The explosive proliferation of AI agents is triggering an unexpected computing crunch—but this time, the shortage isn’t in graphics processing units (GPUs). It’s in the central processing units (CPUs) that have long powered the internet era. Market reports indicate that Amazon Web Services (AWS) management recently issued stringent cost-saving directives to internal engineers, with some engineers now waiting days—rather than hours—for CPU server resources. The squeeze on computing capacity has now spread comprehensively to traditional general-purpose computing infrastructure.
According to a report by tech publication The Information, AWS management convened engineers in May to deliver a warning signal: to ensure that the core EC2 cloud server business can meet all customer demand going forward, engineers must do everything possible to conserve computing resources. This directive covers not only the chronically supply-constrained AI-specific chips but also explicitly targets traditional CPU servers.
Sources familiar with the matter revealed that AWS has set deadlines later this year for teams to reduce their computing resource footprint. Engineers are progressively shutting down EC2 virtual server instances previously used for software development that now sit idle, reallocating that capacity to paying external customers. One AWS engineer with years of tenure at the company said that CPU server resources that once took just hours to provision now require queuing for days—something unprecedented in their career—and warned that such delays are highly likely to slow project delivery timelines.
In response to external scrutiny, AWS officially stated that even amid heavy demand, the company can still meet the computing needs of the vast majority of internal and external customers, and is working closely with internal teams to ensure EC2 resources are utilized as efficiently as ever. AWS also clarified that its resource usage guidelines for employees have not changed due to the current tightness in memory chip supply.
While the impact on external customers remains relatively limited for now, warning signs have emerged in AWS’s Spot Instances. Spot Instances—surplus server capacity sold at steep discounts but reclaimable with just two minutes’ notice—have become increasingly difficult to secure in large batches in recent months. Analysts view this as evidence that the gap between AWS’s overall computing supply and demand is narrowing, compressing the buffer. For enterprise customers reliant on low-cost Spot Instances to run large-scale workloads, this may signal impending upward pressure on cloud computing costs.
Agentic AI Adoption Upends Computing Demand Structure
The core driver behind this CPU shortage is the rapid adoption of Agentic AI. Jing Xie, co-founder and managing director of AI consultancy Elendil Labs, analyzed that AI agents typically require more cloud computing resources during execution, generating sustained demand for CPUs. She observed that per-capita IT spending among clients has doubled precisely because of large-scale AI agent deployment.
“Essentially, a lot of the work being built, produced, and run now requires more CPUs than before,” Xie said. Even within the AI development pipeline itself, CPUs play an indispensable role. For example, when AI companies prepare data for model training, they need CPUs to read raw data from files, images, and videos.
Chip Giants Confirm CPU-to-GPU Usage Nearing 1:1
Statements from major chipmakers further corroborate this trend. Intel CEO Lip-Bu Tan disclosed during the company’s April earnings call that the CPU-to-GPU usage ratio in AI inference—model execution—stood at 1:4. By July, however, Intel CFO David Zinsner stated that this ratio had rapidly climbed to nearly 1:1. Executives at AMD and Arm have made similar comments.
Synthesizing the original disclosures from Intel’s two quarterly earnings calls, the timeline of this ratio shift becomes clearer:
Earnings CallSpeakerDisclosureQ1 2026 (April 23)CEO Lip-Bu TanCPU:GPU usage ratio “used to be about 1:8, now has converged to 1:4″Q1 2026 (April 23)CFO David ZinsnerTraining workloads: roughly 7-8 GPUs per CPU; inference workloads: down to 3-4 GPUs per CPUQ2 2026 (July)CFO David ZinsnerRatio “has nearly reached 1:1 parity,” with expectation it may further tilt toward CPU
In other words, statements from two Intel executives over a three-month span point to the same conclusion: inference and multi-agent collaboration scenarios driven by Agentic AI are elevating the CPU from a supporting role to one of equal importance alongside the GPU.
Beyond CPUs themselves, insufficient supply of companion memory chips and physical space constraints in data centers are jointly exacerbating this computing shortage.
Tech Giants’ Battle for Computing Resources
Balancing internal and external computing resource allocation has become a shared challenge for major tech companies over the past two years. According to reports, Google last year established a committee of senior executives responsible for coordinating computing resource allocation among Google Cloud, the DeepMind AI research division, and consumer businesses. Even so, internal friction persists—star Google AI researcher Noam Shazeer departed earlier this summer, reportedly due to dissatisfaction with computing resource access.
Microsoft’s experience illustrates the flip side: improvements in computing management efficiency can directly translate into revenue growth. Microsoft CFO Amy Hood, during last month’s earnings call, attributed part of Azure’s cloud services growth to “efficiency gains” in managing CPU and GPU server fleets. She stated bluntly: “When we can achieve efficiency gains, those benefits are monetized quickly.”
Three Giants, Three Signals of Computing Anxiety
Comparing the recent situations of AWS, Google, and Microsoft side by side reveals the same story of computing resource tightness manifesting in different ways across each company:
CompanyCore SignalSpecific Data / QuotesAWSInternal austerity; external demand still cannot be fully metCEO Andy Jassy stated on the August 7 Q2 earnings call: “We still cannot fully meet overall demand in 2026”; self-developed Graviton CPUs adopted by 98% of top 1,000 EC2 customersGoogleEstablished senior committee to coordinate computing allocation, yet still lost top talentFormed cross-Google Cloud, DeepMind, and consumer business computing allocation committee in 2025; star researcher Noam Shazeer departed for OpenAI this summerMicrosoftConverting computing scheduling efficiency directly into revenue growthCFO Amy Hood stated on the July 29 FY2026 Q4 earnings call that Azure growth accelerated to 43% (beating guidance), and reduced new GPU dock-to-live time by nearly 50%
Notably, AWS released its latest earnings on the same day (August 7) that The Information report surfaced, effectively providing an official footnote to this CPU crunch: CEO Jassy reiterated that the supply-demand gap cannot be fully closed in the short term. The intensity of customer demand for its proprietary computing capacity was already evident in the 2025 shareholder letter—two large customers had inquired about buying out AWS’s entire 2026 Graviton production capacity, a request AWS ultimately declined.
This also corroborates the earlier assessment that upward pressure on cloud computing costs may be looming: AWS has already raised EC2 GPU Capacity Block prices twice recently—15% in January and approximately 20% in July—primarily reflecting tight high-bandwidth memory (HBM) supply. For now, these price increases are limited to GPU reserved capacity; AWS’s proprietary Trainium chips have not yet followed suit.