How do AI infrastructure companies make money?
In our AI infrastructure market deck, you will find everything you need to understand the market
SUMMARY
AI infrastructure companies make money by selling the compute, hardware, powered capacity and software required to train, deploy and operate AI systems.
The market is really a stack of different businesses. NVIDIA can earn hardware-like revenue with software-like margins, while a GPU cloud can post huge EBITDA and still burn cash because depreciation, interest and capex sit underneath the headline number.
The best economics currently appear where a company controls a bottleneck rather than interchangeable capacity. Advanced accelerators, high-speed networking, scarce power and specialized operating software all have more pricing protection than a plain GPU rental.
AI infrastructure demand is increasingly split in two. Hyperscalers and frontier labs sign enormous negotiated commitments, while smaller developers start with a few dollars of serverless usage and may later graduate into reserved capacity or enterprise contracts.
Long-term contracts are doing more than locking in customers. They help infrastructure providers finance new capacity, improve utilization and, when backed by prepayments or customer-owned hardware, transfer part of the capital risk away from the provider.
CoreWeave shows why neocloud economics are both compelling and dangerous. Demand and backlog can be extraordinary, yet the business still has to outrun hardware depreciation, financing costs and the possibility that rental prices fall before the assets are fully paid back.
Data-center economics are increasingly about megawatts and connectivity rather than floor space. When power is scarce and customers are already connected to clouds, networks and partners inside the same facility, renewal pricing and switching costs can become surprisingly strong.
Inference should become a larger recurring infrastructure expense as AI products move into production, especially with agents making repeated model calls. But basic token resale is likely to stay brutally competitive, so platforms are moving toward dedicated capacity, routing, caching, private deployments and guaranteed throughput.
Infrastructure software has a different attraction: it can grow with AI usage without financing the whole physical stack. Observability, retrieval, routing and deployment products may therefore end up with cleaner capital economics than companies that have to keep buying GPUs.
Cheaper compute is unlikely to kill the infrastructure market. More efficient chips and software lower the cost of each unit of AI work, but they also make longer contexts, more users, video, reasoning and agentic workflows economically possible, which can push total consumption higher.
The strongest business models today are the ones attached to the constraint customers keep hitting next. Right now that points to proprietary compute, networking, power, cooling, scarce data-center capacity and software that materially improves how the rest of the stack performs.
This market map, featured in our AI infrastructure market deck, highlights top companies and startups in the AI infrastructure market
What actually counts as an AI infrastructure company?
AI infrastructure companies make money by selling the compute, hardware, data-center capacity and software needed to build and run AI systems.
That definition covers businesses with very different economics. NVIDIA sells accelerators and networking. CoreWeave rents GPU capacity. Digital Realty rents powered data-center space. Vertiv sells power and cooling equipment. Baseten and Together AI charge companies to run models. LangSmith charges for the software used to monitor and operate AI applications.
The useful boundary is straightforward: customers are paying these companies for something required to train, deploy, serve, connect, store or operate AI.
That creates several different ways to make money. Hardware vendors sell systems. Data-center companies collect rent. Clouds charge by GPU-hour or reserved capacity. Inference platforms charge by token, throughput or dedicated GPU. Software companies charge by trace, request, storage, deployment or seat.
Calling all of this “AI infrastructure” only works if we keep those economics separate.
| AI infrastructure layer | What customers buy | How companies charge |
|---|---|---|
| Chips and networking | GPUs, accelerators, switches | Hardware and systems sales |
| Data centers | Power, space, connectivity | Long-term leases and interconnection |
| AI clouds | GPU capacity | GPU-hours and reserved clusters |
| Inference platforms | Model execution | Tokens, GPU-time and guaranteed throughput |
| Infrastructure software | Retrieval, monitoring, routing, deployment | Usage fees, subscriptions and enterprise contracts |
Why is AI infrastructure such a huge business right now?
AI infrastructure is growing so fast because the biggest technology companies are now spending hundreds of billions of dollars a year on the physical systems behind AI.
