Grok AI Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: August 10, 2026 | License: CC BY 4.0
Grok reached approximately 117 million monthly active users as of March 31, 2026, according to Space Exploration Technologies Corp.’s IPO prospectus filed with the SEC. That figure is the first audited Grok user number ever published. It is also roughly double what independent trackers were estimating for the same period, and the gap is the most important fact in this dataset.
Quick Answer: What Are the Key Grok AI Statistics for 2026?
Grok had approximately 117 million monthly active users who used its AI features as of March 31, 2026, up from approximately 89 million on December 31, 2025 — a 31.5% increase in one quarter (SpaceX Form 424(b)(4), June 11, 2026). The AI segment that contains Grok and X generated $818 million of revenue in Q1 2026 against a $2,469 million loss from operations and $7,723 million of capital expenditure. According to Axis Intelligence Research, that works out to $348.44 of annualized capital and operating cost per Grok monthly active user — our GUCI reading, up 62.5% from the 2025 figure of $214.40.
Key Findings
- According to SpaceX’s Form 424(b)(4) filed June 11, 2026, approximately 117 million monthly active users used Grok’s AI features as of March 31, 2026, compared with approximately 89 million as of December 31, 2025.
- According to Axis Intelligence Research, Grok’s share of the combined Grok-and-X monthly active user base rose from 17.1% in December 2025 to 21.3% in March 2026, a gain of 4.2 percentage points in one quarter.
- SpaceX’s AI segment recorded a $6,355 million loss from operations on $3,201 million of revenue in 2025, and a $2,469 million loss on $818 million of revenue in the three months ended March 31, 2026.
- According to Axis Intelligence Research, SpaceX has contracted approximately $74.1 billion of external compute revenue across its Anthropic and Google cloud agreements — a figure within 1.2% of the $75.0 billion gross proceeds of its IPO.
- According to Axis Intelligence Research, the two compute agreements imply list prices of about $3,846 and $8,364 per NVIDIA GPU per month respectively, a 2.17x spread between two contracts signed weeks apart on the same infrastructure.
How Many People Use Grok in 2026?
The honest answer, until May 2026, was that nobody outside xAI knew. Every Grok user figure in circulation was a management claim, a leak, or a panel-based estimate from a traffic tracker. That changed when SpaceX filed its registration statement, because xAI became a wholly owned subsidiary and its operating metrics entered the SEC record. The final prospectus, SpaceX’s Form 424(b)(4) dated June 11, 2026, is now the only user figure for Grok that carries an auditor’s involvement and an issuer’s liability.
Grok Monthly Active Users, Audited Disclosures
| Metric | Dec 31, 2025 | Mar 31, 2026 | Change | Source |
|---|---|---|---|---|
| MAUs using Grok’s AI features | ~89 million | ~117 million | +31.5% | SpaceX Form 424(b)(4) |
| Combined Grok + X MAUs | ~520 million | ~550 million | +5.8% | SpaceX Form 424(b)(4) |
| Grok share of combined MAUs | 17.1% | 21.3% | +4.2 pts | Axis Intelligence Research calculation |
| Supported accounts (trailing 12 months) | ~1.3 billion | ~1.3 billion | flat | SpaceX Form 424(b)(4) |
The share row is ours. According to Axis Intelligence Research, Grok penetration of the X ecosystem moved from 17.1% (89 ÷ 520) to 21.3% (117 ÷ 550) between the two disclosed dates. Both inputs come from the same paragraph of the same filing, measured on the same methodology, which is the only condition under which a ratio like this means anything.
That methodology is worth reading rather than assuming. The prospectus defines an MAU as a registered user who interacted with Grok or X through a browser or mobile app at least once in the trailing 30 days, states that combined figures are deduplicated using sign-in traffic, and says the company distinguishes Grok users from non-Grok users based on the source of server requests. It also concedes that the deduplication may not catch every instance. A “supported account” is a separate and looser concept: any account that logged in, explicitly including bots and possibly fake accounts. Conflating the 1.3 billion supported-account figure with the 550 million MAU figure — which several summaries of the filing did — inflates the platform by a factor of 2.4.
