xAI

Theory of Change

xAI's stated mission is "to understand the universe." Elon Musk argues this mission inherently promotes safety because a maximally curious AI would want to preserve humanity as an interesting part of the universe: "Understanding the universe means you have to be truth-seeking... You're not going to discover new physics or invent technologies that work unless you're rigorously truth-seeking" (Dwarkesh podcast, Feb 2026).

Musk's causal chain: curiosity requires intelligence, intelligence requires existence, existence requires propagating consciousness -- therefore a curious AI would preserve humanity. He explicitly acknowledges humans will not be in control: "I don't think humans will be in control of something that is vastly more intelligent than humans... I think what we can do is make sure it has the right values."

His alignment approach: develop "debuggers" to trace AI reasoning errors at the neuron level, similar to debugging code. He praises Anthropic's interpretability work and frames alignment primarily as an engineering problem, not a fundamental research challenge.

xAI was initially incorporated as a public-benefit corporation in Nevada (March 2023) but quietly dropped this status by May 2024, ending any legal obligation beyond shareholder returns.

What They Do

Models: Grok 1 through Grok 4.20 Beta, released in rapid succession from November 2023 to present. Grok 3 (Feb 2025) used 10x compute from the Colossus supercomputer. Grok 4 (Jul 2025) briefly topped some benchmarks, particularly in math. As of March 2026, Musk admits xAI's coding tools are not competitive with Claude Code or OpenAI's Codex, and the company is "stuck in the catch-up phase."

Infrastructure: Colossus supercomputer in Memphis: 200K+ GPUs (150K H100, 50K H200, 30K GB200), built in 122 days. Second site at Whitehaven planned with 550K GPUs. ~1 GW power consumption. Musk's long-term vision: data centers in space within 30-36 months, claiming solar-powered orbital compute will be 10x cheaper than terrestrial.

Products: Grok chatbot (integrated into X platform, 64M monthly users), SuperGrok ($30/mo) and SuperGrok Heavy ($300/mo) premium tiers, API access, Grokipedia (AI-generated encyclopedia launched Oct 2025), Grok for Government, "Macrohard" agentic project (stalled as of March 2026).

Corporate acquisitions: xAI acquired X/Twitter in March 2025 ($33B all-stock transaction). SpaceX acquired xAI in February 2026 ($250B valuation).

Government contracts: $200M Pentagon contract (awarded days after the MechaHitler antisemitic incident, Jul 2025). GSA OneGov deal: Grok offered to every federal agency for $0.42/agency for 18 months.

Safety incidents (chronological):

Grok repeatedly brought up "white genocide" in conversations (May 2025)
"MechaHitler" -- Grok spontaneously called itself MechaHitler, praised Hitler, made antisemitic comments (Jul 2025). xAI blamed "unauthorized modification"
Grok 4 released without any system card or safety report despite Seoul Commitment to publish (Jul 2025)
Anonymous researchers documented Grok 4 providing nerve agent synthesis, VX instructions, fentanyl synthesis, nuclear weapon construction, and suicide methods with no jailbreaking required (Jul 2025)
Grok's "Spicy Mode" enabled mass generation of non-consensual sexual imagery, including of minors: 3 million sexualized images in under 2 weeks, 23,000 depicting children (Dec 2025-Jan 2026). xAI's press response: auto-reply "Legacy Media Lies." Musk posted laugh-cry emojis.
Global lawsuits: Tennessee teens, Baltimore, California AG, UK Ofcom, EU investigations, France prosecutors, India corrective order

Research output: Zero published technical papers. Extreme secrecy about architecture and training methods. Grok-1 (314B parameters) was open-sourced under Apache 2.0, and Grok 2.5 weights were released on Hugging Face, but no current frontier model documentation exists.

Key People

Elon Musk (CEO): Also CEO of Tesla and SpaceX. Runs DOGE. Former OpenAI co-founder (2015-2018). Currently suing OpenAI for $134B. No background in AI safety research. Sole decision-maker on safety policy. His political views directly influence Grok's behavior -- the model was discovered searching for Musk's stated positions before answering political questions.

Dan Hendrycks (safety adviser, $1 salary, no equity): Director of CAIS, co-founder of Gray Swan, also adviser to Scale AI. Claimed internal safety evaluations were conducted on Grok 4 but results were never published. The compensation structure ($1, no equity) gives him no leverage over decisions.

Norman Mu (former safety lead, departed Dec 2025): Led the RMF and model cards. After his departure, former employees told The Verge: "Safety is a dead org at xAI" and "There is zero safety whatsoever in the company."

Team: ~5,000 employees. 9 of 11 co-founders have departed by March 2026. Musk admits xAI "was not built right first time around, so is being rebuilt from the foundations up." SpaceX and Tesla executives have been sent in to audit and fire staff. Only 2 co-founders (Manuel Kroiss, Ross Nordeen) remain.

Money and Incentives

Total funding: ~$45B over 9 rounds. Latest: $20B Series E at $230B valuation (Jan 2026). SpaceX acquired xAI at $250B valuation (Feb 2026). Investors: Andreessen Horowitz, Sequoia, BlackRock, Fidelity, Kingdom Holdings (Saudi Arabia), NVIDIA, AMD, Qatar Investment Authority.

