Hyperscaler AI Capex 2026: The $820B Buildout and What It Does to Cloud Economics

Hyperscaler AI Capex 2026: The $820B Buildout and What It Does to Cloud Economics

Hyperscaler AI Capex 2026: The $820B Buildout and What It Does to Cloud Economics

Last Updated: November 17, 2026

Six months ago this post estimated that Microsoft, Alphabet, Amazon, Meta and Oracle would spend roughly $330–355 billion on capital expenditure in 2026. That estimate was badly wrong. After the late-October 2026 earnings round, the companies’ own guidance puts hyperscaler AI capex at roughly $810–840 billion for the 2026 spending year, more than double what the consensus assumed in spring. The money is now so large that it no longer fits inside operating cash flow: Amazon’s trailing free cash flow has turned negative, Meta burned cash in Q3, and Oracle is funding its build with bonds, equity and supplier credit. For cloud buyers, this is not an abstract Wall Street story. It sets GPU prices, decides who gets capacity first, and fixes the depreciation clock that every AI price list must recover. This rewrite replaces every figure with numbers from the Q3 2026 earnings releases and calls.

What this covers: verified Q3 2026 capex and guidance for all five companies, where the dollars go, GPU total cost of ownership and useful-life math, why power is now the binding constraint, and a buyer’s playbook for 2027.

Context and Background

The AI infrastructure cycle has gone through three phases since late 2022. In the first, from 2023 into 2024, capex rose because hyperscalers bought every NVIDIA H100 they could get. The binding constraint was silicon supply, specifically TSMC CoWoS packaging and HBM memory. In the second, during 2025, spending roughly doubled again as Blackwell systems shipped and the industry moved from air-cooled 30–40 kW racks to liquid-cooled racks above 100 kW. In the third, which began in 2026, capex keeps climbing even though chips are easier to get. The limit is now energized floor space: grid interconnects, transformers, and the months it takes to turn a shell into a powered, cooled hall.

The incumbents are the same five companies, but their positions differ. Microsoft, Amazon and Alphabet run public clouds and rent capacity to others, including the frontier labs. Meta builds almost entirely for its own ranking, recommendation and generative AI workloads. Oracle is the outlier, a smaller company that has signed enormous multi-year AI capacity contracts and is building to fulfil them. Neoclouds such as CoreWeave, Nebius and Lambda sit around the edges, both competing with the hyperscalers and supplying them with leased capacity.

What changed between May and November is scale and financing. In spring, the market still debated whether 2026 would be a peak year. By October every company that gave forward commentary said 2027 spend would go up again. Alphabet said capex would “increase significantly” in 2027, and Amazon said it would spend “meaningfully more.” That shifts the question from “is this a bubble year” to “what does a permanently higher capital base do to cloud pricing and margins.”

If you want the chip-level view of the same cycle, our NVIDIA Blackwell Ultra Q2 2026 earnings analysis covers the supplier side. For the energy context, the International Energy Agency’s Energy and AI report estimated that data centres used about 415 TWh of electricity in 2024 and could reach roughly 945 TWh by 2030 in its base case, with the United States accounting for close to half of the growth.

Hyperscaler AI Capex by the Numbers: Q3 2026 Earnings

Hyperscaler AI capex for 2026 now totals roughly $810–840 billion across Microsoft (about $170B, calendar 2026), Alphabet ($195–205B), Amazon (about $220B), Meta ($135–145B) and Oracle ($90–95B, fiscal year to May 2027). Every company that commented on 2027 said spending will rise again, and backlogs exceed $2.4 trillion.

Hyperscaler AI capex 2026 capital flow from cash, debt and leases into short-lived and long-lived assets, capacity and backlog

Figure 1: How hyperscaler AI capex flows from funding sources through asset classes and depreciation into usable capacity, revenue and backlog.

The diagram traces one loop. Operating cash flow, debt and equity raises, and lease or joint-venture structures fund the capex line. That spend splits into short-lived assets (accelerators, CPUs, network gear), depreciated over five to six years, and long-lived assets (land, shells, electrical plant), depreciated over seven to forty years. Neither becomes usable capacity until grid power arrives. Capacity turns into cloud revenue and contracted backlog, which then feeds operating cash flow. The loop breaks if any link stalls: power delays, low utilization, or backlog that does not convert.

