xAI (SpaceXAI)

Last updated: 2026-10-16 ~04:00 (Europe/Zurich)

Ops 16 Oct ~04:00: Quiet — no material new open-web primary in window (Grok Voice 3.0 was Wed 14 — prior digest).

Ops 15 Oct ~04:00: Grok Voice 3.0 (x.ai/news Wed 14 Oct) — speech-to-speech; Artificial Analysis S2S Index 86.9 (#1; Grok Voice 2.0 81.3; GPT-Live-1 Astra 83.2; Gemini 3.8 Live 82.6); Big Bench Audio 98.5%; Full Duplex Bench 97%; τ-Voice 71.6% (~GPT-Live-1 at ~40% lower latency); EVA-Bench (ServiceNow) top accuracy 74%; sub-1 s latency. Powers millions of Tesla in-car conversations + Grok Bot voice mode; Starlink call centres ~20,000 calls/day, ~70% no-human. API $0.08/min (unchanged); grok-voice-latest auto-switched (x.ai/news; The Verge). The X Subscription held.

Ops 14 Oct ~04:00: The X Subscription launched Tue 13 Oct (x.ai/news) — one consumer plan bundling Grok (chat/voice/image/video) + Grok Bot + Cursor + X Premium; shared Grok-model usage pool across Grok and Cursor. Tiers: Free / Plus $8 / Premium $30 / Super $100 / Ultra $200 (Premium was ~$90 combined SuperGrok+Cursor Pro+X Premium+). Premium+ needed for latest Grok models. Existing SuperGrok/Cursor/X Premium can keep current plan or switch (prorated refund; gradual rollout). Teams/Enterprise: same per-seat + SSO/admin; third-party Cursor models still Premium+ at API list. (x.ai/news; TechCrunch).

Ops 12 Oct ~04:00: Grok Bot enabled in all Cybercabs + rest of Model Y fleet (Tsai; product step into Tesla cabins) (TeslaNorth). SpaceXSI rename held. No new x.ai/news post in-window.

Ops 11 Oct ~04:00: X account @SpaceXAI → @SpaceXSI (“SpaceX Super Intelligence”) live (TeslaNorth). Digital Optimus Diablo progress (Musk Sat) noted under Tesla/agent path.

Ops 8 Oct ~04:00: Grok Bot model routing — Musk Wed 7: “Important note regarding Grok @Bot: Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs. Whatever is most likely to give you the best outcome.” Per Gizmodo: simple questions → small fast models, complex → large. Grok Bot engineer Lauren (@poteto): “all bots will be powered by opus 5.5, and they can spawn cloud agents of whatever cursor models we have.” No product page; timing, model visibility, pricing, data terms unstated. Followed a ~2.5 h Grok Bot degradation Tue 6 (TNW; City AM; Gizmodo; TechCityAuthority). Colossus tenant: Musk “Yup” — Reflection AI’s Beam (501B / 23B active, open weights; pretrain 6,144 GB300, RL 10,500) trained on Colossus; Reflection pays $150M/month Jul 2026–2029, 90-day exit (The Inference; Reflection). Terafab: Musk “we will build and run the fab” (see tesla.md). No new x.ai/news post.

Ops 7 Oct ~04:00: No new x.ai/news post. SpaceX ~$40B chip-debt talks (Apollo/Pimco; 2027 close; FT 6 Oct). Gap 2 Oct: Grok in XChat group chats.

Ops 6 Oct ~04:00: No new x.ai/news post in window. Morgan Stanley (Jonas, Sun 4 note): SPCX price implies ~$32/share for AI (~3× 2028 sales, a neocloud multiple); SOTP $165/share enterprise AI after a 50% execution haircut; X + Grok $8; recent short-term neocloud (compute rental) contracts est. $30–50/W; Q3 earnings could give a first look at Cursor and Grok Bot economics; expects Grok 4.8, 4.9 and 5.0 before YE (analyst expectation, not company guidance) (Benzinga). CNBC: SpaceX earns heavily renting capacity to Google and Anthropic (CNBC 5 Oct). Terafab–TSMC: Musk Sat 3 “Just discussions, but something may come of it.” Gap 28 Sep: x.ai/news “Team Bots: shared AI teammates that learn as they work” (x.ai/news).

