For two decades the semiconductor supply chain arranged itself around Apple. Foundries courted its orders, memory makers tuned production to its launch cycles, and the industry’s biggest capacity decisions were made with Cupertino in mind. The first half of 2026 produced the clearest evidence yet that this era is over: the four largest cloud companies have accumulated roughly $2 trillion in combined purchase commitments, and the balance of power in the chip industry has moved with the money.
The figures come from estimates by independent analyst Claus Aasholm, who tracks purchase commitments disclosed in SEC filings, as reported by Tom’s Hardware. Alphabet leads with about $811 billion, up from roughly $140-150 billion in the third quarter of 2025 and by far the steepest increase of any buyer. Microsoft follows at about $678 billion, Meta at about $349 billion, and Amazon at about $130 billion. Apple, by contrast, sits at about $57 billion, nearly flat, with $56.2 billion payable within twelve months. Nvidia’s own commitments, about $119 billion, are offered as a reference point; the chip designer out-commits the company that used to be the industry’s marquee customer.
These are total purchase obligations spanning several years, covering foundry capacity, DRAM and 3D NAND, contract manufacturers, and custom silicon, not memory-specific figures, and Aasholm cautions that they are approximate. Even allowing for imprecision, the shape of the market is unambiguous: procurement power has consolidated in the hands of a few cloud providers whose forward buying dwarfs traditional consumer electronics companies.
The commitments are a response to scarcity. AI infrastructure consumes enormous volumes of accelerators, high-bandwidth memory, and NAND, and every link in the chain is constrained: HBM by DRAM fab capacity, accelerators by wafer capacity, foundry capacity, and advanced packaging. Locking in supply years ahead has therefore become a competitive necessity, and the contracts that make it possible are now disclosed in quarterly filings where analysts can read them.
That is what makes the shift structural rather than cyclical. A buyer guaranteeing hundreds of billions in future purchases can underwrite a foundry expansion or a new memory fab and in return secure priority access to scarce products and future process technologies. TSMC can fund capacity on the strength of hyperscaler money and serve its largest customers first. Access to DDR5, HBM, and 3D NAND stops being a procurement exercise and becomes part of competitive advantage, and memory, long treated as a commodity, has become what the analyst describes as a strategic asset and arguably a competition weapon.
The suppliers are responding. DRAM and 3D NAND makers have gained pricing power, and analysts expect major capacity expansion from Micron, Samsung, and SK Hynix, which until now have been unusually disciplined about adding fabs. Apple’s flat commitments are the most telling data point in the analysis: the company remains one of the world’s largest semiconductor buyers in absolute terms, but it may no longer be the customer that foundries and memory makers plan expansions around. Aasholm’s framing, as reported by Tom’s Hardware, is that suppliers that used to buzz around Apple have found larger commitments elsewhere, and that Apple’s reluctance to follow the new market rules has left it outside the circle. Apple has reportedly lobbied the US government for access to memory chips from blacklisted Chinese vendor CXMT, an indication of how its procurement position has changed.
The open question is what the concentration does to the supply side. Suppliers can either prioritize the customers who can fund future capacity, or stay disciplined with expansions to avoid losses when demand eventually cools. Either way, the semiconductor industry’s customer hierarchy has been rewritten: the companies that build the world’s phones no longer set the agenda, the companies that run the world’s clouds do.
Sources: Hyperscalers commit nearly $2 trillion to secure AI hardware and memory (Tom’s Hardware, Aug 10, 2026); Memory will consume 30% of hyperscaler AI data center spending this year (Tom’s Hardware, 2026); Hyperscaler AI Capex Spending 2026 (Build MVP Fast, mid-2026)