Is the AI Infrastructure Market growing now?
In our AI infrastructure market deck, you will find everything you need to understand the market
SUMMARY
Yes. The AI Infrastructure Market is growing very fast right now, with hyperscaler spending, cloud revenue, chip sales, contracted capacity, data-center absorption and electricity demand all moving higher at the same time.
The strongest evidence is not one giant capex number. It is that several parts of the stack that are measured independently are expanding together, which makes the current buildout much harder to dismiss as a spending bubble with no customer demand behind it.
Hyperscalers are still raising or sustaining infrastructure budgets at extraordinary levels. Amazon now expects roughly $220 billion of cash capex this year, Alphabet is guiding to $195 billion-$205 billion, Meta to $130 billion-$145 billion, and Microsoft says an accounting change lowered its reported capex figure without reducing the underlying investment plan.
The revenue side is keeping up better than many skeptics expected. AWS is growing 37%, Azure 43%, and Google Cloud 82%, while cloud backlogs at Google and Microsoft have climbed into the hundreds of billions of dollars.
The hardware cycle is also getting broader. Nvidia remains central, but AMD, Broadcom, Arista, custom accelerators, CPUs and AI networking are all growing quickly, so the market is no longer just a story about buying more GPUs from one supplier.
Neoclouds have crossed an important threshold. CoreWeave and Nebius are now converting large amounts of power and contracted GPU capacity into billions of dollars of revenue or annualized run-rate revenue, which makes them real infrastructure businesses rather than just financing structures built around scarce chips.
Physical supply is still tight despite the construction boom. North American data-center inventory grew 33% year over year in CBRE's latest study, yet vacancy still fell, including to just 0.3% in Northern Virginia.
The workload mix is shifting toward inference. That is important because training demand arrives in giant waves, while inference happens continuously whenever users, applications and agents call models, making AI infrastructure look more like a recurring utility workload.
Cheaper AI has not reduced infrastructure spending so far. Lower cost per token is being overwhelmed by much higher usage, more production deployments and heavier agentic workflows, which is a classic rebound effect in computing.
The clearest constraint is moving from chips toward electricity. Data-center projects are increasingly discussed in megawatts and gigawatts because accelerators can be ordered faster than transmission lines, substations and generation capacity can be built.
The main weakness is concentration. A small group of hyperscalers and frontier AI companies still controls a huge share of marginal demand, while neocloud financing increasingly depends on long contracts, customer prepayments and debt secured against infrastructure.
So the growth itself is hard to dispute. The more difficult question is whether usage and AI revenue can keep rising fast enough to justify the hundreds of billions of dollars being committed before the full economics of the buildout are known.
This market map, featured in our AI infrastructure market deck, highlights top companies and startups in the AI infrastructure market
What do we actually mean by the AI infrastructure market?
For this article, the AI infrastructure market means the compute stack that trains and runs AI: accelerators, servers, networking, cloud capacity, data centers, cooling and the power systems directly needed to operate them.
We exclude AI applications, consulting and software subscriptions unless the spending directly buys compute infrastructure. That keeps us from calling every dollar linked to AI an infrastructure dollar.
We also avoid adding every layer together as though they were separate markets. When Microsoft buys Nvidia systems, installs them in an Azure data center and rents the capacity to a customer, Nvidia revenue, Microsoft capex and the customer's Azure bill describe different parts of the same economic chain.
So we are looking for several independent measures moving in the same direction: infrastructure spending, hardware sales, cloud consumption, contracted capacity, data-center absorption and power demand. Right now, they are.
Is AI infrastructure spending still accelerating right now?
Yes. AI infrastructure spending is still accelerating right now, with Amazon, Alphabet, Meta and Microsoft all operating at spending levels that would have looked extreme even a year ago.
Amazon now expects roughly $220 billion of cash capital expenditure this year, up from its previous estimate of about $200 billion. Andy Jassy said during Amazon's latest earnings call that higher memory costs contributed to the increase, but he also said the company still expects to have less capacity than customers want.
Alphabet has raised its own full-year capex range to $195 billion to $205 billion after spending $44.9 billion in its latest quarter alone. That quarterly figure was more than double the level from one year earlier. Alphabet says the increase is mainly about bringing compute capacity online faster.
Meta spent $31.1 billion in its latest quarter and now expects $130 billion to $145 billion for the year. Its guidance was already raised earlier this year before being narrowed around that higher range.
