Half the money goes into hardware
Gartner forecasts world AI spending of $2.59 trillion in 2026, 47% more than the $1.76 trillion of 2025. For 2027 the forecast is $3.49 trillion.
In 2026, infrastructure takes $1.43 trillion: data centers, servers, chips and networks. By our calculation from the Gartner forecast, that is more than half of all AI spending.
The reason is simple: models cannot be trained or run without computing power. Companies build capacity first, and the programs and services on top bring revenue later.
For a company that rolls out AI, there is a practical result. A large part of the price of an AI service you buy pays for someone else’s hardware, and it can change as demand grows.
If you plan an AI budget for 2027, remember that the price of computing depends on this demand. Fix the price in the supplier contract for a year, not a month, and keep a reserve for growth.
Hyperscalers
Hyperscalers are companies with huge clouds that rent computing to others. The four largest will direct $700–725 billion to capital expenditure in 2026, against $410 billion in 2025, about +75%.
By company: Amazon about $200 billion, Microsoft about $190 billion, Alphabet $180–190 billion, Meta $125–145 billion.
Most of this money goes to data centers and AI equipment. Amazon, Microsoft and Alphabet then rent this capacity to clients, while Meta uses it for its own products.
Growth from $410 billion to $700–725 billion in one year shows that the largest companies spend on capacity before they earn revenue from it. How analysts see AI spending ahead is covered in the forecast for 2027.
The figures for each company and what exactly the money goes to are on the page about the capex of the largest companies.
Infrastructure startups
Venture money also goes into hardware. In the third quarter of 2026, Crusoe and Nscale, which build data centers and rent computing for AI, raised more than $3 billion each, according to Crunchbase.
Crusoe and Nscale sell the same thing as hyperscaler clouds: chip time to train and run models. For a model developer, such companies are a second supplier next to the large clouds.
In the same quarter, Databricks and Safe Superintelligence received $5 billion each. A large part of this money will also go to computing, only from other suppliers.
A Roscongress report describes circular financing in this chain: chip makers invest in clients who buy their equipment. In such a scheme, part of the demand for chips comes from the chip makers’ own money.
For an investor, this is a question for any infrastructure startup: who its clients are and whose money they pay with. If the answer comes down to one chip supplier, the risk is higher than the revenue suggests.
Power and networks
A data center for AI depends not only on chips but also on power, cooling and communication networks. Without a grid connection, a new data center cannot start, even when the equipment is already bought.
Substations, power lines and cooling systems are part of the cost of every new data center. Gartner forecasts total AI spending to grow to $3.49 trillion in 2027, and the demand for power for computing will grow with it.
The Roscongress report put the chance of a crash of the AI company market in 2026 at above 50%. If demand for computing falls short of expectations, the capacity built and the power brought to it will stand idle.
For a company that sets up its own AI server, the question is the same on a small scale: is there enough grid power in the server room, and who is responsible for cooling. You check this before you buy equipment.
Russia: data centers and the industry’s request
According to RBC, large Russian AI developers asked the state for $4.8–$5.4 billion a year for AI development. In return, they offered to give 5% of their computing capacity each to education and state bodies.
There is no decision on this request, and it is still under negotiation. Companies that need AI now solve the capacity question themselves: they rent a cloud or install equipment on site.
If your AI works with the personal data of clients or employees, 152-FZ, Russia’s personal data law, requires that data to be stored in Russia. So for a Russian company, the choice of infrastructure starts with the question of where the data lives.
For a Russian company that wants to keep AI on its own servers, the infrastructure question is solved on a small scale: which tasks the model handles inside the company and what equipment it needs. The options and the calculation are on the page about your own AI server.
Keeping data inside the company removes the 152-FZ questions and the dependence on an outside cloud, but it needs your own server room and a person responsible for it.