The five companies, line by line

The table below uses only company disclosures from the most recent quarterly report. Dates are the release dates.

Company Report (date) Latest quarter capex Full-year guidance 2027 signal Contracted backlog
Microsoft Q1 FY27, Oct 28, 2026 $45.2B incl. finance leases ($39.4B cash PP&E) About $170B for calendar 2026; Dec quarter above $50B Not quantified Commercial RPO $677B, up 73%
Alphabet Q3 2026, Oct 28, 2026 $51.4B $195–205B, unchanged “Increase significantly” Google Cloud backlog $561B
Amazon Q3 2026, Oct 29, 2026 $77.3B purchases of property and equipment About $220B cash capex “Meaningfully more” AWS backlog $578B
Meta Q3 2026, Oct 21, 2026 $34.91B incl. finance lease principal $135–145B, narrowed from $130–145B Deferred to Q4 call Not applicable
Oracle Q1 FY27, Sept 10, 2026 Reported about $28.5B gross $90–95B for FY27, no more than $70B net cash FY27 is the guide RPO reported at about $664B

Adding the midpoints gives about $822 billion. Treat that as an order-of-magnitude figure, not an audited total. The companies define capex differently: Microsoft includes finance leases, Amazon reports cash purchases, Meta adds finance lease principal, and Oracle reports on a June-to-May fiscal year. Some of the spend also covers non-AI workloads. Even with those caveats, the combined number is more than twice the spring consensus, and the old version of this post understated it by more than half.

Microsoft: demand above supply, capex below plan

Microsoft’s fiscal Q1 2027 results showed revenue of $92.3 billion, up 19%. Microsoft Cloud reached $63.9 billion, and Azure grew 48% in constant currency, ahead of the company’s own guidance. CFO Amy Hood guided Azure to 49–50% growth for the December quarter and said the company still has “demand that exceeds supply.”

The interesting detail is that capex came in below plan. Quarterly capex of $45.2 billion, including finance leases, missed the earlier outlook of more than $50 billion because some third-party capacity deliveries slipped. Microsoft now points to about $170 billion for calendar 2026, down from higher figures earlier in the year, partly because of a shift from finance leases to operating leases. The lesson for buyers is direct: even the company with the most cash cannot place capacity as fast as it wants. The constraint is delivery, not willingness to spend.

Alphabet: the cloud that doubled

Alphabet reported Q3 revenue of $127.6 billion, up 25%. Google Cloud revenue rose 103% to $30.7 billion, with operating income of $11.1 billion and a 36.2% margin. Part of that jump came from TPU system sales, where customers buy TPU hardware rather than renting it. CFO Anat Ashkenazi said nearly 100 customers now use TPU systems and that more TPU system revenue is expected in 2027.

Q3 capex was $51.4 billion. Alphabet said about 65% went to servers and networking, depreciated over six years, and about 35% to data centers, with useful lives of seven to forty years. Guidance for 2026 stayed at $195–205 billion. The stock reportedly fell about 10% the next day, largely because investors wanted the 2027 figure and got only “increase significantly.” Management also warned that third-party compute capacity, which began ramping at the end of Q3, will pressure Cloud margins in Q4. Even Google, with its own chips, is renting capacity from others to meet demand.

Amazon: the largest absolute spender

Amazon’s Q3 net sales were $201.9 billion, up 12%. AWS revenue rose 39% to $46.0 billion, its fastest growth in nearly five years, and AWS operating income reached $17.8 billion. Purchases of property and equipment were $77.3 billion in the quarter and $215.2 billion over the trailing twelve months. As a result, trailing free cash flow was negative $42.4 billion. Amazon kept its 2026 cash capex guide at about $220 billion.

On the call, CEO Andy Jassy said AWS backlog stood at $578 billion and that AWS is “installing an AI accelerator about every 14 seconds.” He said revenue commitments for Trainium and Graviton had each risen more than 20x year over year, and that Amazon intends to make Trainium available outside AWS late in 2027. He also said Amazon expects to use Trainium chips for five to six years, which matters for the depreciation discussion below.