Microsoft spent $41 billion in its latest quarter. About two-thirds went into shorter-lived assets such as CPUs and GPUs. Microsoft recently reduced its reported calendar-year capex expectation from roughly $190 billion to about $175 billion, but the company explained that the change came from how certain data-center leases are classified. Management said the underlying investment plan had not been reduced.
These figures use slightly different accounting definitions, so adding them into one supposedly precise market number would be misleading. The direction is much easier to read: every major hyperscaler is spending at an unusually high level, and several have increased their plans during the year.
| Company | Current infrastructure spending indication | What changed recently |
|---|---|---|
| Amazon | ~$220B cash capex | Raised from about $200B |
| Alphabet | $195B-$205B capex | Raised from $180B-$190B |
| Meta | $130B-$145B capex | Higher than initial 2026 guidance |
| Microsoft | ~$175B reported calendar-year capex | Accounting change lowered reported figure; underlying investment plan unchanged |
As this chart shows, and as featured in our AI infrastructure market deck, search interest in AI infrastructure has risen sharply
Is all that AI infrastructure spending actually producing revenue?
Yes. Cloud revenue is currently rising fast enough that the AI infrastructure buildout already has a large commercial business behind it.
AWS generated $42.2 billion in its latest quarter, up 37% year over year. That was its fastest growth rate in 18 quarters, despite AWS already operating at a $169 billion annualized revenue run rate. Amazon also says its AI business inside AWS has passed a $25 billion annual revenue run rate and is still growing at triple-digit percentages.
Azure and other cloud services grew 43% in Microsoft's latest quarter. Azure also passed $100 billion of annual revenue for the first time.
Google Cloud is growing even faster. Its latest quarterly revenue reached $24.8 billion, up 82% from one year earlier, compared with 63% growth in the previous quarter. Google said AI infrastructure and enterprise AI products were major contributors. Cloud operating income reached $8.8 billion, giving the division a 35.6% operating margin.
Google Cloud's backlog also climbed to $514 billion, up by more than $50 billion in one quarter. According to Alphabet, a little over half of that backlog should become revenue within the next two years.
We therefore have spending growth and customer revenue growth happening at the same time. Some individual projects can still turn out badly, but describing today's buildout as infrastructure being constructed without customers no longer fits the numbers.
Are AI chips still growing this fast?
Yes. AI hardware sales are still growing extremely fast today, although the growth is spreading across more products than GPUs alone.
As of now, Nvidia's latest reported Data Center revenue is $75.2 billion for one quarter, up 92% year over year and 21% from the previous quarter. Nvidia's next quarterly results are due later today, so we are using the newest confirmed figure rather than an estimate.
AMD's latest Data Center revenue reached $6.7 billion, up 107% year over year and 16% sequentially. Its Data Center operating income reached $2.1 billion, compared with a loss one year earlier.
Broadcom provides an important third datapoint because its exposure is different. Its latest AI semiconductor revenue reached $10.8 billion, up 143%. Networking accounted for almost 40% of Broadcom's AI revenue, and management said it had received more than $30 billion of AI semiconductor bookings during the quarter while shipping $10.8 billion.
Arista Networks has also just reported its first quarter above $3 billion of revenue, up 37.7%, while rolling out 1.6 Tbps networking systems for larger AI clusters.
The current hardware cycle is wider than a single accelerator company. GPUs remain huge, but networking, CPUs, custom accelerators and complete rack systems are growing alongside them.
| Company / segment | Latest quarterly revenue | YoY growth |
|---|---|---|
| Nvidia Data Center | $75.2B | +92% |
| AMD Data Center | $6.7B | +107% |
| Broadcom AI semiconductors | $10.8B | +143% |
| Arista Networks total revenue | $3.0B | +37.7% |
This chart, included in our AI infrastructure market deck, shows annual VC investment in AI infrastructure startups
Is the AI infrastructure market starting to move beyond Nvidia?
Yes. Nvidia still sits at the center of AI infrastructure, but growth is now spreading quickly into custom chips, AMD systems and AI networking.
Broadcom's latest results are probably the clearest evidence. Its AI semiconductor revenue rose 143%, and the company expects that business to reach about $16 billion in its current quarter, more than three times the level from one year earlier. Custom accelerators and networking are driving the increase.
AMD is gaining another part of the workload. Anthropic has agreed to deploy up to 2 GW of AMD Instinct MI450 systems, while Microsoft plans to deploy AMD's Helios rack-scale platform and next-generation EPYC processors across Azure.
