The 2026 DRAM Supercycle Is Permanent: What 'Not Cyclical' Means for Every Self-Hosted Hardware Bet

A 32 GB DDR5 kit that bottomed below $80 in mid-2025 now starts at $375. Conventional DRAM contract prices rose 90–95% quarter-over-quarter in Q1 2026 — the largest single-quarter jump on record — and another 58–63% in Q2. NAND flash, which had been the laggard, is now projected to rise 70–75% in Q2, outpacing DRAM for the first time in this cycle.

If you run a self-hosted PaaS — or you are evaluating whether to — those numbers land differently than they do for a consumer shopping for RAM. Every "self-host on a €10 Hetzner box" cost comparison this site (and every other one) has published was built on a hardware price that just moved. Hetzner has already repriced twice in 2026. The question is not whether the math changed. It is how much survives, and whether the answer is still "self-hosting wins" once you recompute with real numbers instead of a frozen baseline.

This post does that recomputation. It traces why the shortage is structural, what "permanent, not cyclical" actually means in IDC's framing, and then runs the same three representative workloads through both the old and new pricing on each side of the comparison — flat-rate Hetzner hardware and metered PaaS billing — so you can see the absolute increase and the preserved delta side by side.

TL;DR — The supercycle in five numbers, and one table​

The table that matters — three workloads, priced before and after the surge on both sides — is in Recomputing the math. The short version: absolute costs are up 30–175% on the self-hosted side depending on instance family, and 15–40% on the metered PaaS side (they already passed DRAM costs through plus margin). The gap between the two narrows in euros but widens in percentage terms for bandwidth-heavy workloads, because Hetzner's 20 TB included bandwidth didn't get repriced at all.

Why this isn't a normal shortage — the HBM zero-sum game​

DRAM shortages are not new. What is new is the mechanism.

Samsung, SK Hynix, and Micron together control more than 95% of global DRAM production. All three have systematically reallocated fabrication capacity away from conventional DDR4/DDR5 toward high-bandwidth memory (HBM) — the vertically stacked memory that sits directly on top of every Nvidia Blackwell GPU and every hyperscaler AI accelerator. The economics make the choice obvious for the manufacturers: HBM sells at a steep premium, is sold out through 2026, and is projected to grow 58% to $54.6 billion in 2026, accounting for nearly 40% of the entire DRAM market.

The conversion math is brutal. Micron has disclosed a 3:1 wafer conversion ratio — each HBM wafer displaces roughly three conventional DRAM wafers. Some analyses put the ratio as high as 4:1 when the logic die is included. Every additional wafer shifted to HBM is three wafers of DDR5 that simply do not get made. SEMI China's president noted that even though the three majors directed 70% of new capacity toward HBM in 2026, the overall capacity shortfall still runs 50–60%.

That is why AI data centers now consume roughly 70% of all memory chips produced worldwide, up from 20–30% as recently as 2022, and why TrendForce's DRAM output growth forecast of just over 20% for 2026 is still not enough to close the gap. It is also why the shortage hits every memory tier simultaneously. HBM is not a niche product siphoning a small slice — it is a reallocation large enough to starve the mainstream.

For a self-hosting cost model, this matters because there is no "unaffected" hardware tier to retreat to. DDR4 is not spared. LPDDR5X for mobile is projected to jump 78–83% in Q2 alone. NAND flash — the SSDs in every Hetzner box — is now rising faster than DRAM. Even the ARM-based lines that saw milder Hetzner increases (~30% vs. 113–175% on CCX/CPX) are built from the same constrained commodity.

"Permanent, not cyclical" — what IDC actually said and when relief comes​

IDC's framing, first published in December 2025 and widely cited through 2026, is precise:

"Not just a cyclical shortage driven by a mismatch in supply and demand, but a potentially permanent, strategic reallocation of the world's silicon wafer capacity."

The distinction matters. A cyclical shortage is a timing problem: demand overshoots, manufacturers add capacity, prices mean-revert within 12–18 months. A structural reallocation is a priority problem: manufacturers chose to serve a higher-margin customer (AI infrastructure) and will keep choosing that customer as long as the margin gap persists. The shortage is not an accident to be fixed — it is a business model working as intended.