Ops 5 Oct ~04:00: Musk on X early Sun 4 Oct: “No more AI” / “SI” / “SpaceX is a super intelligence company”; asked whether SpaceXAI becomes SpaceXSI: “Yes, we will make that change.” No timeline; no formal company announcement; unclear whether legal or branding, or effect on @SpaceXAI / Grok naming. Would be third name since Feb (Teslarati 4 Oct).

Ops 3 Oct ~04:00: Experimental TypeScript SDK @xai-official/sdk announced Fri 2 Oct (pre-1.0; text/voice/image/video + hosted tools; Node ≥22.13) (RuntimeWire).

Status

xAI is now a SpaceX subsidiary. SpaceX announced the acquisition on 2 February 2026; the SpaceX S-1 later names the effective date of the xAI Merger as 2 February 2026 and the earlier xAI–X merger as 28 March 2025. Public product branding on x.ai is SpaceXAI. The current shipping flagship text model is Grok 4.7, released 21 September 2026 (https://x.ai/news/grok-4-7). Prior flagship Grok 4.6 shipped 12 August 2026.

Engineering snapshot (primary sources): - Grok 4.6: 500,000-token context; text + image in, text out; reasoning effort low / medium / high (default) / xhigh; knowledge cutoff 1 February 2026. API pricing $2 / $0.50 / $6 per 1M tokens (input / cached input / output) below 200k prompt tokens; $4 / $1 / $12 above. Rate limits: 150 rps, 50M tokens/min. Regions: us-east-1, us-west-2. Fast variant is twice the price (company announcement). - Grok 4.5 (16 July 2026): trained on “tens of thousands of NVIDIA GB300 GPUs”; coded/agentic/knowledge-work focus; trained alongside Cursor; company-claimed ~80 TPS and ~2× token efficiency vs leading rivals on SWE-bench Pro-style work. - Grok Bot (11 August 2026, early beta): persistent cloud-VM agents (browser, filesystem, terminal, computer-use). 26 August 2026: now included with SuperGrok / Plus / Heavy and Cursor Pro / Pro+ / Ultra / Teams (Standard and Premium); Bot usage separate from Grok/Cursor quotas. Enterprise waitlist. https://x.ai/news/grok-bot-more-plans - NVIDIA Vera / Starmind (24 Aug 2026, NVIDIA Newsroom; not on x.ai/news): SpaceXAI will deploy NVIDIA Vera CPUs for agentic orchestration; Grok factories on Vera Rubin “toward gigawatts”; planned first-generation Starmind AI satellite on an optimized Vera Rubin NVL72. NVIDIA release has no date/MW. Mike Nicolls quoted as SpaceXAI president. Musk on X 24 Aug 2026, quoting NVIDIA (TeslaNorth embed): space-optimized Vera Rubin NVL72 “for launch to orbit in Q4 next year, with significant scale in 2028” — Q4 2027 first orbit / scale 2028. https://nvidianews.nvidia.com/news/spacexai-adopts-nvidia-vera-cpu-to-accelerate-agentic-ai-at-massive-scale · https://teslanorth.com/2026/08/24/spacexai-nvidia-vera-cpu/ - NVIDIA FQ2 FY27 (26 Aug 2026): rev $96.2B (+106% YoY), Data Center $89.0B, non-GAAP EPS $2.22; Q3 guide $108.0B ±2% (no China DC compute). Earnings highlight restates SpaceXAI Vera CPUs. Kress on the call: Vera shipments already underway to lead partners including SpaceX AI (with OCI; AWS starting this quarter). Call did not restate Q4 2027 / 2028 orbit dates. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027 - Grok Build: coding agent (CLI/TUI, later web/mobile); grok-build-0.1 on API (May 2026, early access). Grok Build harness open-sourced 15 July 2026. - Imagine: Video 1.5 GA 16 June 2026 (grok-imagine-video-1.5); 1080p T2V/I2V later in July. Image 2.0 announced 7 August 2026. API list prices (docs): imagine-image $0.02, image-2.0 $0.04, image-quality $0.05; video $0.050/sec, video-1.5 $0.080/sec. - Voice: Speech-to-speech grok-voice-think-fast-2.0 (July 2026); grok-voice-latest routed to it from 5 August 2026. Custom voices (April 2026). - Grok on X: 4.5 rolled to grok.com, X, iOS, Android on 22 July 2026. Server-side x_search / web_search / code_execution tools on the xAI API. - Robotics: xAI does not ship a robot. Tesla has described Grok as a language/voice layer for Optimus / “Digital Optimus” / Macrohard; that is a Tesla product path, not an xAI hardware program. SpaceX S-1 defines Macrohard as an agentic digital-workflow platform in development with Tesla.

