Tesla’s purpose-built Cybercab is hours from a livestreamed Austin launch, SpaceX has filed the first Starship mission that actually goes to orbit, and Grok is now a Pentagon desktop tool. The week was about putting existing stacks into live service—not new manifestos.
Cybercab joins a Robotaxi service that is already scaling in Texas
Source: WIRED, 2 Sep 2026, The Cybercab Is Almost Here; Teslarati, 31 Aug 2026, unsupervised fleet growth; Not a Tesla App, 27 Aug 2026, 79 new Texas registrations
Tesla will hold an invite-only Cybercab event at Gigafactory Texas on 3 September, livestreamed, framed as exclusive access for high-volume Robotaxi riders. The car is a two-seat, steering-wheel-free, pedal-free vehicle that uses the same Full Self-Driving stack already running on Model Y robotaxis. What changed this week is the supporting fleet, not a new driving model: Texas registrations jumped from about 191 to 270 Model Y robotaxis on 27 August, and Tesla began registering purpose-built Cybercabs with the Texas DMV, climbing from seven toward the low-to-mid forties by week’s end. Crowdsourced Robotaxi Tracker data and Tesla’s own “unsupervised fleet is a lot bigger” note point to a sharp rise in no-monitor rides across Austin, Dallas, and Houston, with service hours extended to 6 a.m.–10 p.m. Tesla has cited more than 380,000 unsupervised miles and “zero notable incidents” on the Q2 call; a separately filmed Austin robotaxi still shoved through plastic bollards. The event folds Cybercab into an existing commercial service in Texas and Florida. It is not NHTSA approval, not a consumer-sale start, and not a Bay Area driverless launch.
Why it matters: This is the first time Tesla puts a vehicle with no manual controls into a paying ride-hail product. The software is the same stack already on Model Ys; the hardware removes the last human fallback inside the cabin. Anyone tracking Tesla as an autonomy company should treat Thursday as a product-surface change, not a regulatory finish line. Scale still depends on geofence size, wait times, and whether Cybercab-specific data lets Tesla add units faster than the Model Y fleet.
Horizon: NOW
Evidence grade: demo / filing
Read or watch: FULL
Caveat: Registration is not the same as daily active cars. Many Cybercab test miles still used a safety monitor in the passenger seat. An NHTSA FSD crash probe opened in 2024 remains open.
Company tag: Tesla
Starship Flight 14 is filed as the program’s first orbital mission
Source: TeslaNorth, 1 Sep 2026, SpaceX Files Starship Flight 14 for First Orbit; KeepTrack X Report, 2 Sep 2026, Flight 14 orbital profile; Space.com, 2 Sep 2026, Falcon 9 Starlink 15-23
SpaceX’s FCC experimental paperwork for Starship Flight 14 describes an “orbital second stage” launching from Starbase, Texas, with a requested communications window starting 15 September 2026. Every prior Starship flight, including Flight 13’s 24 July Indian Ocean splashdown, flew a suborbital arc. The planned payload is roughly 20 operational Starlink V3 satellites—the same generation Flight 13 carried but could not leave in orbit because the ship never inserted. SpaceX rates each V3 unit on the order of 1 Tbps of downlink. Elon Musk told investors on the 4 August earnings call that Flight 14 would be the first operational-orbit attempt; he later walked back a same-flight tower catch of the ship to “a few months,” so this profile still points at an Indian Ocean splashdown rather than chopsticks. Separately, Falcon 9 continues the workhorse cadence: booster B1063 flew its 35th mission from Vandenberg on 2 September with 27 Starlink satellites and landed on Of Course I Still Love You. The constellation trackers put launched Starlinks near 12,900 and working satellites a little above 11,000.
Why it matters: Orbital insertion is the line between a reusable test article and a launcher that can actually grow Starlink. V3 satellites are too large for Falcon 9; only Starship can put that generation into service. If Flight 14 holds, Starlink capacity stops being limited by how many Falcon 9s SpaceX can fly and starts being limited by how fast Starship can turn around.
Horizon: NEXT (3–12 months)
Evidence grade: filing
Read or watch: FULL
Caveat: An FCC window is not an FAA launch license and is not a firm countdown. The ship-catch is explicitly deferred.
Company tag: SpaceX | Starlink | Starship
Grok for Government is now on the Pentagon’s desktop
Source: U.S. Department of War, 1 Sep 2026, ChatGPT, Grok Added to War Department’s AI System Options; ClearanceJobs, 2 Sep 2026, Pentagon Expands AI Arsenal
On 31 August the Department of War added Starshield AI’s Grok for Government to GenAI.mil, the department’s secure generative-AI portal that launched in December 2025 with Google’s Gemini for Government. OpenAI’s ChatGPT Mil landed the same day. Both tools are accredited for Controlled Unclassified Information at Impact Level 5. The department says GenAI.mil already has about 1.7 million users inside a roughly 3 million-person military-and-civilian workforce. Pentagon language assigns Grok workspaces, persistent projects, reusable “playbooks,” and adaptive reasoning modes (Auto, Fast, Expert) to tasks such as acquisition research and supply-chain management; ChatGPT Mil is described as the document-heavy planning, policy, logistics, and admin layer. This follows earlier defense contracts of up to $200 million each to OpenAI and SpaceXAI. The tools run in isolated environments that are not supposed to feed commercial training sets.
Why it matters: Grok is no longer only a consumer chatbot on X and in Teslas. Through Starshield it is now an enterprise assistant on the U.S. defense desktop, sitting next to Gemini and ChatGPT. That is a product-surface and distribution change for the SpaceX / xAI stack, and it is the first time Grok is framed as institutional knowledge infrastructure rather than a public feed.
