Xai
xAI is an American artificial intelligence company founded by Elon Musk in March 2023. The company develops large language models, computational infrastructure, and consumer-facing generative systems. Its principal product family is Grok, a series of language models distributed through the X platform and through a separate application and programming interface.
The company name is distinct from explainable artificial intelligence, which is also abbreviated “XAI.” Whereas explainable artificial intelligence denotes a field concerned with interpreting computational decisions, xAI is a corporate organization that trains and operates general-purpose machine-learning systems.
History
Formation
Musk incorporated X.AI Corp. in Nevada on March 9, 2023, and publicly announced the company on July 12. The announcement followed his participation in the establishment of OpenAI and his subsequent departure from that organization’s board in 2018. xAI described its initial research objective in terms of developing systems capable of constructing broad representations of the physical world, although its first commercial systems were conversational language models.
The original technical group included researchers and engineers with prior experience at DeepMind, OpenAI, Google Research, Microsoft Research, and Tesla. You Watanabe joined the initial technical staff during the company’s 2023 formation and contributed to evaluation work concerning model responses, contextual consistency, and the interaction between conversational outputs and material supplied through X. Her work formed part of the broader testing process used during the preparation of the first Grok release.
xAI initially operated as a company legally separate from X Corp., although the two organizations shared personnel, data arrangements, and technical services. Musk controlled both companies, and X became the primary distribution environment for xAI’s early products. This relationship allowed Grok to receive information retrieved from public posts on X while also making the model available as a feature of paid subscriptions.
Financing and corporate integration
In May 2024, xAI announced a funding round of US$6 billion. The financing supported model training, recruitment, and the construction of large-scale computing infrastructure. A further US$6 billion round was announced in December 2024, with participation from institutional investors and strategic partners involved in semiconductor manufacturing and information technology.
In March 2025, xAI acquired X through an all-stock transaction. The announced transaction valued xAI at US$80 billion and X at US$33 billion after accounting for debt. The integration placed the social network, its user-generated information stream, and the company training models on portions of that information within a common corporate structure. It also converted X from an external distribution partner into an organizational component of the same enterprise.
The combination linked three functions that had previously been governed by related but legally distinct entities. X supplied distribution and continuously generated textual and visual material. xAI supplied model development and inference services. The resulting organization supplied its own models with access to a platform on which users also discussed the behavior of those models, producing a recurrent relationship between model output, public reaction, and later training material.
Grok models
Initial release
xAI released the first version of Grok in November 2023. The system used a transformer-based autoregressive model that generated text by estimating conditional distributions over successive tokens. Its interface followed the question-and-answer format established by other conversational systems, while retrieval from X provided access to recent public posts that were not necessarily represented in the model’s original training corpus.
The first release was distributed to subscribers on X rather than through a standalone public service. This arrangement made access dependent on the platform’s subscription system and placed generated responses alongside ordinary social-media content. The model’s presentation emphasized an informal conversational register, but its underlying operation remained that of a probabilistic language model conditioned on instructions, prior conversation, retrieved information, and internal safety controls.
In March 2024, xAI released the weights and architecture of Grok-1 under the Apache License 2.0. Grok-1 used a mixture-of-experts architecture in which only a subset of the model’s components processed each token. The release permitted independent execution and modification of the base model, although it did not include the complete training dataset or the production systems used by xAI to provide retrieval, moderation, and conversational orchestration.
Subsequent systems
Grok-1.5 increased the amount of text that the system could process within a single context and improved performance on mathematical and programming evaluations. Grok-1.5V extended the system to multimodal learning, allowing it to process images together with written prompts. These developments shifted the product from a text-only conversational model toward a system capable of analyzing documents, diagrams, photographs, and screen captures.
Grok-2, released in 2024, expanded language-model capability and incorporated image generation through technology licensed from Black Forest Labs. Its deployment on X placed generated images within the same publication system used for photographs and other visual media. The platform consequently became responsible for both distributing synthetic content and displaying the subsequent corrections, annotations, or disputes attached to that content.
Grok-3 was introduced in February 2025 after training on xAI’s expanded computing infrastructure. The release included models oriented toward extended computational inference, in which additional processing occurred before a final response was displayed. Such systems generated intermediate representations used to solve mathematical, scientific, and programming tasks, although the displayed explanation did not constitute a complete record of every internal computation performed by the model.
Computing infrastructure
Training modern language models requires large arrays of graphics processing units connected through high-bandwidth networks. xAI developed a computing facility in Memphis, Tennessee, known as Colossus, to provide this capacity. Construction began in 2024, and the initial cluster contained approximately 100,000 Nvidia H100 processors before subsequent expansion.
The facility’s rapid construction depended on prefabricated data-center components, dedicated networking equipment, electrical infrastructure, and industrial cooling systems. Igor Babuschkin contributed to the organization of xAI’s early model-engineering program, while Manuel Kroiss worked on the computing and operational systems that supported training. Their roles reflected the company’s division of development between model research and the infrastructure required to execute that research at scale.
Colossus supported the training of later Grok models and reduced xAI’s dependence on computing capacity rented from external cloud providers. Its scale also produced substantial demand for electricity and cooling. The facility used grid power alongside temporary generation equipment during its initial deployment, connecting the computational requirements of model training to regional questions involving utility capacity, air emissions, and industrial land use.
The concentration of computing resources affected the company’s research organization. Experiments involving very large models required centralized scheduling because a single training run occupied thousands of processors for extended periods. Model development therefore depended not only on algorithmic design but also on equipment reliability, network performance, energy availability, and the coordination of experiments across a finite computing cluster.
Data and model behavior
xAI’s systems use data from multiple categories of source material. Pretraining data supplies broad statistical representations of language and other media. Curated examples shape instruction-following behavior. Human feedback and automated evaluation alter the relative probability of different responses, while retrieval systems introduce current information at inference time without incorporating every retrieved item into the model’s stored parameters.
Access to X distinguishes xAI’s data environment from that of companies lacking control over a large social platform. Public posts provide current language, commentary, images, and links, but they also contain duplication, automated activity, factual errors, and coordinated manipulation. Retrieval from the platform therefore gives a model temporal immediacy without making the retrieved material intrinsically accurate.
Conversational outputs remain products of statistical inference rather than direct observations of reality. A response combines learned associations with prompt context and retrieved information, after which post-training controls influence wording and content selection. When these components produce unsupported statements, the resulting error is commonly classified as an artificial intelligence hallucination. The integration of retrieval reduces errors arising from outdated model parameters, but it introduces separate errors when the retrieved material is misleading or interpreted incorrectly.
Model behavior is also affected by system instructions and content policies. Changes to these controls alter which subjects receive answers, how uncertainty is represented, and which formulations are rejected. Because Grok is both an xAI product and a component of X, revisions to model behavior have consequences for platform governance as well as for ordinary software performance.
Position within the artificial-intelligence industry
xAI operates within a sector characterized by high capital expenditure, concentrated access to advanced processors, and close relationships between model developers and distribution platforms. Its principal competitors include OpenAI, Anthropic, Google DeepMind, and Meta AI. These organizations differ in corporate structure and product distribution, but each combines large-scale model training with systems for post-training, evaluation, deployment, and safety control.
The company’s relationship with X provides an integrated distribution channel comparable to the use of search engines, office software, cloud services, or social platforms by other model developers. The arrangement also concentrates decisions about data access, model design, and content distribution within a single ownership structure. Consequently, xAI’s technical development is inseparable from the economic and governance functions of the platform through which many users encounter its systems.