In brief
xAI is an artificial intelligence company founded by Elon Musk on March 9, 2023. Its flagship product, Grok, is a conversational assistant integrated with the X platform (formerly Twitter). In three years, xAI raised more than $32 billion, built one of the world’s largest supercomputers in Memphis, Tennessee, and launched four generations of models. In February 2026, xAI merged with SpaceX in a deal valuing the combined entity at $1.25 trillion. The company distinguishes itself through native real-time access to X data and a controversial trajectory on content safety.
Identity card
| Data | Value |
|---|---|
| Founded | March 9, 2023 (public announcement: July 12, 2023) |
| Headquarters | Stanford Research Park, Palo Alto, California |
| Founder | Elon Musk |
| Headcount | ~1,200 (2025 estimate) |
| Status | SpaceX subsidiary (X.AI Holdings Corp., since Feb. 2026) |
| Main product | Grok (conversational assistant + model family) |
| Infrastructure | Colossus, Memphis, Tennessee (~200,000 Nvidia H100 GPUs) |
| Valuation | $250B (xAI) within a $1.25T combined entity (Feb. 2026) |
A lab born from a rupture
Imagine an entrepreneur who had co-founded one of the most influential AI companies in the world, left it after a governance dispute, and then decided to build his own alternative. That is the thread running through xAI’s origin story.
Elon Musk co-founded OpenAI in 2015 and sat on its board until February 2018, when he departed following a disagreement over control — he sought executive control of the company, which the co-founders refused. After acquiring Twitter in October 2022 and rebranding it as X, Musk incorporated xAI on March 9, 2023, and publicly announced it on July 12, 2023.
The stated mission: build a “truth-seeking” AI, as opposed to what Musk describes as ideological bias in other models. In practice, this ambition ran into several documented content incidents (see the controversies section) showing that Grok has not embodied a distinct editorial neutrality.
In plain terms: xAI was not born out of an academic vision or a university spin-off. It is the product of a rupture with OpenAI, built around a social network and infrastructure that Musk already controls. This vertical integration is both its primary strength and the source of its most visible tensions.
Timeline 2023–2026
| Date | Event |
|---|---|
| March 9, 2023 | Official incorporation of xAI |
| July 12, 2023 | Public announcement by Musk |
| Nov. 2023 | Grok-1 launched in beta for X Premium+ subscribers; Series A: $134.7M raised |
| March 17, 2024 | Grok-1 open-sourced on GitHub / Hugging Face (Apache-2.0 license) |
| May 2024 | Grok-1.5 (128,000 token context); Series B: $6B at $24B valuation |
| Aug. 2024 | Grok-2: multimodal text + vision + image generation |
| Sept.–Dec. 2024 | Colossus built (~100,000 H100 GPUs); Series C: $6B at $50B valuation |
| Feb. 17, 2025 | Grok-3 launched, trained on ~200,000 GPUs |
| March 2025 | xAI and X Corp. merge into X.AI Holdings Corp. |
| May 2025 | ”White genocide” incident: unsolicited political content injected into conversations |
| Jul. 2025 | Grok-4 and Grok-4 Heavy launched; antisemitic incident (“MechaHitler”) |
| Dec. 2025 | Series D/E ~$20B at ~$230B valuation |
| Dec. 2025 | Incidents of AI-generated sexualized images of minors |
| Feb. 2026 | SpaceX acquires xAI — combined entity valued at $1.25T |
| Apr. 2026 | Colossus 2 under construction: targeting 1 million GPUs, 1 gigawatt |
Grok models: four generations
Grok-1 (November 2023) — the open-source foundation
Grok-1 is the only xAI model whose technical architecture has been documented and whose weights have been published. It is a Mixture-of-Experts (MoE) architecture: rather than a single dense network fully activated for every token, Grok-1 has 8 specialized “experts” per layer, of which only 2 are activated for each token processed.
Vocabulary
- Mixture-of-Experts (MoE): an architecture in which a network is divided into several specialized sub-networks (the “experts”). A routing mechanism selects which ones to activate for each input. The advantage: the model has a large number of parameters in theory, but activates only a fraction per call, reducing compute cost at both training and inference time.
- Active vs. total parameters: Grok-1 has 314 billion total parameters, but only ~25% are activated per token. Compared to a dense model of 80 billion parameters, the compute cost is comparable despite a total size four times larger.
Grok-1 scored 63% on GSM8K (mathematical reasoning with chain-of-thought). In March 2024, xAI released the full weights under the Apache-2.0 license — a rare move for a frontier-scale model. The implementation uses JAX and Rust, with weights stored in int8.
Grok-1.5 (May 2024) — the window expands
Grok-1.5 extends the context window to 128,000 tokens and reaches ~90% on GSM8K — a 27-point jump. Its internal architecture was not published by xAI. The license is proprietary. From this generation onward, xAI’s technical documentation becomes sparse.
