On Monday, Meta and Nvidia announced a multi-year deal that puts millions of Blackwell and Rubin GPUs into Meta’s data centers. Not thousands. Millions. Along with Grace and Vera CPUs and Spectrum-X networking — a full-stack Nvidia buildout that positions Meta as one of the largest single buyers of AI compute on the planet.
The deal is massive. But it is not unusual. It is the latest move in an infrastructure arms race that has pushed combined AI capex from five US tech companies past $690 billion for 2026. Nearly double what they spent in 2025.
We tracked the spending commitments of every major player in this race — from the hyperscalers pouring hundreds of billions into data centers to the AI startups burning through revenue faster than they earn it. Here is where each company stands.
The $690 Billion Year
Five companies account for nearly all of it. Their 2026 capital expenditure plans, most of it directed at AI compute and data centers:
Amazon leads at $200 billion — a number that shocked even bullish analysts. Consensus expectations had been closer to $147 billion. CEO Andy Jassy defended the figure by pointing to AWS revenue hitting a $142 billion annualized run rate with growth accelerating to 24% year-over-year. The stock still dropped 8-10% on the announcement.
Alphabet‘s $175-185 billion plan has been revised upward three times from an initial $71-73 billion target for 2025. CEO Sundar Pichai acknowledged the scale causes concern internally. But the cloud backlog surged 55% sequentially to over $240 billion. The demand is real.
Microsoft spent $37.5 billion in a single quarter. Extrapolate that and you get $150 billion annualized, though the company is guiding toward $120 billion for the full fiscal year. Most of it goes to Azure AI infrastructure. OpenAI‘s growing compute requirements are a significant driver.
The Nvidia Dependency
Every company on this list depends on Nvidia. That is the story Meta’s deal makes explicit. Nvidia’s revenue hit $130.5 billion in its most recent fiscal year — nearly all of it from AI chips that these hyperscalers are competing to buy.
The Meta deal specifically covers Blackwell (current generation), Rubin (next generation), Grace and Vera CPUs, and Spectrum-X networking. This is not just a GPU purchase. It is a full infrastructure partnership. Meta is building its AI stack almost entirely on Nvidia silicon.
AMD and Intel compete for a fraction of this market. AMD’s MI300X has gained traction in some workloads, but Nvidia’s CUDA ecosystem and software moat keep it dominant. Intel’s Gaudi accelerators have struggled to find significant adoption outside of specific enterprise use cases.
The Surprise Entry: Saudi Arabia
The biggest wildcard in the AI infrastructure race is not a tech company. Saudi Arabia’s state-backed AI firm Humain invested $3 billion in xAI — Elon Musk‘s AI venture. That single investment exceeds the annual AI budgets of most publicly traded technology companies.
Middle Eastern sovereign wealth funds are quietly becoming major AI infrastructure financiers. The Stargate project — a $500 billion initiative involving OpenAI, SoftBank, and Oracle — draws significant backing from this region.
The AI Startup Revenue Reality
While the hyperscalers spend hundreds of billions building infrastructure, the AI companies using that infrastructure are growing fast — but from a much smaller base.
OpenAI ended 2025 with roughly $20 billion in annual recurring revenue — a threefold increase from the prior year. Anthropic‘s revenue run rate passed $9 billion in January 2026, up from approximately $1 billion at the end of 2024. A ninefold increase in twelve months.
But here is the math that should make everyone nervous: the hyperscalers are spending $690 billion to serve an AI market that is currently generating tens of billions in revenue. The gap between infrastructure investment and revenue return has never been wider in the history of technology.
The Supply Chain Chokepoint
Every hyperscaler reports the same thing: their markets are supply-constrained, not demand-constrained. They would spend more if they could get more chips. TSMC — which manufactures virtually all of Nvidia’s GPUs — is the single biggest bottleneck in the global AI supply chain.
TSMC’s advanced packaging capacity (CoWoS) is allocated years in advance. Nvidia, AMD, and every AI chip designer in the world are competing for the same limited fab capacity. The companies that secured supply agreements early — like Meta just did — have a structural advantage that money alone cannot replicate.
Samsung is attempting to offer an alternative with its advanced packaging, but yield rates have not matched TSMC’s. The duopoly at the leading edge of chip manufacturing is tighter than at any point in semiconductor history.
Who Is Missing
Apple is the notable absence. The most valuable company in the world has made no significant AI infrastructure announcements. Apple Intelligence runs on-device or through partnerships with third-party providers. Their AI strategy is fundamentally different — they are building for inference on consumer devices, not training massive models in data centers.
Whether that is visionary restraint or a strategic mistake will not be clear for years. But right now, Apple is the only trillion-dollar tech company sitting out the biggest infrastructure buildout since the internet itself.
The model era is over. The chip era has begun. And the companies that locked up supply today are the ones that will still be standing when the dust settles.
Our recommendations are based on independent research and analysis. This article may contain affiliate links, which means we may earn a small commission if you make a purchase through our links, at no extra cost to you. This does not influence our editorial decisions or product ratings.