Alphabet shows how quickly the numbers have moved. Its capital spending was about $31 billion in 2022. After its latest quarterly results, the company raised its 2026 capex guidance to roughly $195 billion to $205 billion, with most of that spending going into technical infrastructure. Its quarterly capex alone reached $44.9 billion.
Amazon is moving at a similar scale, while Meta has also pushed annual infrastructure spending deep into the tens of billions. The combined spending plans of the largest hyperscalers now run into several hundred billion dollars.
That money spreads through a long chain. A new AI cluster creates revenue for chip designers, foundries, memory suppliers, networking vendors, electrical-equipment makers, cooling companies, data-center developers and cloud operators before the eventual AI application earns a dollar.
The clearest confirmation that this spending is turning into business rather than sitting idle comes from cloud revenue. In its latest quarter, Google Cloud revenue jumped 82% year over year to $24.8 billion, helped by enterprise AI infrastructure and AI solutions. AWS grew 37% to $42.2 billion, its fastest growth in 18 quarters.
This is more than a construction boom. The expensive infrastructure being installed today is already producing rapidly growing cloud revenue.
As this chart shows, and as featured in our AI infrastructure market deck, search interest in AI infrastructure has risen sharply
Who actually pays AI infrastructure companies?
The biggest AI infrastructure bills currently come from hyperscalers, frontier AI labs and large enterprises, while smaller developers usually enter through pay-as-you-go services.
CoreWeave makes the concentration especially visible. Its latest reported quarter ended with $104.2 billion of revenue backlog, and the company said it had added more than $25 billion of additional customer commitments shortly after the quarter closed. A handful of customers can therefore generate contracts worth tens of billions of dollars.
At the chip layer, companies such as Meta, Microsoft, Google, Amazon and major AI labs buy huge quantities of compute directly or through partners. At the data-center layer, cloud companies can lease hundreds of megawatts at once. Meanwhile, a startup can open an account with Baseten, Fireworks AI or Together AI and begin spending a few dollars on inference.
These are effectively two markets sharing the same infrastructure. One runs on enormous negotiated contracts; the other runs on thousands of small consumption accounts.
The interesting bit is how one turns into the other. A developer starts with serverless inference, grows into dedicated capacity and eventually negotiates a large enterprise contract. Infrastructure companies increasingly design their pricing around that progression.
How do NVIDIA, Broadcom and AMD make so much money from AI chips?
AI chip companies make exceptional money when their hardware changes the economics of an entire data center, and NVIDIA currently shows just how valuable that position can be.
NVIDIA's latest available quarterly results showed $75.2 billion of Data Center revenue, up 92% year over year. Company gross margin was about 75%. Those are extraordinary economics for a company shipping physical products.
Broadcom shows that the opportunity extends beyond general-purpose GPUs. Its latest reported AI semiconductor revenue reached $10.8 billion, up 143%, as hyperscalers bought more custom accelerators and networking products. Broadcom produced adjusted EBITDA equal to 69% of company revenue and converted 46% of revenue into free cash flow.
AMD is smaller but growing quickly. Its latest Data Center revenue reached $6.7 billion, up 107% year over year, with segment operating income of $2.1 billion. Data Center already represented 58% of AMD's total quarterly revenue.
The common feature is pricing power created by performance. If a better accelerator lets a customer get materially more useful computation from a multibillion-dollar facility, the chip can capture far more value than its manufacturing cost suggests.
Advanced AI chips currently look much more attractive economically than ordinary commodity hardware.
This chart, included in our AI infrastructure market deck, shows annual VC investment in AI infrastructure startups
How do GPU clouds like CoreWeave actually make money?
GPU clouds such as CoreWeave make money by buying large clusters of AI hardware and selling access to that capacity at a higher price than the combined cost of hardware, electricity, facilities, networking, operations and financing.