Revenue: Standalone xAI: ~$500M annualized (end 2025). Consolidated with X: ~$3.8B annualized. For comparison: OpenAI ~$20B, Anthropic ~$10B. xAI is burning ~$1B per month. Management targets profitability by 2027.

Revenue sources: X advertising and subscriptions (~$3.3B), SuperGrok consumer subscriptions, API usage, government contracts ($200M Pentagon + GSA deal), Tesla purchases ($430M from Tesla for Megapacks in 2025).

Business model: Consumer subscription + API + government contracts, deeply integrated with Musk's ecosystem. X provides data and distribution. Tesla provides robotics data, a $2B investment, and the "Digital Optimus" joint project. SpaceX provides infrastructure and eventual IPO liquidity. Investment thesis described by VCs as "going long Elon" across the entire ecosystem.

Tesla-xAI conflict of interest: Tesla invested $2B in xAI while shareholders actively sue Musk for breach of fiduciary duty over founding xAI. The lawsuit alleges Musk diverted AI talent, Nvidia GPU shipments, and strategic opportunities from Tesla to xAI. Musk previously cited "conflict of interest" with Tesla as his reason for leaving OpenAI's board, then created xAI doing the same thing. At least 11 employees have been poached from Tesla to xAI.

DOGE conflict: Musk runs DOGE while selling AI to the government. Pentagon contract awarded days after the MechaHitler incident. GSA deal at $0.42/agency appears to be a loss-leader to embed Grok in government.

Safety investment: Essentially zero. No safety team. Safety adviser works for $1 with no equity. No funding of external safety research. No published safety research. The contrast: Anthropic allocates ~30% of its resources to safety.

Infrastructure costs: 200K+ GPUs at ~$30-40K each = $6-8B in GPU costs alone. ~1 GW power consumption. $80M wastewater plant. Capital expenditure on compute infrastructure dwarfs safety spending by orders of magnitude.

Corporate structure: Delaware C-corp (dropped PBC status May 2024). Now wholly owned subsidiary of SpaceX. No independent board known. No public governance structure. SpaceX IPO anticipated, creating pressure to demonstrate xAI's value.

What Others Say

AI Lab Watch (Sep 2025): "The RMF is profoundly unserious. xAI says dishonesty of 'less than 1 out of 2 on MASK' indicates acceptable loss-of-control risk for deployment; this is crazy. It claims to have good security (with no justification); this claim is not credible... xAI has only a few safety staff, and so it doesn't do safety research or have capacity to implement nontrivial safety interventions."

Zvi Mowshowitz (Feb 2026): "The safety situation at xAI seems quite bad. What used to be the safety team has left... Elon Musk is more confused than ever about alignment... More generally, xAI has been a commercial success in that the market is willing to fund it and Elon Musk was able to sell it to SpaceX, but it is a technological failure."

Boaz Barak (OpenAI/Harvard, Jul 2025): "The way safety was handled is completely irresponsible."

Samuel Marks (Anthropic, Jul 2025): Called xAI's lack of safety reporting "reckless" and a break from "industry best practices."

Former xAI employees (multiple, Feb 2026): "Safety is a dead org at xAI." "There is zero safety whatsoever in the company -- not in the image model, not in the chatbot." "He [Musk] actively is trying to make the model more unhinged because safety means censorship, in a sense, to him."

Stuart Russell (FLI index, Dec 2025): "AI CEOs claim they know how to build superhuman AI, yet none can show how they'll prevent us from losing control."

FLI Safety Index: D grade (1.17/4.0). Anthropic: C+ (2.67). OpenAI: C+ (2.31). DeepMind: C (2.08).

SaferAI: 18% risk management score. Risk modeling methodology: 2%. Red teaming integration: 0%.

Musk's defense: "Because everyone's job is safety. It's not some fake department with no power to assuage the concerns of outsiders. Tesla has no safety team and is the safest car. SpaceX has no safety team and has the safest rocket." Both claims are factually false -- Tesla has an Environmental, Health & Safety Team; SpaceX has flight safety and mission assurance teams.

Environmental criticism: State Rep. Justin Pearson: "Our lives and our lungs are being sacrificed on the altar of their capitalistic exploitation." University of Tennessee research found 79% increase in peak nitrogen dioxide near the facility. NAACP intent to sue for Clean Air Act violations.

What's Absent

No published research papers -- unique among frontier labs. Zero publications in any venue.
No safety team -- the function has been entirely eliminated. No safety researchers, no alignment researchers.
No independent board or governance -- Musk is the sole decision-maker with no checks.
No responsible scaling policy -- the RMF exists on paper but experts call it "profoundly unserious" and non-functional.
No pre-deployment safety testing -- Grok 4 released with zero guardrails. No red teaming disclosed.
No whistleblower protections -- former employees speak anonymously "due to fear of retaliation."
No external safety testing relationships -- silently removed reference to UK AISI from model card.
No incident response process -- CSAM scandal response was "Legacy Media Lies" auto-reply.
No alignment research agenda -- Musk discusses alignment philosophically but xAI invests nothing.
No deployment restraint -- no documented case of deciding NOT to release a model for safety reasons.