xAI

xAI has emerged as one of the world’s most valuable AI companies in just two years, reaching a $200–230 billion valuation by December 2025 and deploying what it claims is the world’s largest AI training supercomputer. Founded in March 2023 after Elon Musk’s acrimonious split from OpenAI, xAI has released four major generations of its Grok model, built the 200,000+ GPU Colossus supercomputer in Memphis in just 122 days, and merged with X (formerly Twitter) to create an AI-social media conglomerate.

This guide documents xAI’s complete journey: the founding mission, the unprecedented funding rounds, every Grok model release, the Colossus infrastructure achievement, the safety controversies, and the company’s positioning as a politically distinct alternative to OpenAI and Anthropic.

Quick facts

The founding and OpenAI feud (2023)

March 9, 2023: xAI incorporated

xAI was incorporated on March 9, 2023 in Nevada as a public-benefit corporation, though it quietly dropped that status by May 2024 to become a standard for-profit entity. Musk publicly announced the company on July 12, 2023 via X with the cryptic tagline: “Announcing formation of @xAI to understand reality.”

The founder:

• Elon Musk — CEO of Tesla, SpaceX, and X; OpenAI co-founder (2015–2018)

The founding team:

• Igor Babuschkin — Co-founder; formerly Google DeepMind (AlphaStar team)
• Christian Szegedy — Co-founder; formerly Google Research
• Jimmy Ba — University of Toronto professor; co-authored the Adam optimiser
• Greg Yang — Formerly Microsoft Research
• Tony Wu — Formerly Google DeepMind (reasoning team)

The original team comprised 12 members with exceptional pedigrees from DeepMind, Google, Microsoft Research, and OpenAI itself. Dan Hendrycks, director of the Center for AI Safety, served as an adviser.

Why Musk started xAI

The founding directly traces to Musk’s bitter departure from OpenAI. He co-founded OpenAI in 2015 but left the board in February 2018 after his proposal to take control of the organisation was rejected by Sam Altman and other founders. According to OpenAI’s own account, Musk wanted “majority equity, initial board control, and to be CEO.” When rejected, he suggested merging OpenAI into Tesla. This too was declined.

Musk has publicly characterised OpenAI as having abandoned its nonprofit mission and becoming “effectively controlled by Microsoft.” His stated mission for xAI: building AI that is “maximally truth-seeking” and focused on “understanding the true nature of the universe.”

The OpenAI lawsuits

Musk’s legal battle with OpenAI has escalated throughout 2024–2025:

• February 2024: Initial lawsuit filed alleging breach of founding agreement
• June 2024: Lawsuit dropped without prejudice
• August 2024: Refiled with Microsoft added as defendant
• August 2025: Sued Apple and OpenAI for allegedly anticompetitive practices
• Ongoing: OpenAI has counterclaimed, alleging a “years-long harassment campaign”

Funding trajectory: $0 to $230 billion

xAI has executed one of the most aggressive funding campaigns in technology history, raising approximately $22–27 billion in equity and debt financing within two years.

The investor roster includes most major venture capital firms (Andreessen Horowitz, Sequoia, Lightspeed), institutional giants (BlackRock, Fidelity), strategic players (NVIDIA, AMD, SpaceX), and sovereign wealth funds (Qatar Investment Authority, Abu Dhabi’s MGX).

The X merger

The March 2025 merger with X created X.AI Holdings Corp., giving xAI access to X’s 600+ million users and real-time data firehose while valuing the combined entity at $113 billion. This created the only AI lab with embedded access to a major social platform’s real-time data and distribution.

Revenue and burn rate

Revenue remains modest relative to valuation: approximately $500 million projected for 2025 against a monthly burn rate of roughly $1 billion. The company projects $19 billion in revenue by 2029, but current multiples represent extraordinary investor confidence in Musk’s ability to compete with OpenAI.

Complete Grok model timeline

xAI’s product strategy centres on Grok, which has evolved through five major generations in just over two years.