Sarah Mitchell: The interesting number here is not 117 million. It is the 21.3% penetration rate, because it tells you what Grok actually is in 2026: a feature inside a social platform that a fifth of that platform’s users have touched in a given month, plus a standalone app. Every rival assistant has to buy or earn its distribution. Grok inherits it. The corollary is that Grok’s user curve is partly a function of X’s product decisions rather than model quality — which is why the penetration ratio, not the headline count, is the number to track quarter over quarter.
Why Third-Party Grok User Estimates Disagree by 4x
We are publishing the disagreement rather than resolving it, because the disagreement is the finding.
| Source type | Reported Grok users, early-to-mid 2026 | Basis |
|---|---|---|
| SEC filing (primary) | ~117 million MAU, March 2026 | Server-request classification, issuer-disclosed |
| Company statement via press (secondary) | ~64 million monthly users, September 2025 | Internal figure relayed to staff |
| Panel/traffic trackers (secondary) | roughly 30 million unique visitors | Clickstream panel extrapolation |
Aggregators including DemandSage, SEOProfy and Business of Apps published figures in the 30-to-64 million range for periods overlapping the filing’s 89-to-117 million window. The spread is not evidence that one side is lying. It is evidence that the four sources are counting four different things: web visitors to grok.com, standalone app users, in-X Grok interactions, and the union of all three. Panel-based trackers structurally cannot see Grok invocations that happen inside the X app, and that is precisely where most Grok usage lives.
The practical rule for anyone citing a Grok user number in 2026: state the surface. “117 million monthly active users of Grok’s AI features across Grok and X” is defensible. “117 million Grok app users” is not, and neither is “30 million Grok users.”
What Is the GUCI, and What Does It Say About Grok’s Economics?
GUCI (Grok Unit Cost Intensity) is an Axis Intelligence Research metric measuring the annualized capital and operating cost SpaceX’s AI segment absorbs per Grok monthly active user. It exists because Grok’s user growth and Grok’s cost curve are usually reported in separate paragraphs, and the relationship between them is the whole story.
Formula:
GUCI = (AI segment capital expenditure + AI segment loss from operations, annualized)
÷ Grok AI-feature monthly active users at period end
Inputs and readings, both from SpaceX’s Form 424(b)(4):
| Period | AI capex | AI loss from ops | Annualized total | Grok MAU | GUCI |
|---|---|---|---|---|---|
| FY2025 | $12,727M | $6,355M | $19,082M | 89M (Dec 31, 2025) | $214.40 |
| Q1 2026 | $7,723M | $2,469M | $40,768M | 117M (Mar 31, 2026) | $348.44 |
FY2025: (12,727 + 6,355) ÷ 89 = 214.40 Q1 2026: ((7,723 + 2,469) × 4) ÷ 117 = 348.44
According to Axis Intelligence Research, GUCI rose 62.5% between the two readings while the user base grew 31.5%. Cost intensity is outrunning adoption by roughly two to one. That is the baseline reading of this metric; there is no earlier history because no earlier Grok user figure was audited.
Set that against what the same users generate. AI segment revenue of $818 million across 117 million Grok MAUs is $6.99 per Grok MAU for the quarter — and that figure is generous to Grok, because the AI segment also contains X’s advertising and subscription revenue, which the filing does not break out. Quarterly revenue of $818 million covers 8.0% of the $10,192 million of capital expenditure and operating loss the segment absorbed in the same three months.
Sarah Mitchell: Every frontier lab is running a version of this gap, and every one of them tells you the gap closes when inference scales. What makes SpaceX’s version legible is the segment disclosure: you can see the capex line and the user line in the same document, on the same date, audited. Most labs give you a leaked ARR number and a valuation. Here the denominator is real, and it says each additional Grok user currently arrives attached to more incremental cost than the last one.
What this reading does not capture: GUCI charges the entire AI segment’s capital expenditure to Grok users, while that capex also builds the compute now being rented to third parties (below) and runs X. It is a measure of intensity at the segment level, not a per-user unit cost for Grok in isolation. Segment capex is also lumpy by nature; the Q1 2026 annualization assumes a spending run-rate that a single quarter cannot confirm. Read GUCI as a direction of travel, not a cost accounting result.
