Recommended Free Tools
Short answer: Gartner forecast that hyperscalers would operate approximately $1 trillion worth of AI-optimised servers by 2028. That is an installed-base value, not a prediction that they will collectively spend $1 trillion in 2028. Gartner also forecast $202 billion in worldwide AI-optimised-server spending during 2025, while current company disclosures show annual capital programmes already measured in the hundreds of billions.
The investment case depends on more than buying accelerators. Utilisation, cloud pricing, power availability, custom silicon, depreciation and demand for paid AI services will determine whether the hardware earns an acceptable return.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Nimo AI NAS, Agentic Computer Mini PC and AI Server, AMD Ryzen 7 PRO 8845HS(up to 5.1 GHZ, beat... | $1,999.99 | Buy on Amazon |
What the $1 trillion forecast actually measures
The figure came from a Gartner forecast reported by Computer Weekly on January 21, 2025. Gartner said hyperscalers were expected to operate approximately $1 trillion of AI-optimised servers by 2028. “Operate” describes the value of hardware in service at that point, not a single-year cheque.
The same report forecast $202 billion of worldwide spending on AI-optimised servers in 2025, against $405 billion of total server spending. IT-services companies and hyperscalers together were expected to account for more than 70% of the AI-server spending; that combined category should not be attributed entirely to hyperscalers. Computer Weekly’s report of the Gartner forecast is the source for these figures.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
- [Local AI Inference & 70B Model Ready] Equipped with the AMD Ryzen 7 PRO 8845HS processor, NEXUS is engineered for heavy local AI workloads. With a full-size GPU bay, it runs 70B LLMs natively without an internet connection. Ideal for AI developers and tech enthusiasts who need private environment for coding and model testing.
- [132TB Mass Storage with ZFS Integrity] Features a hybrid storage architecture (3×NVMe + 4×3.5" HDD) supporting up to 132TB. Utilizing the enterprise-grade ZFS file system and ECC memory, it prevents data corruption and bit rot—a must-have for professional photographers and video editors safeguarding 4K/8K RAW footage.
- [OpenClaw-Driven Automation Workflow] The built-in OpenClaw execution layer allows complex automated tasks to be processed locally. Even when offline, your backup schedules and AI file organization continue seamlessly. Say goodbye to monthly cloud subscriptions and high latency.
- [Dual 10GbE & USB4 Ultra-Connectivity] Experience server-class speeds with dual 10GbE ports and a 40Gbps USB4 interface. It enables multi-user real-time collaboration on large project files directly from the NAS, ensuring zero-lag editing for creative studios and production teams.
- [Open-Source ZimaOS for Total Privacy] Running on the fully open-source ZimaOS, NEXUS ensures your data stays physically on-premise with no backdoors. It acts as a "Digital Fortress" for privacy-conscious families and small businesses who demand absolute data sovereignty.
| Measure | Meaning |
|---|---|
| Annual server spending | Hardware purchased during one year. |
| Hyperscaler capex | A broader investment category that can include servers, networking, buildings, land, power equipment, leases and other assets. |
| Installed hardware value | The value of servers operating in a fleet at a particular time. |
| AI-infrastructure investment | A broad category that may include chips, servers, networking, data centres, electricity, cooling and construction. |
Confusing these measures is the main reason the headline sounds larger than the underlying forecast.
Who counts as a hyperscaler?
In its narrowest sense, the term covers very large cloud operators such as Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure. Meta is also a major operator of AI infrastructure, although its business is primarily advertising and consumer internet services rather than selling general-purpose public-cloud capacity. Alibaba, Tencent and other regional providers may be included in worldwide market datasets, depending on the analyst’s methodology.
The Gartner-reported spending-share statement explicitly combined “IT services companies and hyperscalers.” A forecast that uses that grouping does not establish a separate spending total for AWS, Azure, Google Cloud, Meta or any other individual company.
How much are companies committing now?
