Hyperscaler Capex and AI Infrastructure

Worldwide data center capital expenditure increased 92% year over year in Q2 2026, while data center semiconductor and component revenue rose 182%. Dell’Oro Group linked the growth to ongoing AI infrastructure investment and sharply higher memory and storage prices.

Amazon (AMZN), Alphabet (GOOGL), Microsoft (MSFT), and Meta Platforms (META) make the largest commitments. Oracle (ORCL), Alibaba (BABA), Tencent (TCEHY), CoreWeave (CRWV), and Nebius Group (NBIS) are expanding the buyer base. The spending cycle reaches Taiwan Semiconductor Manufacturing (TSM), Nvidia (NVDA), Advanced Micro Devices (AMD), Micron Technology (MU), Samsung Electronics (005930.KS), SK Hynix (000660.KS), Broadcom (AVGO), and ASML Holding (ASML) before a new cloud instance generates its first dollar of revenue.

Key Takeaways:

• Amazon, Alphabet, and Microsoft are expected to spend about 102% of their combined cloud revenue on company-wide capex in 2026, based on UBS estimates reported by 24/7 Wall St.
• Dell’Oro Group reported 92% year-over-year data center capex growth in Q2 2026 and 182% growth in data center semiconductor and component revenue.
• Amazon’s latest 2026 capex target is about $220 billion, Alphabet’s is $195 billion to $205 billion, Microsoft’s is about $175 billion, and Meta’s is $130 billion to $145 billion.
• Inference, storage, and agentic workloads are increasing demand for CPUs, DRAM, storage drives, network adapters, and custom accelerators alongside merchant GPUs.
• The financial test is cloud revenue growth versus depreciation, financing, and power costs, rather than the size of another data center announcement.

The Scale of Hyperscaler AI Capex in 2026

UBS estimates Amazon, Alphabet, and Microsoft will collectively spend about 102% of their cloud revenue on capital expenditure in 2026, according to 24/7 Wall St.’s report on the forecast. The ratio is expected to decrease to roughly 99% in 2027 and 94% in 2028. It compares cloud revenue with company-wide capex, so it should not be interpreted as a claim that every purchased server, building, or network switch supports generative AI.

The ratio reflects reinvestment intensity. Retail, advertising, productivity software, subscriptions, and other businesses help finance infrastructure whose cost is comparable to the cloud revenue generated by the three largest public-cloud operators. The spending draws cash from beyond AWS, Azure, and Google Cloud, even though those platforms are among the main routes through which the resulting capacity will be sold.

UBS’s broader forecast puts total hyperscaler capex at $492 billion in 2025, $1.009 trillion in 2026, $1.447 trillion in 2027, and $1.619 trillion in 2028. Its cumulative estimate for 2026 through 2028 is about $4.1 trillion, more than three times the $1.292 trillion spent during the prior six years. A slower growth rate after 2026 would still leave annual spending rising because each increase applies to a much larger base.

This updates our recent hyperscaler investment tracker. The earlier question was whether the largest providers would keep raising guidance. The Q2 evidence shows spending has also expanded across memory, storage, general-purpose servers, and networking, making the cycle less dependent on GPU unit growth alone.

Amazon, Alphabet, Microsoft, and Meta Compared

The four largest US spenders have different business models and different exposure to the cycle. Amazon funds AWS alongside retail, logistics, and fulfillment assets. Alphabet splits infrastructure across Google Cloud and internal services. Microsoft supports Azure, enterprise software, and consumer workloads while making substantial use of leases. Meta primarily buys capacity for its own products rather than operating a mature public cloud business.

Amazon raised its target from about $200 billion to $220 billion, with higher memory and component costs contributing to the revision. A larger budget can purchase additional capacity, but it can also reflect a higher price for the same planned server fleet.

Alphabet raised its range from an initial $175 billion to $185 billion to a latest $195 billion to $205 billion. It spent $44.9 billion in Q2 2026, with about 60% going to servers and approximately 40% to data centers and networking. First-half capex reached $78.6 billion, leaving a heavy second-half deployment schedule if Alphabet finishes within the raised range.

Google Cloud’s growth is the strongest among the three major public-cloud businesses in the cited quarter. Alphabet also reported negative free cash flow of about $5.9 billion as quarterly capex exceeded operating cash generation. Fast cloud growth and negative free cash flow can coexist because infrastructure consumes cash before depreciation spreads its cost across future periods.

Microsoft’s approximately $175 billion plan sits below Amazon’s and Alphabet’s latest indications. Microsoft’s higher use of leases makes simple cash-capex comparisons incomplete, because financed capacity can become productive before the full cash cost appears in purchases of property and equipment.

Meta’s $130 billion to $145 billion range is smaller in absolute terms but large relative to its operating cash flow. Its Q2 capital spending of $31.1 billion nearly matched the quarter’s $31.9 billion of operating cash generation. Unlike AWS, Azure, and Google Cloud, Meta cannot primarily recover the cost through external infrastructure rental. It must convert compute into better advertising, recommendation, engagement, or new products.