The simple version is hourly GPU rental. Together AI currently lists H100 cluster capacity at $3.99 per GPU-hour on demand, with lower prices for longer reservations. CoreWeave operates at a much larger enterprise-contract scale, where entire clusters can be reserved for years.
CoreWeave's latest quarter shows how large this business has become. Revenue reached $2.58 billion, up 112% year over year, while active power capacity reached roughly 1.5 gigawatts. Management said near-term capacity was effectively sold out.
The company is also trying to make more money from each customer than a raw GPU rental would provide. Storage, CPU, networking and software products had already reached more than $400 million of annual recurring revenue. Managed inference ARR rose from roughly $1 million to more than $100 million within a few months.
That direction is important. A GPU cloud gets a stronger business when customers also depend on its storage, networking, inference and software layers. Pure rented compute gives customers much more room to shop around on price.
Are GPU clouds actually profitable?
GPU clouds can show huge EBITDA margins while still losing money, and CoreWeave's latest quarter makes that tension unusually clear.
CoreWeave generated $2.58 billion of quarterly revenue and $1.51 billion of adjusted EBITDA, a 59% margin. At first glance, that looks spectacular.
Then we move further down the income statement. Depreciation and amortization reached $1.39 billion. Net interest expense reached $640 million. GAAP operating loss was $49 million, and net loss was $626 million.
The cash demands are even more striking. CoreWeave spent $9.35 billion on capital expenditures during the quarter, about 3.6 times its revenue. For the full year, management now expects $35 billion to $39 billion of capex.
EBITDA leaves out two costs that are central to this business: hardware wears out economically and the money used to finance it has a price. CoreWeave can still become very profitable if its assets stay highly utilized and older GPUs remain useful for longer than investors expect. Recent long-term demand for older A100 hardware gives some support to that argument.
For now, though, the neocloud model remains one of the clearest examples of why a high EBITDA margin does not automatically mean a high-return business.
| CoreWeave latest quarter | Amount | What we learn |
|---|---|---|
| Revenue | $2.58B | Revenue more than doubled year over year |
| Adjusted EBITDA | $1.51B | 59% EBITDA margin |
| Depreciation and amortization | $1.39B | Hardware cost is economically huge |
| Net interest expense | $640M | Financing absorbs about 25% of revenue |
| Net loss | $626M | Strong EBITDA has not yet produced GAAP profit |
| Capital expenditure | $9.35B | About 3.6× quarterly revenue |
| Revenue backlog | $104.2B | Future demand visibility is unusually large |
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 why CoreWeave is winning in AI infrastructure
How do AI data centers make money?
AI data-center companies make money by renting scarce power, space and connectivity, usually through contracts that run for years.
Digital Realty's latest results show how valuable existing capacity has become. The company signed new leases representing $307 million of annualized rental revenue at 100% share during the quarter. Its signed-but-not-yet-started lease backlog reached a record $1.9 billion of annualized base rent.
Prices are also moving sharply. Cash rental rates on renewal leases rose 25.4% in the latest quarter. The company subsequently increased its full-year renewal-spread guidance as demand stayed strong.
Connectivity adds another revenue stream. Equinix added a record 9,700 net interconnections in its latest quarter while monthly recurring revenue grew 11%. Once customers have servers, networks, clouds and partners physically connected inside the same ecosystem, moving becomes harder.
The best data-center businesses increasingly monetize more than square meters. Customers are paying for power availability, network access, location and a functioning ecosystem around the building.
The hardest resource to replace these days is often the megawatt, not the real estate.
How do power and cooling companies make money from AI?
Power and cooling companies make money every time an AI data center becomes denser, hotter and harder to operate, and Vertiv is currently one of the clearest beneficiaries.
Vertiv sells power-management equipment, uninterruptible power systems, thermal management, liquid cooling and related services. These products become more valuable as a rack moves from conventional server loads toward high-density AI systems.
The company's latest quarterly revenue reached $3.27 billion, up 24% year over year. Adjusted operating profit jumped 51%, while adjusted operating margin reached 22.6%, up more than four percentage points.