How Much Money Is Grok’s Compute Actually Making?
This is the part of the Grok story that has not been counted anywhere, and it does not come from a chatbot metric at all. It comes from two free writing prospectuses filed weeks apart.
In May 2026, SpaceX entered into Cloud Services Agreements with Anthropic PBC for access to approximately 325,000 NVIDIA GPUs across COLOSSUS and COLOSSUS II, at $1.25 billion per month through May 2029, with May and June 2026 billed at a reduced fee. On June 5, 2026, it signed a Cloud Service Agreement with Google LLC covering approximately 110,000 NVIDIA GPUs at $920 million per month from October 2026 through June 2029, disclosed in a free writing prospectus filed with the SEC.
Contracted External Compute Revenue
| Counterparty | GPUs | Monthly fee | Full-fee window | Contracted total | Source |
|---|---|---|---|---|---|
| Anthropic PBC | ~325,000 | $1,250M | Jul 2026 – May 2029 (35 mo) | ~$43.8B | SpaceX Form 424(b)(4) |
| Google LLC | ~110,000 | $920M | Oct 2026 – Jun 2029 (33 mo) | ~$30.4B | SpaceX free writing prospectus, June 2026 |
| Combined | ~435,000 | $2,170M | — | ~$74.1B | Axis Intelligence Research calculation |
According to Axis Intelligence Research, the two agreements together commit approximately $74.1 billion of external compute revenue — within 1.2% of the $75.0 billion gross proceeds SpaceX raised in the largest IPO on record. At full run-rate the pair produces $26.0 billion of annualized revenue, which is 8.1 times the AI segment’s entire 2025 revenue of $3,201 million.
And here is the number nobody has published. Divide each fee by its GPU count:
- Anthropic: $1,250M ÷ 325,000 = $3,846 per GPU per month
- Google: $920M ÷ 110,000 = $8,364 per GPU per month
- Spread: 2.17x
According to Axis Intelligence Research, the same infrastructure was contracted at implied unit prices differing by a factor of 2.17 in agreements signed roughly one month apart. Neither filing discloses what the bundled CPU, storage, networking or service-level terms are worth, and the Google deal carries a delivery condition — if the committed GPUs are not available by September 30, 2026, Google may terminate after a one-month grace period or accept a pro-rata reduction. Volume, term length, ramp structure and that delivery risk all plausibly explain part of the spread. But a 2.17x gap is large enough that anyone modeling AI compute pricing off a single disclosed contract is modeling a number that does not generalize, and our AI data center financing analysis shows the same variance appearing wherever contract terms become visible.
Sarah Mitchell: Read the two deals together and the strategic picture inverts. Grok is the reason the compute exists, but as of mid-2026 the compute has more contracted revenue than the model does. SpaceX built a gigawatt training cluster to win a model race and is now, in cash terms, primarily a landlord to two competing labs — one of which, Anthropic, it is competing against directly. That is not a criticism. It is the most rational use of stranded training capacity anyone in this industry has found, and it is why the interconnection-queue and power constraints we track in data center electricity consumption by state now show up on an income statement rather than an engineering roadmap.
What Is the Grok API Price List in 2026?
xAI’s published pricing documentation, last updated July 3, 2026 and retrieved August 10, 2026, lists the following text-model rates per one million tokens.
| Model | Context | Input | Cached input | Output | Long-context input | Long-context output |
|---|---|---|---|---|---|---|
| grok-4.5 | 500K | $2.00 | $0.30 | $6.00 | $4.00 | $12.00 |
| grok-4.3 | 1M | $1.25 | $0.20 | $2.50 | $2.50 | $5.00 |
| grok-4.20 (reasoning / non-reasoning / multi-agent) | 1M | $1.25 | $0.20 | $2.50 | $2.50 | $5.00 |
| grok-build-0.1 | 256K | $1.00 | $0.20 | $2.00 | $2.00 | $4.00 |
Source: xAI developer pricing documentation, retrieved August 10, 2026.
