Starship’s first planned orbital hop was pulled off a Sept. 22 FAA airspace window and reset to no earlier than Sept. 28, still carrying 26 production Starlink V3 satellites. In the same six days Samsung started 2 nm trial wafers of Tesla’s AI5 chip at Taylor, Neuralink posted a VOICE-trial clip of reconstructed speech, and xAI put a voice on Grok Bot while Musk mapped Grok 4.8 through Grok 5. The stack’s legal and ground-network tracks moved too: X and SpaceXAI dropped Apple from the App Store antitrust case, Vegas Loop said it will double operating stations by year-end, and Tesla locked an Oct. 1 Roadster reveal in Waco.
\nSource: SpaceX on X, 17 Sep 2026 + SpaceX; FAA advisory coverage, 15 Sep 2026 + TeslaNorth; The Verge, 16 Sep 2026
\nSpaceX said Thursday it is now targeting Starship Flight 14 as early as Monday, Sept. 28, pending regulatory approval — a roughly six-day slip from the Sept. 22 window that had been written into FAA airspace planning. The mission profile did not shrink with the date. Ship 41 is still meant to reach a real orbit for the first time (~275 km, about six revolutions, ~10 hours), deploy 26 operational Starlink V3 satellites (Starlink Group 31-1), then deorbit to a Pacific splashdown west of Chile. Super Heavy (Booster 21) is not attempting a tower catch; it is aimed at a Gulf of Mexico splashdown while SpaceX tests fixes after Flight 13’s landing-burn failure. Each V3 bird is described as ~1 Tbps of downlink, so one flight would add on the order of 26 Tbps — a few percent of constellation capacity, and payload Falcon 9 cannot lift in this form. SpaceX CFO Bret Johnsen has already called this a revenue flight, not another empty dress rehearsal. Limits are still large: license conditions are not closed in public, the ship only circularizes if hardware needed for the later deorbit burn looks healthy, and neither stage is being recovered to the pad.
\nWhy it matters: This is the week Starship stopped being only a test article in the public schedule and became a dated attempt to put paying satellites on a stable orbit. If it works, Starlink V3 cadence is no longer gated solely by Falcon 9 mass and volume. If it slips again or fails the circularization/deorbit logic, the Florida Starship pads and the V3 density story both wait.
\nHorizon: NEXT (days to weeks for the attempt; 3–12 months for cadence)
\nEvidence grade: filing / company schedule (not yet flown)
\nRead or watch: FULL if you track Starlink capacity or HLS; SKIM the date-only posts
\nCaveat: NET dates on this vehicle have already moved through mid-September. An FAA airspace window is not a launch license closeout.
\nCompany tag: SpaceX | Starlink | starship
\nSource: Business Korea, 16 Sep 2026; Not a Tesla App, 17 Sep 2026; Drive Tesla Canada, 17 Sep 2026
\nKorean industry reporting, dated around Sept. 15–16, says Samsung has begun 2-nanometer wafer work at its Taylor, Texas foundry and entered prototype production of Tesla’s AI5 inference chip — months ahead of the plant’s previously advertised November-ish start. The same reports tie the pull-forward to Tesla’s roughly $16.5 billion foundry contract and to packed 2 nm lines in Korea. Samsung’s stated aim is to finish mass-production verification by the end of 2026 and begin earnest supply in 2027. That is not a customer-car milestone. Musk has already said AI4 has the headroom for unsupervised driving, and that AI5 is meant first for Optimus and Tesla training clusters, then vehicles when AI4 is old enough to stop building. Trial wafers test yield, not a shipping computer; a bad yield curve still delays 2027 volume.
\nWhy it matters: Every later Optimus, Cybercab, and in-house training rack in this stack is a function of inference silicon, not just software. Taylor is the U.S. foundry Tesla actually contracted, so moving from tape-out to physical 2 nm lots is the first factory-floor signal that AI5 is a real 2027 part rather than a slide.
\nHorizon: NEXT (yield work now; volume 2027)
\nEvidence grade: rumor / industry sourcing (not a Tesla 8-K)
\nRead or watch: SKIM unless you follow hardware; SKIP the 5×-performance recaps
\nCaveat: No Tesla IR confirmation this week. “Trial production” is yield learning, not parts in Optimus torsos.
\nCompany tag: Tesla | Cross-stack
\nSource: Neuralink on X, 16 Sep 2026; TeslaNorth, 17 Sep 2026
\nNeuralink published a 23-second clinical clip on Sept. 16 under the line “A beautiful surprise,” showing a trial participant produce audible speech — including “I love you” — from intended-speech signals decoded through the implant path. Coverage identifies the VOICE program participant as Kenneth Shock, implanted in January 2026 after ALS took his speaking voice; the company itself treated the clip as a participant moment, not a named case study. The on-screen disclaimer is the operational fact: devices remain investigational, are not FDA-approved, and individual results may not match future outcomes. This is a different product lane from Telepathy cursor control. VOICE is trying to turn motor-speech intention into text or a reconstructed voice. A polished 23-second video does not publish bit-rate, latency, vocabulary size, or session reliability.