Meta: the internal-only spender

Meta reported Q3 revenue of $64.60 billion, up 26%. Capex including finance lease principal was $34.91 billion, and free cash flow was negative $2.12 billion. Meta narrowed its 2026 capex range to $135–145 billion and set total expenses at $175–179 billion. Headcount fell 15% year over year to 66,775, so the company is shifting spend from people to compute.

Meta does not sell cloud capacity, so all of its AI capex must pay back through advertising revenue, engagement and new products. It also uses off-balance-sheet structures heavily. The Hyperion campus in Louisiana is held in a joint venture with funds managed by Blue Owl, where Meta owns a minority stake and leases the capacity back. Those structures keep some spend out of the capex line while still committing Meta to long-dated payments.

Oracle: the leveraged builder

Oracle reports on a fiscal year ending in May, so its latest figures come from fiscal Q1 2027, reported on September 10, 2026. Revenue was $19.35 billion. Oracle guided FY27 capex to $90–95 billion with no more than $70 billion in net cash capex. The gap is meant to be covered by supplier financing, customer prepayments and “bring your own hardware” deals, where a customer funds the chips.

Oracle said it delivered 850 MW of AI capacity and more than 300,000 GPUs since the end of its fiscal fourth quarter. Its RPO is reported at about $664 billion, heavily concentrated in a few AI customers. Oracle has also raised substantial debt and equity during 2026. Of the five, it is the purest example of capex funded by the balance sheet rather than by cash flow.

Compute Economics: GPU TCO, Depreciation and Utilization

Headline hyperscaler AI capex tells you how much was spent. It does not tell you whether a GPU-hour sold at today’s prices earns a return. That depends on three variables: the all-in capital cost per accelerator, the useful life used to depreciate it, and the share of hours it actually does billable work.

GPU total cost of ownership model showing capex, depreciation, power and operations feeding cost per GPU hour and price per token, gated by utilization

Figure 2: A simplified GPU total cost of ownership model. Utilization is the gate that turns a fixed hourly cost into the effective cost of useful work.

The model starts with the accelerator server price, adds a share of networking and facility capex, and spreads that across the chosen useful life. Power at the facility’s PUE (power usage effectiveness) and operating costs add to the hourly figure. Utilization then divides the result: the same hardware is cheap at 90% busy and expensive at 50%. That effective cost per GPU-hour is what eventually shows up in the price per million tokens.

Depreciation schedules are an assumption, not a fact

Every hyperscaler depreciates servers on a straight-line schedule, and the chosen life has a large effect on reported profit. Microsoft and Alphabet use six years for servers and network equipment. Meta moved to 5.5 years at the start of 2025. Amazon went the other way in early 2025, shortening the life of a subset of servers and network gear from six to five years because of the faster pace of AI technology change.

The debate is whether a GPU bought in 2026 can earn money for six years when NVIDIA ships a new generation roughly every year. There are two honest answers. First, older accelerators keep finding work: H100s still rent for several dollars per hour in late 2026, and inference, fine-tuning and batch workloads do not need the newest chip. Second, the economic value of a GPU-hour falls as newer chips deliver more tokens per watt. A six-year life can be correct for accounting and still overstate the margin in years four to six, because the asset is earning less per hour than when it was new.

Jassy’s comment that Amazon expects to use Trainium for five to six years supports the long-life view for custom silicon, where the owner controls the software stack and can keep an older chip fully loaded. Merchant GPUs rented to third parties face more price competition as they age.

Worked example: what useful life and utilization do to cost

The following numbers are illustrative, chosen to show sensitivity. They are not any company’s disclosed cost. Assume an all-in capital cost of $50,000 per accelerator, including its share of server, network and facility. Assume 1.7 kW per accelerator at the facility meter, including cooling overhead, at $0.08 per kWh. Add $0.20 per hour for operations, spares and software. Power is treated as a fixed hourly cost because contracted capacity is paid for whether or not it is busy. Cost of capital is excluded, which flatters every row.

