A record kept in two places
Public discussion of xAI tends to fuse three separate things into one blur: how big a Grok model is, how much compute xAI has, and how fast Grok is improving. These are not the same claim, and they are not even the same kind of claim. One is a specific technical fact that a company either discloses or does not. Another is an infrastructure fact that can be checked against permits, satellite imagery, and utility filings independent of anything the company says. The third is a marketing narrative built from the other two, and it is where the loosest numbers tend to travel.
This article keeps the three apart and dates every claim in it. It covers xAI’s founding in 2023, the release history of the Grok model line from Grok-1 through Grok 4.6 — the most recent documented release at the time of writing — and the construction and expansion of the Colossus datacenter in Memphis, Tennessee, from site selection through its 2026 change of role. Where xAI has made a specific technical disclosure, such as Grok-1’s architecture, it is cited to xAI directly. Where a figure — most persistently, a parameter count for a recent Grok model — has only ever appeared as a leak or an unconfirmed report, it is labelled as such here and never stated as settled fact.
Founding, 2023: “understand the true nature of the universe”
xAI was announced publicly on July 12, 2023, in a post from Elon Musk, with a stated mission “to understand the true nature of the universe” [12]. The date was reportedly chosen deliberately: 7 plus 12 plus 23 sums to 42, the number Douglas Adams’ The Hitchhiker’s Guide to the Galaxy gives as the answer to life, the universe, and everything [12]. The founding team was small — reporting at the time put it at roughly a dozen researchers and engineers, drawn from Google DeepMind, OpenAI, Microsoft Research, Tesla, and academia, with Dan Hendrycks of the Center for AI Safety serving as an advisor [12]. xAI was structured as a company separate from Musk’s other ventures, including Tesla and X (then still branded Twitter), though the announcement noted it would work closely with them [12].
The company’s first public product moved quickly. By November 3, 2023, Musk was previewing an AI chatbot called Grok on X, with early access promised to a limited waitlist of United States users ahead of a wider rollout to X Premium+ subscribers once the product left early beta [13]. xAI itself later described this as “a very early beta product — the best we could do with two months of training,” a characterisation consistent with what the underlying model, Grok-1, actually was: a base model whose pretraining run had concluded in October 2023 [1]. The gap between “announced” and “shipped” matters throughout this history — xAI has consistently previewed products in public before the underlying model was finished, then followed with a separate, later disclosure once training closed.
Grok-1 to Grok-2: an early beta becomes a shipped product
The first model xAI disclosed in any technical detail was Grok-1, and the disclosure came four months after the chatbot’s public debut, not alongside it. On March 17, 2024, xAI open-released Grok-1’s weights and architecture under an Apache 2.0 licence, describing it directly as “a 314 billion parameter Mixture-of-Experts model” with twenty-five percent of the weights active on any given token, trained from scratch on a custom stack built on JAX and Rust [1]. xAI was explicit that this was the raw base-model checkpoint, not fine-tuned for dialogue or any other specific application [1]. This 314-billion-parameter figure is, notably, the only Grok-generation parameter count this article treats as an established fact — because it is the only one xAI has stated in its own technical disclosure rather than left to leaks and secondary reporting, a distinction that becomes important later.
Two weeks after that, on March 29, 2024, xAI announced Grok-1.5, describing “improved reasoning capabilities” and a context window expanded to 128,000 tokens — sixteen times the length of its predecessor — alongside reported results of 50.6% on the MATH benchmark, 90% on GSM8K, 74.1% on HumanEval, and 81.3% on MMLU [2]. These are xAI’s own reported figures rather than independently reproduced ones, and should be read as such.
Grok-2 followed almost five months later. xAI’s announcement, dated August 13, 2024, described Grok-2 as being “currently in beta on X,” with a smaller Grok-2 mini variant releasing alongside it and enterprise API access to follow later that month [3]. xAI’s own announcement claimed that an early version of the model — tested publicly ahead of the formal release — outperformed Anthropic’s Claude 3.5 Sonnet and OpenAI’s GPT-4-Turbo on the community-run LMSYS Chatbot Arena leaderboard [3]. That is xAI’s characterisation of a third-party leaderboard result, worth distinguishing from an independently audited benchmark run by a neutral party.
Grok-3 to Grok 4.6: reasoning, scale, and a widening gap between claim and confirmation
The next model marked a deliberate shift in how xAI talked about compute. Announced on February 17, 2025, and described by xAI as rolling out “in the coming days” as of February 19, xAI said Grok 3 was trained “on our Colossus supercluster with ten times the compute of previous state-of-the-art models,” adding that the company was “preparing to train even larger models on our 200,000 GPU cluster” [4]. This is the point at which Grok releases and the Colossus datacenter’s scale become explicitly linked in xAI’s own communications, rather than only in outside reporting.