Compute (mix of company + S-1 + independent estimates): - Company Colossus page still markets the original 200,000 H100 single-cluster build (122 days, then doubled in 92 days) and a “roadmap to 1M GPUs.” Treat that page as historical marketing, not a live GPU census. - 6 May 2026 Anthropic partnership post: Colossus 1 has over 220,000 NVIDIA GPUs (H100, H200, GB200). - SpaceX S-1 (filed 1 June 2026): COLOSSUS = Paul R. Lowry Rd, Memphis; COLOSSUS II = Memphis + Southaven, MS, part of a “coherent gigawatt-scale AI training cluster”; company claims Colossus + Colossus II collectively provide ~1.0 GW of compute power. First Colossus cluster online in 122 days; first Colossus II cluster in 91 days. AI-segment capex $12.727B (FY2025) and $7.723B (Q1 2026). - Epoch AI tracker (updated 21 August 2026, estimate, not a company count): Colossus 2 ~110k B200 + 330k B300 ≈ 1,112k H100-eq, ~946 MW IT, ~$35.8B capex; Q1 2027 projection ~1.8M H100-eq / 1.53 GW. Independent model — mark as estimate.

Business: - Series E: $20B closed 6 January 2026 (upsized from $15B target). Named investors: Valor Equity Partners, Stepstone, Fidelity, QIA, MGX, Baron Capital; strategic NVIDIA and Cisco Investments. The post said Grok 5 “is currently in training” and Colossus I+II would end the year with over one million H100 GPU equivalents (forward-looking). - SpaceX S-1 AI segment (includes compute rental + Grok + X): FY2025 revenue $3.201B, loss from operations $(6.355)B, Segment Adj. EBITDA $(1.237)B. Q1 2026 revenue $818M, loss from operations $(2.469)B, Segment Adj. EBITDA $(609)M. - Cursor (Anysphere): SpaceX 8-K dated 14 August 2026 — merger effective that day; implied Cursor equity value $60.0B in SpaceX Class A stock (389,289,254 shares of common + additional RSUs/options). OpenAI, 28 August 2026: notified SpaceX it will wind down the contract providing OpenAI models to Cursor, proposed shutoff 12 November 2026 (maximum notice under a change-of-control clause); Cursor will not receive upcoming model Astra. OpenAI: “we cannot be confident that SpaceX will use our technology within our terms of service.” Reuters: Cursor co-founder Michael Truell (now SpaceX) said OpenAI models are ~5% of Cursor user traffic and talks are underway; Anthropic’s Tom Brown said Anthropic will increase compute for Claude in Cursor. Musk (Saturday, via Reuters): “I couldn’t care less.” No x.ai/news post. (OpenAI, 28 Aug) - Anthropic compute lease: company post 6 May 2026 (Colossus 1 access to expand Claude Pro/Max capacity; Anthropic also “expressed interest” in multi-GW orbital compute). Secondary S-1 coverage quotes ~$1.25B/month through May 2029, ~325k GPUs, 90-day termination after an initial period. Dollar terms not in the xAI news post; treat S-1 dollar figures as secondary until quoted from the filing MD&A. Google compute lease (~$920M/month) is secondary reporting, not in the xAI news post. - Combined X+Grok MAU (S-1): ~550M as of 31 March 2026 (520M at 31 Dec 2025); Grok-feature MAU 117M / 89M. Supported accounts ~1.3B TTM. Daily posts on X ~350 million (S-1). Paid subs ~6.3M (see X file).

