SpaceX put 10 megawatts of solar power in space across 3000 gen1 Starlink satellites, then they put 100 megawatts in space with 7000 gen2. soon, they're doing 1000 megawatts with gen3. SpaceX is basically 10xing space solar every few years!

SpaceX’s 100,000-satellite Gen3 Starlink application has officially entered the FCC review process The FCC accepted SpaceX’s application for filing on September 18 SpaceX is requesting authority to deploy up to 100,000 Gen3 satellites across two very-low-Earth-orbit altitude groups: • 323–327.5 km • 473–477.5 km • Multiple orbital inclinations from 26° to 96.9° the scale is insane SpaceX is planning an entirely different scale of satellite internet...with the ambition to deliver far more capacity and lower-latency connectivity around the world Gen1 built Starlink Gen2 scaled it Gen3 is aiming at an entirely different scale

$SIVE looks like both a chokepoint and a bottleneck for CPO next year. Keep seeing information published from nontechnical people who miss any nuances. Here’s the reason why: 1. CW lasers are bottlenecked signaled by $LITE earnings. Laser fabs are heavily allocated to EML likely from former $NVDA contracts. -> Sumitomo/Furukawa = bottleneck -> Win Semi = bottleneck $SIVE does fab-lite, so are they a bottleneck? Yes, $SIVE sits in the laser bottleneck since control output supply of CW lasers from Win Semi and other fabs from allocation way early on (CEO stated they working with more capacity from other players as well). Perfect example is Kioxia/Sandisk. $SNDK controls NAND output, so they’re a bottleneck because they control final pricing. Demand exceeding supply from Ayar, Jabil, other pluggable vendors + Nvidia NVLink CPO ecosystem… final laser supply owned by $SIVE makes Sivers a bottleneck. $SIVE is also likely primary/sole source for Jabil, Gen-1 Ayar, $MRVL Celestial, and other hyperscaler asic/merchant CPO routes. So no way to get around it (can’t hot-swap single channel cw lasers with Sivers) 2. $SIVE is a chokepoint over CPO. $NVDA use $COHR, $LITE (which likely sources external cw capacity from Japanese competitors) $AVGO is likely vertically integrated as well. However: the entire ecosystem around it from ASIC programs (Marvell, AlChip, etc) and merchant programs (Ayar, Lightmatter, Lightelligence) Are all likely designed around $SIVE. Ayar for example, likely tried to multi-source with $MTSI / $LITE back in 2022 but their lasers probably couldn’t match the level of Sivers specification with arrays (removed Lumentum / Macom from their supply chain site recently) If there’s no alternative at least for the initial generations (obviously they’re working to multi-source). That makes $SIVE a structural chokepoint to go through for lasers. Even if you look at the 1.6T LRO $JBL designed, they achieved a “drastic moat” with performance built around $SIVE likely sole source. $SIVE is also the foundry level reference laser design for $GFS, which your hyperscalers use like $AMD (likely using Sivers + maybe Ayar for gen1): If every major player, who hasn’t achieved vertical integration (Nvidia/Broadcom) is using Sivers for CPO… That makes them a chokepoint. Just look at the entire CPO $NVDA NVLink ecosystem partners: every single one are all likely using Sivers. And they all use $GFS as well (where Sivers is default reference). So $SIVE is both a chokepoint and bottleneck when CPO really scales up H2 2027, over one of the biggest architectural shifts of all time (near $0 -> $81B or $91B TAM in the next 1 1/2 years from GS research note) This is why I say $SIVE looks like it could be the next $75B $LITE over the next couple years. All of this should play out next year. And it’s still trading less than a company with $50M in purchase agreements that buys Sivers lasers to repackage them.

Introducing Gen-1 Slides: an open model that matches Claude Opus 5 on slide generation, at ~1/17 of its input-token price. @genspark_ai post-trained it from a @MiniMax_AI M3 base with Fireworks Lab using long-horizon RL. Live today as the default in Genspark AI Slides. 🧵

Generalist AI just released GEN-1.5. It might be robotics' GPT-3 moment. What is one-shot learning via in-context prompting? In the case of language models like GPT-3, including a Q&A example in the prompt before the actual question improved performance across a broad suite of language tasks. For example: Prompt: "Q: Who wrote Romeo and Juliet? A: William Shakespeare Q: Who wrote War and Peace?" [model outputs "Leo Tolstoy"] Similar capabilities are now emerging in GEN-1.5, where a short example (a few seconds of demonstration ) alongside language and sensory inputs can solve tasks the model wasn't explicitly trained for. No gradient updates, just in-context examples unlocking new capabilities. The generalization comes from large-scale pretraining on real-world interaction data. It shows up in other adaptive behaviors too: novel tool use, and learning from human demonstrations. For decades, programming a robot took months and an expert. If showing it once is enough, that changes both how fast a robot becomes useful and who can work with one.