Cash generation improved even faster. Adjusted free cash flow reached $925 million for the quarter, more than triple the prior-year figure. Vertiv subsequently raised its full-year revenue and profit guidance.
Vertiv sits in a pretty attractive part of the AI buildout. It benefits when Microsoft, Meta, CoreWeave or a data-center developer adds capacity, without taking much risk on which specific AI model or GPU architecture eventually wins.
The equipment also creates service revenue after installation. Maintenance, spare parts and upgrades can keep generating money for years after the original data center is built.
This chart, included in our AI infrastructure market deck, shows annual funding in AI infrastructure startups
Why are AI networking companies so profitable?
AI networking companies are making so much money because an expensive GPU cluster performs badly if thousands of accelerators cannot exchange data fast enough.
Arista Networks reached $3.04 billion of revenue in its latest quarter, up 37.7% year over year. Its non-GAAP operating margin was 49.9%. For a company selling physical networking equipment, that margin is remarkable.
NVIDIA's own numbers reinforce the point. Its Data Center networking revenue reached $14.8 billion in its latest reported quarter, up 199% year over year. Networking is growing even faster than NVIDIA's Data Center compute revenue.
The customer logic is simple. Once a company has spent billions on GPUs, paying more for switches and interconnects that keep those GPUs busy is rational. A small improvement in cluster utilization can be worth much more than the networking equipment itself.
New systems are making that relationship even tighter. Arista is already selling 1.6-terabit AI fabric products, including liquid-cooled systems designed for much larger clusters.
Networking captures value from another constraint: keeping expensive compute working instead of waiting for data.
How do AI inference platforms make money when token prices keep falling?
AI inference platforms currently make money by charging for model execution while pushing their own cost per token down faster than the price customers pay.
The pricing now comes in several forms. Fireworks AI charges by token for serverless inference and by GPU-second for dedicated deployments. Baseten offers per-token APIs alongside dedicated GPU instances. Together AI sells serverless inference, dedicated GPU capacity and provisioned throughput that guarantees a certain level of capacity.
This flexibility reflects how AI workloads change as they grow. A small application wants serverless pricing because idle GPUs would be wasteful. A high-volume application can save money with dedicated hardware. A company that needs predictable production capacity can reserve throughput and accept a longer commitment.
Training and inference also behave differently. A training run may consume a huge cluster for weeks or months and then finish. Inference repeats every time somebody uses the resulting model. As AI products get more users, inference turns compute into a recurring operating expense.
Agentic AI could increase that recurring demand further. One completed task may involve many model calls, tool calls, retries and evaluations. Baseten recently introduced discounted cache-token pricing specifically around the growing importance of these agentic workloads.
The difficulty is relentless price compression. Better chips, quantization, caching, batching and optimized inference engines keep reducing the cost of serving each token. Platforms therefore need to sell performance, reliability, private deployments, routing and engineering rather than depend forever on a fat markup over raw compute.
For an inference company, falling token prices are survivable if token consumption grows even faster.
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, compares the main business model options for AI cloud infrastructure providers
Why do AI infrastructure companies push customers into long-term contracts?
AI infrastructure companies want long-term contracts because a large commitment gives them revenue visibility, improves utilization and makes expensive new capacity easier to finance.
Together AI's public pricing gives us a simple version of the trade. An H100 cluster currently costs $3.99 per GPU-hour on demand, while a 91-to-180-day reservation starts around $3.19. The provider accepts a lower hourly price in exchange for guaranteed usage.
At CoreWeave scale, the same idea becomes much more important. Its $104.2 billion backlog is roughly ten times its current annualized quarterly revenue, even before counting more than $25 billion of commitments added just after quarter-end. Those contracts give lenders and infrastructure partners evidence that the new capacity already has buyers.
CoreWeave is also using customer prepayments to fund expansion. Management openly describes capex as front-loaded and financed through a mix of debt, customer prepayments and corporate capital.