Amazon's own chips have also become a major business. The company says revenue from Trainium, Graviton and Nitro has passed a $25 billion annual run rate, with triple-digit growth. Anthropic has committed to as much as 5 GW of Trainium capacity, while OpenAI has committed to roughly 2 GW.
Google has gone one step further and started selling TPU systems for installation inside customer data centers. Alphabet began recognizing revenue from those external TPU sales in its latest reported quarter.
AI compute is becoming a multi-architecture market. That could gradually reduce Nvidia's share of each incremental dollar without reducing the amount of infrastructure being built. Cheaper and more specialized hardware can actually make more AI workloads economical to run.
If you want more recent data on this point, please see our latest AI infrastructure market report.
Are CoreWeave and Nebius becoming real AI infrastructure businesses?
Yes. CoreWeave and Nebius are now producing enough revenue, contracted capacity and cash generation to treat neoclouds as a real layer of the AI infrastructure market.
CoreWeave generated $2.58 billion of revenue in its latest quarter, up 112% year over year. Its revenue backlog reached about $104 billion, and the company said it signed more than $25 billion of additional net customer commitments shortly after the quarter ended.
The physical buildout is moving just as quickly. CoreWeave added almost 500 MW of active power during the quarter, reaching roughly 1.5 GW, while total contracted power rose to around 3.7 GW.
Nebius is smaller but currently growing faster. Its AI cloud generated $575 million in its latest quarter, up 514% year over year. Annualized run-rate revenue reached $3 billion, compared with $1.9 billion only one quarter earlier.
Nebius also reported a 50% adjusted EBITDA margin for its AI cloud business. Management said it continues to sell capacity almost as quickly as it brings that capacity online.
A year ago, it was reasonable to wonder whether neoclouds would remain mostly financing vehicles built around expensive GPU fleets. At least two of them have now crossed into multi-billion-dollar revenue scale or run-rate revenue, with large contracted customer books behind the expansion.
| Company | Latest AI/cloud revenue | Growth | Other useful measure |
|---|---|---|---|
| CoreWeave | $2.58B quarterly | +112% YoY | ~$104B backlog |
| Nebius AI Cloud | $575M quarterly | +514% YoY | $3B annualized run rate |
This chart, included in our AI infrastructure market deck, shows why CoreWeave is winning in AI infrastructure
Is AI infrastructure demand still bigger than available supply?
Yes. Available AI infrastructure is still tight today even after an enormous increase in data-center construction.
CBRE's latest global data-center study gives us a useful physical test. North American inventory across the largest markets increased 33% year over year, yet vacancy still fell.
Northern Virginia added more than 1.1 GW of inventory in one year and still reached a vacancy rate of only 0.3%. Atlanta fell to 1%, Dallas-Fort Worth to 1.8% and Chicago to 2.2%.
Across all 16 major markets followed by CBRE, average vacancy fell from 8.3% to 6.7% despite new supply being delivered. Net absorption in the main U.S. markets rose 34%.
Hyperscalers are also spending aggressively. The physical-market data makes the demand story stronger because rapidly rising construction would normally push vacancy upward if customers were disappearing. Instead, new capacity is being absorbed very quickly.
Amazon's management recently said that even around $220 billion of cash capex will leave the company short of the capacity customers want. Google says demand still exceeds the large amount of capacity it has added over the past three years.
There will eventually be places where developers build too much. We can already see different conditions across individual European and Asian markets. At the market level, however, widespread excess capacity has yet to show up.
Is enterprise AI usage big enough to support all this infrastructure?
Enterprise AI usage is now large enough to contribute meaningfully to infrastructure demand, although the biggest individual contracts still come from a small group of hyperscalers and frontier AI companies.
Amazon's AI revenue inside AWS has crossed a $25 billion annual run rate. Hundreds of thousands of customers now use Amazon Bedrock, and Amazon says customer spending on Bedrock during its latest quarter exceeded the spending from all previous quarters combined.
Google says nearly 90% of the Fortune 100 use Gemini Enterprise. More importantly for infrastructure demand, existing Google Cloud customers are using more than their committed amounts at a rate more than 50% higher than their original commitments, according to Alphabet's latest earnings discussion.
Microsoft gives us another useful usage measure. Azure grew 43% in its latest quarter while Microsoft Cloud's contracted backlog reached $678 billion.
These customers are obviously buying more than generative AI compute, so we should not classify all of those cloud numbers as AI revenue. Still, AI is now showing up inside very large cloud businesses rather than living mainly in experimental budgets.
