Verdict: AI infrastructure demand is driving up memory prices, and the cost shock is finally reaching the budget cloud providers that small teams rely on. Hetzner raised prices up to 37% in April 2026, then again up to 176% for some dedicated-vCPU plans in June 2026; OVHcloud has warned of further hikes in 2026. The good news: with a few architectural moves, most teams can offset a large part of the increase.
Last verified: 2026-06-17 · TL;DR: AI needs HBM → DRAM supply gets squeezed → server RAM costs more → budget hosts raise prices. Big hyperscalers absorb this more easily because their margins on egress and managed services are much higher.
What is actually happening
For years, Hetzner, OVHcloud, and similar budget hosts offered bare compute at a tiny fraction of AWS, Google Cloud, or Azure prices. In 2026, those price gaps are shrinking.
- Hetzner raised prices across its cloud portfolio on 1 April 2026, citing "drastic price increases" in infrastructure operating costs and new hardware procurement. The increases applied to both new orders and existing products, with no grandfathering Hetzner statement on price adjustment.
- A second adjustment on 15 June 2026 hit dedicated-vCPU lines harder: CCX13 rose from €15.99 to €42.99/month (+169%) and CCX23 from €31.49 to €85.99/month (+173%) for new orders and rescales Hetzner Docs price adjustment.
- OVHcloud warned in late 2025 that AI-fuelled memory demand could lift cloud prices 5–10% between April and September 2026, with memory and storage costs the main driver OVHcloud warns AI-fuelled memory surge will lift cloud prices.
The headline many creators have used — that AI is "killing cheap hosting" — is dramatic, but the mechanism behind it is real and worth understanding.
Why AI is pushing memory prices up
AI models need enormous amounts of fast memory. Training and running large models such as GPT-4 class systems requires loading billions of parameters into memory with extremely high bandwidth. That is done with High Bandwidth Memory (HBM), a stacked DRAM technology built using the same wafer capacity and production lines as conventional DDR5 server RAM.
Only three companies produce HBM at scale: SK Hynix, Samsung, and Micron. As they allocate more wafer capacity to HBM for AI accelerators like NVIDIA's H100/H200 and the next-generation Blackwell platform, the available supply of standard DDR5 server memory tightens SK hynix was the initial exclusive supplier of HBM3 to NVIDIA, High Bandwidth Memory bottleneck analysis.
Industry analysts report HBM demand is growing at roughly 80–100% annually while supply expands at 50–60% annually, leaving a structural shortfall through at least 2027 GPU shortage 2026: the HBM memory crisis explained. Memory makers have responded with contract price increases for DDR5, which eventually flows through to the cost of every server that cloud hosts buy.
This is not a conspiracy. It is a straightforward supply-chain trade-off: the same fabs can make either high-margin HBM or lower-margin commodity RAM, and AI demand is tilting that balance.
Why hyperscalers are not raising prices as visibly
AWS, Google Cloud, and Azure have not announced headline price increases of the same magnitude. That does not mean they are unaffected; it means their business model is different.
- Higher starting margins. Budget hosts priced compute close to hardware cost, so a 30–50% jump in memory prices forces them to pass it on. Hyperscalers have much larger markups on compute, egress, and managed services, giving them room to absorb component volatility.
- Egress is the hidden profit engine. Cloudflare estimated in 2021 that AWS charges for data egress at markups of several hundred to over 7,000 percent above wholesale transit costs in some regions, with North American and European egress prices unchanged since 2018 even as wholesale bandwidth fell more than 50% AWS's Egregious Egress. Those outsized margins help subsidize other parts of the stack.
- Long-term contracts and scale. Hyperscalers negotiate directly with chip makers and memory suppliers at volumes small hosts cannot match, so their input costs and timing differ.
The practical takeaway: do not assume AWS is "the good guy" just because it is not raising list prices as sharply. Its compute was already far more expensive at baseline.