Two structural facts are worth extracting. First, the long-context multiplier is exactly 2.0x on every model and every token type, and it triggers at 200,000 tokens for all four — once a prompt crosses that threshold, the long-context rate applies to every token in the request, not just the excess. A 1M-token context window advertised at $1.25 is in practice a 200K window at $1.25 and a 1M window at $2.50. Second, cached input runs 84% to 85% below the cache-miss rate across the lineup, which makes prompt structure a larger cost lever than model choice for any repeated-context workload.
Grok Imagine, Voice and Tool Rates
| Service | Rate |
|---|---|
| grok-imagine-image output, 1K / 2K | $0.02 / $0.02 per image |
| grok-imagine-image-quality output, 1K / 2K | $0.05 / $0.07 per image |
| grok-imagine-video-1.5, 480p / 720p / 1080p | $0.08 / $0.14 / $0.25 per second |
| Speech to Speech (grok-voice-think-fast-2.0) | $0.08 per minute ($4.80/hr) |
| Text to Speech | $15.00 per 1M characters |
| Speech to Text (REST / streaming) | $0.10 / $0.20 per hour |
| web_search, x_search, code_execution | $5.00 per 1,000 calls |
| attachment_search | $10.00 per 1,000 calls |
| collections_search | $2.50 per 1,000 calls |
| Batch API discount (grok-4.3, grok-4.20 family) | 20% |
| Priority processing | 2x standard rates |
Source: xAI developer pricing documentation, retrieved August 10, 2026.
How Big Is Grok Imagine, and What Would It Cost at List Price?
SpaceX’s IPO factsheet, filed as a free writing prospectus on June 3, 2026, discloses that Imagine produces approximately 10 billion images and more than 2 billion videos per month, alongside the ~117 million Grok MAU figure.
Apply the published API list rate to the disclosed consumer volume and the scale becomes legible. According to Axis Intelligence Research, 10 billion images per month at the grok-imagine-image output rate of $0.02 per 1K image would represent $200 million per month, or $2.4 billion per year, of generation delivered at list value — against an AI segment that booked $3,201 million of total revenue across all of 2025.
Formula: 10,000,000,000 images × $0.02 = $200,000,000 per month.
This is an estimate of list-price equivalence, not a cost or a revenue figure. Consumer generations are not billed at API rates, xAI’s marginal cost of serving an image is not disclosed anywhere, and the video volume cannot be priced at all without a duration assumption the filing does not provide (at 480p and $0.08 per second, a six-second clip prices at $0.48, but the clip length is ours, not xAI’s, so we are not publishing a video total). What the calculation does establish is order of magnitude: image generation alone is running at a list-value scale comparable to the segment’s entire disclosed revenue.
How Fast Is xAI Shipping Models?
Since Grok-1 launched in November 2023, SpaceX states it has released four major model versions and notable variations thereof, and describes the cadence as one of the fastest in the industry. The prospectus also discloses that Grok-5 is currently being trained at COLOSSUS II and that next-generation training is expected to scale toward multiple trillions of parameters.
On capability, the filing makes one specific and narrow claim: within two years of its initial release, Grok achieved frontier-level performance in scientific reasoning as measured by GPQA Diamond, on a faster timeline than reported by other leading model providers. It names no score. Any Grok benchmark number circulating with a decimal point attached did not come from this filing, and our AI model vulnerability tracker documents how quickly published capability rankings decay between releases.
Compute Infrastructure, Disclosed
| Item | Disclosed value | Source |
|---|---|---|
| COLOSSUS + COLOSSUS II compute power | ~1.0 gigawatt | SpaceX Form 424(b)(4) |
| COLOSSUS first cluster, time to online | 122 days | SpaceX Form 424(b)(4) |
| COLOSSUS II first cluster, time to online | 91 days | SpaceX Form 424(b)(4) |
| Stated industry benchmark, 100MW greenfield data center | ~2 years | SpaceX Form 424(b)(4) |
| Next phase COLOSSUS II | 400MW+, 220,000+ additional GB300s | SpaceX free writing prospectus, June 2026 |
| Orbital AI compute satellites, target start | as early as 2028 | SpaceX Form 424(b)(4) |
The 122-day and 91-day figures are the company’s own and are cited against the company’s own stated industry benchmark, which is exactly the kind of comparison a registration statement is built to make favorable. The verifiable part is the direction: two clusters, the second faster than the first, both far inside conventional greenfield timelines that our AI data center construction cost analysis puts at multi-year norms.