The market remains top-heavy at the extreme end. OpenAI, Anthropic, Meta, Google, Microsoft and a handful of other companies can each absorb gigawatts of infrastructure. Enterprise adoption is broadening the base underneath those buyers, but it has not replaced them.
This chart, included in our AI infrastructure market deck, shows annual funding in AI infrastructure startups
Is inference becoming the biggest AI infrastructure workload?
Yes. Inference is currently becoming the larger commercial infrastructure workload, which changes the market from occasional giant training jobs toward continuous everyday compute.
Gartner's newest AI infrastructure forecast estimates $23.3 billion of AI-optimized cloud infrastructure spending on inference this year, compared with about $19 billion for training. Inference would therefore represent roughly 55% of AI-optimized IaaS spending.
Gartner expects total AI-optimized IaaS spending to reach about $42.3 billion this year, up 96%, and then rise to roughly $66.1 billion next year.
The difference between training and inference is important for the durability of the market. A model might be trained periodically, while inference happens every time someone asks a model a question, generates code, creates an image or lets an agent complete a multi-step task.
Amazon says enterprises are still early in moving inference into production. Google now processes 22 billion Gemini API tokens per minute across its model ecosystem. Agentic workflows can make the infrastructure requirement even heavier because one user request can trigger several model calls, tool calls and verification steps.
Training created the first giant wave of AI compute demand. Inference is now giving the market a much larger recurring workload.
If you want more recent data on this point, please see our latest AI infrastructure market report.
If AI keeps getting cheaper, why is infrastructure demand still rising?
AI infrastructure demand is still rising because cheaper computation is being consumed in much larger quantities.
The price of producing a useful model response has fallen sharply as chips improve, models get smaller and software becomes better at using each accelerator. Hyperscalers are also building their own chips specifically to lower inference costs.
Normally, that would reduce the amount of hardware needed for a fixed workload. The workload is nowhere near fixed.
Gartner currently expects AI-optimized IaaS spending to almost double this year even while inference economics improve. The same forecast has inference spending growing to $23.3 billion as companies move AI systems into everyday production.
Google's 22 billion Gemini API tokens per minute gives a sense of the scale that cheap inference can create. Amazon says its AI revenue has already passed a $25 billion annual run rate while growing at triple-digit percentages.
We are seeing something close to the classic rebound effect in computing: lower unit costs make more uses affordable, and the resulting increase in usage can overwhelm the efficiency gains.
That relationship will not continue automatically forever. If token growth slows while hardware efficiency keeps improving, infrastructure demand could soften quickly. So far, usage is still growing fast enough to keep the total compute bill moving upward.
This chart, included in our AI infrastructure market deck, compares the main business model options for AI cloud infrastructure providers
Has electricity become the biggest constraint on AI infrastructure growth?
Electricity has become one of the hardest constraints on AI infrastructure growth today, especially for projects measured in hundreds of megawatts or several gigawatts.
The latest U.S. data from Lawrence Berkeley National Laboratory estimates that data centers could consume about 11.8% of total U.S. electricity by 2030 under its reference case. Its scenario range runs from 9.5% to 15.3%.
The reference case works out to roughly 649 TWh of U.S. data-center electricity consumption. That is an enormous change for a sector that historically occupied a much smaller part of national electricity demand.
The IEA reaches a similar conclusion globally. It expects data-center electricity consumption to move from about 485 TWh in 2025 to roughly 950 TWh by 2030. Electricity consumed by AI-focused data centers is expected to triple over that period.
Accelerated servers are the main driver. The IEA expects their electricity consumption to grow about 30% per year through 2030.
That is why AI infrastructure companies increasingly talk about gigawatts, substations, transmission connections and generation agreements alongside GPUs. A company can order accelerators faster than a utility can build a transmission line.
Power availability is also changing where AI infrastructure gets built. Developers are moving toward markets where land and electricity can be secured several years ahead, even when those locations are less convenient than traditional data-center hubs.
Is the AI infrastructure boom actually profitable?
Parts of the AI infrastructure market are already extremely profitable, while other parts still need huge amounts of capital before we know how good the economics really are.
AWS generated $16.6 billion of operating income on $42.2 billion of quarterly revenue, giving it an operating margin of roughly 39%.
Google Cloud reached a 35.6% operating margin in its latest quarter while growing revenue 82%. That is a striking combination for a business now running at almost $100 billion of annualized revenue.