Grok 4 arrived on July 9, 2025, again trained on the 200,000-GPU Colossus cluster, with xAI stating a sixfold improvement in compute efficiency achieved through what it described only in general terms as “new infrastructure and algorithmic work,” and reinforcement-learning training run at “over an order of magnitude more compute than had been used previously” for that purpose [5]. As with Grok 3, no parameter count accompanied the release.
Grok 4.1 followed on November 17, 2025, positioned by xAI primarily as an improvement to real-world usability rather than raw capability: the company reported a drop in hallucination rate on sampled production prompts from 12.09% to 4.22%, an LMArena Elo of 1483 for the reasoning (“Thinking”) variant and 1465 for the non-reasoning variant, and a 64.78% blind pairwise preference rate over the prior production model in xAI’s own evaluations [6]. Grok 4.1 became the default model across xAI’s consumer-facing surfaces on release.
Grok 4.5 shipped on July 16, 2026, trained, per xAI’s own announcement, “across tens of thousands of NVIDIA GB300 GPUs” and developed in part alongside Cursor’s coding-agent platform; xAI’s announcement again disclosed no parameter count [7]. Grok 4.6 followed on August 12, 2026 — the most recent documented release as of this writing — described by xAI as building on Grok 4.5 with a particular focus on long-running agentic work, coding, and knowledge tasks, priced from two dollars per million input tokens through Cursor, Grok Build, and the xAI API [8]. Independent benchmark trackers reported Grok 4.6 scoring 61 on the Artificial Analysis Intelligence Index, matching OpenAI’s GPT-5.6 Sol, though at least one outlet covering the release explicitly flagged that a separately cited 1753 Elo figure — attributed to Musk rather than to a public leaderboard — had not yet been independently confirmed on arenas such as LMSYS at the time of reporting [22]. That caveat, made in the reporting itself, is a useful model for how the figures in this section should generally be read: as claims with a named source, not as settled measurements, until a second, independent party reproduces them.
Site selection and the 122-day claim
Colossus’s location was decided quickly, following a search that was not. According to Forbes’ reporting, xAI evaluated seven or eight other sites before settling on Memphis, and Musk reportedly took roughly a week to decide following a period of intensive negotiation in March 2024 [14]. A corporate entity linked to the project was incorporated in Tennessee that same month, and by June 2024 the location was publicly confirmed: a former Electrolux manufacturing facility in southwest Memphis’s Boxtown neighbourhood, a plant Electrolux had opened in 2012 and closed in 2020, leased from real-estate firm Phoenix Investors [14]. Musk stated in May 2024 that he would personally oversee the supercomputer’s development [14].
What happened next is the claim most associated with Colossus: that the facility went from an empty industrial building to a functioning, single-fabric AI training cluster in 122 days. Musk stated this publicly on September 2, 2024, announcing that the system — 100,000 Nvidia H100 GPUs on a single RDMA fabric — had been brought online [16]. That claim did not rest on Musk’s word alone. NVIDIA, whose Spectrum-X Ethernet networking (SN5600 switches paired with BlueField-3 SuperNICs) underpinned the cluster’s fabric, corroborated the timeline in its own press materials dated October 28, 2024, describing the “supporting facility and state-of-the-art supercomputer” as having been “built by xAI and NVIDIA in just 122 days,” with only 19 days elapsing between the first rack rolling onto the floor and the start of training [15]. It is worth being precise about what kind of corroboration that is: NVIDIA was a commercial partner in the buildout, supplying and marketing the networking gear, not a disinterested auditor — its account should be read as a strong, technically detailed second source rather than a neutral third-party verification.
Doubling, funding, and a second Memphis site
xAI announced a doubling of the original cluster to 200,000 GPUs within roughly ninety days of the initial buildout, adding Nvidia H200s alongside the original H100s [16]. That kind of headline — “doubled in ninety-two days” — is easy to read two different ways, and the difference matters for judging how the next expansion phase should be interpreted. Read as a constant installation rate, doubling 100,000 GPUs to 200,000 over ninety-two days implies roughly
Read instead as a constant compounding rate — the assumption implicit whenever a “doubled again” claim is extrapolated forward — the same ninety-two-day doubling implies a daily multiplicative growth rate
about 0.76% per day. The two models diverge sharply once projected beyond the period they were actually measured over, and neither xAI’s announcements nor most press coverage specifies which one a given “doubling” claim is meant to imply. That ambiguity is worth naming explicitly rather than silently resolving in whichever direction makes the number more dramatic.