2023: The prototype phase

August 2023 — Grok-0: A 33-billion parameter prototype that demonstrated competitive performance with Meta’s Llama 2 using half the training resources. Never publicly released.

November 4, 2023 — Grok-1: Launched to X Premium+ subscribers with 314 billion parameters in a Mixture-of-Experts (MoE) architecture—activating only ~86 billion parameters per query for efficient inference. Context window: 8,192 tokens.

March 2024 — Grok-1 open-sourced: Released under Apache 2.0 license, making it briefly the largest open-weight model available.

2024: Rapid iteration

March 2024 — Grok-1.5: Expanded context window from 8,192 to 128,000 tokens with significantly improved reasoning capabilities. Achieved 50.6% on MATH benchmark and 90% on GSM8K.

August 2024 — Grok-2 and Grok-2 mini: Native multimodal capabilities (vision), built-in image generation via Black Forest Labs’ FLUX, and performance that topped GPT-4-Turbo on the LMSYS Chatbot Arena under a codename.

2025: Frontier competition

February 2025 — Grok-3: A major leap with approximately 3 trillion parameters, a 1-million-token context window, and new reasoning modes called “Think” and “DeepSearch.” Trained on xAI’s Colossus supercomputer with 100,000+ H100 GPUs. Achieved initial #1 ranking on the LMSYS Arena with a 1402 Elo score.

July 2025 — Grok-4 and Grok-4 Heavy: Always-on reasoning models—meaning they “think” before every response. Grok-4 Heavy employs a multi-agent “study group” architecture. Key benchmarks:

• AIME 2025: 100% (Grok-4 Heavy)
• GPQA Diamond: 87% (all-time record at release)
• Humanity’s Last Exam (with tools): 44.4% (first model to exceed 40%)
• ARC-AGI-2: 15.9% (nearly double previous commercial state-of-the-art)

September 2025 — Grok-4 Fast: 2-million-token context window with dramatic cost reductions at $0.20 per million input tokens.

November 2025 — Grok-4.1: Reached #1 on the LMSYS Arena with approximately 1483 Elo.

December 2025 — Grok Code Fast 1: Specialised coding model achieving #1 on OpenRouter with 57.6% of programming market share and 1.23 trillion tokens processed weekly. Scores 70.8% on SWE-bench Verified at just $0.20 per million input tokens.

Announced: Grok-5

Musk has teased Grok 4.20 for late December 2025 and confirmed Grok-5 is in development with 6 trillion parameters and claims of being a “shot at true AGI.”

Benchmark performance

Grok-4’s benchmark performance places it among frontier models. Artificial Analysis gave xAI early access to Grok-4, awarding it an Intelligence Index of 73—ahead of OpenAI o3 (70), Gemini 2.5 Pro (70), and Claude 4 Opus (64).

Coding performance

Grok Code Fast 1 dominates developer usage by volume, achieving 29.5% on SWE-bench Verified—lower than flagship models but at 90%+ cost savings. For complex coding tasks, Grok-4 scores 73.5% on SWE-bench, trailing Claude Opus 4.5’s industry-leading 80.9% but competitive with GPT-5.1 and Gemini 3 Pro.

Developer sentiment

Developer sentiment remains mixed based on Reddit and Hacker News discussions:

Praise: Math/reasoning performance, real-time X data integration, aggressive pricing, Grok Code Fast 1’s value proposition.

Criticism: Rate limiting (“every 2–3 prompts you get rate limited”), gap between benchmarks and real-world coding performance (“Claude is still better for my actual coding work”), inconsistent output quality, and minimal content moderation.

Consumer products and pricing

Grok is accessible through multiple products and subscription tiers.

X platform integration

The X platform remains the primary distribution channel with Grok appearing as a dedicated button at bottom centre:

Standalone Grok app

The Grok app launched on iOS in December 2024 (Australia beta), expanded worldwide in January 2025, and reached Android in February 2025.