Nebius says its AI cloud reached a 50% adjusted EBITDA margin as utilization improved. Broadcom's economics are stronger again: company-wide adjusted EBITDA reached 69% of revenue, helped by its rapidly growing AI semiconductor business.
CoreWeave shows why we still need to be careful. Revenue doubled, but the company is carrying very large financing costs and infrastructure obligations while continuing to build several gigawatts of capacity.
The same pressure is visible at the hyperscalers. Amazon's trailing twelve-month free cash flow moved to a $7.6 billion outflow after property and equipment purchases increased by $66.1 billion, primarily because of AI investment. Alphabet's latest quarterly free cash flow also turned negative as capex more than doubled year over year.
Profitability depends heavily on where we look. Semiconductor suppliers and mature cloud platforms are already earning large profits. New infrastructure owners have much less room for error because they are financing expensive equipment years before the full revenue arrives.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, featured in our AI infrastructure market deck, shows the share of revenue generated by each customer segment in the AI infrastructure market
Is debt making AI infrastructure growth look healthier than it really is?
Debt is helping AI infrastructure grow faster, and it also raises the cost of being wrong about future demand.
Nebius recently arranged its first $775 million secured debt facility, backed by deployed GPU infrastructure and contracted cash flows. The company says it expects more than $9 billion of customer prepayments during the year.
That combination is becoming common. Infrastructure companies sign a long-term customer contract, use the contract to finance GPUs or a data center, and sometimes collect part of the customer's commitment before the equipment is deployed.
CoreWeave is using the same basic model at a much larger scale. Its $104 billion backlog provides a large base of contracted demand, but the company still has to finance and build the infrastructure needed to serve much of that revenue.
Traditional data-center companies are being pulled into the structure too. Core Scientific now has roughly 590 MW leased to CoreWeave under contracts representing more than $10 billion of potential revenue, with some infrastructure spending funded directly by the customer.
We would describe financing as an amplifier rather than evidence of fake demand. Customers are signing large commitments and sometimes putting cash down. The risk comes from the length of those commitments, the debt used to fulfill them and the possibility that hardware prices or customer needs change faster than expected.
Today's AI infrastructure boom has real customers behind it. It also has a lot more leverage behind it than the chip-revenue numbers alone suggest.
Is the AI infrastructure market dangerously dependent on a few buyers?
Yes. Customer concentration is still one of the clearest weaknesses in the AI infrastructure market right now.
CoreWeave's latest filing shows that its three largest customers represented roughly 36%, 26% and 10% of quarterly revenue. About 72% of its revenue therefore came from three customers.
The concentration becomes even more important when we follow the money through the stack. A data-center operator can lease capacity to CoreWeave, which in turn leases compute to a frontier lab or hyperscaler. A lender then finances the servers against those contracts. Several apparently separate businesses can ultimately depend on the spending decision of the same end customer.
The hyperscaler capex numbers show the same concentration from another direction. Amazon, Microsoft, Alphabet and Meta alone are currently responsible for hundreds of billions of dollars of annual infrastructure investment.
Large customers are becoming more numerous, and enterprise usage is spreading. Even so, today's marginal gigawatt is still much more likely to be paid for by a giant technology company than by thousands of ordinary enterprises.
We would treat the current market as broad in technologies but concentrated in purchasing power. That distinction becomes crucial if even one major buyer slows spending.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows how GPU cloud infrastructure technology has evolved over time
What would actually show that the AI infrastructure boom is slowing?
The AI infrastructure market will look meaningfully weaker when several things deteriorate together: cloud growth, hardware orders, capacity absorption and contract economics.
One delayed data center would tell us very little. Big infrastructure projects move constantly because of permitting, electricity, construction schedules and financing.
Cloud revenue would be harder to dismiss. If AWS, Azure and Google Cloud all started slowing sharply while management stopped talking about capacity constraints, the demand picture would look different.
Physical data centers give us another clean test. CBRE currently sees record-low vacancy in several large U.S. markets despite 33% North American supply growth. A sustained rise in available capacity would tell us that construction had finally caught up with demand.
Hardware orders would probably weaken around the same time. Broadcom currently has AI semiconductor bookings running far above shipments, AMD has doubled Data Center revenue, and Nvidia's latest confirmed Data Center quarter still showed 92% growth. Broad weakness across those suppliers would carry far more weight than a slowdown at one company.
Finally, we would watch the financing terms offered to neoclouds. Smaller customer prepayments, shorter contracts, weaker pricing or difficulty financing GPUs would suggest buyers and lenders were becoming less confident.
