Reading time: approximately 9 minutes
You wanted to build a PC this spring. The market had other plans.
You opened PCPartPicker, dropped in a mid-range CPU, a respectable GPU, 32GB of DDR5, a 2TB NVMe drive, a decent PSU, and a 240mm AIO. You hit the total button. The number that came back was not the number you were expecting. It was, in fact, the number you would have expected if you had specified a professional workstation and accidentally summoned a demon.
Welcome to 2026. The hyperscalers ate the silicon supply chain, and all you got was this overpriced SSD and a water-cooling unit assembled from what used to be affordable copper.
This is not panic. This is not a Reddit thread. This is a structured autopsy of exactly what happened, with the actual numbers, the actual companies responsible, and a realistic timeline for when — or whether — any of this gets better.
The CoWoS Carve-Up: NVIDIA Took the Table and Left You the Crumbs
Everything traces back to packaging. Specifically, TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging technology — the process that lets NVIDIA stack HBM memory directly alongside its GPU dies on a silicon interposer, producing the dense, power-efficient compute tiles that make H100s and H200s worth $25,000 to $40,000 per unit to the people buying them by the rack.
NVIDIA has booked somewhere between 800,000 and 850,000 CoWoS wafers for 2026 . That represents more than 60% of TSMC’s total CoWoS capacity for the year. Sixty percent. Of an advanced packaging node that every other chipmaker on the planet also needs. AMD needs it for MI300X derivatives. Apple needs adjacent capacity for leading-edge SoCs. Qualcomm, MediaTek, and Broadcom all have programs queued. NVIDIA essentially called dibs on the entire buffet and sent everyone else a doggy bag.
TSMC itself is not complaining. HPC and AI now account for 55% of TSMC’s total revenue , a figure that would have seemed delusional five years ago when smartphones were the assumed forever-growth category. TSMC is plowing $52 to $56 billion in CapEx into 2026 specifically to expand capacity — but that expansion feeds AI demand first, last, and in between. Consumer electronics ride the wake, if there is one.
The downstream effect on consumer GPUs is stark. The RTX 5090 launched at $1,999 MSRP . It is currently trading on the secondary market between $3,000 and $3,500 , with some analyst price targets putting the ceiling near $5,000 if NVIDIA’s allocation strategy for the second half of 2026 remains as constricted as it has been in Q1. This is not a scalper story. Scalpers are a symptom. The disease is that NVIDIA physically cannot produce enough cards when it is dedicating its TSMC packaging allocation to H200s that sell for twenty times the price at twenty times the margin.
Your RAM Bill: Micron Killed Crucial and Raised Your Prices on the Way Out
In December 2025 , Micron announced the end of the Crucial consumer brand, effective February 2026 . If you have been buying memory for more than a decade, you know Crucial as the reliable, reasonably priced house brand that Micron ran as a retail front for its consumer DRAM and NAND. It is gone. Micron redirected those production lines — and that sales infrastructure — toward AI and enterprise customers. The stated rationale was margin optimization. The unstated rationale was that a single H100 requires HBM that costs more than several hundred Crucial kits, and Micron’s shareholders have opinions about that math.
The immediate market consequence arrived on schedule. DDR5 contract pricing rose more than 100% in the span of a single quarter. A 32GB DDR5 kit from Samsung that was priced at roughly $149 is now selling for approximately $239 — and that is on a good day when stock is actually available. The pricing move was not subtle or gradual. It was a step function, the kind you see when supply gets pulled rather than when demand spikes.
The structural reason is worse than the pricing story. SK Hynix has confirmed that HBM consumes three times the wafer capacity per gigabyte compared to standard DRAM . Three times. That means every HBM3E stack that ships in an H200 or an AMD MI300X represents three gigabytes of conventional DRAM that did not get made. Multiply that across the hundreds of thousands of AI accelerators being produced monthly, and you get a very clean explanation for why your memory upgrade quote looks like a small car payment. SK Hynix has also provided the timeline that nobody wanted to hear: meaningful consumer DRAM supply relief is not expected before 2028 .