Oracle, Alibaba, Tencent, and Neocloud Expansion

Oracle occupies a more financing-sensitive position than the Big Four. Oracle Cloud Infrastructure can use contracted demand to support construction, but a thinner cash-flow base leaves less room for delays between ordering equipment, energizing a site, and recognizing cloud revenue.

CoreWeave and Nebius expand the market beyond diversified technology companies. UBS projects roughly $130 billion of cumulative capex for CoreWeave and $93 billion for Nebius from 2026 through 2028. These specialized providers give customers an alternative route to accelerated compute, but their economics are more directly exposed to hardware use, financing costs, and the resale value of older accelerators.

Alibaba and Tencent are building under a different set of constraints. Alibaba indicated it would exceed its original $56 billion three-year capex target and issued an HK$80 billion, or $10.2 billion, Hong Kong placement to fund the buildout, according to AP’s earnings coverage.

Tencent reported Q2 2026 revenue of RMB204.8 billion, up 11% year over year, while net profit was nearly unchanged amid higher spending. Chinese providers face restricted access to advanced imported accelerators, so more of their budgets flow toward domestic processors and a separate hardware stack. The financial pressure is recognizable, but the supplier beneficiaries differ from those serving US hyperscalers.

Data Center Capex Growth and Component Inflation

Dell’Oro’s 92% Q2 capex increase reflects both additional equipment and higher equipment prices. Accelerated servers remained the main spending driver, while higher memory and storage prices raised average server selling prices. Dell led server OEM revenue, followed by Supermicro and Lenovo, and white-box server revenue reached a record high.

A separate Dell’Oro semiconductor and component report found revenue for parts used in servers and storage systems rose 182% in Q2. DRAM and storage-drive revenue grew at triple-digit rates, with average selling prices per bit more than doubling year over year. Accelerated server deployments also raised demand for HBM, CPUs, memory, and high-speed back-end network interface cards.

These numbers complicate comparisons with previous years. A dollar spent in 2026 buys a mix with more accelerators, more HBM, faster networking, and more expensive storage than an earlier general-purpose server deployment. Calling the change a simple increase in server count misses the shift in system value and power density.

Dell’Oro expects global data center capex to exceed $3 trillion by 2030. Its July 2026 forecast was nearly twice its January outlook, reflecting higher cloud-provider guidance, larger projections for data center power capacity, and higher commodity costs. The firm expects high-end accelerators to remain the largest spending category, but it also sees general-purpose servers benefiting from inference, agentic, and storage workloads.

TSMC, Memory, and ASML Capture the Upstream Spend

TSMC is the main foundry link between hyperscaler budgets and deployable accelerators. Its August 2026 revenue reached NT$514.8 billion, up 53.3% year over year and 10.1% from July, according to its monthly revenue disclosure and related coverage. TrendForce put TSMC’s Q2 foundry share at 72.5%, compared with 5.9% for Samsung Foundry and 5.4% for SMIC.

TSMC’s 5-, 4-, and 3-nanometer nodes ran at full capacity during Q2 on strong orders for AI server chips. CoWoS advanced-packaging capacity was also reported as sold out through 2026, with lead times of 52 to 78 weeks. A hyperscaler can design a custom processor and still face the same foundry and packaging bottlenecks as a merchant GPU vendor.

Memory has become another budget-setting input. The global DRAM industry generated $154.73 billion in Q2 2026 revenue, a 59.5% sequential increase. HBM consumes more wafer capacity and carries higher revenue per wafer than conventional DRAM, encouraging suppliers to shift production toward data center products and tightening supply elsewhere.

ASML sits one step further upstream. The lithography supplier raised its 2026 sales outlook to about EUR44 billion as chipmakers added advanced logic and memory capacity. EUV demand connects cloud budgets to semiconductor capital equipment: hyperscalers order systems, chip designers reserve foundry capacity, foundries expand advanced nodes, and ASML receives tool orders. Alternatives such as Canon’s nanoimprint technology target selected wafer segments, but the 2026 buildout still depends heavily on ASML’s EUV systems for leading-edge production.

The Training-to-Inference Shift Changes the Chip Mix

Worldwide spending on AI-optimized infrastructure as a service is forecast to reach $42 billion in 2026, up 96% from $21.5 billion in 2025. Inference accounts for $23.3 billion of the 2026 total, compared with $19 billion for training. The crossover changes what cloud operators need from each server generation.

Training favors large, tightly connected accelerator clusters that maximize throughput during scheduled model runs. Inference runs continuously and is more sensitive to memory capacity, memory bandwidth, latency, storage access, and the cost of serving each accepted output. Agentic workloads also create demand for general-purpose CPUs and storage because a task can involve retrieval, tool calls, intermediate state, and repeated model execution.

Dell’Oro observed this broadening directly in Q2. Nvidia’s Blackwell platform and custom accelerators from Google and Amazon continued gaining momentum, while general-purpose server demand remained strong. Its forecast expects the AI-specialized cloud segment, which includes model builders and neoclouds, to grow at nearly a 60% compound annual rate through the forecast period.