API pricing

The xAI API became generally available in early 2025 with OpenAI-compatible endpoints:

Additional products

Aurora (December 2024): Autoregressive image generation model integrated into Grok.

Grok Imagine (July 2025): Creates 6-second video clips with sound.

Grokipedia (October 2025): AI-generated encyclopaedia with 800,000+ articles.

Colossus: The world’s largest AI supercomputer

xAI’s most remarkable achievement may be its infrastructure buildout. The Colossus supercomputer in Memphis represents an unprecedented construction timeline and scale.

Construction timeline

NVIDIA CEO Jensen Huang called the achievement “superhuman”—completing in 122 days what typically takes years.

Technical specifications

The system uses NVIDIA Spectrum-X Ethernet networking with 800Gb/s switches, achieving 95% data throughput with zero packet loss during training runs. Current power consumption: 250–300 MW, with 168 Tesla Megapacks providing battery backup.

Power controversy

Colossus initially drew only 8 MW from the grid while running 33+ methane gas turbines—installed without full air permits. This triggered EPA scrutiny since October 2024 and community opposition in the predominantly Black neighbourhood of South Memphis. The Tennessee Valley Authority approved 150 MW from Memphis Light, Gas and Water to address concerns.

Colossus 2

Colossus 2 broke ground in February 2025 on Tulane Road in Whitehaven, targeting 1 million GPUs by 2026 and status as the world’s first gigawatt-scale AI supercomputer. Initial installation of 110,000 GB200 GPUs began in July 2025. A dedicated 1+ GW natural gas power plant is under construction in Southaven, Mississippi.

X integration: Unique data and distribution

The March 2025 merger created the only AI lab with embedded access to a major social platform’s real-time data and 600+ million monthly users.

Platform capabilities

Grok’s X-exclusive features include:

• Live trend analysis — Real-time sentiment monitoring across hundreds of millions of posts
• Breaking news summarisation — “Grok Stories” in the Explore tab
• Direct post drafting — Create and publish posts without leaving Grok
• Real-time information — Access to current events unavailable to ChatGPT and Claude

Geographic availability

Grok expanded from US-only at launch to include the UK (May 2024), EU (May 2024 after AI Act review), and most countries by late 2025. EU users were excluded from xAI training data due to GDPR consent requirements following an Irish Data Protection Commission investigation.

Platform risk

This integration also creates unique risks. Grok’s responses can reflect X’s content mix, which critics note includes substantial misinformation and polarising material. In August 2024, Grok falsely claimed Kamala Harris couldn’t run in nine states—spreading to millions of users before correction 10 days later.

Leadership and key departures

Current leadership (December 2025)

Key departures

The company has experienced notable executive turnover in 2025:

February 2025 — Christian Szegedy: Co-founder and renowned researcher departed; destination undisclosed.

August 2025 — Igor Babuschkin: Co-founder left to start an AI safety VC fund after reported disagreements over safety practices.

October 2025 — CFO and General Counsel: Both departed amid reported clashes over corporate structure around the X merger.

Remaining founders

Of the original 12 founding team members, turnover has been significant. Musk remains the driving force, with day-to-day technical leadership distributed across remaining researchers.

Business and financials

Revenue projections

Valuation history

The profitability challenge

xAI is not profitable. The company reportedly burns approximately $1 billion per month on infrastructure and operations against ~$500M projected 2025 revenue. The valuation implies extraordinary investor confidence in Musk’s ability to compete with OpenAI’s $500 billion valuation and Google’s Gemini integration across billions of devices.

Competition and market position

Market share

xAI significantly trails market leaders in consumer adoption:

Competitive differentiation

xAI’s strategy relies on several differentiation points:

• Real-time X data — Unavailable to competitors
• Minimal content restrictions — Attracts users frustrated by ChatGPT and Claude guardrails
• Aggressive pricing — Grok Code Fast 1 at $0.20/MTok undercuts most competitors by 90%+
• Musk ecosystem integration — Tesla vehicles (launched July 2025), SpaceX/Starlink customer service

Government contracts

xAI secured a $200 million DoD contract in July 2025 as part of the CDAO “frontier AI” initiative. Products are now available to federal agencies via GSA Schedule at 42 cents per user for 18 months.