Product — Grok 4.7 (21 Sep 2026)

SpaceXAI posted Introducing Grok 4.7 on x.ai/news (21 Sep 2026). Larger base vs 4.6; longer RL on multi-hour tasks; stronger self-verification and safeguards. Same list price as 4.6: $2 / $6 per 1M input/output (fast variant 2× price / 2× output speed). Available in Cursor, Grok Build, Grok API, harnesses, and cloud routers. Company table (selected): CursorBench 4.0 46.3% (4.6 40.4%; Fable 5.1 Max 51.8%); DeepSWE v1.1 71.0% high effort; EEBench 64.0%; AA Briefcase 1,657; Terminal-Bench 4.0 38.0%; Harvey Legal 19.6%; HealthBench Professional 56.7%*. Not Grok 5. Primary: https://x.ai/news/grok-4-7.

Product — Memory in Grok Build (16 Sep 2026)

SpaceXAI posted Memory in Grok Build on x.ai/news (Product · 16 Sep 2026). Grok Build keeps cross-session markdown notes (conventions, decisions, project facts); capture after completed turns; /memory, /dream, /flush. Off by default (GROK_MEMORY=1 or config). Not a Grok 5 / Colossus / orbital-compute date move. Primary: https://x.ai/news/grok-build-memory.

Also listed on the same index (gap since Aug 30 digest, not this 24h): Grok Bot for Enterprise (3 Sep), Grok Bot now works with X (1 Sep).

Product — Grok Bot in Tesla vehicles (22 Sep 2026)

Tesla rolled Connectors and Grok Bot into vehicles (22 Sep 2026): hands-free inbox/calendar/files/tasks plus multi-step errands. Agent features limited to SuperGrok Heavy (~$300/mo). Secondary: The Verge / Teslarati summarizing Tesla’s post; Bloomberg via Verge cited ~400k weekly Grok Bot users ~a month after Bot launch. Not an x.ai/news model ship. Sources: https://www.theverge.com/transportation/999043/tesla-gets-grok-bot · https://www.teslarati.com/tesla-integrates-grok-bot-ultimate-personal-assistant/.

Talent — Luhman brothers (23 Sep 2026)

Musk confirmed Wed 23 Sep 2026 that Troy Luhman and Eric Luhman (ex-OpenAI Sora research; Troy led Sora 2 audio; both co-authored the Sora world-simulators report) are joining SpaceXAI. Troy: long-term research. Secondary: TeslaNorth 23 Sep. Not an x.ai/news model ship.

Compute — Colossus +660k GB300 path (25 Sep 2026)

Musk 25 Sep 2026 (via Tom’s Hardware): Colossus 1 = 150k H100 + 50k H200 + 30k GB200; Colossus 2 = 110k GB200 + 440k GB300. Another 220k GB300 “fully operational next week,” 220k in November, 220k by late December “if we get lucky” → path to ~1.44M GPUs. Firm building 1.2 GW plant. Secondary coverage of X post — not an x.ai/news model ship. Source: https://www.tomshardware.com/tech-industry/data-centers/elon-musks-spacexai-to-add-another-660-000-ai-gpus-this-year-nearing-a-total-of-1-44-million-in-operation-firm-is-building-1-2-gigawatt-power-plant-to-bring-systems-fully-online.

