TikTok U.S. hunts for multiple gigawatts of computing capacity, sources tell CNBC

TikTok U.S. is seeking four gigawatts of computing capacity by the early 2030s, according to one source familiar with the discussions. The company is exploring ways to secure that capacity over the coming years. The request comes as businesses compete for infrastructure needed to run increasingly demanding digital services.

The need is connected to TikTok’s artificial-intelligence plans. The company is hiring for machine-learning roles and describes one product as a flagship business-to-business AI system focused on large-model inference and creative automation. It also plans to retrain, test, and update its recommendation algorithm using U.S. user data.

TikTok U.S. currently runs primarily on Oracle Cloud Infrastructure, and Oracle is expected to remain its main cloud provider. The article does not say exactly how much capacity TikTok already uses or whether Oracle alone will supply all four gigawatts. It identifies the target as a possible requirement by the early 2030s.

In this context, a gigawatt describes the power capacity associated with computing infrastructure. One gigawatt equals one billion watts. It is closer to measuring the electricity that facilities and equipment can draw than to measuring the number of servers, storage space, or software tasks directly.

A computing system’s useful output depends on more than its power rating. The hardware design, chips, cooling systems, networking, and workload all affect how much work that electricity supports. AI workloads can use substantial power because they may involve many specialized processors operating together. Four gigawatts therefore signals a very large infrastructure requirement.

The article uses “four gigawatts of compute power” as TikTok U.S.’s reported target for the early 2030s. It does not define the term technically or explain whether the figure refers to continuous facility power, total planned capacity, or another industry measure. Those details would be needed for a precise engineering comparison.

Four gigawatts is enormous when treated as electrical capacity for computing. As a broad industry comparison, a major technology company’s individual data-center campus may be measured in hundreds of megawatts, while the largest planned clusters can approach or exceed a gigawatt. Four gigawatts could therefore represent several large campuses operating together.

The key mechanism is aggregation. A campus combines computing equipment, networking, cooling, power systems, and backup infrastructure. Multiple campuses can be connected through networks and managed as one broader computing platform. The more capacity a company secures, the more AI jobs it can run concurrently, although efficiency and chip availability still matter.

The comparison is indicative, not exact. The article gives TikTok U.S.’s four-gigawatt target but provides no figures for a typical technology company or data-center campus. Industry reporting also uses different definitions, including utility power, facility power, and installed computing capacity, so direct comparisons require caution.

AI workloads increase competition for computing capacity because they often require many processors working together. Training a model can involve repeated calculations across huge datasets. Running the finished model for users also consumes capacity, especially when responses must be generated quickly and at large scale. Companies therefore need both large systems and reliable access to them.

TikTok U.S. offers a clear example. Its job postings mention machine learning and a business-to-business AI product focused on large-model inference and creative automation. The company also plans to retrain, test, and update its recommendation algorithm using U.S. user data. Those activities can create continuing demand for computing resources.

The article places TikTok’s plans within a broader global scramble for digital infrastructure deals. It does not quantify the total industry shortage or identify every competing company. Still, its reported four-gigawatt target shows how AI ambitions can turn computing access into a long-term strategic requirement.

TikTok U.S. has several possible sources of future computing demand. Its website lists numerous machine-learning roles, and a job advertisement describes a flagship business-to-business AI product. That product is being built around large-model inference and creative automation, both of which could require substantial, continuing infrastructure.

The recommendation system is another major operation. TikTok U.S. said it would retrain, test, and update its machine-learning recommendation algorithm on U.S. user data. Those steps require systems that process data, run training jobs, evaluate results, and serve recommendations to users. Each activity can add to the company’s overall computing needs.

The article does not state which specific product will consume the reported four gigawatts. It also does not provide a timetable for each project. It does show a clear direction: localized algorithm work, new AI products, and expanding machine-learning operations could make additional capacity increasingly important.

Relying on Oracle Cloud Infrastructure gives TikTok U.S. access to computing resources without requiring the company to own every physical data center. A cloud provider supplies facilities, hardware, networking, and related systems. The customer uses that capacity through a service arrangement, while the provider manages much of the underlying infrastructure.

The article says TikTok U.S. runs primarily on Oracle Cloud Infrastructure, or OCI. It also says OCI typically operates in facilities that Oracle leases. This means Oracle can arrange and manage physical space and equipment, while TikTok U.S. uses the computing environment for its platform and planned AI work. The arrangement separates service use from facility ownership.

Cloud reliance does not remove the need to secure capacity. TikTok U.S. still needs enough available computing power for its workloads, and Oracle may need to expand or reserve infrastructure to meet demand. The company says Oracle will continue as its primary provider, but the article does not say Oracle will supply every future gigawatt.

Training is the learning phase. An AI system processes examples and adjusts its internal parameters so it can recognize patterns or generate outputs. Training often involves repeated calculations across large datasets. It can be especially demanding because the system may run for long periods while improving the model.

Inference is the use phase. A trained model receives a request, such as a text prompt or creative task, and produces an answer or result. TikTok’s job advertisement specifically mentions large-model inference and creative automation. Storage is different: it keeps models, user data, and other files. Transmission moves that information between users, servers, and processing systems.

These functions can use separate or shared infrastructure. Training may need concentrated processing, inference needs responsive capacity for live requests, and storage or transmission depends on data systems and networks. The article mentions retraining, testing, updating, and inference, but it does not describe TikTok’s detailed hardware architecture.