Who Is Investing in AI Infrastructure

The four largest US hyperscalers are now projecting combined $725 billion in capital expenditure for 2026, a 77% increase from about $410 billion in 2025, based on figures from Goldman Sachs via Yahoo Finance. This single figure has become the central focus of the AI infrastructure market, but it hides important differences. These companies are spending differently, on different items, and with varying financial flexibility. Amazon (AMZN) has already reached a point the others have not: its trailing capex surpassed its operating cash flow, while Alphabet (GOOGL) reported its first quarterly cash burn since its 2004 IPO.

Key Takeaways

• Microsoft, Alphabet, Amazon, and Meta plan roughly $725 billion in combined 2026 capex, up 77% from 2025, with Morgan Stanley stating the 2027 consensus of $1.2 trillion still falls $200 billion short.
• Amazon increased its 2026 target to $220 billion due to memory-cost inflation, while its AWS backlog grew by $132 billion in one quarter to $496 billion.
• The supply bottleneck has shifted from GPUs to memory and power: TSMC raised its 2026 capex budget to $60-64 billion, and SK Hynix controls about 58% of the HBM market.
• China is now a separate market, with Alibaba expecting to exceed its $56 billion three-year target and Tencent’s Q2 profit stalling amid rising AI spending.
• Custom silicon (Trainium, TPU, Maia, MTIA) is expanding at roughly three times the rate of merchant GPUs, but none of it is available for external customers.

The $725 Billion Baseline and Guidance Increase

The headline figure keeps rising because each quarterly report updates it upward. Amazon began 2026 with guidance near $200 billion, then raised it to $220 billion during its Q2 call, citing memory cost inflation as the main factor. Alphabet increased its 2026 range to $195-205 billion from $180-190 billion. Microsoft maintained its underlying spend around $190 billion, with about $25 billion due to higher memory and component prices. Meta narrowed its range to $130-145 billion, effectively raising the minimum. Together, these four companies are now expected to spend over $700 billion in 2026 alone, a 77% rise from 2025’s roughly $410 billion and nearly triple the $226 billion spent in 2024.

The quarterly trend confirms this sharply. Combined purchases of property and equipment across the four rose from $88.2 billion in 2Q25 to $97.3 billion in 3Q25, $118.6 billion in 4Q25, and $129.8 billion in 1Q26, according to company filings compiled by SiliconAnalysts. This represents a $41.6 billion increase in quarterly spending over nine months, which is why Morgan Stanley now believes analysts are still underestimating the cycle.

Morgan Stanley’s concern is based on calculations, not optimism. The firm’s $1.4 trillion 2027 estimate is 17% higher than a consensus that itself rose $170 billion in one quarter to $1.2 trillion. The firm points out that consensus growth of 29% for 2027 implies non-AI cloud infrastructure will grow only 7% year over year, which conflicts with the visible committed pipeline of database services, SaaS, streaming, and enterprise workloads that have grown at double-digit rates for years. The infrastructure under construction supports both AI and non-AI workloads, so assuming the entire guidance increase is AI capex while keeping other spending flat results in a forecast that does not match what companies are actually building.

Who Is Spending Where: Differences in Financial Position

The four companies are not interchangeable, and this shows in the ratio of capex to operating cash flow. As explained in our August 20 hyperscaler spending analysis, Amazon’s trailing capex reached 102% of operating cash flow, compared with 63% at Alphabet, 61% at Meta, and 57% at Microsoft. Amazon’s $44.2 billion in first-quarter purchases exceeded its $26.0 billion in first-quarter operating cash flow, a ratio of $1.70 of capex for each $1.00 of cash generated.

This difference affects which companies can handle timing mismatches. Amazon’s retail, logistics, and AWS divisions all compete for the same corporate cash, and the company’s free cash flow shifted to a trailing outflow of $7.6 billion in Q2 from an $18.2 billion inflow a year earlier. Alphabet posted negative free cash flow of about $5.9 billion in Q2, its first since 2004, as $44.9 billion of capex exceeded $39.1 billion of operating cash flow. Microsoft and Meta, by contrast, retain operating cash after capital spending, giving them more flexibility to handle delays or component price increases.