\nWhy it matters: For anyone tracking Neuralink as more than a cursor demo, speech is the first capability that changes daily life for people who can no longer talk. The clip is evidence the decoding stack can carry a short, high-stakes phrase in a controlled setting. It is not evidence the system is a product, a therapy, or a general speech prosthesis.
\nHorizon: NEXT (more trial footage and enrollment); SPECULATIVE for general availability
\nEvidence grade: demo (trial video)
\nRead or watch: FULL the 23-second clip; SKIP the “mind reading” writeups
\nCaveat: Investigational only. One emotional phrase is not a speech benchmark.
\nCompany tag: Neuralink
\nSource: Elon Musk on X, 14 Sep 2026 + TeslaNorth; Elon Musk on X, 17 Sep 2026 (“Grok @Bot now has a voice”); Grok Build notes, 16 Sep 2026
\nOn Sept. 14 Musk said Grok 4.8 — described as a 2.5-trillion-parameter model trained on xAI’s new C++ stack — would finish training that week and start reinforcement learning. He ranked Grok 4.7 as roughly Opus 5.0 class with multimodal still weak, called 4.8 a noticeable step, put 4.9 in “Astra/Fable” territory, and answered an AGI question by pointing at Grok 5. Grok 4.7 itself had already missed a Sept. 12 public slot; Musk had blamed RL that punished response length and made the model quit hard tasks early. Separately, on Sept. 16–17 Grok Build shipped memory as generally available and a reliability pass (MCP structured JSON, cancelled tools actually stop, session checkpoints kept), and Musk confirmed Grok Bot now speaks. None of that is a public Grok 4.8 API card. Training-complete is not eval-complete, and AGI talk remains a calendar placeholder.
\nWhy it matters: After SpaceX absorbed xAI, Grok is no longer a side chatbot. It is the language layer being wired into Tesla cabins, X, and enterprise agents. A voice-capable Bot plus a dated 4.8 training handoff is the product-surface change this week; the 4.9/5 ladder is only a map.
\nHorizon: NOW for Bot voice and Build memory; NEXT for 4.8 after RL
\nEvidence grade: shipped (Bot voice, Build); company claim (4.8 training)
\nRead or watch: SKIM Musk’s ranking thread; SKIP AGI commentary
\nCaveat: No independent evals on 4.8. 4.7 still had not shipped as a public model ID when 4.8 was named.
\nCompany tag: xAI/Grok | X | Cross-stack
\nSource: Reuters, 14 Sep 2026; 9to5Mac, 14 Sep 2026; POLITICO, 15 Sep 2026
\nOn Sept. 14, X Corp and SpaceXAI asked a Fort Worth federal court to dismiss their claims against Apple with prejudice, saying those claims had been “resolved,” while expressly continuing the case against OpenAI. The 2025 suit alleged Apple’s ChatGPT integration into Apple Intelligence and App Store treatment shut out rivals such as Grok. Terms of the Apple resolution were not filed. By Sept. 15–16 Judge Mark Pittman ordered the Musk companies to produce any Apple agreements after OpenAI argued the hidden deal could affect its defense. A with-prejudice drop means those Apple claims cannot be refiled. It does not put Grok inside Siri, and it does not end the OpenAI fight, which still has a 2027 trial setting on the public docket.
\nWhy it matters: Grok’s consumer surface still runs through phone OS defaults and store ranking. Clearing Apple while leaving OpenAI on the hook is a distribution and legal split: one platform fight closed, the model-competitor fight not.
\nHorizon: NOW for Apple; NEXT for whatever the judge forces into the record
\nEvidence grade: filing
\nRead or watch: SKIM the motion; SKIP settlement speculation
\nCaveat: No public term sheet. “Resolved” is not the same as an announced Grok-on-iPhone integration.
\nCompany tag: X | xAI/Grok | Cross-stack
\nSource: Tesla on X via TeslaNorth, 14 Sep 2026
\nOn Sept. 14 Tesla amplified an owner clip of Grok voice commands plus Full Self-Driving Supervised running an errand and returning home without a hand on the wheel, writing that “FSD Supervised w/ Grok is a lot like having your own Robotaxi.” That is a product-positioning change, not a regulatory one. Hardware 4 cars were on FSD Supervised v14.3.7; Hardware 3 remained on a lighter v14.1 stack, so the combo is build- and market-gated. Supervised still means a responsible human in the seat. It is not the Austin Cybercab service (no wheel, no pedals, geofenced, already under an NHTSA certification audit opened Sept. 3). Tesla is documenting more Grok jobs in the cabin — climate, media, now text messaging in the Model Y manual — which is how the company wants owned cars to feel like robotaxis before unsupervised status widens.
