Michael Søndergaard

Have I not managed to convince you yet? Come hang out in our Discord to keep the debate going. It's where we talk compilers, GPU programming, and other nerdy things with our dev team: scale-lang.com/s/discord?utm_…

Generality is free when your abstraction sits at the right level. Dive into Part 4 of my series, "Why hardware-agnostic isn't the same as lowest-common-denominator": scale-lang.com/posts/2026-06-…

You can't have portability AND peak performance. That's the assumption people make. They're right if you rely on portability by abstraction at runtime. But that assumes a cross-vendor system must be a "lowest-common-denominator" system.

@VaibhavSisinty Yeah because everybody likes more space junk filling orbits, and the physics of radiating heat into space tell us to go down this road. Hint: no they don’t, and we also didn’t get hyperloop for similar physics 101 reasons. Elon does a lot of great stuff. This is not it.

Okay this is genuinely insane. SpaceX just unveiled a satellite whose only job is to run AI. Not internet. Not GPS. Just compute, floating in orbit. It's called AI1, and the reason behind it breaks your brain. AI data centers on Earth are hitting a wall, not a chip wall, a physics wall. They need staggering amounts of power and water just to stay cool, and we're running out of grid and land to build them. So Musk's answer is: stop building them on Earth. In orbit, the sun never sets. Free power, 24/7. No water for cooling, you just radiate heat into the vacuum of space. The two things choking AI on the ground barely exist up there. And here's the wild part: Musk says it's easier to build than a Starlink satellite. Strip out the complex antennas and it's "a lot of solar cells, a radiator, and some laser links." One AI1 carries the compute of an Nvidia GB300 rack, the same hardware data centers fight over down here. AI1 is just the first one. The plan is a constellation of up to a million of them. And the timing isn't an accident, SpaceX goes public this week at a ~$1.75 trillion target. This isn't a rocket company anymore. It's positioning itself as the power grid for AI, in space. The race for AI compute just left the planet. Literally. @SpaceX

@jtatarchuk Well done Jeff

I turn 40 today. A decade ago I was still trying to figure it all out. Today, I'm grateful to be surrounded by an incredible team, customers, partners, investors, friends, and family who helped make this journey possible. Today we're announcing TensorWave's Series B: $350M raised at a $1.55B valuation. We're still early. The mission remains the same: build the infrastructure powering the future of Al and help accelerate what's possible. Thank you to everyone who believed in us. The momentum is building. Accelerate responsibility.

@AtlasInference @SpectralCom @AIatAMD This is so cool!

Atlas Inference is running Qwen3.6-27B on AMD Strix Halo 🥳 Using @SpectralCom's SCALE ROCm backend, our CUDA kernels compile and run on RDNA⚙️ Cross-architecture inference from ONE codebase 🗣️ Thank you @AIatAMD for the gift 🙏 POC ✅ excited to keep tuning performance⚡️

𝗦𝗽𝗲𝗰𝘁𝗿𝗮𝗹 𝗖𝗼𝗺𝗽𝘂𝘁𝗲 𝗶𝘀 𝗻𝗼𝘄 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗡𝗩𝗜𝗗𝗜𝗔 𝗜𝗻𝗰𝗲𝗽𝘁𝗶𝗼𝗻 𝗽𝗿𝗼𝗴𝗿𝗮𝗺. 🟩 Inception is @nvidia's program for AI startups - a membership that gives access to technical resources, preferred pricing on NVIDIA hardware and software, and exposure to a global network of investors and partners. CUDA is the de-facto standard for AI developers, and we’re honored to play our part in growing the ecosystem.

From CUDA portability to real-world performance on @AMD GPUs, @SpectralMichael, CEO of @SpectralCom shares why open, portable AI infrastructure matters more than ever. Watch the short interview from Beyond Summit 2026 on our YouTube channel: youtube.com/watch?v=cwmiGY…

Everyone says CUDA can't target TPUs. What they mean is nobody has written the compiler that raises CUDA code to something a systolic backend can consume. Those are very different sentences. Full post — Part 3 of why @SpectralCom exists: tinyurl.com/mtnxcjsw

𝗨𝗽 𝘁𝗼 𝟮𝟱.𝟳× 𝗳𝗮𝘀𝘁𝗲𝗿. Unmodified CUDA. AMD silicon. Thanks to the @tensorwave team for benchmarking SCALE on MI355X and publishing the numbers. Port to AMD used to mean a rewrite. Now it means a recompile.

@HotAisle @tbpn @modal @bernhardsson @SpectralCom Yeah the modal guy needs to read my blog: scale-lang.com/posts/2026-05-…

@tbpn @modal @bernhardsson We are already there thanks to @SpectralCom and @SpectralMichael. Oh and AI can just do the conversion too. Oh and a token doesn’t care where it was produced. I have been calling the moat dead for months now, nice job catching up.

The CUDA moat is real, but probably not for long, says CEO of AI infrastructure platform Modal @bernhardsson. He says he's bullish on alternative accelerators over the 2-3 year timeline, even though there's currently zero demand from his customers for TPUs, etc. "The cost today of rewriting your software to run on those stacks is very high... But the cost is going to go down." "You're going to have software that basically lets you take CUDA-compatible stuff and run it on alternative accelerators."

Obvious take: if agents can write native code for any GPU, who needs a portable CUDA toolchain? Just point the model at each GPU. I think that's exactly backwards — and wrote up why. tinyurl.com/5ac6y5jt

CUDA isn't a standard. No consortium, no ratified spec. It's whatever NVIDIA ships next. That's supposed to be a weakness. It's why CUDA wins. The cross-vendor layer can't be a consortium either. It has to be a company. That's what we're building at @SpectralCom. SCALE compiles unmodified CUDA on AMD and NVIDIA today. No PDF to argue about — a toolchain. Full piece: [bit.ly/3PBsYsU](bit.ly/3PBsYsU)

Compiler Explorer godbolt.org via @CompileExplore would you like to add scale-lang.com to godbolt?

AI compute is bottlenecked. When the hardware mix inevitably diversifies, how will production clusters actually operate? Our CEO @SpectralMichael is joining @AkashBajwa96's Gradient Descending roundtable this Wednesday to dig into the evolving hardware landscape.

What does the AI cluster of the future look like? Hint: it won't just be a wall of B200s. Our CEO Michael Søndergaard is joining @AkashBajwa96 for the next Gradient Descending roundtable to discuss ASICs, non-NVIDIA chips, and breaking the software lock-in. 🗓️ Feb 25 | 8:30 AM GMT 🎟️ Request invite: luma.com/f7s69jj2 #AI #HPC #DeepTech #AMD #NVIDIA #ASIC

@jtatarchuk @SpectralCom Will save one for you! (Ps fixed my handle now)

@SpectralCom @MichaelSnd43892 i need one of these shirts!

We told Michael to pretend the pins were #GPU vendor lock-in.

@HotAisle What if only the ports weren’t necessary and the upstream versions just worked without needing to be seperately maintained?

6 months ago, AMD moved a bunch of NVIDIA -> ROCm ported projects to a different location on GitHub. It showed promise that things were going to be improved on. Since then, the NVIDIA upstream projects have moved on to newer versions (ex: rmm fork is still 25.02 and upstream is now 25.10+). Those improvements don't appear to have been fully ported back to the forks. Today, the ports still don't have CI / CD setup or individual releases, and it was all just declared "production ready." cc: @AnushElangovan