Horizon: NOW
Evidence grade: shipped
Read or watch: SKIM
Caveat: This is unclassified / CUI work, not weapons control. “Starshield AI” is the defense wrapper; it does not mean Grok is flying missiles.
Company tag: xAI/Grok | SpaceX | Cross-stack
OpenAI will cut Cursor off its models after the SpaceX deal
Source: OpenAI, 28 Aug 2026, Our decision on Cursor following its acquisition by SpaceX; CNBC, 29 Aug 2026, OpenAI to end model access to Cursor
OpenAI notified SpaceX that it intends to wind down the contract that supplies OpenAI models to Cursor, with a proposed shutoff of 12 November 2026—the maximum notice in the change-of-control clause. SpaceX closed its acquisition of Cursor’s parent on 14 August in a deal reported at $60 billion. OpenAI’s stated reason is that it “cannot be confident that SpaceX will use our technology within our terms of service,” citing X’s post-acquisition contract breach and Musk’s sworn admission that xAI violated OpenAI’s terms. Future models, including Astra, will not go to Cursor. Cursor CEO Michael Truell said OpenAI models are about 5 percent of Cursor user traffic and that the companies are talking. Cursor remains the coding surface tied to Grok, Grok Build, Grok Bot, and SpaceXAI compute.
Why it matters: The $60 billion coding-tool purchase was supposed to give the Musk stack a default IDE and a path to other labs’ models. OpenAI is forcing that path closed. Cursor users keep current access into mid-November; after that the product is a Grok / SpaceXAI house tool unless another frontier lab backfills. The feud is now a procurement fact, not a quote-tweet.
Horizon: NEXT (3–12 months)
Evidence grade: shipped / filing
Read or watch: FULL
Caveat: The November date is proposed, not a completed outage. Five percent of traffic is small in aggregate and large for teams that standardized on GPT inside Cursor.
Company tag: xAI/Grok | SpaceX | Cross-stack
SpaceX will cast its own gas-turbine blades to pull AI power forward
Source: TechCrunch, 30 Aug 2026, Musk’s faster path to more gas turbines; Business Insider, 31 Aug 2026, SpaceX will build gas turbine parts
After The Information reported a “blades and vanes foundry” and an 830-acre land assembly near the Starlink factory in Bastrop, Texas, Musk confirmed the plan on 30 August. SpaceX and Tesla are each “building 100 GW/year of solar production capacity as fast as possible,” he wrote, “but natural gas will still be needed to supplement and bootstrap solar for several years.” The bottleneck is casting turbine blades and vanes. In-house casting at SpaceX, he said, can bring turbines online “up to 18 months” sooner. GE Vernova has said it is effectively sold out of gas-turbine production through 2030. SpaceX has already spent heavily on mobile and permanent turbines for Colossus-class compute, including the APR Energy acquisition and Doosan units. The foundry is a supply-chain play, not a utility.
Why it matters: Colossus, Grok training, Tesla inference, and eventual orbital-compute talk all run into the same constraint: watts on a schedule. Making the scarce metal parts in Bastrop is how the stack tries to buy time while solar factories ramp. It also locks Tesla energy and SpaceX manufacturing into the same power thesis.
Horizon: NEXT (3–12 months)
Evidence grade: rumor / demo (confirmed intent; foundry not producing)
Read or watch: SKIM
Caveat: Confirming a foundry is not the same as qualifying aerospace-grade turbine blades. Pollution and local permitting are the other constraint Musk did not solve in the post.
Company tag: SpaceX | Tesla | Cross-stack | xAI/Grok
Grok can now run the Tesla cabin—on Ryzen cars, over the network
Source: Tesla via Not a Tesla App, 1 Sep 2026, Grok Think Fast 2.0; Motor1, 2 Sep 2026, 116 voice commands; We Talk Tesla, 29 Aug 2026, FSD reunited with Summer Update
Firmware 2026.26.6.5 is bundling the Summer Update with FSD v14.3.8 (HW4) and v14.1 Lite (HW3). Tesla confirmed the in-car assistant is now Grok Think Fast 2.0. Release notes and Tesla’s support page say Grok can place calls, play music, adjust climate, search Controls, open the glovebox, and answer vehicle questions. Community tests logged 116 of 170 spoken commands without a tap. Chained requests work: one sentence can fold mirrors, raise wipers, and drop temperature. The feature needs the AMD Ryzen infotainment computer; Intel Atom cars are locked out. Grok in the car is rate-limited against the signed-in Grok account, requires Premium Connectivity or Wi-Fi, and has no on-device fallback—audio leaves the car for SpaceXAI inference. Grok does not command Full Self-Driving.
Why it matters: This is the first shipped cross-stack product where Grok actuates a Tesla, not just talks about one. It makes the cabin another Grok surface next to X, the API, Cursor, and GenAI.mil. The limit is hardware generation plus the network: older cars and dead connectivity get none of it.
Horizon: NOW
Evidence grade: shipped
Read or watch: SKIM
Caveat: Actuating climate and a glovebox is not vehicle-bus control of driving. Usage caps and connectivity make it a cloud feature bolted to the car.
Company tag: Tesla | xAI/Grok | Cross-stack
Watch next
- 3 Sep 2026 — Tesla Cybercab launch event, Gigafactory Texas (livestream).
- Mid-Sep 2026 (FCC window from 15 Sep) — Starship Flight 14 first orbital attempt / ~20 Starlink V3, pending FAA licensing.
- 12 Nov 2026 — Proposed OpenAI model cutoff inside Cursor.