Meta planned to spend $115bn to $135bn on data centre infrastructure this year. In April, the company raised that to $125bn to $145bn. The reason it cited was blunt: higher memory prices. Microsoft blamed the same. Together with Google and Amazon, the four hyperscalers will spend $725bn on capex in 2026, up 77% from $410bn the year before. Nearly all of it is infrastructure for AI. And nearly all of it is hitting the capex line harder than expected because memory is expensive.
Into that landscape comes Vistara, Meta's custom CXL ASIC for reusing DDR4 memory. The company will present the design at ISCA, the top architecture conference, on Monday. On the surface it is an engineering accomplishment. The deeper read is that it solves a unit economics problem. Meta can no longer afford to discard old memory and buy new.
The maths behind the repricing
Data centre memory is a bulk commodity. When one hyperscaler's demand moves, prices move for everyone. Meta's data centres now consume 70% of the world's memory output. Google and Microsoft consume most of the rest. The four of them have created an artificial scarcity. Every dollar added to the cost of a DDR4 or DDR5 DIMM multiplies across millions of modules. At that scale, recycling becomes economics, not environmentalism.
A typical data centre server might hold 10 to 20 DIMMs. Multiply that by the millions of machines across Meta's footprint. A $10 per module price rise becomes a multi-billion-dollar problem in annual capex. The company is not choosing to recycle memory out of principle. It is choosing to recycle memory because it has no other way to meet its capacity targets while staying within budget guidance.
The accelerator is scarcity
Three years ago, when GPU costs were the limiting factor, no one talked about memory reuse. The accelerator was the constraint. Now the constraint is something older and more fundamental. The CPU needs more memory to feed the GPU. The GPU needs more memory for embeddings and KV caches. Traditional memory expansion—just buying more DDR5—is no longer a viable path. So Meta engineered its way around it.
Vistara takes DDR4 modules from servers that have been retired, attaches them via a CXL ASIC to newer servers, and makes them look like local memory to the operating system. The trick is doing that without the latency penalty that normally comes from disaggregated memory. Meta solved it with custom silicon and careful software tuning. Workloads that cannot tolerate the latency are simply not run against the expanded pool. The rest gain capacity at near-zero cost.
Not the first, won't be the last
Meta is not unique in recognising this problem. Microsoft, Google, and others are pursuing similar paths. Marvell, the memory controller vendor, is shipping CXL expanders designed for DDR4 reuse. But Meta's Vistara is significant because it is the first production deployment at meaningful scale. The company has deployed it across millions of servers. That is the engineering story.
But the real story is the one that made the engineering necessary. Meta guided capex up by $10bn in a single quarter and blamed memory prices. That guidance cut will propagate through their entire technology roadmap. Every dollar spent on old memory is a dollar not spent on new capacity, new chips, or new facilities. Vistara buys the company time by making expensive budget go further.
The innovation solves a problem that should not exist. In a normal market, memory prices would signal supply constraints, and supply would respond. But in 2026, supply cannot respond because demand from four companies has outrun the world's production capacity. So Meta built a system to use what it already owns.
That is not innovation driven by ambition. It is innovation driven by necessity in a constrained world.