Seeing a hype wave around GEN-1.5, and rightfully so. Lots of respect to Pete & Andy for executing so well. The secret is in the naturally repetitive motions in human-collected data. There're 2 main sources for such repetitions: (1) Symmetric patterns. Sorting, tidying, and assembling almost never finish in one motion. Open any assembly manual from IKEA, and you find most objects symmetrical. You drive one bolt, then its twin, then the next pair. Every {bolt A, bolt B} pair is a natural continuation in context, and the second instance is a free training signal that imitates the first ("prompt"). (2) Recovery. Humans drop things all the time, but we pick them up so fast, we don’t even notice. That reflex to fix is half of our physical competence. The key insight is to keep the failed first half instead of trimming it away. If the model consumes the full arc, fumble, catch, continue, then recovery shows up organically at test time. It's funny that in-context improvement results from *NOT* over-sanitizing your data. The other critical ingredient is UMI. I've been saying for a while that teleop will not last, and GEN-1.5 is driving the final nail in the coffin. UMI is essentially a human wearing the robot gripper to collect data directly (human → data). Teleop inserts a layer of separation: human → VR/skeletal device → robot → data, which bleeds out all the human "physical intuition". The subtle sleight of hand we perform constantly with objects, the micro-adjustments, the feel of a part snapping into place, is nearly impossible to capture when you can't feel the environment directly. Once you have enough data, many behaviors can actually be zero-shot. For example, you don't even need finetuning to pick up a novel object. The model "just knows" what to do given a similar scene in the training distribution. Whether in-context learning truly works or not also depends on how far away the test is from training. Currently, the demos are still a bit too simple to conclude. I'm cautiously optimistic. Still, it's a great day in robotics.

STARLINK REAPPLIES FOR INDIA 🇮🇳 GEN 2 APPROVAL Starlink is seeking authorization for nearly 30,000 Gen 2 LEO satellites operating at 340-615 km, including direct-to-device connectivity. India previously approved only Starlink’s 4,408-satellite Gen 1 network after rejecting the earlier Gen 2 application. D2D rules and satellite spectrum pricing are still being finalized, with Jio, Amazon Leo and OneWeb also competing for the market. Source: ET

The Low-DK Fabric Roadmap for AI Hardware (Gen 1 to Gen 4): • Gen 1 (Glass): NV Compute Tray • Gen 2 (Glass): NV Switch Tray • Gen 2.5 (Extreme Glass / NEZ-Grade): The physical limit of glass fiber • Gen 3 (1st Gen Quartz / Q-Glass): NV Next-Gen LPU (52-layer) • Gen 4 (2nd Gen Quartz): M10-grade CCL Gen 1–2.5 push traditional glass to its limit. Gen 3–4 leap into pure Quartz for the AI race. ⚡️ #PCB# #CCL# #NVIDIA# #Hardware# #SupplyChain# #QGlass#

BREAKING: @GeneralistAI just shipped a one-shot model for robots! Called GEN-1.5, it is a robot foundation model from Generalist AI that takes in video (a 30-second memory window) plus language, proprioception and other sensors, and outputs 100 Hz action trajectories for manipulation. It was pretrained for over eight months on the company's proprietary data engine of real activities captured in homes, warehouses and factories (1,891,392 scenes), with no simulation data. It is the successor to GEN-0 and GEN-1, and the blog's claim is that GEN-1.5 shows one-shot and few-shot adaptation, compositional chaining, human-to-robot imitation and zero-shot sim transfer, none of which it was explicitly trained to do. Unfortunately, model size, robot hardware, inference latency, model weights, and pretraining data volume are not disclosed. Generalist states plainly that it made no architectural changes to promote in-context learning, no inner or outer meta-learning loop, and yet the model adapts from a single 3-to-12-second demonstration (59% one-shot average across 10 tasks). That is the actual GPT-3 moment ported to control: rather than engineering few-shot adaptation (the entire meta-learning literature), you scale pretraining and the adaptation falls out. It is the maximalist counter to the Patch Policy / @LerrelPinto / @ylecun "engineer the representation, keep it small" camp -> Generalist bets on emergence from scale, not architectural cleverness. Few-shot adaptation changes the weights by less than 0.15%, which means the model already contains the skill. Ten gradient steps on five minutes of data lift success to 83%, but move the weights under 0.15%, so adaptation is retrieving a capability the prior already holds, not learning a new one. That is the strongest version of the thesis: the prior is the moat and real robot data barely addresses it. Generalist's twist is that the nudge needed is almost nothing, which inverts the "we need robot data to adapt the model" premise under the VLA stacks. It pretrains on real data only, with zero simulation, yet claims zero-shot transfer to sim, the reverse of the usual direction! The data-strategy claim underneath is that broad-enough real pretraining subsumes sim as just another domain, rather than sim bootstrapping real, real-to-sim instead of sim-to-real. I wonder how did they built their data pipeline, about which they unfortunately do not provide details.

$SPCX Starlink is taking another shot at India with a new application for nearly 30,000 Gen 2 satellites including D2D service. India previously approved only its Gen 1 network as spectrum pricing and D2D rules are still being finalized with $AMZN Leo also competing.