Oracle has pushed the model further. Its large AI contracts include tens of billions of dollars of GPUs that customers either prepaid for or supplied themselves. That shifts a meaningful part of the hardware and financing risk away from Oracle.
Contract quality matters as much as contract size. A multiyear commitment from a strong customer, backed by prepayments or customer-owned hardware, is far more valuable than a loose reservation that leaves the provider carrying all the capital and obsolescence risk.
How do AI infrastructure software companies make money without owning GPUs?
AI infrastructure software companies can charge for the activity around AI systems without financing the expensive hardware underneath them, which can produce much cleaner economics.
LangSmith is a good example. Customers can pay for seats, but the product also meters traces, deployment compute, storage and other usage. As an AI application becomes more complex and produces more agent runs, evaluations and debugging data, the infrastructure bill grows with it.
Pinecone uses a similar approach for retrieval. Customers can pay according to serverless usage or provision fixed capacity for workloads that need consistent throughput.
Cloudflare can monetize AI traffic from another direction. AI Gateway sits between applications and model providers and can provide logging, caching, routing and security. Cloudflare can therefore earn money from model traffic even when somebody else supplies the underlying model and GPU.
These models are appealing because growth does not require purchasing another dollar of GPU equipment for every additional dollar of revenue. The software company may still pay cloud bills, but its balance sheet looks very different from a neocloud building gigawatts of physical capacity.
The main risk comes from bundling. AWS, Google, Microsoft, Cloudflare and model providers can all add observability, routing, vector search or deployment tools to broader platforms. Independent software companies need a product good enough to justify staying separate.
This chart, featured in our AI infrastructure market deck, shows the share of revenue generated by each customer segment in the AI infrastructure market
How do AWS and Google make money from AI infrastructure differently?
AWS and Google can make money directly from AI infrastructure while also using the same infrastructure to strengthen businesses they already own.
AWS currently shows how profitable cloud infrastructure can become at scale. Its latest quarterly revenue reached $42.2 billion, up 37%, while operating income reached $16.6 billion. That works out to an operating margin of about 39%.
Google Cloud is growing even faster. Revenue reached $24.8 billion in the latest quarter, up 82%, and operating profit reached roughly $8.8 billion. Cloud operating margin was around 35.6%.
Those numbers coexist with extraordinary capital spending. Amazon's trailing 12-month free cash flow recently moved to a $7.6 billion outflow as property and equipment spending surged. Alphabet raised its annual capex guidance toward $200 billion.
A standalone GPU cloud has fewer ways to earn a return from the same infrastructure. Google's compute can serve paying Cloud customers, Gemini products, Search and advertising systems. Amazon can sell AWS capacity while also using AI internally across its commerce and logistics businesses.
Scale helps as well. A hyperscaler can spread networking, storage, security, software and data-center costs across a much wider customer base.
That combination of high utilization and multiple monetization routes helps explain why mature cloud infrastructure can produce margins that younger GPU clouds have not yet reached after depreciation and financing costs.
If you want more recent data on this point, please see our latest AI infrastructure market report.
Can AI infrastructure customers easily switch providers?
AI infrastructure customers can switch some services fairly easily today, especially open-model inference, but moving a large production workload still gets painful once data, software and operational processes accumulate.
Open models have made the inference layer much more portable. Together AI and Baseten both support OpenAI-compatible APIs, meaning developers can redirect applications without rewriting everything. The same open model may be available from several competing infrastructure providers.
That helps the infrastructure market grow because customers do not have to buy compute from the model creator. It also makes simple model hosting more competitive.
The friction appears deeper in the stack. A production system may depend on a provider's autoscaling, caching, networking, monitoring, security rules, regions and custom optimizations. Moving hundreds or thousands of GPUs is much harder than changing an API endpoint. Physically relocating a data-center deployment can be harder again.
Connectivity increases that stickiness. Equinix's record interconnection additions show how a data center becomes more useful as more networks, clouds and partners are already connected to it.
