TSMC's CoWoS advanced packaging line, the single industrial stage that assembles every high-density NVIDIA AI GPU with SK Hynix and Samsung HBM stacks, is fully allocated through mid-2027, with demand running 1.4 to 1.6 times supply over the next 18 to 24 months and hardware lead times stretching to 36 to 52 weeks [S3].
Against that upstream ceiling, the downstream value pool is splitting in two: a primary lease-and-finance market where residual-value assumptions are now the central underwriting variable on GPU collateral, and a secondary resale market where a six-year-old RTX 3090 still benchmarks as the best dollar-per-token value for LLM workloads [S1][S2]. The global GPU market is valued at US$ 102.8 billion in 2026 on its way to US$ 652 billion by 2033, a 30.2% CAGR [S4].
Where the Bottleneck Actually Sits in 2026
The shortage stopped being a wafer problem more than a year ago. NVIDIA and TSMC have materially expanded both Hopper and Blackwell production since late 2024, but the gate is now TSMC's CoWoS line, a Chip-on-Wafer-on-Substrate process that bonds the compute die to HBM stacks on a shared silicon interposer, and TSMC's CoWoS capacity is fully allocated through mid-2027, with demand exceeding supply by a factor of 1.4 to 1.6 over the next 18 to 24 months [S3].
That second doubling is still not enough. Microsoft, Google, Amazon, and Meta placed multi-year forward contracts in 2024 and 2025 that absorb most of NVIDIA's allocation through the end of 2027, so neocloud operators without 2024-dated orders now see a spot market that is effectively a residual-allocation drip feed [S3]. For buyers specifying any industrial pressure transmitter or flow meter into the same AI-build datacenters, the practical effect is that GPU delivery dates now drive the entire rack-level bill of materials, not the other way around.
Upstream Value Pool: Who Captures the Packaging Margin
Upstream, value concentrates in three nodes: TSMC at CoWoS, the HBM suppliers (SK Hynix, Samsung) and NVIDIA at the system design layer. Because the compute die and the HBM stacks cannot be sold separately as a finished AI product, every high-density AI GPU effectively routes through one supplier at one stage, a structural feature SemiAnalysis and Silicon Analysts both cite when modeling the 1.4–1.6 demand-to-supply ratio through 2027 [S3].
Hyperscalers locked in Blackwell allocation in 2024, paying forward-contract premia that effectively transfer scarcity rent upstream. The closer analogue is the 1970s IBM mainframe era that American Compute uses as a historical parallel for GPU residual risk: aggressive forward bookings are profitable until the supplier cuts price, at which point lessors carrying inflated residuals took the loss [S1]. For 2026 procurement, the practical signal is that any datacenter bill of materials, including the lamps and light fittings and the construction machinery and equipment for the building shell, has to be sequenced backwards from a confirmed CoWoS slot, not from a vendor quote.
Downstream Value Pool: Residual as the Underwriting Variable
For the lender, lessor, and equity investor, the residual is the number that decides whether a GPU deal closes. American Compute's 2026 report, drawing on transaction data covering more than 600,000 secondary-market units, treats residuals not as a predicted depreciation curve but as a band, and ranks NVIDIA margin compression and an AI bubble pop as high-probability adverse events, with bullwhip over-ordering and hyperscaler refresh waves at medium, and pure obsolescence at low [S1].
The deal-shaping range is narrow: most commercial lenders require a 36-month residual above 5% of original cost, while the more aggressive end of the market, and the lessor side of the FMV lease, is underwriting residuals north of 10% of equipment cost over five years, premised on a useful life American Compute places at up to eight years [S1]. The reason this matters downstream is that overbooked residuals hit lessors first, exactly as they did in the 1970s IBM precedent, because the FMV lease books the residual on day one and must realize it as cash at term end [S1].
The Secondary Market as a Parallel Value Pool
Where the primary market is rationed by CoWoS, the secondary market is setting real prices for compute. The 2026 value observation that a 2020-era RTX 3090 still benchmarks as the best dollar-per-token card for many LLM workloads is not nostalgia, it is a direct consequence of new-card pricing being driven by macro factors (the US federal debt crossed US$ 40 trillion in August 2026 per [S2]) rather than by die-area cost, with 2026 secondary-market supply including large-scale recycling of datacenter cards whose first-life workloads have ended.
That secondary pool is also what makes the residual bankable: American Compute's 600,000-unit secondary-market dataset is the empirical anchor for the residual bands, and the same supply is what lets a lessor with a maturing 36-month FMV lease actually clear the residual instead of taking a writedown [S1]. A useful sanity check when sizing a GPU financing bid is to ask whether the residual assumption clears against observable secondary-market comps, not against vendor list price.
Application Mix and Where the Volume Actually Lands
By application, Gaming is the leading 2026 segment and AI and Machine Learning is the fastest-growing segment through 2033, and that mix shift is what re-primes the upstream/downstream split every quarter [S4]. North America carries the largest 2026 regional share, while Asia Pacific is the fastest-growing regional block on the back of Japan and South Korea AI and cloud investment [S4].
For procurement teams, the practical reading is that the consumer-grade tail (gaming, content creation, integrated GPUs) keeps volume moving at the package level, but the high-ASP, high-residual AI and inference cards are where the upstream packaging constraint binds and where residual-value discipline decides deal economics. The geographic spread of AI build-out is widening beyond the US, with regional data-center investments in India, Japan, and South Korea indicating that the AI segment is no longer a US-only story [S4].
Limits, Failure Modes, and What to Watch Next
Three failure modes are explicit in the research: overbooked residuals (the IBM-mainframe precedent) [S1], an AI bubble pop that compresses secondary prices faster than lease terms mature [S1], and a packaging normalization before 2028 that invalidates forward contracts priced on scarcity [S3]. Two signals worth tracking: TSMC's next CoWoS capacity update, which will set the earliest credible 2028 supply normalization date, and secondary-market clearing prices for H100 and RTX 3090 cohorts, which feed directly back into residual bands on new deployments.
This topic is covered further in Roller Chain Selection for Agriculture Machinery: Pitch, Duty Class, and Drive Fit.