None of those changes needs to happen first. We would become much more cautious when several start happening together.
Is the AI Infrastructure Market growing now?
Yes. The AI Infrastructure Market is growing very fast right now, and the newest evidence shows growth across spending, cloud revenue, chips, networking, neoclouds, data-center capacity and electricity demand.
The strongest part of the case is how many independent datasets point in the same direction.
Hyperscalers are still raising infrastructure budgets. AWS has just recorded its fastest growth in 18 quarters. Google Cloud growth has accelerated to 82%. AMD has doubled Data Center revenue. Broadcom's AI semiconductor business has more than doubled again. CoreWeave and Nebius are converting gigawatts of infrastructure into billions of dollars of revenue and contracts.
Physical capacity is also being absorbed quickly. Data-center vacancy has fallen even while developers add large amounts of new supply. At the same time, electricity demand has become large enough for national laboratories and energy agencies to model data centers as a material part of future power consumption.
The shape of the market is changing, though. Training created much of the original spending wave, while inference is now becoming the larger recurring workload. Nvidia remains central, but custom chips, AMD, networking equipment and alternative AI clouds are taking a bigger part of incremental spending.
We are less confident about the eventual return on all this capital than we are about the growth itself. Hundreds of billions of dollars are being committed before the full revenue arrives, customer concentration remains high, and debt is increasingly important to the buildout. A slowdown in usage could therefore hurt infrastructure owners much faster than it hurts companies selling the scarce components today.
For now, though, the conclusion is hard to argue with: the AI Infrastructure Market is still expanding at an exceptional pace. The useful question has moved on from whether growth exists. The real uncertainty is how long demand can keep absorbing infrastructure this quickly before supply, financing costs and AI revenue finally catch up with one another.
If you want more recent data on this point, please see our latest AI infrastructure market report.
In our AI infrastructure market deck, we identify pain points entrepreneurs should prioritize
OUR METHODOLOGY
This analysis tests whether the AI Infrastructure Market is growing now by looking at the compute stack that trains and runs AI: accelerators, servers, networking, cloud capacity, data centers, cooling and the power systems directly needed to operate them.
We exclude AI applications, consulting and software subscriptions unless the spending directly buys compute infrastructure. We also avoid adding Nvidia revenue, hyperscaler capex and customer cloud bills into one market total when they can describe different parts of the same economic chain.
Instead, we break the market into separate analytical dimensions and test them independently. The main ones are hyperscaler investment, cloud revenue and AI usage, semiconductor and networking sales, neocloud revenue and contracted capacity, physical data-center supply and absorption, inference demand, electricity constraints, profitability, financing and customer concentration.
We prioritized the freshest reported evidence we could use: latest-quarter revenue, current capex guidance, bookings, backlog, contracted power, active capacity, vacancy, net absorption, customer commitments and current electricity-demand forecasts. No single datapoint decides the conclusion.
We also separate growth from quality. A market can be expanding very quickly while still carrying serious risks around customer concentration, leverage, financing costs and returns on capital. Those risks affect how durable the boom may be, but they do not erase current growth if spending, revenue and physical demand are still rising.
Key company sources include Amazon's Q2 2026 results, Alphabet's Q2 2026 results, Meta's Q2 2026 results, and Microsoft's FY2026 Q4 results for hyperscaler spending, cloud growth, backlog and AI usage.
For chips and networking, we used Nvidia's fiscal Q1 2027 results, AMD's Q2 2026 results, Broadcom's fiscal Q2 2026 results, and Arista Networks' Q2 2026 results.
For neoclouds, we relied on CoreWeave's Q2 2026 results and Nebius' Q2 2026 results, including revenue, backlog, annualized run-rate revenue, capacity and financing information.
For physical infrastructure and power, the main external sources are CBRE's Global Data Center Trends 2026 for inventory, vacancy and absorption, Gartner's 2026 AI-optimized IaaS forecast for training versus inference spending, Lawrence Berkeley National Laboratory's United States Data Center Energy Usage Report: 2025 Update, and the International Energy Agency's Key Questions on Energy and AI.
The conclusion comes from the accumulated weight of that evidence. We become more confident when independent measures such as spending, cloud revenue, hardware sales, contracted capacity, physical absorption and electricity demand all point in the same direction, and more cautious when several of them begin to weaken together.
This chart, included in our AI infrastructure market deck, shows the share of revenue by region across Europe, Asia, North America, Africa, and South America in the AI infrastructure market
Related blog posts
- AI Infrastructure M&A: what is happening now?
