Elon Musk Weekly Brief — 29 Oct–4 Nov 2026
\nOver the past week the connected Musk stack showed both concrete service expansion and the first clear public admission that agent-scale inference economics remain difficult. Tesla opened unsupervised Robotaxi service in a new city, SpaceX detailed the next large Colossus compute sites, Neuralink named its first two products while releasing early speech numbers, and Grok Bot users encountered sudden usage ceilings as Musk cited unsustainable serving costs relative to revenue.
\nSource: Not a Tesla App / Tesla Robotaxi account, 2 Nov 2026, https://www.notateslaapp.com/news/4932/tesla-expands-robotaxi-service-launches-unsupervised-rides-in-fort-worth
\nTesla’s official Robotaxi account announced that fully unsupervised rides (no safety monitor) became available in Fort Worth starting 6 a.m. local time on 2 November. The geofence covers roughly 143 square miles centered on downtown and extending to Saginaw, Benbrook, White Settlement and parts of the north and west sides. This is Tesla’s seventh city with unsupervised service and its fourth in Texas, after Austin, Dallas and Houston. As of early November the Texas DMV automated-vehicle registry listed approximately 1,244 Tesla vehicles (roughly 573 Cybercabs and the balance mostly Model Y), with Cybercabs now a substantial fraction of the registered fleet. Service remains geofenced, hours-limited in practice, and far smaller in absolute miles and fleet size than Waymo’s multi-city network.
\nWhy it matters: Each new city is incremental evidence that Tesla’s unsupervised stack can be replicated beyond the original Austin geofence. The growing Cybercab registration count shows the purpose-built vehicle is moving from prototype to fleet asset, even if most current miles are still logged on Model Y. Readers tracking the autonomy transition should note both the geographic expansion and the still-modest absolute scale relative to the company’s longer-term claims.
\nHorizon: NOW
\nEvidence grade: shipped
\nRead or watch: SKIM
\nCaveat: Registration does not equal active passenger-carrying vehicles; geofence boundaries and actual utilization rates are still limited.
\nCompany tag: Tesla
\nSource: Elon Musk on X / TechCrunch, 1–2 Nov 2026, https://techcrunch.com/2026/11/02/elon-musk-throttles-grok-bot-limits-due-to-unsustainably-high-costs/
\nOn 1 November Musk replied to users complaining that their weekly Grok Bot allowances were exhausting days early: “The cost of serving Grok Bot was unsustainably high relative to revenue.” Users on Premium and higher tiers reported meters jumping unexpectedly (sometimes overnight or within hours) and hitting 100 %. Grok Bot is the agentic product available on the $30+ plans; it is distinct from the base Grok chatbot. SpaceX’s most recent public figures showed strong AI revenue growth and a profitable Super Intelligence division, yet the company also disclosed heavy AI capital expenditure. No specific cost or revenue numbers for Grok Bot itself were released, nor was a clear statement given on whether limits were deliberately lowered or whether affected users would receive resets.
\nWhy it matters: This is the first direct acknowledgment from Musk that the unit economics of the flagship agent product are underwater at current pricing and usage intensity. Because agents consume more tokens than simple chat, rapid adoption can widen the gap between revenue and serving cost even while top-line AI numbers look strong. The episode is relevant to anyone watching how the SpaceX AI division (post-xAI absorption) balances growth against compute expense.
\nHorizon: NOW
\nEvidence grade: shipped
\nRead or watch: FULL
\nCaveat: No quantitative breakdown was provided; future pricing, advertising, or tier changes remain unspecified.
\nCompany tag: xAI/Grok | SpaceX
\nSource: Elon Musk on X / multiple reports, 31 Oct–1 Nov 2026, https://www.datacenterdynamics.com/en/news/spacexai-launching-two-additional-data-center-campuses-in-mississippi-to-further-expand-colossus-2/
\nMusk described two new Colossus 2 clusters in the Memphis/Mississippi area. Gun Range (formerly a police/fire training site) is planned for roughly 220 000 Nvidia GB300 GPUs configured for training with high-bandwidth networking, targeting operations by the end of November. Spam (a former Hormel warehouse) is planned for about 200 000 GB300s optimized for inference with simpler networking, targeting December. He stated that the broader Colossus 2 complex is on track for around 1.2 million Blackwell-class chips by year-end. The sites join existing Macrohard / Macroharder / Minihard clusters. Local air-quality litigation over turbine power generation at the complex continues.
