Just tested DeepSeek V4 Pro + OpenCode, already better than 5.3-Codex in engineering design, performance advice, and code quality. High KV cache hit rate also cuts costs significantly. @deepseek_ai@OpenAI
There's an underrated part of the AI Sat Mini design that nobody seems to have picked up on.
Up until now, the 2 major critiques to Orbital Datacenters have fundamentally been:
1. The Solar Irradiance hitting the satellite in a Dawn-Dusk SSO will make running GPUs on it thermally impractical
2. The Earth Albedo hitting the radiators will make it challenging to dissipate heat
The AI Sat Mini config takes care of both.
- The Flatsat Bus is pointed perpendicular to the sun load edge-first, minimizing Solar Load.
- The Radiator is pointed perpendicular to both the Earth and the Sun Loads, able to dissipate heat to free space on both sides
- The Solar Array mounting seems to be fairly thermally isolated, and the massive span allows for a ton of radiative dissipation from the dark rear-face (as is often the case with space SPAs, they don't store that much)
These are fairly clever orbit & conops driven geometry choices that become apparent once you start mission design, and de-constrain the design quite simply.
Your flatellite + radiator is almost permanently shaded, and the only incentive becomes to run your compute as hot as possible to leverage the Eb = σ• T⁴ scaling. Passive heatpipes conducting heat from the chip pedestal to the radiator may be sufficient, without engaging in the insane exercise of mass producing 0g coolant pumps.
Now there seem to be 2 variants here between the images posted, but that may just be render inaccuracy.
Vertical Config: Image 1 (with the Starship for scale) shows the solar arrays along the edge of the flatellite.
Horizontal Config: Image 4 (the on-orbit render) shows the solar arrays perpendicular to the face of the flatellite.
Either scenario supports the earlier observations, as the solar array plane is free to rotate. However as others have noted, the Vertical config may offer a fair degree of gravity gradient stabilization.
This is another interesting bit - Orbital Compute will require very little pointing accuracy, compared to most other space applications like EO or Comms that track a target (Imaging AOI or Receivers). As long as the satellite points roughly at the sun, and the gimbaled ISL terminals can maintain a stable connection to Starlink, the pointing requirement is negligibly coarse. You could probably forgo reaction wheels and control this whole thing with magnetorquer pulses.
SpaceX has a habit of eliminating broad constraints at architecture level quite cleverly, and I was looking forward to how they think about Orbital Compute. Safe to say I was quite delighted by this design.
First view of the 100kw AI Mini Sat with solar panels and heat radiator to scale. “And that’s just the Mini version. We expect future versions to go to the megawatt range.” — Elon
The key missing ingredient is a terawatt of AI compute. Fully integrated fab with recursive
ModelGate
An OpenAI-compatible API Gateway
for local LLM deployment.
Multi-backend.
Quota control.
Rate limiting.
Hybrid routing.
Open source ↓
github.com/derekwin/Model…
1/ 🚀 Merged 7 small PRs into ModelGate today! Major improvements on security, stability, and maintainability. Let's dive in! 🎉 #Go#LLM#APIGateway#OpenSource
Local LLMs are easy.
Production LLM infrastructure is not.
If you’re serious about AI infra,
you need a model-aware gateway layer.
Open source ↓
github.com/derekwin/Model…
Would love feedback from infra builders.
This unlocks 3 powerful patterns:
1️⃣ Enterprise internal AI governance
2️⃣ SaaS tiered AI monetization
3️⃣ Vertical model API productization
Without rebuilding auth + billing every time.
Most people think running a local LLM = docker run ollama.
That’s cute.
Running it in production — with multiple users, quotas, routing, and hybrid cloud — is a completely different architecture problem.
Let’s break it down 👇
1K Followers 2K FollowingStupid, but try to keep Learning, curiosity and Courage | Creator https://t.co/UWe3rFdsqC | eBPF & AI | Bring new orders and chaos! | 🐟好吃
333K Followers 5K FollowingYour #1 News source on everything X + xAI and now SpaceX | Contributors: @HXMnCK, @xUpdatesRadar and @swak_12 I | Hit Follow and sign up for notifications! 🔔
69K Followers 7K FollowingGRAMMY® Nominated EMMY Winning World Record Holding musician/teacher/hubby/f̶o̶s̶t̶e̶r̶ parent🥰 TXT me at +1 (858) 758-2493 🎶album: Buddy's Magic Tree House
11K Followers 118 Following@saranormous and @eladgil host your podcast guide to the AI revolution. weekly interviews with the builders (on Apple, Spotify and YouTube)
154 Followers 254 FollowingPhD CompSci at @cambridge_uni doing security research. Maintain the web version of @obey_love. Publish games to mobiles/switch under @mazette_games
29K Followers 172 FollowingThe Robotic Systems Lab designs machines, creates actuation principles, and builds up control technologies for autonomous operation in challenging environments.
4K Followers 2K FollowingBuilding transparent startups at https://t.co/gRbPzPfZgV
Work with me:
$3000 to build a startup with you for one week
$10000 to build a startup with you for one month
328K Followers 5K FollowingWe are the LinkedIn Social Support Team and avail 7 days a week! If you need help with your LI account, DM us your email address that is listed on your profile.
16K Followers 199 FollowingLarge Model Systems Organization: Join our Slack: https://t.co/vzYOTP4w6C. We developed SGLang https://t.co/OjwQadINKU, Chatbot Arena (now @arena), and Vicuna!
73K Followers 3K FollowingThe world's largest professional organization advancing #computing as a science and profession. Also @mastodon.acm.org
Likes & shares ≠endorsement
32K Followers 442 FollowingCo-founder, Recursive. Professor, CS, U. British Columbia. CIFAR AI Chair, Vector Institute. | ML, AI, deep RL, deep learning, AI-Generating Algorithms (AI-GAs)
3K Followers 438 FollowingSNIA is an industry organization that develops global standards and delivers education on all technologies related to data. We are the experts on data.