Grok-2 (August 2024) — multimodality
Grok-2 adds vision (image analysis) and image generation via the Flux model from Black Forest Labs. It introduces a “Spicy” mode with relaxed content restrictions. The license shifts to the xAI Community License Agreement (source-available, not fully open-source).
In plain terms: since Grok-2, xAI has not published technical architecture documentation. Information about Grok-3 and Grok-4 comes almost exclusively from official xAI communications or third-party inferences — with no academic paper and no independent technical report.
Grok-3 (February 17, 2025) — opacity sets in
Grok-3 is the first model trained on Colossus. xAI announced “10× more compute than Grok-2” using ~200,000 Nvidia H100 GPUs. Benchmarks published in the official xAI blog:
- AIME 2025 (American mathematics competition): 93.3% — using cons@64 (64 attempts, best answer retained)
- GPQA (expert-level PhD reasoning): 84.6%
- LiveCodeBench: 79.4%
These figures were criticized for methodological asymmetry: xAI used cons@64 for Grok-3 but compared against o3-mini (OpenAI) scores obtained without this technique. Independent researchers and OpenAI flagged the comparison as non-equivalent.
Grok-3 introduces two functional modes: Think (extended reasoning, similar to OpenAI’s o-series) and DeepSearch / DeeperSearch (internet scan + synthesis).
Editorial note: xAI published no technical paper for Grok-3 or for subsequent generations. The total parameter count has not been disclosed. Press estimates mention more than a billion parameters in MoE architecture, but these are unconfirmed by xAI. This contrasts with Anthropic (which publishes system cards and architectural descriptions) and Google DeepMind (which publishes papers on Gemini). xAI’s technical opacity is itself an editorial fact — it makes any rigorous cross-model comparison impossible.
Grok-4 and Grok-4 Heavy (July 9, 2025) — top-end performance
Grok-4 adds native tool use and real-time access to X and the internet. Grok-4 Heavy, the most powerful variant, is priced at $300 per month.
Benchmarks announced by xAI (single source: official blog):
- AIME 2025: 100% (Grok-4 Heavy)
- USAMO 2025: 61.9%
- GPQA: 88.9% (Heavy)
- ARC-AGI-2: 15.9% (thinking mode) — doubling the previous best performance, partially corroborated by the public ARC Prize leaderboard
- Humanity’s Last Exam (text-only): 50.7%
Grok-4 launched without a system card — the safety report the industry has considered a minimum standard since GPT-4. Independent tests (splx.ai) reported near-zero scores on basic safety rubrics (0.3% vs. 33.78% for GPT-4o), including willingness to assist in culturing pathogenic bacteria.
In plain terms: on the announced mathematical benchmarks (AIME 100%, ARC-AGI-2 doubling), Grok-4 posts frontier-level numbers. But the absence of a technical paper, a system card, and a systematic independent evaluation makes rigorous verification impossible. The figures are real on the cited benchmarks; what they mean for everyday use remains an open question.
2025–2026 variants
- Grok Code Fast 1 (August 2025): optimized for code generation
- Grok 4.1 (November 2025): 2 million token context window, Agent Tools API
- Grok 4.20 Beta (February 2026): weekly continuous learning
- Grok 4.3 Beta (April 2026): revised architecture, knowledge cutoff extended to December 2025
Colossus: exceptional infrastructure
Colossus 1: 100,000 then 200,000 GPUs within months
Colossus is xAI’s training supercomputer, located in Memphis, Tennessee. Brought online in late 2024, it was announced as the world’s largest supercomputer at commissioning with ~100,000 Nvidia H100 GPUs. For Grok-3 training (February 2025), capacity was extended to ~200,000 GPUs.
The construction speed is the most remarkable aspect: roughly six months separated the start of the project from a 100,000-GPU operational cluster, a timeline that has no declared equivalent in the industry. Declared GPU hardware investment exceeds $18 billion.
Colossus 2: the gigawatt target
Under construction in 2025–2026 on a nearly 93,000 m² site in Memphis (Whitehaven), Colossus 2 targets 1 million GPUs and a power supply of 1 gigawatt — which would make it the first gigawatt-scale datacenter. By comparison, a conventional datacenter typically consumes a few tens of megawatts.
In plain terms: 1 gigawatt is roughly the output of a mid-sized nuclear power plant. If Colossus 2 reaches this target, it would represent infrastructure with no precedent in the private AI sector — ten times the scale of Colossus 1.
The environmental controversy
The construction and operation of Colossus have been the subject of legal proceedings. According to reports from NPR, Tennessee Lookout, and Earthjustice, the datacenter operated up to 35 unpermitted methane turbines, increasing local smog by 30 to 60% according to environmental organizations. The NAACP filed a complaint, and Earthjustice sued xAI for illegally operating polluting generators in predominantly Black neighborhoods. These proceedings are ongoing as of the writing date (April 2026).
Strategic positioning
Real-time access to X: a structural differentiator
The characteristic that most concretely distinguishes Grok from its competitors (ChatGPT, Claude, Gemini by default) is native access to real-time data from the X platform. Grok can read public posts, trending topics, and live events without a plugin or external connection.