Useful life Depreciation per hour at 100% Effective cost at 90% utilization At 70% At 50%
4 years $1.43 $1.96 $2.52 $3.53
5 years $1.14 $1.64 $2.11 $2.96
6 years $0.95 $1.43 $1.84 $2.57

Two conclusions follow. Moving from a six-year to a four-year life raises the effective cost by about 37% at 90% utilization. Dropping from 90% to 50% utilization raises it by about 80% at any life. In other words, utilization matters more than the accounting life, and both matter more than small differences in chip price.

Compare that with what the market charges. Our November 2026 cloud GPU pricing survey found on-demand H100 prices of roughly $4.39–6.16 per GPU-hour at neoclouds and $6.88–12.29 at hyperscalers, with the median H100 rate up about 31% year over year. Those prices sit well above the illustrative cost floor, which explains why hyperscalers keep building. The margin disappears only if prices fall sharply, utilization drops, or useful life turns out much shorter than the books assume.

Utilization is the real competitive weapon

Utilization explains why the five companies behave differently. Meta can keep internal clusters busy because it controls the workload queue: training jobs, ranking models and inference all come from one owner. Oracle said its GPU utilization reached 97.9% in the quarter, which is consistent with selling large blocks to a few anchor tenants on long contracts. Public clouds face more fragmented demand and must hold some headroom for on-demand customers, which is one reason on-demand prices stay high.

Custom silicon adds a second lever. TPUs and Trainium let Alphabet and Amazon keep their own software stack and pack internal and external workloads together. Alphabet’s Q3 commentary that newer TPU generations offer up to 80% better inference price performance versus the prior generation is a claim about cost per token, not peak FLOPS. That is the metric that matters once the hardware is depreciating.

For a practitioner view of how per-hour cost becomes per-token margin, see our guide to LLM token economics and gross margin of AI features.

Where Hyperscaler AI Capex Goes and How It Is Financed

The capex line mixes assets with very different lives and risks. Alphabet’s disclosure that about 65% of Q3 capex went to servers and networking and 35% to data centers is the clearest public split. Microsoft has described a similar majority share going to short-lived assets such as CPUs and GPUs. The short-lived share is what drives the depreciation wave over the next five to six years.

The depreciation wave

Hyperscaler AI capex hits the income statement slowly. A dollar spent on GPUs in 2026 shows up as roughly one-sixth of a dollar of depreciation per year through 2031 or 2032. Because spending has roughly doubled each year since 2023, depreciation lags capex by several years and is now accelerating. Alphabet said higher infrastructure investment will raise depreciation and data center operating costs, and its other cost of revenues rose 32% in Q3. Meta’s total-expense guide of $175–179 billion reflects the same pressure.

This creates a predictable squeeze. Even if capex flattened in 2028, depreciation would keep rising until the 2026–2027 cohorts finished their schedules. Cloud margins can only hold if revenue grows at least as fast as depreciation, which is why every company emphasised backlog.

Backlog as collateral

Combined, Microsoft’s commercial RPO ($677B), Google Cloud’s backlog ($561B), AWS’s backlog ($578B) and Oracle’s reported RPO (about $664B) come to roughly $2.5 trillion. That backlog is the main justification for the capex. But conversion timing varies. Microsoft said about 30% of its commercial RPO converts within 12 months. Alphabet said just over 55% of its backlog converts within 24 months. Oracle’s backlog is longer-dated and more concentrated.

Concentration is the hidden risk. A meaningful share of these backlogs reportedly comes from a small group of frontier AI labs. Microsoft warned that large OpenAI contracts make bookings and RPO growth volatile from quarter to quarter. If one large customer slows, the backlog shrinks faster than the depreciation schedule does.

Funding outside cash flow

Through 2024, the hyperscalers funded capex from operating cash flow and still bought back stock. In 2026 that stopped being true for some of them. Amazon’s trailing free cash flow is negative $42.4 billion. Meta’s Q3 free cash flow was negative. Oracle reported negative free cash flow and has raised debt and equity. Off-balance-sheet structures, such as Meta’s joint ventures and the long-dated leases that Microsoft and Oracle sign with third-party developers, add commitments that do not appear in the capex line.