The Storage Apocalypse: Every NAND Vendor Sold 2026 Twice
NAND flash pricing has done something genuinely unusual: it overcorrected from a two-year glut and then kept going. The canonical data point that is making SSD builders reach for antacids: a 1TB TLC NAND chip was priced at approximately $4.80 in July 2025 . By early 2026, that same chip was trading at $10.70 . That is a 123% increase in roughly six months . In NAND market terms, this is what a supply reset looks like when it meets structural demand that was not in the model.
“We are heading into a pricing apocalypse through 2027. Every major NAND manufacturer has already committed their 2026 output. There is no spot market to speak of.” — Phison CEO S.C. Huang , industry briefing, Q1 2026
Phison is one of the largest SSD controller manufacturers in the industry. When the CEO of the company that makes the chips managing your SSD uses the word “apocalypse,” it is reasonable to take the statement at face value rather than as a marketing quirk.
The mechanism here is enterprise SSDs and QLC NAND for hyperscale storage deployments. Data centers need storage too — enormous, sustained amounts of it. AI training runs generate checkpoint data at scales that require petabytes of fast flash. Inference infrastructure needs low-latency local storage at every node. The hyperscalers placed forward contracts on 2026 NAND production months ago. Consumer and DIY channels are purchasing from whatever is left after those contracts are filled, which in practical terms means premium pricing on constrained availability. All 2026 production capacity is effectively spoken for. If you need a 4TB drive for a new build today, you are paying 2026 prices whether you like it or not.
Even Your PSU and Cooler Got Expensive: The Copper Problem Nobody Warned You About
The silicon story is well-documented. The materials story is less discussed but equally real. Data center construction — not just the chips inside, but the physical infrastructure — consumes copper, tin, and silver at industrial scale. Power delivery systems for a single hyperscale facility can require thousands of tons of copper wiring and busbars. Cooling infrastructure — whether that is immersion cooling, direct liquid cooling, or conventional CRAC systems — adds more. Data center construction in 2025 and 2026 has been running at a pace that is straining commodity markets for all three metals.
The result for PC builders: PSU prices are up 6 to 10% across most major SKUs in the first quarter of 2026. CPU coolers and AIOs are up 6 to 8% . These are not massive price swings individually, but they compound. Add 7% to your PSU, 7% to your cooler, and those line items push your build total in the wrong direction at a time when your GPU and RAM are already doing their own damage.
Corsair, be Quiet!, and Noctua have all adjusted retail pricing on flagship products in the past two quarters. None of them issued press releases about it. You notice when you go to buy the thing.
The Spending Machine: Where $690 Billion Actually Goes
To understand why none of this is accidental or temporary, look at where the money is flowing. The five major hyperscalers have collectively committed the following to infrastructure spending in 2026:
| Company | 2026 CapEx Estimate | Primary Driver |
|---|---|---|
| Amazon (AWS) | ~$200B | AI inference infrastructure, regional data centers |
| Google (Alphabet) | $175–185B | TPU v5 deployment, Gemini training clusters |
| Microsoft (Azure) | ~$120B+ | OpenAI partnership infrastructure, Copilot scale |
| Meta | $115–135B | Llama training, recommendation model acceleration |
| Oracle | ~$50B | OCI GPU cluster expansion, sovereign AI contracts |
| Combined Total | ~$660–690B |
To put $690 billion in context: that is roughly the annual GDP of Switzerland , deployed in a single year, almost entirely into silicon, copper, power infrastructure, and cooling. Every dollar of that CapEx is competing for the same fabrication capacity, the same materials, and the same logistics infrastructure that PC components use. The PC market is not small — but it is not $690 billion, either.
NVIDIA’s H100 at $25,000 to $30,000 per unit and the H200 at $30,000 to $40,000 per unit represent a pricing tier that makes consumer GPU economics look like rounding errors. NVIDIA does not need to optimize margins on a $1,999 RTX 5090 when it can sell an H200 for twenty times that price to a customer who will take a thousand of them with no complaints about availability.
When Does It Actually Get Better?
The honest answer is: not soon, not uniformly, and not without conditions.
