Memory, Not GPUs, Is Now the Real Bottleneck
AI infrastructure spending refers to the long-term capital outlays that technology giants commit to build and expand data centers, networking, power, and specialized chips required to run large-scale artificial intelligence workloads. Today, soaring memory chip costs are turning that spending from a strategic choice into a survival requirement as leading cloud and AI players race to secure scarce capacity and avoid being left behind.
Amazon’s decision to lift its 2026 capital spending plan to about $220 billion is the clearest signal yet that memory has become the chokepoint for AI infrastructure spending. CEO Andy Jassy has been explicit: higher costs for artificial intelligence memory chips are a main reason the company raised its estimate from around $200 billion, a jump of roughly $20 billion. In other words, the economics of AI are now being set less by headline-grabbing GPUs and more by the expensive high-bandwidth memory needed to keep those chips busy.
Demand for those advanced memory chips has surged alongside GPU purchases, creating shortages and pushing prices higher. That dynamic is not temporary friction; it is reshaping what hyperscaler capex looks like for the rest of the decade.
Amazon’s Capex Shock: Paying Up for Scarce Memory
The brutal truth behind Amazon’s $220 billion capex plan is that the company is paying more and getting less capacity than it expected. Rising memory chip costs are driving higher infrastructure spending, even as the company concedes it still will not meet customer demand. That should alarm anyone who assumed AI would quickly enjoy cloud-like economies of scale.
Jassy has warned that, even with the raised budget, Amazon is unlikely to have enough computing capacity to satisfy demand in 2026 and expects the same dynamic in 2027, while calling existing demand for 2028 “striking”. This is not a measured, optional buildout; it is a race to catch up. Most of that spending targets AI infrastructure: more data centers, stronger networking, and custom chips to power cloud AI services. Yet memory is acting like a tax on every rack the company deploys.
The implication is stark: hyperscaler capex is no longer only about scale, it is about securing enough high-cost memory to stay relevant in the AI platform wars.
SoftBank’s Power and Data Center Bet Shows How Deep This Goes
SoftBank’s recent moves show that the memory squeeze does not exist in isolation; it ripples into power, real estate, and financial engineering. The group spent approximately ¥968.9 billion in the first quarter of fiscal 2026 on assets linked to power generation and data centers as it builds the physical infrastructure for its broader AI strategy. That is not a side project; it is a wholesale reshaping of its balance sheet around AI capacity.
Property, plant and equipment jumped by about ¥725.9 billion over just one quarter, mainly because of those power and data center acquisitions. The investments, concentrated in its Energy Global subsidiary, are paired with complex warrant structures that are already producing large derivative losses and higher share-based compensation expenses. In plain terms, SoftBank is willing to take accounting pain today to lock in the power and data center footprint it believes AI will demand tomorrow.
That willingness underscores a broader shift: data center investment is now both an energy story and a financing story, all driven by the same memory-constrained AI buildout.
Persistent Memory Constraints Are Locking Hyperscalers into a Capex Arms Race
The most worrying part for investors is that this spending spiral does not look short-lived. Demand for advanced memory chips has surged so much that it has created supply shortages and driven prices higher. Jassy’s admission that Amazon will not have enough capacity to meet all its demand in 2026 or 2027, combined with strong demand already visible for 2028, points to memory supply constraints that will outlast any single upgrade cycle.
This leaves hyperscalers trapped in a rising capex cycle. AI workloads require more memory per unit of compute, but memory supply cannot scale as quickly, so providers feel compelled to over-commit capital to secure inventory. The competition is not only for customers; it is for the components that make AI capacity possible. Even as the largest technology firms collectively steer hundreds of billions into AI infrastructure, capacity remains tight and expensive.
The result is an uneasy equilibrium: huge AI infrastructure spending, persistent AI capacity constraints, and a structural advantage for those that can prepay for memory and power years in advance.
Conclusion: Memory Economics Will Decide the Winners of the AI Era
The AI story has been sold as a triumph of clever models and powerful GPUs, but the real battle is increasingly about memory, data centers, and power. Amazon’s higher capex plan, driven largely by memory chip costs, and SoftBank’s heavy spending on power and data center infrastructure show that AI infrastructure spending is being reshaped from the ground up.
As long as memory chip costs stay high and supply lags demand, AI capacity constraints will persist, and hyperscaler capex will remain elevated. The firms willing and able to lock in memory, power, and sites ahead of time will not just run larger models; they will define the price and availability of AI for everyone else. In this phase of the AI race, balance sheets and supply contracts matter as much as algorithms.
