Most coverage of the AI boom focuses on models and capabilities. The less glamorous story — and the one that will actually show up on your invoices — is physical. The infrastructure underneath AI is straining, and the cost is spilling over onto companies that aren't running a single large model of their own.
Two squeezes happening at once
Memory. Soaring demand for high-bandwidth memory from AI data centres is driving sharp price increases and supply constraints for the DRAM and NAND used in ordinary devices. Manufacturers are prioritising high-performance memory for AI training and inference, squeezing supply for everything else — and industry reporting indicates some segments have seen dramatic increases, with Samsung and Apple expected to pass higher component costs into upcoming flagship devices. Some in the industry have started calling it "RAMageddon." Whatever the label, it flows straight through to the price of laptops, servers, and phones your business buys.
Power. Google's 2025 environmental report disclosed a 37% year-over-year jump in electricity consumption — the largest in the company's history — attributed largely to AI data-centre expansion. Coverage of the report noted Google's data centres alone consumed over 42 million megawatt-hours, comparable to the annual usage of an entire country like New Zealand or Denmark. That demand pressures grids, energy prices, and ultimately the cost of the cloud compute you rent.
Why this reaches companies that "don't do AI"
You may not train models, but you buy the same DRAM, rent the same cloud regions, and refresh the same laptop fleets as everyone else. When AI infrastructure consumes a disproportionate share of memory production and grid capacity, three costs move:
- Hardware refreshes get more expensive. Server and endpoint purchases priced on cheap, abundant memory need re-forecasting.
- Cloud pricing feels upward pressure. Providers passing through energy and hardware costs, plus capacity contention in the hottest regions, can mean higher prices or scarcer instances.
- Lead times stretch. Supply constraints mean the hardware you order may take longer to arrive — a planning problem as much as a budget one.
How to plan around it
- Re-forecast hardware budgets with headroom. If your refresh assumes historical memory prices, add margin and revisit quotes closer to purchase.
- Right-size before you scale. The cheapest megawatt-hour is the one you don't consume. Audit over-provisioned cloud resources and idle instances before AI-era pricing makes waste expensive.
- Value efficiency in architecture. Efficient code, sensible caching, and appropriately sized models aren't just good engineering now — they're direct cost control as compute gets pricier.
- Extend refresh cycles where it's safe. If memory prices spike, holding serviceable hardware a little longer (without compromising security or support) can ride out the peak.
- Watch region and lead-time risk. For cloud-heavy workloads, track capacity and pricing in your regions; for hardware, order earlier and confirm lead times.
The bigger picture
This is a reminder that AI is not weightless. It runs on chips that must be manufactured and electricity that must be generated, and when demand outpaces supply, the cost lands on everyone sharing the same suppliers and grids. The winners won't be the companies that ignore this in favour of the model headlines — they'll be the ones that plan for the physical reality underneath.
Where Educatifu fits
We help companies control infrastructure cost — cloud right-sizing, efficient architecture, and hardware planning that accounts for where prices are heading, not just where they've been. If AI-era pricing is starting to show up in your bills, get in touch for a practical efficiency review.