AI Data Centre Power Demand, GPU/PSU Shortages & Memory Prices (2026)

AI data centres will use 20% of US power by 2035. Here's how that's driving GPU, PSU and memory shortages and what to do about it in 2026.

AI Data Centre Power Demand Is Breaking the PC Market: GPU, PSU and Memory Shortages Explained (2026)

AI data centres are forecast to consume 20% of US electricity by 2035, up from roughly 6% today. This surge is putting pressure on GPUs, PC memory and high-wattage power supplies, pushing up prices and making budget hardware harder to find.

If you’re building a gaming PC, a 3D-printing farm or a home AI lab in 2026, you’re competing directly with hyperscalers for silicon, memory and power hardware. This guide explains the demand shock, the supply-chain bottlenecks and what you can do about them.

Table of Contents

Data Centre Power Demand: The 194 GW Shock

AI is turning electricity into one of the technology sector’s most important resources. New projections suggest that US data centres could need 194 GW of power capacity by 2035: enough to account for around one in every five kilowatt-hours generated in the country.

That figure represents an 83% increase on an earlier forecast released only months before it. In other words, estimates of how much power AI infrastructure will require are rising far faster than grid planners expected.

Key Numbers at a Glance

  • Current position: Data centres account for roughly 6% of US electricity use.
  • 2030 projection: Data centres could account for around 12% of US electricity use.
  • 2035 projection: Capacity could reach 194 GW, or about 20% of national electricity use.
  • AI’s role: Training and operating AI models are expected to make up a major share of new data centre demand.
  • Global impact: AI-led data centre expansion is adding electricity demand on the scale of an entire major economy.

To put 194 GW into perspective, a conventional nuclear power station is often rated at around 1 GW. Supplying that much additional data centre load would therefore require power equivalent to nearly 200 such reactors, alongside major upgrades to transmission, substations and local distribution networks.

Why the Forecast Changed So Fast

The problem is not simply that there are more data centres. It is that the type of work inside them has changed. Traditional cloud workloads such as storage, websites and business software use power, but modern AI training clusters can run thousands of high-end accelerators continuously.

Large language models, image and video generators, AI agents and multi-modal systems require huge amounts of computation. Every extra model parameter, user query, training run and inference request needs chips, memory, cooling and electricity.

Even if the US repeated its strongest recent year for connecting new data centres to the grid every year for a decade, forecasts suggest there would still be a sizeable gap between available power and expected demand. That mismatch is now rippling through the wider technology supply chain.

Where Demand Is Concentrated

  • Mid-Atlantic states: Regions served by PJM are becoming major data centre hubs, creating intense competition for local grid capacity.
  • Texas: ERCOT is attracting AI data centres with available land, renewable generation and a comparatively flexible electricity market.
  • Cloud hubs: Data centre expansion clusters around fibre routes, cheap power, favourable planning rules and proximity to major users.

The important point for PC builders is simple: the same companies buying tens of thousands of AI servers also influence demand for memory, networking hardware, power electronics and advanced chip packaging.

GPU Shortages: Why Budget Cards Are Hit Hardest

When people hear the phrase “GPU shortage”, they often picture expensive flagship graphics cards. In 2026, however, budget and entry-level cards are increasingly at risk of becoming the hardest products to find at sensible prices.

The Budget GPU Problem

Major graphics-card manufacturers have warned that entry-level GPUs may face especially severe shortages through the second half of 2026. That matters because these are the cards used by ordinary PC gamers, students, entry-level content creators and budget workstation builders.

The cause is not just GPU silicon. Graphics cards also need memory, voltage-regulation components, PCBs, coolers and power connectors. If any one of those parts becomes scarce or expensive, the finished card costs more to build.

AI Is Absorbing Memory Output

AI data centres are consuming a growing share of the world’s memory production. High-end AI accelerators rely on high-bandwidth memory, while server CPUs and storage systems require huge quantities of DRAM and NAND flash.

Consumer graphics cards use GDDR6 or GDDR7 memory rather than HBM, but they still compete for manufacturing capacity, engineering attention and shared supply-chain resources. When AI demand rises, memory makers naturally prioritise the products with the highest margins and the largest long-term contracts.

Why GPU Prices Rise Even When Supply Exists

It is possible for a graphics card to be technically available but still overpriced. Board partners buy GPU-and-memory kits, then add the PCB, VRMs, cooling system, packaging, distribution and retailer margin. When the price of the kit rises, the final retail price follows.

  • GDDR6 and GDDR7: Higher graphics-memory costs make every GPU tier more expensive.
  • Budget products: Low-margin cards have less room to absorb component increases.
  • Flagship products: Premium buyers can sometimes tolerate price rises, so manufacturers may prioritise them.
  • Second-hand market: Used cards can become inflated when new options are scarce.

This creates a painful situation for buyers: a mid-range card can cost more than expected, while the lower-cost GPU that was supposed to be the value option becomes unavailable or poor value.

What to Expect From GPU Prices

Prices may stabilise at higher levels before they fall. New fabrication plants, packaging capacity and memory output take years to build, validate and ramp. Unless AI investment slows sharply, meaningful relief may not arrive until late 2027 or beyond.

That does not mean every card will rise forever. Promotions, regional stock levels and product launches will still create buying opportunities. But it does mean you should judge hardware on actual price-to-performance today rather than waiting for the market to magically return to older price levels.

Memory Prices: DRAM, NAND and HBM at Record Highs

Memory is the less visible part of the AI hardware crunch. GPUs receive most of the attention, yet shortages of DRAM, NAND flash and HBM can be just as disruptive.

Why DRAM Is Getting More Expensive

For years, PC memory became cheaper, faster and more plentiful. That pattern has broken down. AI servers need enormous amounts of system memory, and each high-end accelerator also requires specialised, extremely fast memory close to the GPU die.

When Samsung, SK Hynix and Micron choose where to allocate wafer capacity, HBM and server-grade memory can be more profitable than commodity desktop RAM. The result is less available capacity for standard DDR4 and DDR5 products.

What This Means for PC Builders

  • DDR4: Older platforms are not immune; legacy memory can jump in price when suppliers reduce production.
  • DDR5: New-build RAM is faster and more future-proof, but capacity upgrades can be costly.
  • 32 GB: A practical minimum for modern gaming, browser-heavy work and general content creation.
  • 64 GB or more: Worth considering for video editing, CAD, 3D rendering, virtual machines and local AI work.

For anyone building a workstation, memory should no longer be an afterthought. If your work genuinely benefits from more RAM, buying enough at the time of the build may be better value than planning to upgrade later.

NAND Flash and SSD Pricing

NAND flash powers SSDs, memory cards and USB storage. AI workloads create heavy demand for fast storage because models, training datasets, checkpoints and logs must be read and written continuously.

As a result, SSD prices are under pressure too. High-capacity NVMe drives are particularly desirable for creators, gamers and AI users, so a 2 TB or 4 TB drive is no longer a casual purchase. If you need bulk fast storage for footage, games, project files or 3D assets, consider buying during a genuine sale rather than assuming a better one is around the corner.

HBM: The AI Accelerator Bottleneck

High-Bandwidth Memory, or HBM, is the memory technology used beside modern AI accelerators. It is built by stacking memory dies vertically and connecting them with extremely dense links. This delivers enormous bandwidth in a small footprint, but it is difficult and expensive to manufacture.

HBM is a major reason AI GPUs remain scarce. Producing it requires not only DRAM

Trending now

WhatsApp Safety 2026: What Every Parent Should Know
A Parent’s Calm Guide to Kids, Phones and the Internet