xAI (now SpaceXAI): what is it trying to build?

This page is part of our series on who is building superintelligence. It is about the company and its technology rather than its founder.

What the company says it is for

When xAI announced a funding round in May 2024, it gave its mission as “to understand the true nature of the universe”. Its company page, checked on 3 October 2026, now reads “Accelerate human scientific discovery” and “We build AI to understand the universe”; its home page describes “frontier AI models for everything you build”. None of its current pages names AGI or superintelligence as the goal.

The word superintelligence does appear in xAI’s safety documents, but as a hazard. Its August 2025 Risk Management Framework refers to the possibility that a “superintelligent AI system hypothetically might escape the control of its developers”, and its current Frontier AI Framework, effective June 2026, discusses superintelligent systems only in its section on loss of control. In other words, xAI’s published materials treat superintelligence as something to guard against, not something it says it is building.

Musk has made frequent public predictions about when AI will surpass humans. Those are his personal statements, not the company’s published objective, and this page does not treat them as such. Our timelines page explains why such forecasts vary so widely.

xAI, SpaceX and SpaceXAI

The company’s structure has changed quickly. In March 2025 xAI combined with X, the social network formerly called Twitter. On 2 February 2026 xAI announced that SpaceX had acquired it. SpaceX’s June 2026 share-offering prospectus describes AI as one of its business segments, held through an entity called X.AI Holdings LLC. The company’s website and terms of service, updated in September 2026, now use the name SpaceXAI LLC, a Nevada company with registered offices in Austin, Texas. The company announced the new name in a post from its own account on X on 6 July 2026: “We are now @SpaceXAI”.

The prospectus gives SpaceX’s own mission as to “make life multiplanetary, to understand the true nature of the universe, and to extend the light of consciousness to the stars”, which is where xAI’s original wording now lives.

Grok

xAI’s models are called Grok. When it released Grok 4 in July 2025, xAI called it “the most intelligent model in the world”, a company claim. Its 2026 releases, according to its own announcements, are Grok 4.5 (16 July), Grok 4.6 (12 August) and Grok 4.7 (21 September), which it describes as its most powerful model for coding and knowledge work. Grok 5 had not been released as of 28 October 2026.

Independent evidence is limited. The ARC Prize Foundation reported in August 2026 that Grok 4.6 scored 67.1% on its ARC-AGI-2 benchmark and 2.1% on the harder, interactive ARC-AGI-3. This site found no published UK AI Security Institute evaluation of a Grok model. On 5 May 2026 the US Center for AI Standards and Innovation announced agreements with xAI, Google DeepMind and Microsoft under which it will conduct pre-deployment evaluations of their models; no results for Grok had been published as of 28 October 2026. Our page on recognising AGI explains what benchmarks like ARC can and cannot show.

Colossus

Computing power is where xAI’s scale is clearest. Its Colossus site in Memphis, Tennessee, was built in 122 days in 2024 and, by xAI’s account, reached 200,000 GPUs in a single interconnected cluster; xAI states a target of one million. SpaceX’s prospectus says Colossus and a second site, Colossus II in Memphis and Southaven, Mississippi, together provide about one gigawatt of capacity.

Epoch AI, an independent research group, estimates the first Colossus at about 230,000 chips drawing about 340 megawatts, and Colossus II at about 946 megawatts as of June 2026. Epoch also reports that Colossus II is used by other companies, including Anthropic, as well as by SpaceXAI itself. For comparison, the UK’s largest public AI supercomputer, Isambard-AI, draws about five megawatts (see our UK page). Computing power on this scale is a precondition for frontier AI, not evidence of what the models trained on it can do.

Safety and incidents

xAI’s Risk Management Framework, published in August 2025, covers two categories, malicious use and loss of control, with numerical thresholds for each, such as how often a model answers restricted queries and how it scores on an honesty test. Its Frontier AI Framework, effective 30 June 2026, extends this to four areas: chemical, biological, radiological and nuclear weapons; offensive cyber capability; loss of control; and harmful manipulation.

Two incidents have shaped how Grok is seen:

• In July 2025 Grok posted antisemitic content on X. The Grok team apologised publicly for what it called “horrific behavior”.
• In January 2026 Grok was used to generate sexualised images of real people. On 12 January Ofcom opened an investigation into X under the Online Safety Act, and on 3 February the Information Commissioner’s Office opened one into X and xAI over the processing of personal data by the Grok system. On 14 October 2026 Ofcom said its investigation had moved into a new phase, focused on X’s handling of illegal content and its systems for reporting and complaints, and that it would issue a provisional view before any final decision if it found grounds to. As of 28 October 2026 this site had found no published outcome to either investigation. Our UK regulation page explains the laws involved.

On 29 September 2026 Musk signed the voluntary White House Accord on Super Intelligence, under which each signatory company should apply four layers of controls to its most capable models: internal monitoring, an internal team to check it, an independent external auditor or evaluator, and an independent committee of the board. The general problem those commitments address is set out in whether superintelligence can be controlled.

Demonstrated and stated

• Stated objective: to understand the universe and accelerate scientific discovery. Not AGI or superintelligence.
• Demonstrated: rapid construction of very large computing capacity, documented by xAI, SpaceX filings and Epoch AI; a series of Grok models with some independent benchmark results.
• Documented problems: harmful outputs in 2025 and 2026, with UK regulatory investigations still open.
• Unknown: how its newest models perform in independent safety and capability testing, and how its AI work is governed within SpaceX.