Microsoft pins $25B to memory costs

Immense AI datacenters are turning into memory machines as much as GPU farms. That is the real shift hiding inside Microsoft’s latest spending plan. The company told investors on April 29 that fiscal 2026 capital spending will reach about $190 billion, and roughly $25 billion of that increase comes from higher component costs rather than simply buying more capacity. Microsoft is still building hard. But the surprise is where the inflation is coming from — memory and storage are now big enough cost centers to move the whole budget. ### Why does memory suddenly matter this much? AI servers need absurd amounts of memory. Not just ordinary DRAM, but high-bandwidth memory — HBM — stacked right next to accelerators so models can be fed data fast enough. A modern AI cluster is basically a balancing act between compute, networking, power, and memory. For a while, GPUs got all the attention. But once hyperscalers started ordering at full tilt, memory turned into the bottleneck that could actually choke the whole system. TrendForce says server-memory suppliers now have “absolute pricing power,” with contract prices still moving sharply upward because cloud demand keeps the supply gap open. ### What did Microsoft actually say? The cleanest takeaway is this: Microsoft’s budget went up by $25 billion because parts got more expensive. The company still expects to spend around $40 billion in the current quarter alone, and Amy Hood said Microsoft expects capacity constraints to last through 2026. That matters because it means the extra money is not a simple story about more datacenters or more boxes. A meaningful chunk is just the same AI buildout costing more than expected. ### Why are prices rising so fast? Because the memory makers can sell everything. Micron has already locked up its entire calendar-2026 HBM supply under price-and-volume agreements. SK hynix is signaling similar scarcity, with HBM effectively sold out well ahead of delivery windows. When suppliers know the biggest buyers need the parts no matter what, pricing gets reset upward fast. And because fabs are shifting capacity toward premium AI memory, conventional server DRAM and enterprise storage get tighter too. ### Is this just a Microsoft problem? Not at all. The bigger story is that hyperscaler capex numbers are starting to overstate how much fresh physical capacity is really being added. Tom’s Hardware pegged combined 2026 capex plans for Google, Amazon, Microsoft, and Meta at about $725 billion, up 77% from 2025. But if a bigger share of that money is being absorbed by pricier memory, storage, and other components, then headline spending growth exaggerates the underlying expansion in racks and servers. Basically — more dollars does not automatically mean proportionally more compute. ### Why does that matter to investors? Because investors have been reading capex as a proxy for AI momentum. Turns out that proxy is getting noisier. Microsoft’s AI annual recurring revenue is now above $37 billion, up 123% year over year, which is real growth. But if component inflation is eating a larger slice of infrastructure budgets, returns take longer to show up and the procurement game gets harsher. The winners are the cloud giants that can pre-buy supply and the memory vendors with scarce product. Everyone else waits in line. ### So what is the real signal here? The AI buildout has entered a new phase. Last year the constraint people obsessed over was accelerators. This year the constraint is the full bill of materials around them — especially memory. That changes how to read every giant capex number from here on out. Some of the spending boom is real expansion. Some of it is just the price of admission going up. ### Bottom line Microsoft did not just reveal a bigger AI budget. It revealed that memory has become expensive enough to bend the budget itself. That is a sign of demand strength, but also of a tighter, more inflationary AI supply chain than the market had been assuming.