I’ve been idly constructing the specs my subsequent desktop PC – a Linux gaming and AI rig – for the higher a part of the previous 12 months. It’s beginning to really feel like a doomed train.
The plan was modest sufficient: sufficient reminiscence to run an honest open-weight massive language mannequin regionally – as much as 128GB – a current-generation CPU with a good thread rely and a GPU that wouldn’t embarrass itself when requested to load a 70 billion-parameter mannequin into VRAM.
There’s only one drawback. Constructing a rig like that at the moment can price as a lot as a second-hand automobile. The explanation, in a phrase, is that hyperscalers have eaten the buyer {hardware} market alive.
Reuters reported this week that Korean reminiscence maker SK Hynix is being aggressively courted by world tech companies providing to take a position instantly in its manufacturing strains. One supply described the chip maker’s accessible manufacturing capability as “basically zero”.
SK Hynix’s shares are up roughly 154% this 12 months. Microsoft alone has guided the market in direction of $190-billion in capital expenditure in 2026, of which an estimated $25-billion the corporate has attributed to rising chip and element prices. Meta Platforms, Google and Amazon are spending equally eye-watering quantities. Now the little man can’t catch a break.
Hey, large spenders
The size of the dedication is beginning to present on the hyperscalers’ personal stability sheets. The Monetary Occasions reported on Friday (paywall) that mixed free money move at Amazon, Google dad or mum Alphabet, Microsoft and Meta is forecast to fall to roughly $4-billion within the third quarter – down from a post-pandemic quarterly common of $45-billion – with full-year free money move on monitor for its lowest degree since 2014, a 12 months when the 4 had mixed income round a seventh of at the moment’s.
Amazon is predicted to take a position $200-billion in 2026 and burn money for the 12 months; Microsoft will burn money in a minimum of one quarter, Meta within the second half, and Alphabet’s full-year free money move will hit its lowest in additional than a decade. Each Alphabet and Meta have paused share buybacks – Alphabet’s first such pause since 2015 – and have between them issued greater than $100-billion of recent debt in latest months to maintain the build-out funded. Mixed, the 4 are committing roughly $725-billion to AI infrastructure this 12 months, up from $600-billion only a few months in the past.
In consequence, reminiscence chip makers – principally SK Hynix, Samsung Electronics and Micron Know-how – have reallocated wafer capability away from commodity DDR5 and shopper DRAM in direction of the high-margin, high-bandwidth reminiscence utilized in AI accelerators: HBM3e in current-generation Nvidia and AMD elements, with HBM4 now ramping for next-generation silicon together with Nvidia’s Rubin platform.
The outcome has been described, solely half-jokingly, because the “RAMpocalypse”. DRAM costs rose by roughly 172% in 2025. HP advised buyers on its most up-to-date earnings name that reminiscence now accounts for 35% of the invoice of supplies in a PC, up from 15% to 18% the earlier quarter.
A 32GB DDR5 equipment that bought for beneath $90 a 12 months in the past will set you again as musch as $530 at the moment. The 128GB-plus configurations I’d truly must run a helpful native LLM are beginning to demand the type of cash that may require a dialogue with a financial institution supervisor.
For South African consumers, the image is bleaker. As TechCentral reported in March, native retailers Evetech, Dreamware Know-how and Tech.co.za had been already describing eye-watering worth strikes: DDR5 up by as a lot as 230% in a single quarter, DDR4 modules up 150% to 200%, SSDs up 35% to 50%. The image has worsened since then.
No matter cushion South Africans might need hoped for from a firming rand has been worn out by the worldwide crunch, and the nation sits effectively down the worldwide allocation queue no matter what the foreign money does on a given day.
Graphics playing cards have been worse for longer. Nvidia’s GeForce RTX 5090 launched in 2025 at US$1 999. Customized variants at the moment are retailing for $3 000 or extra. Manufacturing of the 5090 and 5080 has reportedly been lower by 30-40% as Nvidia diverts GDDR7 reminiscence and packaging capability to its knowledge centre enterprise, which now accounts for the overwhelming majority of group income. A single high-end AI accelerator can promote for round $40 000; a flagship gaming card can’t. Nvidia can be derelict in its obligation to shareholders if it did the rest. Nonetheless, it hurts – and gaming YouTubers are spilling bile concerning the firm due to it.
The CPU is subsequent
What’s now threatening to make an terrible scenario even worse is what is going on to CPUs.
For many of the previous two years, the AI increase was a GPU and reminiscence story. CPUs weren’t the place the actual workloads ran and costs remained comparatively sane. That narrative is now flipping. Agentic AI – the transfer from chat-style interactions in direction of autonomous software program brokers that plan, name instruments and execute multi-step duties – seems to lean on CPUs greater than GPUs.
AI brokers want precisely the mixed-throughput the place CPUs excel. Even chat inference, whereas GPU-heavy on the token era aspect, requires vital CPU assets for orchestration. Scale that throughout the volumes the hyperscalers at the moment are planning to deploy these chips, and the CPU aspect of the info centre is changing into the brand new bottleneck.
Nowhere has this shift been extra seen than at Intel, which a 12 months in the past seemed like a candidate for a strategic break-up. The share worth has greater than quintupled since, briefly touching $114 this week and lifting the corporate’s market worth previous $570-billion.
Intel’s long-promised 18A course of node – a technical return to kind for an organization that had fallen behind its principal rival, Taiwan’s TSMC – is now in quantity manufacturing at its new Arizona fabrication plant. The primary industrial Panther Lake processors constructed on 18A shipped earlier this 12 months and are getting good evaluations within the technical press.
Extra vital nonetheless, Apple, which dropped Intel as a silicon provider in late 2023, is reportedly contemplating shifting some chip manufacturing again to the corporate.
For Intel, that is vindication. For the remainder of us, it’s doubtlessly horrible information. Capability that may as soon as have been pointed on the shopper market is being dedicated to knowledge centre and hyperscaler clients years prematurely. AMD has its personal pricing energy drawback and each incentive to carry the road. Its share worth has accelerated quickly, too, in latest months.
What ties the threads collectively is that the reminiscence scarcity and the CPU squeeze have a standard trigger: the complete world semiconductor provide chain is now optimising for a small variety of very massive clients constructing knowledge centres. And there’s no aid in sight – AI infrastructure might but transform one of many greatest speculative bubbles in historical past, however there’s no motive to imagine that it’s going to burst anytime quickly.
So, know-how hobbyists like me, who would most wish to play with AI fashions on our personal machines and revel in highly effective {hardware} to play the most recent video games, are confronted with a vexing drawback. The economics of on-device computing have, for the primary time within the PC’s lengthy historical past, been damaged by the info centre. And there’s no apparent repair. — (c) 2026 NewsCentral Media
- Duncan McLeod is editor of TechCentral
Get breaking information from TechCentral on WhatsApp. Join right here.
