Alphabet and Amazon are on track to pour roughly $420 billion into artificial intelligence infrastructure this year, a spending wave that is flowing straight to four semiconductor companies and rewriting the economics of the memory market. Alphabet guided for $195 billion to $205 billion of capital expenditure in 2026, while Amazon raised its own plan from $200 billion to $220 billion, citing the surge in memory-chip prices. The combined outlay makes the two cloud giants the single largest source of demand for AI compute and storage silicon.
The spending baseline
The $420 billion figure is not a forecast, it is the sum of two explicit guidance ranges issued by the hyperscalers themselves. Amazon’s $20 billion upward revision came after memory costs climbed, and Alphabet’s range suggests a similar pressure. Both companies have signaled that 2027 spending will be higher still, with Nvidia telling investors it expects total AI hyperscaler outlay to exceed $1 trillion next year. The money is not discretionary; it is the cost of keeping cloud capacity ahead of customer workloads that have no precedent in scale.
Nvidia and Broadcom split the compute pie
Nvidia’s graphics processors remain the default choice for AI training and inference, giving the company a de facto standard that lets cloud clients move workloads between providers without rewriting code. Broadcom occupies the alternative lane: it co-developed the Tensor Processing Unit with Alphabet, a custom chip that delivers better cost-performance for workloads tuned to it. The trade-off is lock-in, TPUs run on Google Cloud, so adopting them makes switching providers harder. Broadcom’s custom AI semiconductor division booked $10.8 billion in the second quarter and management expects the business to surpass $100 billion in annual sales, a trajectory that implies the division alone could soon rival the entire company’s current revenue base.
Memory shortage lifts Micron and Sandisk
Micron and Sandisk are benefiting from a different dynamic: a structural shortage of DRAM and NAND that has pushed selling prices up while input costs stay flat. Sandisk disclosed that two-thirds of its recent revenue growth came from price increases, only one-third from higher volumes. Micron’s management says the supply deficit will persist into 2028, giving both companies a multi-year window of expanding margins. The shortage is a direct function of hyperscaler demand, Amazon explicitly tied its spending increase to memory costs, so the memory makers are effectively capturing a slice of the cloud giants’ capex before it reaches the compute vendors.
The trillion-dollar signal
Nvidia’s $1 trillion hyperscaler spending estimate for 2027 frames the current year as the early innings of a cycle that has already rewritten the income statements of four chipmakers. The compute side is a duopoly with a lock-in twist; the memory side is a duopoly with a supply constraint that management teams say will not break for years. The only variable is whether the cloud buyers’ revenue growth can sustain capital intensity at this level, a question the source does not answer, and one that will determine whether the $420 billion becomes a floor or a ceiling.