Next milestone

Item	What is known	Confidence
Grok 4.7	Shipped 21 Sep 2026. Same $2/$6 as 4.6.	High
Grok 5	Confirmed “currently in training” on 6 Jan 2026. S-1 says next-gen training is expected to scale “toward multiple trillions of parameters.” No official release date, model card, or parameter count. Q1/Q2 2026 windows cited in secondary coverage have passed.	High that it is in training; low on date/size
Grok 4.6 distribution	Bedrock GA 19 August 2026; Microsoft Foundry public preview 26 Aug 2026 (500k context; low/medium/high/xhigh). Copilot / Gemini Enterprise Agent Platform / Vertex already on the news page. https://x.ai/news/grok-4-6-microsoft-foundry	High
Grok Bot GA / enterprise	26 Aug: included on SuperGrok + Cursor Pro/Teams plans; still early beta; enterprise waitlist. Chris Park (~18 Aug) X API reads still “launching soon” — no date.	Medium
Colossus 2 / 1M H100-eq	Series E YE target. 25 Sep Musk: +220k GB300 next week / Nov / late Dec → ~1.44M GPUs path (Tom’s Hardware).	Medium (Musk schedule vs energized count)
Orbital AI compute	Musk 24 Aug 2026 (quoting NVIDIA, TeslaNorth): space-optimized Vera Rubin NVL72 for Q4 2027 first orbit / significant scale 2028. NVIDIA release itself has no date. S-1 earlier intent: “as early as 2028.” Anthropic expressed interest. Still not a 2026 product.	Musk date (via TeslaNorth); S-1 still intent
Macrohard / Tesla overlap	S-1: in development with Tesla. Not an xAI ship date.	Uncertain timing

Competition

xAI’s own 12 August 2026 Grok 4.6 post compares against GPT-5.6 Sol Max (OpenAI) and Fable 5 Max (Anthropic) on AA Intelligence Index, GDPVal-AA, CursorBench 3.2, DeepSWE 1.1, FrontierCode 1.1, APEX-Agents, Terminal-Bench v3.0. Company chart: Grok 4.6 61 vs Fable 5 Max 62 vs GPT-5.6 Sol Max 61 vs Grok 4.5 High 56 on AA Intelligence Index. These are xAI-published figures using third-party or self-reported competitor scores — not an independent bake-off.

Rival	2026 posture vs xAI
OpenAI	Still the consumer/API incumbent. xAI charts put GPT-5.6 Sol Max roughly tied on the composite index, ahead on some SWE/agent benches (DeepSWE, Terminal-Bench).
Anthropic	Coding/agent leader in xAI’s own DeepSWE/FrontierCode/APEX numbers (Fable 5 Max). Unusual customer of Colossus 1 — competitor and tenant.
Google DeepMind	Distribution via Vertex / Google Enterprise Agent Platform (xAI headlines). Gemini family not in the 4.6 score table.
Meta	Open-weights Llama line; not in xAI’s 4.6 published table. Compete on open models + infra, not on Grok consumer UX.

Positioning: xAI is competing on (1) GB300-scale training + cheap frontier API ($2/$6), (2) Cursor-native coding agents after the $60B close, (3) multimodal Imagine/Voice, (4) X real-time data, (5) selling spare Colossus capacity. Weakness the S-1 itself flags: AI segment is heavily loss-making and capex-intensive; Grok 5 has slipped relative to early-2026 chatter.

Key people

Person	Role	Notes
Elon Musk	Founder; SpaceX CEO/CTO/Chairman	Controls the parent after the Feb 2026 merger.
Michael Nicolls	President of xAI/SpaceXAI (reported Apr 2026)	Former Starlink exec. Title from Bloomberg / press, not an x.ai leadership page.
Bret Johnsen	SpaceX CFO	Signed the 14 Aug 2026 Cursor 8-K; S-1 finance.
Gwynne Shotwell	SpaceX President & COO	Press reports ops overlap with the AI division after the merger. Not independently confirmed on x.ai.
Michael Truell	Cursor co-founder; now SpaceX	Reuters 28–29 Aug: OpenAI models ~5% of Cursor traffic; talks with OpenAI on the wind-down.

Several xAI research co-founders (Igor Babuschkin and others) are reported to have left in 2026. Personnel churn is secondary reporting.

Sources

Secondary / estimates (do not treat as company numbers) - Epoch AI Colossus 2 tracker (updated 21 Aug 2026): https://epoch.ai/data/ai-data-centers/directory/colossus-2 - Bloomberg (Nicolls; Colossus 1 lease rationale); TechCrunch 15 Aug 2026 (Cursor close) - TeslaNorth 24 Aug 2026 (embeds Musk quoting NVIDIA; not a company filing): https://teslanorth.com/2026/08/24/spacexai-nvidia-vera-cpu/ - Community Grok 5 6T/10T parameter claims — unconfirmed