Three data points make the "permanent" label concrete:

Supply growth is capped. IDC projects 2026 DRAM supply growth of just 16% year-over-year and NAND growth of 17% — both below historical norms and both dwarfed by AI-driven demand growth. You cannot grow your way out of a reallocation when the reallocation itself determines how much growth goes to the constrained product.

New fabs are years away. Micron's first new Idaho fab is not expected to begin DRAM production until late 2027 at the earliest. Samsung's P4L and SK Hynix's M15X — the mega-fabs that could materially add supply — are on similar timelines. Even when they come online, their initial output is already largely committed to HBM4, the next-generation stack slated for mid-2026 with 2.0+ TB/s bandwidth.

Forecasts keep sliding right. Micron's CEO projected in June 2026 that supply would remain tight through 2027, with gradual improvement beginning in 2028. Kearney's PERLab put the horizon at 2030. SK Hynix has warned the shortage could persist past 2030. Bloomberg reported DRAM spot prices up nearly 700% year-over-year by July 2026 — a spot-market overshoot that signals how far contract prices still have to run to clear.

For anyone amortizing hardware, the implication is direct: do not budget against a price that assumes mean reversion in two quarters. The honest planning assumption for late 2026 is a trend line that keeps climbing, not a frozen price point that will come back down.

Recomputing the math — 3 workloads, before and after the surge​

The core deliverable. Same three workloads, priced on both sides of the comparison using actual 2026 numbers — not the pre-hike baselines most cost comparisons (including this site's own earlier ones) were built on.

The workloads​

Self-hosted on Hetzner — pre-surge vs. post-surge​

Hetzner repriced cloud servers in two waves in 2026. The April 1 increase hit all regions (+30–37% in Germany/Finland, up to 38% in the US/Singapore). The June 15 "portfolio standardization" hit dedicated-vCPU lines hardest: CCX (dedicated) up 2.1–2.73x and CPX (shared AMD) up 2.4–2.75x in EU regions, with ARM and Intel-shared lines rising a comparatively mild ~30%. Setup fees were cut to partly offset the monthly increase, and existing contracts were grandfathered — but every new MachineDeployment or rescale pays the new rate.

Applying that to the three workloads (cheapest fitting instance per workload):

The pattern is stark: memory-light workloads on ARM/Intel-shared lines are up ~30–37%, while memory-heavy workloads that need CCX/CPX dedicated cores are up 2–2.75x. Your instance family is now a cost decision, not just a performance one.

Metered PaaS — same workloads (2026 pricing)​

PaaS vendors already passed DRAM and hardware cost increases through in their own 2026 repricings — plus their platform margin on top.

PaaS estimates use each vendor's published 2026 unit pricing. "Metered" totals include the platform's own hardware-cost passthrough — these vendors buy the same DRAM Hetzner does, then add margin. Egress is the dominant variable on Render; RAM/vCPU on Railway; the long tail of sub-meters (NAT, snapshots, private networking) on Fly.io.

The comparison — does self-hosting still win?​

Reading the table honestly:

• Absolute costs are up on both sides. Self-hosting is not immune — a memory-heavy workload on CCX hardware costs 2x what it did a year ago. Anyone quoting a frozen pre-surge Hetzner price is stale.
• The percentage advantage narrows but does not close. The gap shrinks most on memory-heavy workloads forced onto CCX/CPX (where Hetzner's hike was steepest) and least on ARM-friendly workloads. Even the worst case (workload C on CCX) is still roughly half the cheapest metered alternative.
• Bandwidth is where the gap widens under stress. Every row above includes egress in the PaaS total but not in the Hetzner total — because Hetzner's 20 TB included bandwidth was not repriced. For workload C at 1 TB egress, Render alone charges ~$146 in overage; Hetzner's included 20 TB absorbs it at zero marginal cost. As egress grows, the flat-rate advantage compounds rather than erodes.
• Sensitivity — what if DRAM doubles again? If contract prices add another 50–60% (one more TrendForce-style quarter), expect Hetzner ARM/Intel-shared lines to rise another ~20–30% and CCX/CPX another ~40–60% (manufacturers pass through with a lag and Hetzner with another). The metered PaaS side would rise similarly plus the platform's margin on the increase. The absolute gap grows; the percentage gap stays roughly stable or widens slightly for bandwidth-heavy shapes.