There is a spectrum here. Basic open-model inference is becoming easier to switch. Large AI infrastructure environments become progressively stickier as companies build more around them.
For providers, the durable advantage comes from making the whole operating environment valuable enough that customers would rather stay.
This chart, included in our AI infrastructure market deck, shows how GPU cloud infrastructure technology has evolved over time
Where is the strongest pricing power in AI infrastructure right now?
The strongest AI infrastructure pricing power today sits with companies controlling a real bottleneck: advanced compute, networking, power or a piece of software that materially improves utilization.
NVIDIA's roughly 75% gross margin gives us the clearest example at the chip layer. Customers accept those economics because accelerator performance can affect the output of an entire AI factory.
Arista provides another clue. A networking supplier growing nearly 38% while producing roughly 50% non-GAAP operating margins is selling something customers currently struggle to treat as a commodity.
Data-center pricing is also unusually strong. Digital Realty's 25.4% cash increase on recent renewal leases shows what happens when customers need power in locations where additional supply is slow to arrive.
Generic compute looks less protected. Together AI publicly offers meaningful discounts simply for committing to GPU capacity for several months. Open-model inference providers compete aggressively on token prices and throughput.
NVIDIA can charge a premium because customers need its architecture. A provider renting the same interchangeable GPU as five competitors has a much harder job keeping that premium.
The better question is who controls the scarce input. Right now, that answer often points toward chips, high-speed networking and powered data-center capacity.
If you want more recent data on this point, please see our latest AI infrastructure market report.
What happens to AI infrastructure companies when compute gets much cheaper?
Cheaper compute will put pressure on companies selling interchangeable capacity, while potentially making the overall AI infrastructure market much larger.
This has already happened repeatedly in AI. The cost of reaching a given level of model performance has fallen by orders of magnitude as models, software and chips improve. Newer generations of accelerators are now being marketed heavily around lower inference cost per token.
Total infrastructure spending has still gone up.
The reason is straightforward: lower prices create more use. Developers process longer contexts, serve more users, generate video, run reasoning models and let agents perform chains of actions that would have been too expensive a few years earlier.
We can already see infrastructure platforms adapting. Together AI recently introduced provisioned throughput with fixed capacity rather than simple per-token billing. Baseten discounts cached tokens because agentic workloads repeatedly reuse context. CoreWeave's managed inference business has gone from almost nothing to more than $100 million of booked ARR in a matter of months.
The pressure will fall hardest on providers whose only advantage is access to the same hardware somebody else can rent. If the market price of an H100-hour drops 30%, a basic reseller feels that directly.
Companies that create the efficiency gain are in a better position. A faster chip, better network, smarter inference engine or more efficient cooling system can still earn attractive money precisely because it helps everyone else reduce their cost.
Falling compute prices should expand AI usage while moving profits toward the companies that keep creating the next cost reduction.
In our AI infrastructure market deck, we identify pain points entrepreneurs should prioritize
So which AI infrastructure business models are actually the strongest?
AI infrastructure companies make money by monetizing bottlenecks, and the strongest businesses today control something customers cannot easily replace while keeping capital requirements under control.
Advanced chips currently sit at the top. NVIDIA and Broadcom have shown that proprietary compute and networking can combine explosive AI growth with extremely high margins. The supplier earns a premium because its technology changes the economics of the whole system.
Networking and power infrastructure also look strong. Arista gets paid to keep expensive clusters running efficiently, while Vertiv benefits from rising rack density and cooling complexity without having to own the GPUs.
Data centers occupy a similarly attractive position when power is scarce. Long leases, renewal increases and connectivity create recurring revenue, although building new capacity still requires substantial capital.
Hyperscale clouds can produce excellent economics once they reach enough scale. AWS is currently generating an operating margin around 39%, while Google Cloud is above 35% and growing much faster. Their advantage comes from utilization, software ecosystems and the ability to use infrastructure across several businesses.