Oracle (ORCL) is in a more vulnerable position than the four. Reuters reported that Oracle’s fiscal 2026 capex was 174% of operating cash flow, the highest ratio among the group, and its shares have underperformed as investors consider the thinner cash-flow base alongside aggressive cloud expansion. The difference between order visibility and return visibility is clearest here: suppliers like Nvidia (NVDA) and Broadcom (AVGO) have strong order visibility because budgets are announced, while infrastructure owners need return visibility that depends on usage, pricing, useful life, and financing cost, all of which take years to assess.

The Supply Side: TSMC, HBM, and Memory Pricing

Spending is now moving upstream into a supply chain with limited capacity, and the bottleneck has shifted from logic to memory. TSMC raised its full-year 2026 capital budget to between $60 billion and $64 billion, allocating roughly 70% to 80% to advanced process technologies and 10% to 20% to advanced packaging, testing, and mask-making. The foundry generated nearly NT$783 billion in cash from operations in Q2 2026, spent NT$496 billion on capex, and still increased its cash balance to NT$3.1 trillion. Its July revenue of NT$467.58 billion was up 44.7% year over year, exceeding its own 40% full-year growth target.

Memory pricing has become the main source of pricing power. SK Hynix controls about 58% of the high-bandwidth memory market in Q1 2026, according to GlobalData’s analysis reported by InvestorIdeas, giving it significant influence over how quickly premium GPUs can ship. Samsung regained the top spot in overall DRAM shipments with 39% share during Q2 2026, according to Counterpoint Research, while Micron narrowed the gap with SK Hynix. Memory manufacturers earn an estimated three to five times more revenue per wafer producing HBM than conventional DRAM, which explains why manufacturing capacity has shifted toward AI data center use and pushed memory contract prices sharply higher, adding roughly $20 billion to Amazon’s projected infrastructure bill alone.

ASML (ASML) benefits quietly at the top of the supply chain. The lithography supplier raised its 2026 sales guidance again to about €44 billion, driven by demand from chipmakers adding capacity for both AI logic and memory. The firm’s EUV monopoly on advanced logic and HBM production means the hyperscaler capex cycle ultimately results in bookings at a single Dutch company, concentrating the entire buildout’s risk on one supplier’s performance.

The Shift from Training to Inference and Chip Mix

The spending composition is changing as much as the total amount. Gartner projects worldwide spending on AI-optimized infrastructure-as-a-service will reach $42 billion in 2026, up 96% from $21.5 billion in 2025, with inference at $23.3 billion compared to $19 billion for training. This crossover matters because inference and training require different hardware. Training demands raw compute throughput across large clusters, while token generation depends heavily on memory capacity and bandwidth, since model weights and cached context must be available for every generation step. This explains why the memory bottleneck affects inference most and why HBM pricing now drives hyperscaler budgets.

The clearest structural change is the growth of custom silicon. Amazon’s Trainium, Google’s TPU, Microsoft’s Maia, and Meta’s MTIA are all expanding, and TrendForce estimates custom chip shipments grew 44.6% in 2026, about three times the 16.1% growth rate for Nvidia’s merchant GPUs. Amazon claims its Trainium3 delivers roughly 50% lower total cost of ownership than Nvidia’s Blackwell NVL72 at rack scale, and AWS’s AI services and chips businesses each passed $25 billion annual revenue run rate in Q2, both growing at triple-digit rates.

The trade-off is clear. None of these four custom chips are available for external customers to rent; they are exclusive to their builders. Custom silicon lowers each hyperscaler’s internal compute costs but does not reduce demand for Nvidia hardware in the broader market, which is why Nvidia’s revenue has not declined even as custom silicon grows in unit terms. The cost to migrate off CUDA remains the barrier that keeps the merchant GPU market intact. As noted in our weekend chip news analysis, Nvidia is passing memory cost increases to customers with server price hikes exceeding 15% on early-2027 Vera Rubin and Grace Blackwell systems, a move that confirms strong demand and shows that even a company with a 75% gross margin cannot absorb rising memory costs.

