Research / revised primer
Capital-light is not capital-free.
The fabless model did not disappear. Its manufacturing capital migrated into commitments, prepayments, inventory, guarantees, and strategic relationships, while the owners of scarce assets kept the residual upside.
1. The old fabless bargain
Specialist foundries let designers avoid direct factory ownership while retaining access to increasingly capable processes. That specialization widened participation, improved focus, and transferred a large class of asset risk to manufacturers. The model remains productive.
The hidden option was flexibility: a designer could rely on a supplier’s capacity without financing every future expansion or carrying the residual asset through every cycle. When the qualified bottleneck is abundant, that option is powerful. When it becomes scarce, allocation, duration, and downside sharing move onto the customer’s balance sheet.
The old CapEx-only proof of fabless capital efficiency is incomplete because material capacity exposure can appear as commitments, prepayments, inventory, guarantees, and investments.
Caveat. Do not state that every fabless company is more capital-intensive than every IDM.
2. Shadow Capital
“Shadow Capital” is a research taxonomy for manufacturing exposure that reported CapEx alone does not capture. It includes contractual capacity, firm purchases, prepayments, strategic inventory, facility guarantees, and manufacturing investments. The categories must never be collapsed into a fabricated asset total.
The useful comparison is not “commitments equal factories.” It is that a low owned-asset line can coexist with large, long-duration obligations required to ship the product. The customer may carry meaningful capital and downside risk without receiving the physical asset, learning curve, subsidy, or residual option.
NVIDIA disclosed $119 billion of manufacturing, supply, and capacity commitments at April 26, 2026.
Caveat. Keep separate from cloud, investment, and other vendor commitments.
Open source
Of NVIDIA’s $119 billion manufacturing, supply, and capacity commitments, $95 billion was payable during the remainder of fiscal 2027.
Caveat. Use the filing’s fiscal-period wording.
Open source
NVIDIA reported $10.383 billion of PP&E at January 25, 2026.
Caveat. Period differs by one quarter from the $119B commitment balance.
Open source
NVIDIA spent about $6.042 billion on property, equipment, and intangible assets in fiscal 2026.
Caveat. Label the exact cash-flow line rather than implying all spending was fab-related.
Open source
Purchase commitments are not automatically prepaid cash and should not be treated as economically identical to fab construction.
Caveat. Core caveat for all charts.
3. Downside reveals the structure
Capacity control is easiest to celebrate in a demand surge. The contract becomes legible when regulation changes, a product slips, or demand falls. Product-specific inventory, non-cancellable supply, and excess purchase obligations show where the risk actually landed.
NVIDIA recorded a $4.5 billion fiscal 2026 charge associated with excess H20 inventory and purchase obligations after export restrictions.
Caveat. Do not repeat the earlier $5.5B estimate as the actual charge.
Open source
NVIDIA reported $7.2 billion of provisions for inventory and excess purchase obligations in fiscal 2026.
Caveat. Annual figure, not solely H20.
Open source
AMD recorded a $440 million net inventory and related charge associated with MI308 export restrictions.
Caveat. Use company filing.
Open source
4. The expanded fab
A working system depends on leading-edge and mature-node wafers, memory, packaging, substrates, test, optics, power, cooling, and integration. The strategic “fab” is therefore the narrowest qualified layer that can stop the complete product.
That is why customer-backed supplier expansion and public support for packaging matter to the thesis. They show capital following the bottleneck even when the buyer does not become the asset owner.
Apple agreed to spend more than $30 billion on Broadcom FBAR filters through 2031.
Registry exception. The supplied registry labels this claim Corroborated but provides one source. Independent corroboration remains pending; the original state is preserved rather than silently rewritten.
Caveat. Long-term supply agreement, not Apple-owned PP&E.
Open source
Broadcom plans to invest $1.5 billion to expand its Fort Collins factory under the Apple agreement.
Registry exception. The supplied registry labels this claim Corroborated but provides one source. Independent corroboration remains pending; the original state is preserved rather than silently rewritten.
Caveat. Customer-backed supplier expansion.
Open source
SK Hynix’s CEO said in July 2026 that the memory industry could face its worst shortage in 2027 and that demand could exceed supply beyond 2030.
Registry exception. The supplied registry labels this claim Corroborated but provides one source. Independent corroboration remains pending; the original state is preserved rather than silently rewritten.
Caveat. Time-sensitive executive forecast, not an established future fact.
Open source
ASE expected its advanced-packaging business to double to about $3.2 billion in 2026.
Registry exception. The supplied registry labels this claim Corroborated but provides one source. Independent corroboration remains pending; the original state is preserved rather than silently rewritten.
Caveat. Company outlook.
Open source
Amkor’s Arizona advanced-packaging project was supported by up to $400 million of U.S. CHIPS funding and was expected to cost about $2 billion.
Registry exception. The supplied registry labels this claim Corroborated but provides one source. Independent corroboration remains pending; the original state is preserved rather than silently rewritten.
Caveat. Use final award document when available.
Open source
GM and GlobalFoundries entered a long-term agreement for dedicated U.S. semiconductor capacity.
Registry exception. The supplied registry labels this claim Corroborated but provides one source. Independent corroboration remains pending; the original state is preserved rather than silently rewritten.
Caveat. Illustrates vertical integration by contract.
Open source
5. Ownership, control, and the counterexample
Ownership is the strongest form of control because it can preserve residual value, learning, allocation, and public-policy standing. It is also the most unforgiving form of capital exposure. An uncompetitive or underutilized fab can destroy enormous value.
The conclusion is not that every designer should become an IDM. It is that the company must deliberately own, finance, or contractually control the bottleneck that governs its product, and must price the downside of that control honestly.
Texas Instruments announced plans to invest more than $60 billion across seven U.S. fabs.
Registry exception. The supplied registry labels this claim Corroborated but provides one source. Independent corroboration remains pending; the original state is preserved rather than silently rewritten.
Caveat. Investment plan; execution, timing, and utilization remain uncertain.
Open source
Intel’s experience shows that fab ownership without process competitiveness and utilization can destroy value.
Caveat. Editorial inference supported by losses, restructuring, and project changes.
Open sources
The strongest semiconductor businesses own the bottleneck, finance it, or control it through contracts that survive scarcity.
Caveat. Editorial synthesis, not a quoted fact.
Intel is a caution, while TSMC is evidence that competitive, highly utilized manufacturing ownership can produce superior economics.
Caveat. Avoid claiming that all owned capacity creates value.
Open sources
6. A falsifiable program
The thesis should be judged by what happens next: whether commitments remain elevated, contract duration lengthens, customers finance more expansion, strategic inventory persists, packaging attracts capital, and hybrid control models multiply. The prediction tracker starts at baseline because the supplied pack does not include current status evidence.
- Baseline · too earlyManufacturing commitments remain elevated
- Baseline · too earlyCommitment duration lengthens
- Baseline · too earlyMore customer-financed expansion
- Baseline · too earlyInventory remains strategic
- Baseline · too earlyPackaging attracts more subsidy and capital
- Baseline · too earlyThe fabless middle tier consolidates
- Baseline · too earlyHyperscalers increase direct supply-chain participation
- Baseline · too earlyAsset owners obtain better terms before building
- Baseline · too earlySecond-source programs become more expensive and explicit
- Baseline · too earlyInvestor metrics change
- Baseline · too earlyGovernments attach access conditions
- Baseline · too earlyThe distinction between fabless and IDM weakens