Hyperscale Companies Commit Nearly $2 Trillion: AI is Changing Who Buys Chips

For years, Apple was one of the buyers with enough volume and financial muscle to secure components through large, long-term contracts. Artificial intelligence has altered that hierarchy. Alphabet, Microsoft, Meta, and Amazon collectively hold contractual commitments nearing $2 trillion, according to an analysis by Claus Aasholm based on their public financial reports. The figure requires some nuance, but it reflects a hard-to-ignore shift: big tech buyers are no longer just securing servers for the next quarter—they’re locking in compute capacity, memory, data centers, and energy for multiple years.

The key points of the AI hardware race in 30 seconds

• Alphabet, Microsoft, Meta, and Amazon’s contractual commitments are approaching $2 trillion, according to Aasholm’s analysis.
• Alphabets leads with $811 billion committed as of June’s close.
• Not all is chips or memory: commitments also include data centers, energy, cloud services, inventory, licenses, and other contracts.
• The scale shows how hyperscalers are reserving future capacity before needing it.
• For memory, servers, and data centers, AI is changing who takes priority in the supply chain.

It’s important to clarify this precisely. Talking about “$2 trillion in AI hardware” would be inaccurate.

These aggregated figures stem from different categories of liabilities and commitments companies disclose in their financial documents. They include purchases of technical infrastructure, inventory, servers, cloud contracts, data centers, leases, energy, and, depending on the company, other obligations. They do not equal capital expenditures (CapEx) that will be spent within a single fiscal year.

This snapshot remains extraordinary because it reveals how much capital companies competing to dominate AI infrastructure are committing in advance.

And Alphabet probably provides the clearest example.

Google’s commitments soar from $149.1 billion to $811 billion in six months

At the end of 2025, Alphabet reported $149.1 billion in purchase commitments and other contractual obligations. Of that, $113 billion was short-term and mainly related to technical infrastructure and inventory.

Just three months later, the figure had changed dramatically.

In its March 2026 documentation, Alphabet reported $332.4 billion, with $138 billion short-term. The company explained that these obligations mainly involved technical infrastructure costs and inventory through long-term supply agreements and open purchase orders, as well as content licenses and energy contracts take-or-pay.

By the end of June, another leap occurred.

Alphabet already had $811 billion in future commitments, nearly $500 billion more than three months earlier. About $200.7 billion of that was short-term. These obligations include everything from chips and data centers to electricity, inventory, and content licenses.

The progression is more revealing than any individual figure:

In roughly six months, these commitments have increased more than fivefold.

This is happening even as Alphabet plans capital expenditures of between $180 billion and $190 billion in 2026, about six times what it spent in 2022. The company anticipates increasing that amount significantly again in 2027, emphasizing that the vast majority of current CapEx is aimed at technical infrastructure.

This begins to reveal the true scale of the shift.

It’s no longer enough to have the money for GPU purchases

The initial phase of generative AI explosion was straightforward: everyone wanted Nvidia GPUs, and there weren’t enough.

The situation in 2026 is far more complicated.

A GPU requires HBM memory. Thousands of GPUs need switches, high-speed networking, and storage. Racks demand power and cooling. Buildings require transformers, substations, and electrical connections. And all this must be available roughly at the same time.

The bottleneck has shifted from a single component to the entire industrial chain needed to convert electricity into computing capacity.

That’s why hyperscalers are doing something completely rational from their perspective: reserving capacity ahead of time.

The problem is that Amazon, Microsoft, Google, and Meta are not typical buyers.

They have enormous balance sheets, generate tens of billions in cash, and can sign multi-year contracts to secure future production. When several do this simultaneously, the rest of the market ends up competing for the remaining capacity.

Memory illustrates this well.

Memory has become a strategic asset in AI

For decades, memory was one of the most cyclical components in computing.

Manufacturers expanded production when prices rose. The new capacity led to oversupply, prices dropped, and a new cycle began.

AI is temporarily disrupting that dynamic.

Accelerators need large amounts of High Bandwidth Memory (HBM), but data centers also consume enormous amounts of conventional DRAM and NAND storage.

And prices are reacting.

TrendForce forecasts for Q3 2026 indicate quarterly increases of between 13% and 18% for conventional DRAM and between 10% and 15% for NAND Flash. These are significant increases, though less than the near 60% jumps seen in some segments during Q2.

The phenomenon has even gained a name: chipflation.

Major cloud providers are signing long-term supply contracts to secure memory and other components, increasing pressure on an industry whose capacity cannot expand as quickly as purchase orders.

Building a new DRAM factory takes years. Increasing HBM is even more complex, involving manufacturing, stacking, TSVs, packaging, and additional processes.

Hence, reserving memory for 2027 or 2028 can be as crucial as designing the accelerators that will use it.

Microsoft and Meta are also investing in the future

Alphabet currently leads with the most impressive figure, but it’s far from an isolated case.

Microsoft has been showing the same trend for a while.

By the end of FY 2025, Microsoft reported $109.953 billion in purchase commitments, plus $32.149 billion in construction, and $178.701 billion related to operational and financial leases.

The scale has continued to grow in 2026. Some obligations relate to AI hardware, others to data center contracts extending over several years.

Meta provides an even more evident example of how the infrastructure model is changing.

As of March 31, 2026, Meta had $237.67 billion in non-cancellable contractual commitments, mainly related to cloud capacity from third parties, servers, networks, data centers, and certain hardware products from Reality Labs.