Company disclosures show the scale of the buildout, but they are not directly interchangeable with Gartner’s server-only estimate.
| Company and date | Reported figure | What it includes |
|---|---|---|
| Microsoft, FY26 Q3 call (2026) | Approximately $190 billion of calendar-year 2026 capex guidance. | Company-wide capital spending. Management said about two-thirds of the latest-quarter capex went to short-lived assets, primarily GPUs and CPUs, and that capacity remained constrained through at least 2026. |
| Alphabet, 2025 results and 2026 guidance | $91.4 billion of 2025 capex; $175 billion–$185 billion expected in 2026. | Alphabet said roughly 60% of 2025 capex was servers and 40% data centres and networking. The 2026 range also includes those broader technical assets. |
| Gartner forecast reported January 2025 | $202 billion worldwide AI-optimised-server spending in 2025. | A market forecast, not an audited final result, and not a total for all AI infrastructure. |
Microsoft’s guidance is discussed in its FY26 Q3 earnings call. Alphabet’s figures are from its 2025 Q4 earnings call. Adding these numbers would double-count different accounting scopes and reporting periods; neither company’s figure is an AI-server-only budget.
What is an AI-optimised server?
The category is broader than a server containing a Nvidia GPU. It can include:
- GPU servers and integrated GPU systems.
- Custom training and inference accelerators or ASICs.
- Host CPUs, high-bandwidth memory and high-speed network interfaces.
- Switches, storage and rack-scale systems that keep training and inference clusters supplied with data.
- Liquid-cooling and power-delivery equipment attached to the computing system.
Buildings, substations, grid connections and land are essential to an AI cluster, but Gartner’s server forecast should not automatically be expanded to include all of those assets.
Why the buildout is accelerating
Training and inference
Frontier-model training consumes large clusters for concentrated periods. Inference creates a continuing requirement to answer user prompts, run agents and generate content. Inference can eventually be the steadier source of utilisation, but its economics depend on token prices, model efficiency and workload volume.
Cloud customers and internal products
Customers rent accelerators for training, fine-tuning and serving models. Hyperscalers also use the same capacity for search, advertising, recommendations, productivity software, coding tools and internal research. Microsoft has linked its spending to cloud demand, first-party applications, AI solutions, research and development, and replacement of older servers. Alphabet has cited Google DeepMind, Google Services, Google Cloud demand, model development and AI compute capacity.
Capacity reservations and replacement
Long-term customer commitments and supply reservations encourage providers to order ahead. Hardware is also replaced as newer systems deliver more performance per watt, even when older equipment still functions.
The hardware stack and the custom-chip strategy
Hyperscaler spending flows through an ecosystem that includes accelerator designers such as Nvidia and AMD; custom-ASIC developers; semiconductor manufacturers and HBM suppliers; server makers; networking and optical-component vendors; storage providers; cooling and power-equipment companies; and data-centre operators.
First-party silicon
Google develops TPU systems; Amazon offers Trainium and Inferentia; Microsoft has Maia accelerators and Cobalt CPUs; and Meta has developed its Training and Inference Accelerator (MTIA). Microsoft said in its FY26 Q3 call that Maia 200 was live in selected data centres and Cobalt CPUs were deployed across nearly half of its data-centre regions. Its fleet also includes custom networking, security and virtualisation silicon.
Custom chips can lower cost per token, improve power efficiency, provide supply and roadmap control, and fit tightly with a provider’s software. They require substantial compiler, framework, model-compatibility and developer-tool investment, however. General-purpose GPUs remain valuable where workloads change quickly, portability matters or the surrounding software ecosystem is decisive. First-party silicon is a diversification and optimisation strategy, not proof that GPUs will disappear.
Who benefits from the spending?
- Accelerator suppliers: provide merchant GPUs, CPUs and interconnect products.
- Custom-chip and semiconductor partners: design and manufacture provider-specific processors.
- Memory and systems suppliers: provide HBM, servers, racks, storage and networking.
- Infrastructure providers: supply buildings, colocation, power equipment, cooling and construction.
- Hyperscalers: convert the installed fleet into cloud revenue and higher-value software products.
