Hardware · Market
GPU Shortage 2026: Why It’s Happening and When It Ends
The 2026 GPU shortage isn’t really a shortage of GPUs — it’s a shortage of the two things needed to assemble them. The binding constraint is TSMC’s CoWoS advanced packaging, which bonds the GPU die to its high-bandwidth memory, and which has been sold out through 2026 with NVIDIA holding roughly 60% of capacity. The second constraint is HBM memory itself, where the three suppliers have sold out their entire 2025 and 2026 production. On top of that, hyperscalers committed around $650 billion in AI capital spending for 2026 and placed forward orders that locked up most of NVIDIA’s allocation. The result is data-center GPU lead times of 36 to 52 weeks and Blackwell deliveries slipping into 2027. Relief begins to arrive in late 2026 as new HBM and packaging capacity comes online, but the constraints are structural and persist into 2027.
Key takeaways
- It’s packaging, not silicon. TSMC’s CoWoS capacity to bond GPUs to memory is the single tightest bottleneck.
- HBM is sold out. All three memory makers have booked their entire 2025–2026 high-bandwidth memory production.
- Hyperscalers locked the supply. Around $650B in 2026 AI capex and forward orders consumed most of NVIDIA’s allocation.
- Lead times are 36–52 weeks. Blackwell deliveries have slipped into 2027, making this a structural constraint, not a blip.
- Relief starts late 2026. New HBM and packaging capacity eases pressure in Q4, but tightness persists into 2027.
If you’ve tried to buy a high-end GPU in 2026, you’ve met a queue measured in quarters rather than days. The reflex is to blame raw chip production, but that’s not where the bottleneck is — and understanding the real cause changes how you plan around it. This guide explains why the shortage exists, why it’s different from past chip crunches, what it does to lead times and prices, and when the pressure realistically eases.
What is actually in short supply?
The most important thing to understand is that this isn’t a shortage of GPU silicon in the way past chip crunches were. Modern AI accelerators need three things to become a finished product: a leading-edge logic die, stacks of high-bandwidth memory, and an advanced packaging step that bonds them together. The fabs can produce the logic dies; what they can’t do fast enough is package them and feed them enough memory. So the shortage is really a shortage of packaging capacity and memory, with the GPU die itself the least constrained part of the chain.
This is why the 2026 shortage is structural rather than a temporary demand spike like the pandemic-era crunch. Three distinct forces converge: TSMC’s advanced packaging is the hard bottleneck, high-bandwidth memory is fully allocated, and the generational shift from Hopper to Blackwell compressed years of demand into a narrow window. Each one alone would tighten supply; together they produce a constraint that capacity expansion can’t quickly resolve. The diagram traces where a GPU gets stuck.
Why this isn’t the pandemic chip crunch
It’s tempting to file the 2026 shortage alongside the pandemic-era chip crunch of a few years earlier, but they’re fundamentally different, and the difference matters for how long it lasts. The earlier crunch was a demand spike meeting a temporarily disrupted supply chain — once factories caught up and panic-buying subsided, it resolved. The 2026 shortage has no such self-correcting mechanism, because it’s driven by structural limits in advanced packaging and memory that take years and tens of billions in capital to expand, against demand that is still accelerating rather than normalising.
The scale underneath it is what makes it durable. AI chips represented less than a fraction of a percent of total wafer starts a couple of years ago yet already generate a large share of semiconductor revenue, and the capital spending behind them doubled to hundreds of billions annually in just two years. When demand grows that fast and the binding constraints are the slowest parts of the supply chain to build, the result isn’t a blip you wait out — it’s a constraint that reshapes IT planning cycles, invalidating the assumption that you can simply order hardware when you need it.
The CoWoS packaging bottleneck
CoWoS, which stands for Chip-on-Wafer-on-Substrate, is TSMC’s advanced packaging process that integrates the GPU compute die with its memory stacks on a single silicon interposer. That interposer is what enables the terabytes-per-second of memory bandwidth modern AI models need — something a traditional package-on-board design can’t approach. The critical point is that without this packaging step, even wafers built on TSMC’s most advanced nodes are just loose dies, not functional accelerators. Packaging, not wafer fabrication, has become the binding constraint on AI hardware.
The capacity numbers explain the squeeze. CoWoS output sat around 75,000 to 80,000 wafers a month in 2025 and is expanding toward 120,000 to 135,000 by the end of 2026 — a steep ramp, but against 2026 demand estimated near a million wafers, up from roughly 370,000 in 2024. TSMC’s CEO described the capacity as sold out through 2026, and NVIDIA alone holds an estimated 60 percent of it, with the top three customers accounting for over 85 percent. Capacity is being pre-booked faster than it comes online, which is why lead times stay long even as raw output nearly doubles.
HBM memory is sold out
The second constraint is high-bandwidth memory, and it’s nearly as tight as packaging. HBM is supplied by just three companies — SK Hynix with roughly half the market, Samsung with around 40 percent, and Micron with the remainder — and all three have sold out their entire 2025 and 2026 production. SK Hynix’s CFO stated plainly that the company has sold its whole 2026 HBM supply, while Micron has said it can meet only 55 to 60 percent of its customers’ demand. When the suppliers are this booked, no amount of GPU die production helps.
HBM is hard to scale for structural reasons. Each stack integrates multiple memory dies bonded with through-silicon vias, a specialised process with low yield tolerance that can’t simply be converted from standard DRAM production lines. It also consumes far more wafer capacity per gigabyte than ordinary memory. The result is something unusual for the memory industry, which normally sees prices fall over time: HBM3E contract prices rose around 20 percent for 2026. That price increase is a clear signal that demand is structurally outpacing what suppliers can build, the kind of pattern our AI infrastructure trends guide tracks across the sector.