SK Hynix’s public guidance puts consumer DRAM supply normalization at 2028 or later. That is not a typo. Two years. The HBM capacity build is a multi-year capital commitment — fabs do not get retooled in a quarter. Samsung, SK Hynix, and Micron all invested heavily in HBM3 and HBM3E capacity in 2024 and 2025. That capacity is locked into production runs with forward contracts. Repivoting a portion of it back to DDR5/LPDDR5 for consumer markets would require both technical retooling and a willingness to walk away from margins that are structurally better in HBM. No public statements from any of the three suggest that is being considered before 2027 at the earliest.
TSMC’s pricing is a separate problem. The company has implemented multi-year price escalators locked in through at least 2029 , tied to the capital commitments its hyperscaler and AI customers have made. These are not annual renegotiations — they are structured agreements. Even if AI demand cooled tomorrow, TSMC’s contracted pricing would continue to flow through to every fabbed component for years. AMD, Intel, and GPU vendors all source from TSMC at prices that get renegotiated up, not down.
NAND has a slightly shorter fuse. The Phison CEO’s “apocalypse through 2027” framing implies that 2028 could see meaningful relief as new capacity from Samsung’s P3 expansion in Pyeongtaek and SK Hynix’s M15X facility comes online. But those facilities are also partly committed to enterprise-grade QLC and storage class memory for hyperscalers. Relief in consumer TLC pricing depends on where those new wafer starts get allocated — and the hyperscalers are not passive in those conversations.
What You Should Actually Do Right Now
If you are planning a build or an upgrade in the next 12 months, here is the practical guidance that follows from everything above — absent wishful thinking about market corrections that are not coming on the schedule you want.
- Buy RAM now if you need it, but do not overbuy. DDR5 pricing has already repriced. There is no evidence of near-term relief, and waiting is unlikely to reward you. Get exactly what your build requires. Do not speculate on 64GB “just in case” unless you have a concrete use case.
- On storage: buy 2TB, skip 4TB. The per-GB pricing on 4TB consumer SSDs is still carrying the worst of the NAND premium. 2TB drives have a slightly better price-per-GB right now and are more widely in stock. If you need 4TB, consider two 2TB drives rather than one 4TB.
- On the GPU question: consider last-gen. The RTX 4080 Super and RX 7900 XTX are not the GPUs of the moment, but they are available at something approaching rational prices and will not make your games unplayable. An RTX 5090 at $3,500 on the secondary market is not a good deal — it is a premium on scarcity, not performance.
- Lock in PSU pricing on known-good units immediately. Copper pricing is unlikely to retrace. The Seasonic, Corsair, and be Quiet! units you have been watching are not going to get cheaper. This is a “buy the one you want now” category.
- Do not wait for TSMC pricing to normalize. It will not, on any timeline relevant to a build you are planning this year. Factor current GPU and CPU die costs into your budget as a floor, not a ceiling.
- If you are on a tight budget: AM5 + B650 + last-gen GPU is still a strong platform. Ryzen 7000-series CPUs have not been caught up in the same CoWoS crunch. A Ryzen 7 7800X3D and a B650 motherboard remain relatively sane on pricing. Build the platform now, upgrade the GPU later when — and if — supply loosens.
The GDDR7 and GDDR6 memory on consumer GPUs now represents more than 80% of the bill of materials on most mid-to-high-end cards. That ratio will only compress when DRAM capacity returns to consumer allocation — which, again, SK Hynix is not promising before 2028. Any GPU you buy between now and then is priced to reflect that constraint.
The Part Nobody Wants to Write
PC building as a hobby has survived GPU mining booms, pandemic supply shocks, and multiple DRAM price-fixing cycles. It will survive this too. But “survive” is doing meaningful work in that sentence.
The 2026 squeeze is structurally different from prior disruptions because it is not a demand spike or a logistics shock. It is a deliberate, well-capitalized reallocation of the global semiconductor supply chain toward a higher-margin customer class. The hyperscalers are not behaving irrationally. They are executing exactly as capital markets reward them to execute. The collateral damage is a hobby market and a consumer electronics supply chain that cannot outbid them.
The next concrete milestone to watch: TSMC’s Q2 2026 earnings call , where any commentary on CoWoS expansion timelines or reallocation signals will be the most credible indicator of when consumer GPU supply begins to normalize. Until then, build what you can afford at current prices, or wait with clear eyes about what you are waiting for.