A sale to a server manufacturer, a hyperscaler’s capital purchase and a customer’s cloud reservation can represent different steps in the same economic chain. Treating each as separate demand can produce double counting.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Physical constraints may matter more than capital
Access to money or chips does not instantly create usable capacity. Projects can be delayed by grid interconnection queues, transformers and switchgear, land and permitting, data-centre construction, network-fabric deployment, water and cooling availability, semiconductor and HBM supply, local opposition and regional power prices.
Microsoft said it added another gigawatt of capacity in its latest quarter but remained constrained despite rapid expansion. Capacity shortages show that demand exceeds available supply at a point in time; they do not prove that every planned facility will earn an adequate return.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Can a trillion-dollar fleet earn its cost?
There is no settled answer. The relevant calculation is not the purchase price alone but the revenue and gross profit generated over the useful life of each system.
- Utilisation: idle accelerators still incur depreciation and facility costs.
- Revenue per accelerator-hour: pricing must cover power, networking, support and hardware depreciation.
- Model economics: lower token prices and more efficient models can increase demand while reducing revenue per unit of compute.
- Obsolescence: a new architecture can shorten the economic life of an older fleet.
- Product monetisation: AI features must increase subscriptions, advertising, cloud consumption or other revenue enough to justify their cost.
- Custom-silicon payback: savings must exceed design, software and deployment costs.
Microsoft has said continued AI-infrastructure investment and rising AI-product usage pressured cloud gross margins, although efficiency gains offset part of that pressure. Its FY26 Q3 performance disclosure provides that management commentary.
What could slow or reverse the buildout?
- Inference demand may grow more slowly than forecasts assume.
- Model compression, smaller models or better algorithms may reduce compute requirements.
- Customers may reject high-priced AI services or move to open-source and self-hosted alternatives.
- Custom accelerators may reduce demand for general-purpose GPU fleets in predictable workloads.
- Fast architecture changes can strand equipment before depreciation is complete.
- Energy, permitting, financing or construction constraints can delay projects or make them uneconomic.
- A supply glut could push down accelerator-rental prices.
- Regulation may restrict data-centre construction, model deployment or cross-border data use.
The Gartner-reported discussion of AI devices included a similar warning: sales could rise without a compelling must-have application that justifies a premium. Hardware enthusiasm can arrive before durable software monetisation.
What this means for enterprise buyers
The right choice is workload-specific rather than a simple decision to buy the fastest accelerator.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →| Workload or priority | Usually worth comparing | Main risks |
|---|---|---|
| Variable experiments and early products | On-demand or reserved cloud capacity. | Regional shortages, changing prices, egress and lock-in. |
| Predictable, high-utilisation inference | Reserved capacity, custom accelerators and private systems. | Hardware obsolescence and under-utilisation if demand falls. |
| Large-scale training | Accelerator availability, interconnect topology, storage throughput and cluster scheduling together. | A cheap chip with inadequate networking can make the whole run uneconomic. |
| Sensitive or regulated data | Private infrastructure or a compliant regional provider. | Power, cooling, specialist staff and procurement lead times. |
| Mixed or portability-sensitive workloads | Hybrid or multi-cloud designs. | Data movement, duplicated operations and framework incompatibility. |
Compare effective cost per training run or per useful output, not just the advertised hourly instance price. Check the actual accelerator model and memory, regional availability, on-demand versus reserved terms, minimum commitments, storage and interconnect charges, egress, support, framework compatibility, data residency and the ability to move workloads elsewhere. Official buying pages include AWS machine learning, Azure AI, Google Cloud AI infrastructure, Oracle Cloud AI infrastructure and CoreWeave. Prices and capacity change by region and contract, so no single current price should be assumed.
Verdict
The defensible interpretation is that hyperscalers are building toward a trillion-dollar installed base of AI-optimised servers by 2028. Annual capital spending is already enormous, but company capex figures include much more than AI servers, while Gartner’s figures are forecasts with a different scope. The central business question is whether utilisation and AI-service revenue can cover depreciation, electricity, facilities, financing and software costs before the hardware becomes obsolete.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