What Regulatory Exposure Does Grok Carry?
Grok is one of seven consumer-facing chatbots under a live Federal Trade Commission study. On September 11, 2025, the FTC issued orders under Section 6(b) of the FTC Act to Alphabet, Character Technologies, Instagram, Meta Platforms, OpenAI OpCo, Snap and X.AI Corp., seeking information on how each company monetizes user engagement, develops and approves characters, tests for negative impacts before and after deployment, and complies with the COPPA Rule. The Commission voted 3-0. A 6(b) study carries regulatory force without initiating enforcement, and its findings can inform later policy or cases.
The prospectus itself discloses additional exposure: an Irish Data Protection Commission inquiry into Grok’s GDPR compliance, the FTC chatbot inquiry, and active class-action litigation involving misuse of AI products for non-consensual explicit imagery. It flags Imagine’s “Spicy” mode and “Unhinged” voice mode as carrying heightened risk. For a consumer assistant with 117 million monthly users, this is the risk register the issuer chose to write down in a document it is liable for — which makes it more informative than most safety commentary. Our AI safety statistics coverage tracks how these disclosures compare across labs.
Grok vs the Assistant Market: What Can and Cannot Be Verified
Third-party market-share figures for Grok circulate widely and none of them are primary. Reuters reported in February 2026, citing research firm Apptopia, that Grok’s U.S. chatbot market share reached 17.8% in January 2026, up from 14% in December 2025 and 1.9% in January 2025. Similarweb-derived estimates for grok.com monthly visits published by various outlets ranged from roughly 283 million to 314 million across early 2026.
We are naming these in plain text and not linking them, because they are panel estimates relayed through secondary reporting rather than published methodology documents. They belong in the record; they do not belong in a headline. According to Axis Intelligence Research, the only Grok usage figures currently defensible for citation in research or journalism are the two SEC-disclosed MAU readings and the ratios derived from them. Everything else is an estimate of a differently-defined population, and our broader AI statistics work applies the same evidence hierarchy across the assistant category.
Methodology
Collection. Every quantitative claim in this article was fetched from a primary document during production on August 10, 2026. The dominant source is SpaceX’s Form 424(b)(4) final prospectus dated June 11, 2026, filed under Registration No. 333-296070 on SEC EDGAR (CIK 1181412), supplemented by two free writing prospectuses filed under Rule 433 in June 2026, xAI’s developer pricing documentation, and the FTC’s September 2025 press release. Third-party tracker estimates are included only in the disagreement section, named in plain text, and flagged in the dataset as secondary estimates.
Formulas. GUCI = (AI segment capital expenditure + AI segment loss from operations, annualized) ÷ Grok AI-feature MAU at period end. Grok penetration = Grok AI-feature MAU ÷ combined Grok-and-X MAU, both measured on the same date and same disclosed methodology. Contracted compute totals = disclosed monthly fee × full-fee months stated in each filing. Implied GPU price = disclosed monthly fee ÷ disclosed GPU count. Imagine list-value = disclosed monthly image volume × published API output rate at 1K resolution.
Currency and dating. All figures in U.S. dollars. All fiscal figures as reported by the issuer. All user figures as of the date stated in the filing. Every row in the accompanying dataset carries its own as-of date and retrieval date.
Scope and caveats. GUCI attributes segment-wide capital expenditure to Grok users, and that capital also serves X and the third-party compute business; it measures intensity, not isolated unit cost. The Q1 2026 reading annualizes a single quarter of lumpy capital spending. Contracted compute totals assume both agreements run their full stated terms, and both contain 90-day termination provisions after an initial period. The Imagine calculation applies API list rates to consumer volumes that are not billed at those rates and is labeled an estimate throughout. Benchmark scores for Grok models are deliberately absent because the filing states a qualitative claim and no number.