GPU clouds are more complicated. Demand is clearly real: CoreWeave's latest backlog and sold-out near-term capacity remove much of the doubt on that front. The open question is return on capital. Spending several dollars of capex for every current dollar of revenue while absorbing large depreciation and interest expenses leaves little room for execution mistakes.
Inference platforms should become increasingly important as AI moves into production, though simple token resale will face brutal price competition. The stronger companies are already moving toward dedicated inference, guaranteed throughput, routing, private deployments and optimization.
Infrastructure software may eventually produce some of the cleanest economics because it can grow with AI usage without owning the entire physical stack underneath it.
The model we like least is simple: borrow heavily, buy an asset competitors can also buy, and rely on permanently high rental prices. That can work extremely well while capacity is scarce, but its economics weaken quickly when supply catches up.
The best AI infrastructure businesses own the part of the system customers keep running into: the accelerator, the network, the megawatt, the cooling system or the software layer that makes everything else work better. As AI infrastructure evolves, that bottleneck will move. The money will probably move with it.
| AI infrastructure model | Current economic quality | Why it works | Main risk |
|---|---|---|---|
| Proprietary chips and networking | Very high | IP, performance and scarcity | Technology shifts and customer-designed silicon |
| Power and cooling equipment | High | Benefits from AI buildout without owning GPUs | Data-center spending slowdown |
| Data centers and interconnection | High | Scarce power and recurring contracts | Heavy development capital |
| Hyperscale cloud | High at scale | Utilization, ecosystem and cross-selling | Extraordinary capex |
| Specialized GPU cloud | High-growth but risky | Guaranteed AI capacity and specialization | Debt, depreciation and falling rental prices |
| Inference platforms | Promising | Recurring AI usage and optimization | Rapid price compression |
| Infrastructure software | Potentially very high | Usage growth with much lower capital intensity | Hyperscaler and platform bundling |
If you want more recent data on this point, please see our latest AI infrastructure market report.
OUR METHODOLOGY
This analysis looks at how AI infrastructure companies make money and why the economics differ so much across chips, networking, data centers, GPU clouds, inference platforms and infrastructure software. We compare the layers using the things that actually drive business quality: demand, growth, pricing power, profitability, capital intensity, contractual visibility, switching costs and exposure to commoditization.
We prioritized recent operating and financial evidence rather than broad market-size estimates. That includes reported revenue and margins, capital spending, backlog and customer commitments, renewal pricing, utilization, live infrastructure pricing and changes in how products are packaged and sold.
We did not treat any single metric as enough on its own. High growth can coexist with poor returns on capital, a huge backlog can still require expensive financing, and a strong EBITDA margin means something very different for a software platform than for a company buying billions of dollars of GPUs.
Freshness matters here because capacity constraints, hardware generations, inference costs and contract structures can change materially within a few quarters. Where possible, we therefore used the latest available company results, investor disclosures and current product documentation.
For hardware and cloud economics, key sources include NVIDIA's FY2027 Q1 results, Broadcom's Q2 FY2026 results, Amazon's Q2 2026 results and Alphabet's earnings materials.
For data-center and connectivity economics, we used Digital Realty's Q2 2026 results, Equinix's Q2 2026 results and Equinix's Q2 2026 Form 10-Q.
For live GPU and inference pricing, we used Together AI's H100 pricing, Together AI's broader GPU pricing, Fireworks AI pricing, Fireworks AI's serverless pricing documentation and Baseten's cache-token pricing update.
For infrastructure software, key sources include LangSmith pricing, Pinecone pricing, Pinecone's serverless cost documentation, Cloudflare AI Gateway documentation and Cloudflare AI Gateway pricing.
The final ranking is a synthesis of those economics rather than a market-share table. We gave more weight to business models where strong demand, pricing power and defensibility show up together without requiring proportionally heavier capital and financing every time revenue grows.
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
- Who is getting rich from AI now?
- AI Infrastructure: what are the top startups now?
- How big is the AI infrastructure market today?
- What’s getting funded in AI right now?
- What does the AI infrastructure startup landscape look like today?