The practical effect is that the AI buildout now depends partly on credit markets. That does not make it fragile on its own, because these are mostly highly rated borrowers. It does mean that higher interest rates or wider credit spreads would raise the true cost of every GPU-hour, a cost our illustrative table leaves out.

Power: The Binding Constraint on AI Capacity

For hyperscaler AI capex, money and chips are no longer the scarcest inputs. Energized capacity is. Microsoft attributed its capex shortfall to slipped third-party deliveries. Alphabet called itself supply constrained and is renting third-party compute. Amazon described customers reserving “up to a gigawatt of capacity.” Oracle reported its progress in megawatts. When companies start describing their business in gigawatts, power has become the unit of competition.

AI data center power constraints from site selection through interconnection queues, substations, behind-the-meter generation and PPAs to energized liquid-cooled halls

Figure 3: The critical path from a chosen site to revenue-producing racks. Grid interconnection, transformer lead times and local permitting usually set the schedule, not chip delivery.

The diagram shows why. A site must clear utility interconnection studies before it can receive large amounts of power. That leads to new transmission and substation work, which depends on transformer and switchgear supply. In parallel, operators add behind-the-meter generation, such as gas turbines or fuel cells, and sign long-term power purchase agreements. Only after the campus is energized can liquid-cooled halls with racks above 100 kW go live and start earning revenue.

Why power, not chips

Three factors make power the bottleneck. First, rack density has risen from around 30–40 kW for air-cooled racks a few years ago to well above 100 kW for current liquid-cooled Blackwell-class systems, with future generations heading higher. A single gigawatt campus is now a normal planning unit. Second, grid interconnection is slow. Utilities must study each large load, and transmission upgrades can take years. Third, electrical equipment such as large power transformers has long lead times, so even approved projects wait for hardware.

The IEA’s estimate of roughly 945 TWh of data centre demand by 2030 implies more than doubling from 2024. In the United States, data centres are expected to account for close to half of electricity demand growth through 2030 in that analysis. That is why hyperscalers are signing nuclear agreements, funding gas generation and exploring higher-voltage DC power distribution inside the hall. We cover those designs in our analysis of behind-the-meter power and nuclear PPAs for AI data centers and in our explainer on 800V DC AI data center power architecture.

What power scarcity does to economics

Power scarcity changes the economics in a way that favours incumbents. A company that already holds energized sites can deploy new chips quickly and earn returns sooner. A company waiting for an interconnect pays depreciation-equivalent carrying costs on idle equipment or leases. It also raises the value of efficiency: if a site is capped at a fixed number of megawatts, the only way to grow revenue there is more tokens per watt. That favours newer accelerators, liquid cooling and custom silicon, and it shortens the useful economic life of older, less efficient chips at power-constrained sites.

Power also brings reliability risk. Larger campuses concentrate more capacity behind fewer substations, and a single electrical failure can take out a significant share of a region. The October 2026 incident covered in our analysis of the Google Cloud europe-west2-b power outage is a reminder that zone design still matters when every rack is drawing more power.

Trade-offs, Gotchas, and What Goes Wrong

The bull case for hyperscaler AI capex rests on three assumptions: demand keeps outrunning supply, backlog converts on schedule, and useful lives hold. Each can fail.

Demand risk. If model efficiency improves faster than usage grows, the same workloads need fewer GPUs. Efficiency gains have so far raised usage rather than cut it, as cheaper tokens unlock new applications. That pattern is not guaranteed to continue, especially for training clusters built for frontier runs that may not repeat at the same scale.

Concentration risk. Much of the $2.5 trillion backlog is concentrated in a few AI labs that are themselves funded by private capital. If funding tightens for those labs, contracts can be renegotiated, delayed or restructured. Microsoft’s warning about RPO volatility is the polite version of this risk.

Depreciation risk. If the economic life of a 2026 GPU is closer to four years than six, reported margins overstate true returns. The accounting would eventually catch up through impairments or shorter schedules, as Amazon partly did in 2025.