How hyperscaler demand locked up supply
The third force is the sheer scale of hyperscaler spending, which absorbed the available supply before anyone else could reach it. The four largest cloud providers — Microsoft, Google, Amazon, and Meta — collectively committed around $650 billion in AI infrastructure capital expenditure for 2026, a roughly 71 percent jump year over year. That capital didn’t sit idle; it went into multi-billion-dollar forward orders for Blackwell GPUs placed throughout 2025, which consumed most of NVIDIA’s allocation capacity through 2026 and into 2027.
This is what turned a tight market into a locked one for everyone else. When the biggest buyers reserve capacity a year or more ahead, the allocation is effectively spoken for before smaller enterprises even enter the queue. It also compounded the generational timing problem: enterprises that sensibly held off on H100 purchases to wait for Blackwell all converged on the market at the same moment as the hyperscalers were upgrading, concentrating demand into a window the supply chain was never built to absorb at once.
The effect on lead times and prices
The practical effects show up first in lead times, which have stretched to a degree that breaks normal IT planning. Data-center GPU lead times now run 36 to 52 weeks, with some configurations quoted at 52 to 78 weeks, and Blackwell-class deliveries have slipped into 2027 for buyers who haven’t already ordered. The metric that actually gates supply isn’t instantaneous manufacturing capacity but the booking window — how far out you have to order — and that window keeps extending even as factories expand. The table summarises the three forces and their effects.
| Constraint | Why it’s tight | Relief timing |
|---|---|---|
| CoWoS packaging | Sold out; ~60% held by NVIDIA | H2 2026 ramp, into 2027 |
| HBM memory | 3 makers, all booked through 2026 | Late 2026 new capacity |
| Hyperscaler demand | ~$650B capex, forward orders | No near-term slowdown |
| N3 logic / power | Industry convergence on 3nm | Constrained into 2028 |
Prices moved in the same direction. Cloud GPU lease rates roughly doubled as on-demand capacity tightened, and the effects rippled beyond the data center — because Samsung and Micron diverted consumer memory lines toward lucrative AI HBM, ordinary DRAM grew scarce too, pushing retail memory prices up sharply and raising the cost of smartphones and PCs. The shortage of AI accelerators became, indirectly, a shortage of the memory in everyday devices.
Is it only GPUs that are constrained?
No, and this is part of why the shortage is so persistent. Beyond packaging and HBM, leading-edge logic wafers themselves have tightened as the entire industry converged on TSMC’s 3-nanometer node — NVIDIA’s next-generation Rubin, AMD’s MI400, Google’s newest TPUs, and Amazon’s Trainium all target the same process at once, while 2-nanometer capacity is already booked into 2028. The competition for a single node compounds the packaging and memory squeeze rather than relieving it.
There’s also a constraint that isn’t silicon at all: data-center power. Even teams that secure GPUs increasingly find that energy and the facilities to house dense, power-hungry systems are their next bottleneck, a reality that connects to broader questions of where compute can physically be built — including the data sovereignty trends shaping where infrastructure gets located. The shortage, in other words, is really a stack of overlapping constraints, which is why no single capacity expansion resolves it.
When will the GPU shortage end?
The honest answer is that relief is gradual, not sudden, and it begins in late 2026 rather than ending there. New HBM3E capacity from Samsung and Micron is expected online in late 2026, which should start easing the on-demand price premiums on H200 and B200, and TSMC’s CoWoS expansion brings meaningful new packaging capacity in the second half of the year. Because packaging gates output, when CoWoS expands, finished GPU supply can ramp faster than raw memory yield improvements alone would allow. So Q4 2026 is the realistic point where pressure starts to reduce.
But starting to ease is not the same as clearing. Industry consensus, echoed by executives across TSMC, the memory makers, and NVIDIA, is that the binding constraints persist through at least the first half of 2027, with some analysts pointing to late 2027. This isn’t a transient disruption you can wait out — it’s better treated as the new baseline for AI hardware planning. The teams that navigate it best stop assuming supply will normalise on their old timelines and build their roadmaps around constrained availability instead.
How should you plan around the shortage?
Given that the shortage is structural, the practical response is to plan around availability rather than wait for it to disappear. The terminal lays out the strategies that work in this market.
# Securing AI compute when GPUs are sold out RENT FIRST … cloud beats a 52-week purchase queue for most teams ORDER EARLY … place purchase orders months ahead, via allocation partners TAKE WHAT SHIPS in-stock H200 beats a backlogged B200 you can’t get RIGHT-SIZE … quantize to fit a smaller, available GPU GO WIDER … evaluate AMD MI-series and neocloud providers DON’T OVER-SPEC capacity you can’t get is worse than a tier down # Plan for constrained supply as the baseline, not the exception.
Two of these matter most. Renting from the cloud sidesteps the purchase queue entirely, which is why it’s the fastest path to compute for most teams right now — and shopping specialty providers rather than hyperscalers can offset the doubled lease rates. And taking what’s actually available beats holding out for the newest silicon: an H200 you can deploy this quarter delivers far more value than a B200 stuck behind a year-long backlog, a trade-off our GPU server buying guide works through in detail. For teams that want owned, dedicated infrastructure sized to what they can actually deploy today, our dedicated servers in Toronto can be configured around currently available GPU tiers — while the discipline of right-sizing and renting for burst capacity keeps you productive without waiting out a shortage that won’t fully clear until 2027.