Memory and chip price increases ate about $25B of Microsoft's AI budget, company says
- Microsoft told investors on April 29 its fiscal 2026 capital spending will hit about $190 billion, with roughly $25 billion tied to higher memory and chip prices. - The striking detail is the mix — about $31.9 billion was spent in the March quarter alone, while Microsoft’s AI business hit a $37 billion run rate. - That matters because AI demand is rising fast, but scarce HBM memory is now inflating hyperscaler build costs.
Microsoft’s AI buildout just got a lot more expensive — not because it changed its plans, but because the parts got pricier. On April 29, Microsoft said fiscal 2026 capital expenditures will reach about $190 billion, and a big chunk of that jump comes from higher memory and semiconductor costs. The rough number making waves is $25 billion. That is money Microsoft says it is effectively absorbing because the hardware inside AI data centers now costs more than expected. (cnbc.com) ### Where did the $25 billion come from? The number showed up around Microsoft’s latest earnings discussion for the quarter ended March 31, 2026. The company posted $82.9 billion in revenue, said its AI business has now surpassed a $37 billion annual revenue run rate, and told investors it expects full-year capex of about $190 billion. The key twist is th(cnbc.com)s. (news.microsoft.com) ### Why are memory and chips the problem? AI servers are brutally memory-hungry. The expensive part is not only the GPU or accelerator — it is also the high-bandwidth memory, or HBM, sitting right next to it. HBM is the fast memory that keeps giant models fed with data. If that supply gets tight, the whole server ge(news.microsoft.com) a redesigned memory system because inference performance depends on feeding the chip fast enough. (blogs.microsoft.com) ### Why is HBM so expensive now? Because almost everyone wants the same thing at once. Micron has said it already locked in price and volume agreements for its entire calendar 2026 HBM supply, including HBM4. SK hynix has also signaled that demand is outrunning available supply well into 2026. Basically, the memory vendors have pricing power, and hyperscalers are paying up to secure slots. (investors.micron.com) ### Is Microsoft still actually adding capacity? Yes — and that is what makes this story important. This is not a company backing away from AI spending. It is a company spending even more to get roughly the same roadmap done on time. Microsoft spent $31.9 billion on capex and finance leases in the March quarter alone, up 49% year over year, (investors.micron.com)d 39% to 40% in constant currency. (cnbc.com) ### Why should anyone outside Microsoft care? Because Microsoft is one of the clearest windows into hyperscaler economics. If Microsoft says $25 billion of its AI budget is getting eaten by component inflation, that tells you this is not just a procurement headache. It is a structural cost issue for the whole industry — especially for companies trying to (cnbc.com)much bigger capex wave across Google, Amazon, Meta, and Microsoft. (tomshardware.com) ### Does this change the AI business math? Probably, yes. The catch is that AI demand is real enough to justify the spending — Microsoft’s $37 billion AI revenue run rate says that clearly. But when memory prices surge, the economics of every token, seat, and inference request get tighter. That pushes cloud companies toward custom silicon, tighter workload scheduling, and product(tomshardware.com)arce. Microsoft’s push with Maia 200 fits that logic. (news.microsoft.com) ### So what is the real signal here? The real signal is not “Microsoft spent a lot.” Microsoft always spends a lot. The signal is that a meaningful slice of AI capex is no longer buying ambition — it is buying around scarcity. When the memory market tightens, even the biggest cloud platforms cannot brute-force their way out cheaply. (cnbc.com) ### Bottom line? AI infrastructure is turning into a supply-chain story as much as a software story. Microsoft’s latest numbers show the bottleneck is not demand. It is the cost of the chips and memory needed to serve that demand at scale. (cnbc.com)