China: A Separate Market with Different Economics

The US four hyperscalers tell only part of the story. China’s hyperscalers are building a parallel AI infrastructure stack under export-control restrictions, and their spending is now a significant macroeconomic event. Alibaba (BABA) reported a 75% drop in quarterly profit to about $1.6 billion as AI investment spending increased, and CEO Eddie Wu told analysts the company is likely to exceed its original $56 billion three-year capex target. The company then issued an HK$80 billion ($10.2 billion) Hong Kong placement, the largest primary follow-on offering by a Hong Kong-listed company, to fund the buildout, and shares fell about 10% in response.

Tencent (TCEHY) reported Q2 2026 revenue of RMB204.8 billion, up 11% year over year, but net profit barely changed as the company increased AI spending. The stock was down 26% so far in 2026 as investors grew concerned about rising expenses amid intense domestic AI competition. The Chinese buildout differs structurally from the US one: it relies on domestic chips and a restricted import pipeline, so its capex does not flow to Nvidia or TSMC’s leading edge in the same way, and its financing depends more on equity dilution than on the investment-grade bond market that funds US hyperscalers. This explains why Alibaba’s placement and Tencent’s profit stall correspond to Amazon’s free-cash-flow swing and Alphabet’s cash burn, reflecting similar pressures through different financial mechanisms.

The Conversion Test: What Determines the Next Phase

The focus has shifted from whether they will spend to whether they can generate returns. Early signs on conversion are mixed. AWS operating margin rose to 39.4% in Q2, up 650 basis points year over year, and Amazon stated that AI infrastructure investments are reaching breakeven in under three years. Google Cloud’s margin increased to 35.6% from 20.7% year over year. These are the strongest indicators so far that the buildout can cover its cost of capital.

The opposing factor is the depreciation wall. Combined trailing depreciation for the four was about $149 billion compared to $433.9 billion of capital expenditure in the four quarters through March 2026, meaning current spending far exceeds the depreciation charge reflected in operating income. Cash leaves during construction, but the income statement records the cost over five to six years for servers and 25 to 40 years for buildings. The companies have reached different conclusions about useful life: Microsoft, Alphabet, and Meta have extended server lifespans to smooth near-term margins, while Amazon shortened some server lives from six to five years, citing rapid AI-driven obsolescence. This disagreement over the same equipment category is a subtle warning in the sector.

Power capacity is now the main physical limit, ahead of chips. Goldman Sachs projects US data center power demand doubling from 31 gigawatts in 2025 to 66 gigawatts by 2027, and only about 50% to 60% of data center capacity planned for 2027 is expected to come online on schedule because power delivery, not equipment, is the bottleneck. That is why Microsoft signed a 20-year power purchase agreement to restart Three Mile Island’s Unit 1, why Google contracted with Kairos Power for small modular reactors, and why hyperscalers are becoming de facto power providers. A facility with allocated accelerators but no grid connection cannot generate cloud revenue, so the order of land acquisition, power availability, and building construction now controls deployment speed more than chip supply does.

Nvidia will report fiscal second-quarter earnings on August 26. Key points to watch include whether data center revenue exceeded $80 billion and whether gross margin stayed near 75% despite memory cost increases. The forward supply commitment rising to $119 billion from $50 billion will indicate whether the order pipeline continues to grow. Beyond individual capex reports, the most important metric is the ratio of cloud revenue growth to depreciation growth, because that ratio, not the headline spending number, will determine whether 2026’s $725 billion spending surge becomes the base of a sustainable compute business or the peak of a cycle that balance sheets cannot support.

Related Reading

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

• Hyperscaler Spending
• Latest Chip News and AI Infrastructure
• Xiaomi New CPU Performance
• What is Nostr? Benefits of Nostr Network
• How to Fix Claude API Problems

Sources and References

Sources cited while researching and writing this article:

• Goldman Sachs via Yahoo Finance
• Jeff Bezos’ Amazon Just Raised Its AI Spending Target to $220 Billion for 2026. Here’s What That Capex Hike Means for Investors.
• over $700 billion in 2026 alone
• SiliconAnalysts
• Street Got Cloud Capex Wrong: Morgan Stanley’s $1.4T Math After Hyperscaler Earnings
• AWS Backlog Hits $496B as Amazon Raises AI Spend to $220B and Capacity Runs Short
• TSMC Raises 2026 CapEx Budget to Capture Long-Term AI, HPC Demand
• InvestorIdeas
• Alibaba quarterly profit drops 75% as AI investment spending grows