They also had $182.88 billion in obligations for leases not yet started, mainly data centers, colocation, and network infrastructure, with contracts starting between 2026 and 2036, some extending up to 30 years.

Just in April, Meta added approximately another $24 billion in multi-year infrastructure contracts.

They are no longer just buying servers.

They are reserving buildings not yet operational, electrical capacity needed, cloud infrastructure for years, and components not yet manufactured.

Amazon is also working on its own chips

Amazon adds another dimension to this race.

While AWS continues to utilize large quantities of Nvidia infrastructure, it has been developing its own processors such as Trainium and Graviton for years.

In Q2 2026, Amazon stated that Anthropic and OpenAI had made multi-year commitments involving several gigawatts around Trainium, with other companies also adopting their accelerators.

This introduces an interesting consequence for the supply chain.

The competition is no longer just Nvidia vs. AMD for orders. Google designs TPU, Amazon develops Trainium, Microsoft has its own accelerators, and Meta is working on custom silicon.

But they all ultimately need many of the same things: advanced wafers, packaging, HBM, DRAM, NAND, servers, fiber, switches, electricity, and data centers.

Switching manufacturers for these components doesn’t eliminate bottlenecks.

In some cases, it simply shifts the problem elsewhere in the supply chain.

Apple is no longer the buyer everyone feared to lose

One of the most interesting consequences of this new hierarchy emerges here.

Apple remains a massive company, buying huge quantities of memory, storage, displays, and semiconductors to produce hundreds of millions of devices.

For years, that volume gave Cupertino significant negotiating power.

A memory supplier had every incentive to secure capacity for Apple because losing an iPhone contract could mean losing billions of dollars.

AI is introducing clients capable of committing even larger amounts, often over longer periods.

Aasholm’s analysis places Apple’s commitments around $57 billion, far below the figures reaching hundreds of billions for Alphabet or Microsoft. That comparison is not a direct indicator of AI investment but reflects a change in scale.

The key takeaway: the primary customer for memory manufacturing no longer necessarily has to be the highest-volume smartphone seller.

It could be the one that needs to fill data centers for the next five years.

True power is shifting from buying cheap to ensuring supply

This might be the most fascinating aspect of this movement.

Traditionally, being a giant buyer meant getting better prices.

In a constrained market, priorities shift.

The goal is no longer just to pay less, but to ensure product availability.

For Google, it may be more damaging to have a data center worth billions waiting months for memory, servers, or electrical equipment than to pay a premium for these components.

This changes negotiations with suppliers.

Samsung, SK Hynix, and Micron can plan capacity via long-term contracts. TSMC can justify new investments with demand visibility. Server and electric equipment manufacturers can organize production with years-closed orders.

But this also impacts all buyers below hyperscalers.

PCs, smartphones, traditional enterprise servers, storage, and other devices are indirectly competing for the same industrial capacity.

Prices are already feeling the pressure. Morgan Stanley estimates that rising component costs could cause increases of up to 15 percentage points in certain PCs and smartphones.

What if in five years, this capacity is excess?

There’s another side to this story.

Committing nearly $2 trillion does not guarantee that future demand will justify all investments.

Contracts span many years, and some include different levels of flexibility. Not all this figure will convert immediately into CapEx or semiconductor purchases.

The industry is making a huge bet based on a core assumption: the demand for AI computing will keep growing enough to use all this capacity.

Alphabet has reasons to believe this.

Google Cloud’s backlog reached $462 billion at the end of Q1, and the company states that AI Solutions has become the main driver of cloud growth.

However, models are also rapidly improving efficiency. Cost per token decreases, quantization techniques emerge, smaller models are developed, and hyperscalers are creating dedicated accelerators to lower inference costs.

The question remains: will this efficiency reduce overall infrastructure needs, or will it backfire, making AI so cheap that its utilization multiplies?

Hyperscalers are clearly betting on the latter.

The AI race is now being decided with 2028 contracts

For a long time, AI competition focused on models: who has the best benchmark, the largest context, or the most capable agent.

Beneath that, a much more physical race is underway.

Google, Microsoft, Amazon, and Meta are trying to secure factories, memory, servers, data centers, and electricity for models that haven’t even been announced yet.

This could be one of the most significant changes AI has brought to the tech industry.

Power no longer belongs only to those designing the best chips; it now also belongs to those who can reserve the necessary industrial capacity before their rivals.

And when four companies collectively commit nearly $2 trillion, the effect extends beyond their own data centers.

It impacts server memory costs, which client gets early access to new HBM generations, how much capacity manufacturers build, and potentially how much the rest of the market pays for computers, smartphones, and storage.

AI is not just consuming more chips; it’s changing who controls the ability to direct the world’s chip production.

Frequently Asked Questions

Will hyperscalers really spend $2 trillion on chips?

No. That figure results from aggregating various contractual commitments from Alphabet, Microsoft, Meta, and Amazon over several years. It includes technical infrastructure, data centers, cloud contracts, energy, inventory, and other obligations—not solely semiconductors.

How much has Alphabet committed?

At the end of June 2026, Alphabet reported approximately $811 billion in future commitments, up from $332.4 billion three months earlier, and $149.1 billion at the end of 2025.

Why is AI affecting memory prices?

AI data centers require large amounts of HBM, DRAM, and storage. Major buyers are locking in long-term supply agreements while expanding manufacturing capacity takes years, creating supply pressures and driving up prices.

Has Apple stopped being an important buyer?

No. Apple remains one of the largest component purchasers globally. What has changed is the scale of commitments related to hyperscalers’ expansions, reserving infrastructure and capacity for AI years in advance.