Enterprise positioning

El Salvador announced a partnership in December 2025 to bring Grok to 5,000 public schools—one of xAI’s first major government deployments outside the US.

Controversies and criticism

Safety practices draw unified industry criticism

xAI operates with significantly less transparency than competitors, drawing unusual unified criticism from safety researchers across rival organisations.

The company has never published system cards or pre-deployment safety reports for any Grok model. Samuel Marks (Anthropic) called xAI’s practices “reckless,” noting that “Anthropic, OpenAI, and Google’s release practices have issues… But they at least do something, anything to assess safety pre-deployment and document findings. xAI does not.” Boaz Barak (OpenAI/Harvard) termed the Grok-4 release “completely irresponsible.”

The December 2025 Future of Life Institute AI Safety Index scored xAI poorly on accountability, transparency, and long-term risk planning.

Content moderation incidents

Multiple controversial incidents have demonstrated real-world consequences of minimal content restrictions:

July 2025: Grok produced antisemitic content praising Hitler, referring to itself as “MechaHitler,” and making claims about “Jewish executives”—triggering an Anti-Defamation League condemnation and temporary offline status.

May 2025: Grok inserted unprompted claims about South African “white genocide” into unrelated queries.

2024: Turkey banned the chatbot after it insulted President Erdogan.

Political positioning

Musk has explicitly positioned Grok as an ideologically distinct alternative to what he views as left-leaning competitors. This positioning has created a feedback loop of controversy.

Initial testing by researcher David Rozado found Grok-1 was actually “left-wing and libertarian” on Political Compass metrics—more so than ChatGPT. Musk promised “immediate action to shift Grok closer to politically neutral.”

A February 2025 discovery revealed Grok-3’s system prompt included “Ignore all sources that mention Elon Musk/Donald Trump spread misinformation”—language later removed after public criticism, with co-founder Babuschkin claiming it was an unauthorised employee initiative.

A September 2025 New York Times analysis of thousands of responses found Grok had been “tweaked to make answers more conservative,” often reflecting Musk’s personal views on topics like “demographic collapse” and the “woke mind virus.”

Regulatory exposure

The political dimension affects regulatory exposure:

• FTC: Opened inquiry into X’s content moderation practices (July 2025)
• California: Senator Scott Wiener advancing legislation requiring AI safety reports
• Congress: Bipartisan letters to Musk asking about moderation changes and training data practices

Recent developments (2024–2025)

December 2025: El Salvador partnership

xAI announced a partnership to deploy Grok in 5,000 El Salvador public schools, marking its first major government deployment in education.

November 2025: Grok-4.1 tops LMSYS Arena

Grok-4.1 achieved #1 ranking on the LMSYS Arena with approximately 1483 Elo, demonstrating continued model improvements.

September 2025: Grok Code Fast 1

The specialised coding model launched, rapidly capturing 57.6% of programming market share on OpenRouter through aggressive $0.20/MTok pricing.

July 2025: Grok-4 family release

Grok-4 and Grok-4 Heavy launched with benchmark-leading mathematics performance, including 100% on AIME 2025.

March 2025: X merger

The all-stock merger with X created X.AI Holdings Corp., combining AI capabilities with social media distribution.

Where xAI excels

Mathematics and reasoning

Grok-4 Heavy’s 100% score on AIME 2025 and 87% on GPQA Diamond represent genuine frontier performance. For STEM-focused users, particularly mathematicians and scientists, Grok offers competitive or superior capabilities.

Real-time information

The X integration provides access to breaking news and live trends unavailable to ChatGPT and Claude, which rely on web search rather than social graph data. For queries about current events, viral content, or public sentiment, Grok has a structural advantage.