Why the advantage survives when both sides climb​

If both sides buy the same DRAM, why does self-hosting still come out ahead? Three structural reasons — none of which depend on DRAM prices going back down.

1. One passthrough vs. two​

When DRAM costs rise, Hetzner raises the box price once. When the same DRAM cost hits a PaaS vendor, the vendor raises its underlying hardware cost and keeps its platform margin — whether that margin is an explicit seat fee (Vercel $20/seat, Qovery $299–$2,099/mo control-plane fee), a usage markup (Railway's $10/GB RAM includes margin over raw hardware), or an egress multiplier ($0.15/GB on Render vs. bulk bandwidth Hetzner already bought). The PaaS tenant pays the commodity increase twice: once in the hardware, once in the margin applied to that hardware. The self-hosted tenant pays it once.

Northflank's BYOC model makes this explicit: you pay Northflank's compute-plan rate plus whatever AWS/GCP/Hetzner charges underneath. Qovery's "agentic infrastructure platform" at $299–$2,099/mo is a control-plane subscription layered on top of a cloud bill you already pay. A Cluster API fleet on owned Hetzner hardware has no second layer — the Hetzner invoice is the invoice.

2. Bandwidth was never repriced​

Hetzner's 20 TB included bandwidth per server survived both 2026 repricings untouched. Every PaaS egress meter — Render's 5GB/25GB included then $0.15/GB, Fly.io's per-GB with a $0.12/GB Africa/India premium, AWS/GCP/Azure's $0.087–0.12/GB — is a separate line that scales with traffic. As apps get more bandwidth-heavy (larger payloads, more API calls, AI-generated content), the flat-rate side's advantage is not just preserved — it grows. The same pattern shows up in object storage: Cloudflare R2 at $0 egress vs. S3 at $0.09/GB is a 127x spread that no DRAM repricing touches.

3. A trend line is plannable; a ratchet is not​

Hetzner's increases, while painful, were announced, dated, and grandfathered for existing contracts. You can provision before a hike, choose a cheaper instance family, or hold reserved capacity. PaaS repricings — Vercel's four events in 20 months, Render's 100 GB → 5 GB egress cut, Fly.io's two new meters in early 2026, Cloudflare Workflows' per-step billing starting August 2026 — change what gets measured, not just the price per unit. Forecasting next quarter from this quarter's invoice requires re-reading pricing pages, not just multiplying by a new unit cost. That predictability tax is invisible in any single price comparison but compounds every quarter the ratchet turns.

What to actually do about it — a hedging checklist for platform operators​

If you run a Cluster API fleet (or plan to), the supercycle is not a reason to abandon the hardware bet. It is a reason to be more deliberate about how you place it.

Pick the right instance family. The June 15 repricing made this the single highest-leverage decision. ARM (CAX) and Intel-shared (CPX Intel, CX) lines rose ~30%. AMD shared (CPX) and dedicated (CCX) lines rose 113–175%. If your workload fits on ARM — and most API/SSR/queue workloads do — the supercycle costs you a third of what it costs on CCX. Benchmark before defaulting to dedicated cores.

Provision before the next hike. Hetzner grandfathered existing contracts at old pricing through both 2026 events. New orders and rescales pay the new rate. If you know you will need capacity in the next two quarters, provisioning it now locks the current rate. The same logic applies to reserved or committed-use discounts on any provider.

Hold headroom, not just-in-time capacity. Hetzner also restricted new cloud-server creation starting June 26, 2026 — a hard capacity ceiling distinct from the price increases, caused by the same component shortages. A MachineDeployment scale-up can now fail with "no capacity in this location" as a distinct error from cost. Multi-location and multi-server-type fallback in your CAPH provider config is no longer optional.