The buildout was capitalised in stages that track closely with the infrastructure milestones. xAI’s Series C round, announced December 23, 2024, raised six billion dollars and named its purpose explicitly: expanding Colossus from 100,000 to 200,000 Nvidia Hopper-generation GPUs and continuing the training run that became Grok 3 [10]. A second Memphis-area site followed. According to independent industry reporting, xAI acquired a one-million-square-foot warehouse plus two adjoining hundred-acre parcels on March 7, 2025, for what became Colossus 2, and by August 22, 2025, had installed 119 air-cooled chillers representing roughly 200 megawatts of cooling capacity on that site [17]. The same reporting characterised the pace as compressing roughly six months of buildout into a span that took competitors such as Oracle, Crusoe, and OpenAI closer to fifteen months for comparable projects — a comparative claim from that outlet’s own analysis, not an xAI disclosure, and one this article repeats as attributed commentary rather than as an independently adjudicated fact [17].
On December 30, 2025, Musk announced the acquisition of a third Memphis-area building, reported by trade press as bringing the combined site’s targeted capacity toward roughly two gigawatts [16]. Data Center Frontier’s reporting — dated January 14, 2026 and citing Mississippi economic-development officials — put the third facility (referred to in coverage by the informal name “MACROHARDRR”) in Southaven, DeSoto County, Mississippi, with operations expected to begin the following month and investment exceeding twenty billion dollars as disclosed by state officials, backed by sales, corporate income, franchise, and property tax incentives [16]. xAI’s Series E round, announced January 6, 2026 and raising twenty billion dollars from investors including Nvidia, Cisco, Fidelity, and Qatar Investment Authority among others, described the company’s aggregate infrastructure at that point as exceeding one million H100-GPU equivalents across its datacenters, with a further model — Grok 5 — then in training [11]. As of this writing, Grok 5 has not been released; it is discussed only as an unreleased, in-training project, not as a shipped model with confirmed specifications.
What independent reporting adds: power, permits, and a change of tenant
Speed claims about datacenter construction generally have a second half that vendor announcements do not volunteer: how the power was actually sourced. Independent and legal reporting on Colossus is where that half of the story is documented. According to Data Center Frontier’s account, xAI’s initial Memphis approval covered 150 megawatts from the Tennessee Valley Authority and Memphis Light, Gas and Water, with natural-gas generation deployed on-site as an interim measure while grid-connected capacity was built out in parallel [16]. That interim measure became a legal dispute. Earthjustice, representing the Mississippi and national NAACP, filed a federal Clean Air Act lawsuit on April 14, 2026, in the Northern District of Mississippi, alleging that xAI and its subsidiary MZX Tech were operating 27 gas turbines without an air permit at the Southaven site powering Colossus 2, with 41 further turbines proposed, following a 60-day notice of intent to sue delivered in February 2026 that the company did not resolve [20]. The suit’s estimated annual emissions figures for the unpermitted turbines include more than 1,700 tons of nitrogen oxides, 180 tons of fine particulate matter, and 19 tons of formaldehyde [20]. This is a litigant’s filing, not a court finding, and should be read as an allegation; it is included here because it documents, with dated specificity that vendor announcements do not provide, one mechanism by which such rapid buildouts secure power ahead of standard permitting timelines. Data Center Frontier’s reporting notes that a similar dispute over unpermitted turbines had already played out at the original Colossus 1 site, where the company reportedly removed unpermitted units after an earlier notice of intent to sue and obtained permits for the turbines it kept running [16].
Colossus 1’s role within xAI’s infrastructure also changed markedly by 2026. On May 6, 2026, xAI announced that Anthropic — a direct competitor in frontier AI — would use capacity at Colossus 1, by then described by xAI as containing “over 220,000” Nvidia GPUs across H100, H200, and GB200 generations, to expand serving capacity for Claude Pro and Claude Max subscribers, with the announcement also noting Anthropic’s interest in a further partnership on orbital AI compute capacity [9]. Independent technology reporting framed this as a consequence of Colossus 1’s mixed-generation hardware being poorly suited to training frontier models efficiently, with Colossus 2 — built on uniform Blackwell-generation hardware — taking over as xAI’s primary training cluster while Colossus 1 shifted toward inference work, including for a competitor’s product [21]. That reporting also cited an estimate that the arrangement could bring xAI close to six billion dollars in annual revenue from the deal, a third-party estimate rather than a disclosed contract figure [21].