Delivery risk. Microsoft’s Q3 miss showed that third-party data center deliveries slip. When they do, capex falls short of guidance while depreciation from earlier cohorts keeps rising, which squeezes margins from both directions.

Cost-inflation risk. Memory prices, transformer prices and construction labour all rose during 2026. Microsoft reportedly attributed a portion of its 2026 spend to higher component pricing. Some of the headline capex growth is inflation, not added capacity.

Measurement gotchas. Comparing companies on headline capex is risky. Microsoft includes finance leases, Amazon reports cash purchases, Meta adds finance lease principal, and Oracle reports gross and net figures on a different fiscal year. Off-balance-sheet leases and joint ventures are excluded everywhere. Always read the definition before comparing two numbers.

For buyers, the main anti-pattern is assuming that hyperscaler overbuild will soon make GPUs cheap. Prices rose in 2026, not fell. Plan for scarcity through at least 2027 and treat any price drop as upside.

What Hyperscaler AI Capex Means for Cloud Buyers

Cloud buyers sit at the end of this capital chain. The hyperscaler AI capex numbers tell you three things about the next 18 months: capacity will stay tight in popular regions, providers will push long commitments that de-risk their backlog, and price competition will come mainly from custom silicon and neoclouds rather than from hyperscaler discounts.

Decision flow for cloud buyers choosing between spot, custom silicon commitments, hyperscaler reserved GPUs and neocloud contracts

Figure 4: A buyer’s decision flow for AI capacity in 2027, from workload shape to procurement model, with a FinOps review as the common end point.

The flow starts with workload shape. Bursty or batch work should stay on on-demand or spot capacity and provider batch APIs, where the provider absorbs idle time. Steady 24×7 demand justifies a commitment. If the workload can run outside CUDA, TPU or Trainium commitments often offer the best cost per token. If it cannot, the choice is between a hyperscaler reservation, which is worth paying for when your data and services already live there, and a neocloud contract, which is usually cheaper but needs careful exit terms. Every path ends in the same FinOps review of utilization and egress.

Commitments are the new currency

Providers with $500 billion-plus backlogs prefer customers who commit. Expect reserved-capacity programs, capacity blocks and multi-year agreements to get the best allocation. That suits buyers with predictable load but punishes those who over-commit. The worked example above applies to you as well: a committed GPU at 50% utilization costs you nearly twice as much per useful hour as one at 90%.

Portability is worth real money

Custom accelerators are where providers have the most pricing freedom, because they own the full cost stack. Workloads that run on JAX, PyTorch with XLA, or AWS Neuron can access that capacity. Teams locked into CUDA-only kernels have fewer options and weaker negotiating power. Our comparison of ROCm vs CUDA on AMD Instinct covers a third option for teams willing to port.

Region and power-aware planning

Because power is the constraint, capacity appears unevenly. New capacity lands where energized campuses exist, which is not always where your users or data are. Plan for the possibility that your preferred region has a waitlist, and price in cross-region data transfer if you train or batch-process elsewhere.

Practical Recommendations

Treat AI capacity as a constrained commodity through 2027 and procure it the way manufacturers buy scarce components: forecast, commit to a baseline, and keep flexible options for the peaks. The aim is to pay for useful work, not reserved hours.

Start by measuring utilization honestly. Many teams reserve GPUs and then run them at 30–50% because of poor batching, idle development clusters or oversized instances. Raising utilization usually saves more than negotiating a lower rate. Next, separate workloads by shape. Steady inference belongs on committed capacity, while training experiments and batch jobs fit spot capacity or batch APIs.

Then test portability. Run a representative model on at least one custom accelerator and one non-NVIDIA GPU, even if you do not switch. The benchmark itself gives you leverage in negotiation. Finally, read the contract like a lender. Long commitments to a provider that is itself leveraged, or to a neocloud with concentrated customers, carry counterparty risk. Ask about capacity delivery dates, service credits for late delivery, and rights to move commitments between regions or chip generations.