\nWhy it matters: Compute capacity is the binding constraint on both Grok training/inference and any broader SpaceX AI services. Naming concrete sites, chip counts, and near-term dates moves the build-out from general aspiration to scheduled infrastructure. The inference-versus-training split also signals how the company intends to allocate capacity between model development and product serving.
\nHorizon: NEXT
\nEvidence grade: filing / primary claim
\nRead or watch: SKIM
\nCaveat: Timelines are Musk’s targets; actual power, cooling, and delivery constraints have previously shifted similar schedules.
\nCompany tag: SpaceX | xAI/Grok | Cross-stack
\nSource: Neuralink / Elon Musk on X, 28–30 Oct 2026, https://driveteslacanada.ca/news/elon-musk-neuralink-blindsight-ready-regulatory-approval/
\nNeuralink posted that one ALS participant in its VOICE study reached 75 words per minute at 5.8 % word-error rate on a 125 000-word vocabulary in the first week of speech-decoding use. The same period included a short engineering visit that prototyped automatic voice-activity detection, phone-call integration, iPhone dictation, and Grok-agent control via imagined speech. On 30 October Musk named the first product “Telepathy” (high-bandwidth thought-to-computer or thought-to-human communication) and the second “Blindsight” (vision restoration, including for those who have never seen). He described Blindsight as “ready from an engineering standpoint” and “just waiting on regulatory approval.” Both devices remain investigational; no commercial clearance has been granted.
\nWhy it matters: Concrete speed and accuracy numbers, even from a single participant, give a clearer picture of near-term clinical capability than earlier qualitative demos. Naming the products and pairing the implant with Grok agents tightens the link between Neuralink and the rest of the AI stack. Blindsight’s regulatory status is the next concrete gate.
\nHorizon: NEXT
\nEvidence grade: demo / primary claim
\nRead or watch: FULL
\nCaveat: Speech data are early, from one participant, and not yet in a peer-reviewed paper; Blindsight has no human vision-restoration results published.
\nCompany tag: Neuralink
\nSource: Tesla DEFA14A filing / VoteTesla.com, 29 Oct 2026, https://driveteslacanada.ca/news/tesla-shareholder-letter-musk-director-reelection-2026/
\nIn a letter filed with the SEC and posted on VoteTesla.com, Tesla’s board asked shareholders to vote for all proposals at the 3 December annual meeting, emphasizing the re-election of Class I directors Elon Musk, Chair Robyn Denholm and JB Straubel. The letter cites trailing-twelve-month revenue above $100 billion, more than 10 million vehicles delivered, 1.74 million paid FSD (Supervised) customers, and a ~75 % FSD attach rate on new North American deliveries in Q3. It does not discuss the separate SpaceX–Tesla special committees that are studying a possible combination.
\nWhy it matters: The meeting is the next formal governance checkpoint. Continuity of the three directors is presented as necessary for the “next phase of transformational growth” that includes Robotaxi, Optimus and energy. Any later merger discussion would also require board-level process.
\nHorizon: NOW
\nEvidence grade: filing
\nRead or watch: SKIM
\nCaveat: The letter is advocacy; voting outcomes and any subsequent strategic announcements are still ahead.
\nCompany tag: Tesla
\nSource: xAI / SpaceXSI developer docs, early November 2026, https://docs.x.ai/developers/advanced-api-usage/watermarking.md
\nFrom 7 November, text responses from grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the grok-4.20 family will include a statistical watermark by default. The watermark is invisible, applied during token selection, and detectable later by SpaceXSI; it does not alter the visible text or require client-side changes. Team admins can opt out per model (with a legal-exception declaration); some models are locked on. An x-watermark response header indicates whether a given completion was watermarked. The change is framed as compliance with EU AI Act transparency requirements.
Why it matters: It is a concrete product-surface change for every API user of the current frontier models. Downstream systems that consume Grok text (agents, code tools, structured outputs) will now carry a detectable provenance signal unless an admin opts out.
\nHorizon: NOW
\nEvidence grade: shipped
\nRead or watch: SKIM
\nCaveat: Opt-out process and enforcement details are still being exercised; watermark robustness against adversarial removal is not fully characterized in public docs.
\nCompany tag: xAI/Grok
\n