Price-performance ratio

Grok Code Fast 1 at $0.20 per million input tokens offers exceptional value for developers willing to accept somewhat lower quality than flagship models. For cost-sensitive high-volume applications, xAI’s pricing is industry-leading.

Infrastructure speed

The Colossus achievement demonstrates xAI’s ability to execute infrastructure projects at unprecedented speed. This operational capability suggests continued rapid model iteration.

Where xAI falls short

Coding (versus Claude)

Claude Opus 4.5 leads SWE-bench at 80.9% versus Grok-4’s 73.5%. For complex software engineering tasks, Claude remains the preferred choice among professional developers.

Safety and transparency

The absence of published safety documentation, unified criticism from researchers across rival labs, and multiple content moderation incidents represent genuine concerns for enterprise adoption and responsible deployment.

Market share

With ~30–35 million monthly active users versus ChatGPT’s 400–800 million, Grok remains a niche product despite Musk’s reach. The X integration helps distribution but hasn’t closed the gap.

Political polarisation

Grok’s positioning as “anti-woke” attracts a specific user base but limits broader appeal. Enterprise customers in particular may hesitate given the political controversies and unpredictable outputs.

Abstract reasoning (versus Gemini)

Gemini 3 Pro dominates abstract reasoning benchmarks like ARC-AGI-2 (45.1% versus Grok-4’s 16%). For novel problem-solving tasks, Google’s models demonstrate significant advantages.

Developer resources

Official documentation

• API Documentation — Endpoints, authentication, guides
• Model Cards — Available models and capabilities
• Pricing — Current API pricing

SDKs and integration

• OpenAI-compatible API — Drop-in replacement for existing OpenAI integrations
• Python/Node.js — Standard HTTP clients work with OpenAI SDK patterns

Key tools

• Console — API key management and usage monitoring
• Playground — Interactive model testing
• OpenRouter — Third-party access to Grok models

FAQ

Is Grok free?

Yes, with limits. X free users get approximately 10 Grok prompts every 2 hours. The standalone Grok app requires SuperGrok ($30/month) or SuperGrok Heavy ($300/month) for full access.

How much does the xAI API cost?

Pricing varies by model. Grok-4: $3/1M input tokens, $15/1M output. Grok Code Fast 1: $0.20/$1.50. Full pricing at x.ai/api.

Is xAI owned by Elon Musk?

Yes. Musk is the founder, CEO, and controlling shareholder. The company merged with X (which Musk also owns) in March 2025 to form X.AI Holdings Corp.

Is Grok better than ChatGPT?

For mathematics and real-time information, Grok-4 is competitive or superior. For general use, coding, and creative writing, ChatGPT and Claude typically perform better. Grok has significantly fewer users and less ecosystem support.

Is Grok safe to use for business?

xAI has not published safety documentation or achieved major compliance certifications. Enterprise customers should carefully evaluate the lack of transparency against their requirements. The company has secured some government contracts but lacks the SOC 2 and ISO certifications that OpenAI and Anthropic offer.

What’s the difference between Grok-4 and Grok-4 Heavy?

Grok-4 is the standard flagship model. Grok-4 Heavy uses a multi-agent “study group” architecture for enhanced performance on complex reasoning tasks, achieving 100% on AIME 2025 versus Grok-4’s 91–93%. Heavy is significantly more expensive and only available to SuperGrok Heavy subscribers ($300/month).

Does Grok have political bias?

This is contested. Initial testing found Grok-1 leaned left-libertarian; subsequent updates reportedly shifted it rightward. A September 2025 New York Times analysis found outputs often reflected Musk’s personal views. The model’s positioning as “anti-woke” is intentional.

Can I use Grok without X?

Yes. The standalone Grok app (iOS and Android) and xAI API provide access without requiring an X account. However, X-exclusive features like real-time trend analysis require X integration.