What this means for your hosting bill
If you run a SaaS, e-commerce site, or any workload on budget cloud infrastructure, the math is changing. Three practical effects stand out:
- Fixed-cost assumptions break. Teams that built pricing around a €30/month server now face €85/month for the same spec. If you charge per seat or per credit rather than passing through infrastructure cost, your margin just got squeezed.
- The "just use Hetzner" default is no longer automatic. Even after the increases, Hetzner can still be cheaper than DigitalOcean, Vultr, or AWS for equivalent specs, but the gap is narrower. You now need to re-run the comparison for each workload rather than relying on a rule of thumb.
- Memory-heavy workloads are hit hardest. Database servers, caching layers, and anything running large language models locally feel the DRAM shortage most directly.
How to keep costs under control
You cannot fix the global memory market, but you can reduce your exposure.
| Tactic | When it helps | Caveat |
|---|---|---|
| Lock in existing instances | You already have Hetzner servers created before the June 2026 rescale window | Existing products keep prior terms; new orders and rescales get new prices |
| Right-size and autoscale | Variable traffic; over-provisioned dev/staging environments | Requires monitoring and discipline; not a one-time fix |
| Use ARM where possible | Cloud-native apps, containers, interpreted languages | Recompile/test; some legacy binaries need x86 |
| Cache aggressively with a CDN | Static assets, media, public content | Cloudflare R2 and similar zero-egress storage can cut transfer bills dramatically |
| Separate compute from storage | Workloads with large data sets but intermittent compute | Object storage plus serverless or batch workers can be cheaper than always-on servers |
| Re-negotiate or multi-cloud | You have steady scale and can commit to reserved capacity | Hyperscaler egress fees still make "all-in on AWS" expensive for public-facing apps |
A concrete example: moving a media-heavy site from AWS S3 to Cloudflare R2 saves the egress line item entirely. For a 10 TB/month public-read workload, AWS S3 Standard can cost roughly $900/month in egress alone after the free tier, while R2 charges $0 for egress Cloudflare R2 vs AWS S3 comparison. That single change can offset multiple memory-driven server price increases.
What this means for you
For most small businesses and builders, this is a reminder to treat infrastructure as a living cost rather than a one-time decision. The era of "set it and forget it on a €5 VPS" is not over, but the margin for error is smaller.
Three actions worth taking now:
- Audit your current bill by workload. Identify which servers are memory-bound, which are bandwidth-heavy, and which are idle most of the time.
- Compare alternatives on total cost, not headline compute price. Include egress, storage, snapshot, and backup costs. A server that looks cheap can become expensive once data transfer is included.
- Build for portability. Containers, infrastructure-as-code, and S3-compatible object stores make it easier to move workloads when one provider's pricing changes again.
FAQ
Is AI really the main reason cloud prices are rising in 2026?
It is a major driver for memory-intensive hosts. Hetzner and OVHcloud both cite rising hardware and memory costs tied to AI demand as the reason for their 2026 adjustments. Energy, supply chain, and currency effects also play a role, but the memory squeeze is the clearest AI-linked factor.
Will prices go back down?
Analysts expect memory prices to remain elevated through at least late 2026 and possibly into 2028, depending on how quickly HBM4 and expanded DRAM capacity come online. Do not plan around a quick return to 2024 price levels.
Should I move everything to AWS or Azure to avoid increases?
Not automatically. Hyperscalers have not raised list prices as sharply, but their baseline compute and egress costs are far higher. Run a total-cost comparison for your actual workload before migrating.
Are smaller providers like DigitalOcean, Vultr, or Linode also affected?
Yes. They may not move at the exact same time, but they buy from the same memory and SSD supply chains. Expect pricing pressure across the industry.
What is the single biggest lever to cut my cloud bill right now?
For most public-facing apps, reducing data transfer cost with a CDN or zero-egress object storage often saves more than downgrading a server. Egress fees compound quickly as traffic grows.
Discussion
0 comments