About This Dataset
Title: Grok AI Statistics 2026 — Users, Segment Financials, Compute Contracts and API Pricing
Publisher: Axis Intelligence Research
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Format: CSV, UTF-8, one observation per row, 15 provenance columns
Temporal coverage: 2023-11-01 to 2026-08-10
Spatial coverage: Global, with United States regulatory scope
Citation: Axis Intelligence Research, Grok AI Statistics 2026, 2026.
APA: Axis Intelligence Research. (2026). Grok AI statistics 2026: Users, revenue, and compute economics from SEC filings. https://axis-intelligence.com/grok-ai-statistics/
MLA: Axis Intelligence Research. “Grok AI Statistics 2026: Users, Revenue, and Compute Economics from SEC Filings.” Axis Intelligence Research, 2026, axis-intelligence.com/grok-ai-statistics/.
Chicago: Axis Intelligence Research. “Grok AI Statistics 2026: Users, Revenue, and Compute Economics from SEC Filings.” Axis Intelligence Research. 2026. https://axis-intelligence.com/grok-ai-statistics/.
Frequently Asked Questions
Is the 117 million figure Grok app users or total Grok users?
Neither, precisely. The prospectus defines it as monthly active users who used Grok’s AI features, identified by the source of their server requests, across both the standalone Grok surfaces and X. It includes someone who invoked Grok inside the X app and never opened grok.com. Citing it as “Grok app users” overstates the standalone product; citing app-store or web-traffic figures as “Grok users” understates the real footprint by a wide margin.
Why do panel trackers report roughly 30 million Grok users when the SEC filing says 117 million?
Because clickstream panels measure web and app sessions on identifiable properties, and the majority of Grok invocations happen inside the X application where a panel cannot separate them from ordinary X activity. The two numbers are measuring different populations with different instruments. The filing figure is the broader definition and the only one with issuer liability attached.
Does the AI segment revenue figure tell you how much Grok earns?
No, and this is the most common misreading of the filing. The AI segment consolidates Grok subscriptions, the Grok API, X advertising and X subscriptions. SpaceX does not break out Grok-specific revenue anywhere in the prospectus. Any figure presented as “Grok revenue” for 2025 or Q1 2026 is a derivation someone made from an undisclosed split, not a disclosed number.
What does the 200,000-token threshold do to real Grok API costs?
It doubles them for the whole request. Once a prompt reaches the long-context threshold, the long-context rate bills every token in that request, not the tokens above 200K. For grok-4.3, a 210,000-token prompt costs $2.50 per million rather than $1.25 per million on all 210,000 tokens. Workloads that hover near the boundary are the ones where the pricing model bites hardest, and prompt compression below 200K is often worth more than switching models.
Why is xAI renting compute to Anthropic and Google if it needs it for Grok-5?
The prospectus states directly that capacity is sufficient for internal training and inference including Grok-5 at COLOSSUS II, and that the agreements permit reallocation of that capacity to internal use in the future. The economic logic is that gigawatt-scale clusters are built ahead of demand and idle capacity has no salvage value, so contracting it at $2.17 billion per month converts a stranded asset into a revenue line while the internal model business scales.
What explains the 2.17x gap between the Anthropic and Google GPU prices?
The filings do not say. Plausible contributors visible in the documents: different GPU counts (325,000 versus 110,000), different term structures, different ramp schedules, and a delivery condition in the Google agreement that lets Google terminate or take a pro-rata fee reduction if committed GPUs are not available by September 30, 2026. Neither filing itemizes the bundled CPU, storage and networking components, so the per-GPU figures are implied unit prices for a bundle, not GPU rental rates.
What are the fastest-moving Grok figures to re-check before citing?
MAU and segment financials, both of which will be restated at each quarterly filing, and API pricing, which xAI has changed repeatedly and versions with a “last updated” date on the docs page. The compute contract terms are the most stable, because they are fixed-fee agreements with disclosed end dates in 2029 — though both carry 90-day termination rights after an initial period.
Which Grok statistics should never be cited without a source date attached?
All of them, but especially model capability claims. The prospectus asserts frontier-level GPQA Diamond performance without publishing a score, and the model lineup has turned over four times since November 2023 with grok-4.5 now flagship and grok-4.3 still in the API. A Grok benchmark figure with no model ID and no date describes a model that may no longer be served.