Custom chips give AWS, Google, Microsoft, and Meta another way to control workload-specific costs. Trainium, TPU, Maia, and MTIA reduce dependence on merchant accelerators for selected workloads, but each increases dependence on the provider’s own software stack and supply commitments. Nvidia remains central because many external customers value CUDA compatibility and the ability to move workloads across a broad installed base.

Financing, Depreciation, and Supply-Glut Risk

The historical cloud model used operating cash flow to fund most infrastructure. That model is changing as construction, power equipment, and hardware purchases accelerate. Debt, leases, customer prepayments, joint ventures, and special-purpose vehicles are filling the gap.

Depreciation creates a second timing issue. Current operating margins therefore reflect only part of the cost associated with the latest construction and hardware cohorts.

The accounting treatment also differs by company. Alphabet and Microsoft use longer lives for parts of their server and networking fleets, which spreads expense across more reporting periods. Amazon shortened the life of a subset of servers and network equipment from six years to five years because of faster technology development, particularly in AI and machine learning. Both approaches can be defensible, but they produce different near-term margins for economically similar assets.

A supply glut would appear first through lower use, weaker rental pricing for older accelerators, or slower backlog conversion. Newer hardware can reduce the cost per output enough to weaken the economics of an older fleet before its book value reaches zero. Buildings and power connections can host several generations of equipment, but the servers inside them carry a shorter economic life.

Prediction Scorecard

My pending Alphabet forecast calls for 2026 full-year capex of at least $195 billion by the company’s full-year report. The latest $195 billion to $205 billion guidance and $78.6 billion of first-half spending keep that call intact, but execution requires a much larger second-half deployment pace.

My pending Google Cloud forecast calls for full-year 2026 operating income above $10 billion. Q2 operating income of $8.8 billion and a 35.6% segment margin support that threshold, although rising depreciation from the current investment cycle will become a larger expense in subsequent periods.

A previous analysis on this site forecast that Microsoft shares would close above $400 by December 31, 2026. That market call remains separate from the infrastructure thesis: Azure demand and a lower capex-to-operating-cash-flow burden support the business case, but valuation and interest rates will also determine the share-price result.

Outlook and the Next Capex Test

The next earnings cycle should be evaluated based on execution rather than another round of headline commitments. Amazon needs to show that AWS growth and backlog conversion can bring capex back below operating cash flow. Alphabet must deploy enough during the second half to reach its raised range without allowing depreciation to exceed cloud operating income. Microsoft must clarify how much new capacity arrives through cash purchases versus leases. Meta must connect internal compute spending to measurable revenue or cost improvements.

The supply-side checks are equally specific. TSMC’s HPC-linked demand and CoWoS availability indicate how many accelerators can ship. SK Hynix, Samsung, and Micron determine HBM availability and memory pricing. ASML’s EUV bookings indicate whether foundries are adding leading-edge capacity rather than relying only on existing fabs. Power interconnections decide when completed equipment can enter service.

Friday’s market close showed investors still rewarding technology exposure, but without a broad risk-on surge. Compared with the September 17 figures in our latest chip and infrastructure update, the Nasdaq added 104.24 points while the Dow reversed part of its prior gain.

The 102% cloud-revenue comparison shows the size of the bet, but the more useful operating measure is cloud revenue growth minus depreciation growth. Rising revenue with faster-rising depreciation would compress margins even while sales expand. Rising revenue with high use and slower unit-cost growth would validate another round of construction.

The 2026 buildout has already moved beyond a GPU purchasing cycle. It now includes CPUs, custom accelerators, HBM, storage, network adapters, optical links, cooling, power equipment, buildings, leases, and debt. The providers that turn those assets into consistently used inference capacity will gain a cost and availability advantage. Providers that energize sites late, buy the wrong chip mix, or carry underused hardware will absorb the same spending through weaker free cash flow and higher depreciation.

Disclosure: This article is for informational and analytical purposes only. It is not investment advice, a price target, or a recommendation to buy, sell, or hold any security.

Related Reading

More in-depth coverage from this blog on closely related topics:

• Hyperscaler Investment Trends for 2026
• Google Open Agentic Orchestrator for AI
• How to Save AI Models from Deletion
• What Does ChatGPT Know About My Web Activity
• What Is RSA-896 and Its Security Implications

Sources and References

Sources cited while researching and writing this article:

• Data Center Capex Grew 92 Percent in 2Q 2026, Driven by Surging AI Demand and Memory Costs, According to Dell’Oro Group
• AI’s Absurd Spending Boom? Hyperscalers Are Spending 102% of Cloud Revenue on Capex
• TechTimes
• MLQ
• CNBC
• FactSet
• Data Center IT Semiconductor and Component Revenue Surged 182 Percent in 2Q 2026, According to Dell’Oro Group
• TSMC August 2026 revenue hits record high on AI chip demand