\nWhy it matters: The stack’s autonomy story this year is split: purpose-built Cybercabs in a few Texas/Florida geofences, and millions of owned cars that still need a supervisor. Wiring Grok into that second fleet is how Tesla tries to make the owned car feel like the network product without waiting for every regulator.
\nHorizon: NOW as a software experience; NEXT for unsupervised expansion
\nEvidence grade: shipped (supervised software + official post)
\nRead or watch: SKIM
\nCaveat: Official slogan ≠ unsupervised operation. NHTSA’s Cybercab audit is a separate constraint on the wheel-less vehicle.
\nCompany tag: Tesla | fsd | xAI/Grok
\nSource: The Boring Company on X, 14 Sep 2026; Teslarati, 14 Sep 2026; FOX5 Vegas, 14 Sep 2026
\nThe Boring Company posted on Sept. 14 that “Vegas Loop is getting bigger – the station count will double by end of year,” and attached urgent listings for Loop drivers and a senior ops manager. Teslarati’s count puts the live system at 14 stations with 28 as the end-2026 operating target; Clark County has already entitled a much larger 123-station map. FOX5 reported tunnels pushing south toward Paradise Road, with Virgin Hotels and a Paradise/Tropicana stop among the near-term openings, timed against November’s Formula 1 weekend. That is a claim about stations that collect fares, not another entitlement vote. The system still uses human-driven Teslas in tunnels. Airport service remains incomplete.
\nWhy it matters: Vegas is the only Boring network that actually moves civilians every day. Doubling live stations is the first operating-scale test of whether Prufrock plus Tesla vehicles can look like transit before Dubai and Nashville are more than holes in the ground.
\nHorizon: NEXT (by 31 Dec 2026)
\nEvidence grade: company claim
\nRead or watch: SKIM
\nCaveat: Entitled stations ≠ open stations. F1 is a hard calendar; the company already said dates can move.
\nCompany tag: Boring
\nSource: Tesla on X, 12 Sep 2026; Reuters, 12 Sep 2026; Electrek, 12 Sep 2026
\nTesla posted a “Go for launch” teaser on Sept. 12 and put a countdown on the Roadster page pointing at Oct. 1. Reservation-holder invites specify Waco, Texas, that evening. Reporting around the invite describes a possible demonstration of the long-promised SpaceX cold-gas thruster package — remotely, with spectators held well back — not a start-of-production date. Musk has already told shareholders production is a 2027–2028 problem. A reveal after nine years of slips is a product-surface event only if something physically leaves the ground or a locked spec sheet ships. Until then it is a dated demo with a cross-company propulsion footnote.
\nWhy it matters: The thruster story is one of the few consumer products that literally welds Tesla and SpaceX hardware together in public. The date is now on a calendar. Production is not.
\nHorizon: NEXT (Oct. 1 event); SPECULATIVE for series production
\nEvidence grade: company event listing / teaser
\nRead or watch: SKIM until the night of
\nCaveat: Tesla has missed Roadster dates since 2020. Hover footage, if any, will be a stunt envelope, not a street-legal flying car.
\nCompany tag: Tesla | Cross-stack
\nSource: Elon Musk on X, 17 Sep 2026 (quoting Starlink); Tesla Megapack on X, 18 Sep 2026, amplified by Musk; Starlink hardware notes from the V5 reveal
\nMusk stated Sept. 17 that the Starlink V5 residential terminal “is in production,” quoting Starlink’s own hardware post. V5 is the smaller, lighter, lower-draw dish already shown in July (about 1.1 kg, roughly half V4’s average power, 375+ Mbps rated, select-area availability). Production is the difference between a reveal kit and a volume SKU; geography still follows the factory, not the tweet. The same 24 hours, Tesla Megapack said Australia’s first 8-hour long-duration battery is operating at Limondale, NSW — 144 Megapacks for RWE, charging twice as fast as they discharge so daytime solar can cover evening and overnight peaks. That is a shipped energy product, not a slide about “long duration someday.”
\nWhy it matters: Starlink’s constraint is no longer only satellites; terminals and power draw decide who can install. Megapack duration is the other half of the same infrastructure thesis: store cheap electrons long enough to run compute, factories, and grids after sunset.
\nHorizon: NOW (first sites and early V5 units); NEXT for wide V5 availability
\nEvidence grade: shipped (Limondale); company claim (V5 production)
\nRead or watch: SKIM
\nCaveat: V5 is not global. One 8-hour Australian site does not make 8-hour duration the default Megapack SKU.
\nCompany tag: Starlink | Tesla
\n