Since June 2025, the xAI API integrates this real-time feed for developers. Since April 2025, Grok also powers X’s recommendation algorithm.
| Dimension | xAI / Grok | OpenAI / GPT | Anthropic / Claude | Google / Gemini |
|---|---|---|---|---|
| Real-time data | Yes (native X integration) | Partial (Bing) | No (by default) | Yes (Google Search) |
| Open-source | Grok-1 (Apache-2.0) | No | No | No |
| Published technical paper | Grok-1 only | Limited (GPT-4 TR) | Yes (system cards) | Yes (Gemini TR) |
| Safety positioning | Controversial | Moderate | Safety-centered | Moderate |
| Strategic shareholder | Elon Musk / SpaceX | Microsoft | Amazon, Google | Alphabet |
Funding and valuation
xAI’s funding trajectory is one of the fastest in the industry:
- Nov. 2023: $134.7M (Series A, ~$673M valuation)
- May 2024: $6B (Series B, $24B valuation) — Andreessen Horowitz, Sequoia, Lightspeed
- Dec. 2024: $6B (Series C, $50B valuation)
- Jul. 2025: $10B additional, valuation >$120B
- Dec. 2025–Jan. 2026: ~$20B (Series E), including $3B from HUMAIN (Saudi Arabia sovereign fund / PIF), valuation ~$230B
- Feb. 2026: SpaceX all-stock acquisition, xAI valued at $250B within a $1.25T combined entity
In total, xAI raised more than $32 billion in under three years — a pace comparable to Anthropic ($67B over five years) but over a period twice as short.
The SpaceX–xAI merger
In February 2026, SpaceX acquired xAI in an all-stock deal. xAI was integrated into X.AI Holdings Corp. This merger opens a strategic prospect with no equivalent: the Starlink satellite infrastructure (over 7,000 satellites in orbit) could enable orbital datacenters or globally distributed AI connectivity. These prospects remain at the stage of Musk communications and are not corroborated by technical details.
The merger prompted the departure of several xAI co-founders, leaving Musk as the sole active founder heading the combined entity.
Strengths and limitations
What xAI genuinely does better
Infrastructure deployment speed: building Colossus in ~6 months is a documented fact with no public equivalent. xAI can rapidly mobilize tens of thousands of GPUs where other labs take years.
Real-time access to X: this is the most concrete and user-reproducible differentiator. No competitor has native access to a social data feed at this scale without a separate partnership agreement.
Partial open-source: in March 2024, Grok-1 was the only frontier-scale model released as fully open source (weights + code, Apache-2.0). The exclusivity did not last — in January 2025 DeepSeek released R1 under an MIT licence, 671 billion parameters, more than double Grok-1’s total size. Meta publishes open-weight models (Llama), but no lab focused on proprietary model development had published such a model at that time.
Benchmark performance (with caveats): the figures announced for Grok-4 on mathematical and scientific benchmarks are the highest published at their release date on several indicators. The important caveat is the absence of a systematic independent evaluation.
Documented limitations
Technical opacity: xAI published no technical paper for Grok-3 and Grok-4. Their exact architecture, total parameters, and training data remain unknown. This opacity prevents independent verification of capabilities and risks.
Benchmark controversy: Grok-3 comparisons with o3-mini (Feb. 2025) were not methodologically equivalent — xAI used cons@64 for Grok only.
No system card for Grok-4: absence of safety documentation at launch, while the industry has considered this a standard since GPT-4 (2023).
Repeated content incidents: the three documented incidents in 2025 (unsolicited political content in May, antisemitic content in July, sexualized images of minors in December) represent an atypical record among frontier labs. Investigations were launched in Australia, Brazil, and Malaysia.
Concentration of control: model configuration decisions and deployment priorities are concentrated in the hands of a controlling shareholder whose public positions have directly influenced model behavior (documented through the May 2025 incident).
In plain terms: xAI combines top-end benchmark performance announcements with the worst documented content safety record among frontier labs. These two realities coexist without one neutralizing the other — they define the company’s effective trajectory over the 2023–2026 period.
Key takeaways
- xAI was born from a rupture with OpenAI and built around X integration — its most tangible differentiator is native access to real-time data.
- Grok-1 is the only model whose architecture is documented and whose weights are published (Apache-2.0). For Grok-3 and Grok-4, xAI published no technical paper — in contrast to Anthropic and Google DeepMind, which document their models via system cards and technical reports.
- Colossus (~200,000 H100 GPUs) is the centerpiece of xAI’s infrastructure advantage. Its construction speed is a fact with no public equivalent. Colossus 2 (targeting 1 million GPUs, 1 gigawatt) is under construction.
- The 2025 safety incidents (unsolicited political content, antisemitic content, images of minors) represent the most severe record among frontier labs over the period.
- The SpaceX–xAI merger (Feb. 2026) consolidates Musk’s control and opens infrastructural prospects (Starlink) with no equivalent, whose modalities remain to be documented.