Diversify the provider baseline. Netcup's RS 1000 G12 (4 dedicated AMD EPYC cores, 8 GB DDR5, 256 GB NVMe for €8.74/mo) now undercuts post-hike Hetzner CCX on raw dedicated cores. It is not a drop-in CAPH replacement today — API maturity, regional footprint, and Cluster API provider tooling all lag — but it is worth tracking as a second data point so "Hetzner price" stops being a single number in your model.

Model the trend, not the point. Replace any frozen "€X Hetzner box" constant in your cost model with a trend line. A simple approach: take your current Hetzner invoice, apply +25–35% for ARM/Intel-shared fleets and +80–120% for CCX/CPX-heavy fleets as the realized 2026 impact, then project +15–25% per additional quarter of DRAM price growth. Compare that trend against the PaaS trend (hardware passthrough + margin + egress growth) rather than a single-month snapshot.

The honest framing for late 2026​

The cheapest self-hosting pitch — "a €4 Hetzner box vs. $50 on a PaaS" — was always a price point. What the DRAM supercycle forces is a price trend line — and the honest version of that trend line slopes up for everyone.

What survives is the structure: one hardware passthrough instead of two, bandwidth included instead of metered per gigabyte, and a pricing model that moves on announced dates with grandfathered contracts instead of on the vendor's changelog. Those properties do not depend on DRAM being cheap. They depend on owning the hardware boundary yourself.

IDC's word choice — "permanent, not cyclical" — is worth taking literally. If the reallocation of silicon toward HBM is structural, then the right question is not "when do prices come back down" but "which cost structure best absorbs a commodity that keeps repricing quarter over quarter." On that question, a flat-rate fleet on owned hardware — even a more expensive flat-rate fleet than a year ago — is still the structure that bends less.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. The cost advantage in this post comes from the same property bex is built on: the Hetzner invoice is the only invoice. No per-GB RAM meter, no egress overage, no control-plane subscription on top. Star the repo on GitHub or deploy your first app today.

Sources​

• TrendForce press releases, Jan–Feb 2026: Q1 2026 conventional DRAM contract price forecast revised from 55–60% to 90–95% QoQ; PC DRAM +105–110%; NAND +33–38% → 55–60%; Q2 forecasts 58–63% DRAM, 70–75% NAND — TrendForce 1Q26 survey (Jan 5), TrendForce revised outlook (Feb 2)
• TrendForce via TechTimes / Electronics Weekly, Jun 2026: 32 GB DDR5 kit floor $374.97 on PCPartPicker; DRAM spot prices +700% YoY (Bloomberg, Jul 2026)
• IDC via Altium / Livemint / Datafloq, Dec 2025–2026: "Not just a cyclical shortage but a potentially permanent, strategic reallocation of the world's silicon wafer capacity"; 2026 DRAM supply growth 16% YoY, NAND 17%; AI DCs ~70% of memory consumption
• Micron disclosures, 2026: 3:1 HBM-to-DRAM wafer conversion ratio; new Idaho fab not before late 2027; CEO tight-through-2027 guidance (Jun 2026)
• SEMI China / Gate / Motley Fool, 2026: HBM market $54.6B in 2026 (+58% YoY), ~40% of DRAM market; 70% of new capacity to HBM yet 50–60% shortfall; OpenAI Stargate ~900K wafers/mo by 2029
• Kearney PERLab / SK Hynix / Bloomberg, 2026: shortage until 2030+ forecasts; SK Hynix warns past 2030
• Hetzner announcements, 2026: April 1 cloud +30–37% (Germany/Finland), up to 38% US/Singapore; June 15 CCX +2.1–2.73x, CPX +2.4–2.75x; setup fees cut, existing contracts grandfathered; Jun 26 capacity restrictions
• Cloud hosting price tracking, 2026: AWS H200 EC2 +15% (DataCenterDynamics); Railway $10/GB RAM + $20/vCPU; Render egress 100GB→5GB/25GB + $0.15/GB; Fly.io NAT/metrics/private-networking new meters (Feb 2026)