency now requires an exporter to prove, order book in hand, that selling abroad will not push a domestic customer’s delivery date. Export control has historically asked who the buyer is. This rule asks whether there is enough to go around — which is a regulator conceding, in the Federal Register, that there is not.
When Will the AI Chip Shortage End?
The honest answer is that the published data does not yet contain the turn, and the components of ACSI say why.
Three things would have to happen, and each is observable in primary sources:
Memory pricing normalises. Micron’s gross margin is the fastest-moving component in the index. Guidance for fiscal Q4 2026 is approximately 86% — higher, not lower. Component C2 will not fall until that guidance does.
Forward commitments flatten. NVIDIA’s $119 billion balance is the strongest scarcity signal in the dataset because it is a cash decision, not a forecast. The next quarterly filing either extends the curve or breaks it. There is no third option.
Capacity catches up. This is the only component already moving in the direction of relief. High-technology capacity is growing at a 13.2% annual rate and utilization sits 3.7 points below its long-run average. Fabs commissioned in 2026 produce revenue-grade output on a multi-year lag, which means the relief is scheduled but not yet delivered — a pattern visible in the factory-construction data behind our US manufacturing statistics.
Anyone quoting a specific end date for the AI chip shortage is quoting a forecast, not a measurement. Axis Intelligence Research will publish the next ACSI reading when NVIDIA files its fiscal Q3 2027 Form 10-Q and Micron reports fiscal Q4 2026 — the two events that can actually move the index.
Methodology
Collection. Every figure in this analysis was retrieved on August 21, 2026 from a primary document opened during production: SEC filings on EDGAR (NVIDIA Form 10-Q; TSMC Form 6-K for Q2 2026 and Q2 2025), Micron’s fiscal Q3 2026 results release, the Federal Reserve’s G.17 statistical release of August 18, 2026, the Semiconductor Industry Association’s August 6, 2026 sales release, and the Bureau of Industry and Security final rule at 91 FR 1684. No figure in this article is drawn from a secondary aggregator, an analyst estimate, or an unattributed market-research summary.
ACSI construction. Four components, each rescaled against its own comparison base = 100, then weighted 0.35 / 0.30 / 0.20 / 0.15 as disclosed in the table above. C1 through C3 use the year-earlier reported period as base. C4 uses the Federal Reserve’s own published 1972–2025 average for the series, because the current G.17 release does not publish a monthly year-earlier figure for NAICS 3344. Arithmetic was independently recomputed in two passes; the component contributions sum to 177.89, reported as 177.9.
ACLR construction. Forward manufacturing, supply and capacity commitments divided by trailing-quarter revenue, with both inputs taken from the same Form 10-Q so that no cross-filing reconciliation is required.
Scope. ACSI measures the merchant AI accelerator supply chain through four listed companies and one national statistical series. It does not capture custom ASIC programmes at hyperscalers, Chinese domestic accelerator production, contract terms inside strategic customer agreements, or data-centre power and interconnect constraints, which bind independently of silicon availability. The index reads supply-chain behaviour, not end-customer wait times; no primary source publishes accelerator lead times, and Axis Intelligence Research does not estimate them. Fiscal calendars differ across the four companies, so components are aligned to each company’s most recently reported period rather than to a common calendar date; the snapshot date of August 21, 2026 refers to compilation, not to a synchronised measurement window.
Comparability note. NVIDIA’s commitments line was labelled “inventory purchase and long-term supply and capacity obligations” in the April 2025 and April 2024 filings and “manufacturing, supply, and capacity commitments” in the April 2026 filing. Axis Intelligence Research treats these as the comparable forward-commitment disclosure and flags the wording change rather than silently splicing the series.
About This Dataset
Contents. 69 observations covering AI accelerator supply commitments, memory pricing and margins, leading-edge foundry node mix, US semiconductor capacity utilization and production, global semiconductor sales by region, US export control thresholds, and every Axis-computed figure in this article — each with source organisation, source document, URL, retrieval date, primary-source flag and, for computed rows, the full formula.
Coverage. Temporal: April 2024 to July 2026. Spatial: global, with United States, China and Macau, Europe, Japan, Americas and Asia Pacific breakouts.
Format. UTF-8 CSV, one header row, snake_case columns, ISO 8601 dates, raw numeric values with units in a separate column.