A corporate epilogue: SpaceX, xAI, and a new name
The company’s corporate identity changed twice more in 2026, and both changes are documented well enough to include briefly, even though they sit slightly outside the Grok-and-Colossus throughline. SpaceX acquired xAI in an all-stock transaction reported as valuing SpaceX at roughly one trillion dollars and xAI at roughly 250 billion dollars, announced in early February 2026 and reported at the time as creating a combined entity worth approximately 1.25 trillion dollars, with Musk framing the deal around eventually developing space-based AI datacenter capacity [18]. Following SpaceX’s own initial public offering in June 2026, which reportedly raised seventy-five billion dollars at a valuation near 1.77 trillion dollars, the merged entity completed a rebrand on July 6, 2026, adopting the name SpaceXAI and a new logo while explicitly retaining the Grok product name for the chatbot, its mobile apps, SuperGrok, and the developer API [19]. For a history of xAI specifically, this is close to an endpoint: the corporate entity that announced itself on July 12, 2023, no longer exists as an independent company, even though the Grok model line and the Colossus infrastructure it trains on continue under the same product names.
What is confirmed and what is still a rumor
Set against this dated record, the single most persistent category of unverified claim about xAI concerns Grok model size. Grok-1’s 314-billion-parameter Mixture-of-Experts architecture is the one figure in this entire history that xAI itself disclosed in a technical announcement, complete with the twenty-five-percent active-weight detail that only makes sense as a genuine architectural disclosure rather than a marketing rounding [1]. No equivalent disclosure exists for Grok-2, Grok-3, Grok 4, Grok 4.1, Grok 4.5, or Grok 4.6 — each of those release announcements, read directly, discusses benchmarks, pricing, and context length, and is silent on parameter count [3] [4] [5] [6] [7] [8].
Into that silence, a specific figure has circulated widely for Grok 4.5 and Grok 4.6: a claim that both are built on a shared internal foundation model, informally referred to in leaks as “V9,” with a parameter count reported by multiple secondary outlets as 1.5 trillion — though at least one of those same reports notes an earlier, conflicting account putting the figure closer to 2 trillion [22]. This figure has never appeared in an xAI system-card-style disclosure of the kind that accompanied Grok-1; it originates in reporting describing brief, unofficial sightings of model names inside third-party developer tools, and in figures Musk has referenced without an accompanying technical disclosure. The inconsistency between the 1.5-trillion and 2-trillion versions of the same rumor is itself evidence that neither has been confirmed by a primary source — genuine disclosures do not usually come in two incompatible sizes. The same caution applies further out: reported parameter figures for the still-unreleased Grok 5, variously cited in trade coverage as six trillion or ten trillion, rest on the same kind of unconfirmed sourcing and are not treated as fact anywhere in this article.
The infrastructure side of the record is comparatively well corroborated, though not uniformly so. The 122-day initial build and 19-day GPU-to-training interval have independent, technically detailed corroboration from NVIDIA, an interested but well-informed party [15]. Site-selection facts — the number of sites considered, the week-long decision, the specific Electrolux building — come from Forbes’ independent reporting rather than xAI’s own account [14]. The clearest gap between claim and confirmation on the infrastructure side is in Colossus 2’s power figures: xAI and Musk’s public statements about gigawatt-scale capacity have been reported by industry trade press largely as targets and announcements, and the underlying legal filings over unpermitted generation are the most concrete documentary evidence available for how, specifically, some of that power was actually being produced at given points in time [20].
What the record actually supports
Three things can be said with confidence from primary sources and independently verified reporting. xAI was founded and announced on July 12, 2023, by a small team assembled largely from other frontier labs, and it shipped its first public chatbot within four months, ahead of any technical disclosure about the model underneath it. The Grok model line has released on a roughly semi-annual to quarterly cadence since November 2023, with xAI disclosing extensive benchmark and pricing detail at each release but a specific parameter count only once, for Grok-1, in March 2024. And the Colossus datacenter went from site selection in March 2024 to a operational 100,000-GPU training cluster by September 2024 on a timeline independently corroborated by a hardware partner, then expanded in stages — funded by named rounds, sited across two states, and partly powered, according to a federal lawsuit, by generation capacity built ahead of the permits required to run it — before its original facility was handed in part to a competitor for inference work in 2026, under a corporate parent that, by the middle of that year, was no longer called xAI at all.
What the record does not support is any specific parameter count for a Grok model released after Grok-1. Every such figure currently circulating is a rumor with a named, attributable, and inconsistent source — not a disclosure. Treating the two categories differently is not a stylistic preference; it is the only way to describe this company’s history without quietly converting someone’s guess into a fact along the way.