  • Measure GPU utilization per team and set a target above 70% for committed capacity.
  • Split workloads into steady, bursty and batch, and procure each differently.
  • Benchmark at least one TPU, Trainium or AMD path before signing a multi-year deal.
  • Negotiate delivery dates, late-delivery credits and generation-swap rights.
  • Budget for flat or rising GPU prices in 2027, not falling ones.
  • Track egress and cross-region transfer, which grow as capacity spreads to power-rich regions.

This article analyses infrastructure strategy and public disclosures. It is not investment advice.

Frequently Asked Questions

How much are hyperscalers spending on AI capex in 2026?

Based on guidance from Q3 2026 earnings, the five largest spenders plan roughly $810–840 billion. Amazon guided to about $220 billion in cash capex, Alphabet to $195–205 billion, Microsoft to about $170 billion for calendar 2026, Meta to $135–145 billion, and Oracle to $90–95 billion for its fiscal year ending May 2027. Definitions differ, since some figures include finance leases and Oracle uses a different fiscal year, so treat the sum as an approximation rather than a precise total.

Will hyperscaler capex keep growing in 2027?

Every company that commented said yes. Alphabet said capex will “increase significantly” in 2027, and Amazon said it will spend “meaningfully more.” Microsoft and Meta deferred detailed 2027 figures, though Microsoft still reports demand above supply. Specific 2027 numbers are expected with Q4 results in early 2027. The main things that could slow growth are power delivery delays, weaker backlog conversion, or tighter credit for large AI customers.

How long do GPUs last for depreciation purposes?

Microsoft and Alphabet depreciate servers and network equipment over six years. Meta uses 5.5 years, and Amazon shortened a subset of servers to five years in 2025. Physically, many GPUs keep working for six years, and Amazon says it expects to use Trainium for five to six years. Economically, the value of each GPU-hour falls as newer chips deliver more output per watt, so later years usually earn lower margins than the straight-line schedule implies.

Why is power the main constraint on AI data centers?

Modern liquid-cooled AI racks draw well over 100 kW each, so new campuses need hundreds of megawatts to a gigawatt or more. Connecting that load requires utility studies, transmission upgrades and large transformers with long lead times. The IEA projects global data centre electricity use could roughly double to about 945 TWh by 2030. Chips can be bought in months, but energized sites take years, so power now sets the pace of capacity growth.

Will the AI capex boom make cloud GPUs cheaper?

Not soon. On-demand H100 prices rose about 31% year over year by November 2026, according to our pricing survey, and providers report demand above supply. Prices may fall for older chips as newer generations arrive, and custom silicon such as TPU and Trainium offers better cost per token for compatible workloads. Buyers should plan for flat or rising prices on current-generation GPUs through 2027.

Is hyperscaler AI capex a bubble?

There are real risks: negative free cash flow at some companies, heavy backlog concentration in a few AI labs, and depreciation lives that may be optimistic. Against that, cloud revenue is accelerating, with Azure up 48%, AWS up 39% and Google Cloud up 103%, and backlogs exceed $2.4 trillion. A fairer framing is that the industry has made a large, leveraged bet that demand keeps outrunning efficiency gains, which has held so far.

Further Reading

References

  1. Amazon.com, Inc., “Amazon.com Announces Third Quarter Results,” October 29, 2026. ir.aboutamazon.com
  2. Amazon Q3 2026 earnings call, October 29, 2026 (transcript via The Motley Fool). fool.com
  3. Meta Platforms, “Meta Reports Third Quarter 2026 Results,” October 21, 2026. investor.atmeta.com
  4. Alphabet Inc., Q3 2026 results and earnings call, October 28, 2026. abc.xyz/investor; call summary via MarketBeat on Yahoo Finance
  5. Microsoft Corporation, FY27 Q1 results and earnings call, October 28, 2026. microsoft.com/investor; coverage via Fortune
  6. Oracle Corporation, fiscal 2027 Q1 results and earnings call, September 10, 2026. investor.oracle.com; call summary via Yahoo Finance
  7. International Energy Agency, Energy and AI, April 2025. iea.org

By Riju — about

1 Comment

Leave a Reply

Your email address will not be published. Required fields are marked *