Licence. CC BY 4.0. Free to reuse, redistribute and build on with attribution.
Citation: Axis Intelligence Research, AI Chip Shortage 2026: Statistics, Supply Commitments and the Axis Chip Scarcity Index, 2026.
Cite This Research
APA Axis Intelligence Research. (2026). AI chip shortage 2026: Statistics, supply commitments and the Axis Chip Scarcity Index. https://axis-intelligence.com/ai-chip-shortage/
MLA Axis Intelligence Research. “AI Chip Shortage 2026: Statistics, Supply Commitments and the Axis Chip Scarcity Index.” Axis Intelligence, 21 Aug. 2026, axis-intelligence.com/ai-chip-shortage/.
Chicago Axis Intelligence Research. “AI Chip Shortage 2026: Statistics, Supply Commitments and the Axis Chip Scarcity Index.” Axis Intelligence, August 21, 2026. https://axis-intelligence.com/ai-chip-shortage/.
Frequently Asked Questions
Does a high fab utilization number mean chips are scarce?
Not on its own, and 2026 is the year that became obvious. Utilization is output divided by capacity, so a fab operator adding capacity faster than it adds output posts a falling utilization rate in the middle of record production. US semiconductor utilization was 75.4% in July 2026 while the production index rose 11.9% year over year, because high-technology capacity is expanding at a 13.2% annual rate. Read the two together or the number misleads.
Why do supply commitments predict scarcity better than lead times?
Because commitments are audited and lead times are anecdote. A manufacturing and capacity commitment is a contractual balance disclosed in a filing, dated, and subject to auditor review. Quoted lead times come from sales conversations, vary by customer tier, and are never published on a consistent basis. When NVIDIA’s commitment balance quadrupled in twelve months, that was a cash decision by a company with perfect visibility into its own allocation.
Is high-bandwidth memory a separate shortage from the accelerator shortage?
They are coupled but not identical. An accelerator without its co-packaged memory stack is not a shippable part, which is why the Bureau of Industry and Security wrote a total DRAM bandwidth threshold of 6,500 GB/s into its January 2026 rule alongside a processing-performance threshold. Micron’s 84.6% gross margin on 2.5% inventory growth indicates the memory side is at least as tight as the logic side, and possibly tighter.
What would falsify the claim that AI chips are supply-constrained?
Three observable reversals: Micron’s gross margin compressing toward historical DRAM norms, NVIDIA’s forward commitment balance flattening or declining across two consecutive filings, and inventories building at either company without a corresponding revenue decline. None of the three appears in the most recent primary data. Any one of them appearing would be a meaningful signal, and the ACSI components are chosen so that it would show up immediately.
Do export controls make the AI chip shortage better or worse for US buyers?
The January 2026 rule was drafted to protect domestic allocation. Exporters seeking case-by-case review must certify that shipments abroad will not delay US orders and will not divert global foundry capacity from US end users, with exports to China and Macau capped at 50% of what ships to US customers. Whether that materially loosens domestic supply is not yet measurable in published data; what is measurable is that NVIDIA’s China and Hong Kong revenue fell from $9.659 billion to $4.550 billion year over year.
Is the 2026 AI chip shortage comparable to the 2021 automotive chip shortage?
The mechanism is different. The 2021 shortage was a mature-node capacity failure with idled assembly lines downstream. The 2026 constraint sits in leading-edge logic, advanced packaging and high-bandwidth memory, while overall fab utilization runs below its long-run average. Comparing the two leads to the wrong policy conclusion, because adding trailing-edge capacity does nothing for a 3-nanometer-plus-HBM bottleneck.
Why does Axis Intelligence Research not publish a CoWoS allocation percentage?
Because no primary source discloses one. Advanced packaging allocation shares circulate widely as precise-sounding percentages, but neither TSMC’s quarterly release nor any customer filing states them. Publishing an unverifiable figure inside a formula would make every number downstream of it unverifiable too. The packaging constraint is real; the specific percentages in circulation are not sourced, and this dataset omits them.
When does the next ACSI reading publish?
When the inputs move. The index is recomputed after NVIDIA files its next quarterly report and Micron reports fiscal Q4 2026, since those two filings drive 65% of the weighting. Interim Federal Reserve G.17 and SIA releases are logged in the dataset but do not on their own trigger a new reading.