You have to be irrationally obsessed with something

Check out the new episode on physics-based computing:

In the new episode I'm breaking down the engineering behind Colossus 2 - the world's largest AI data center:

I met the world's first synthetic human:

⁉️Why it didn’t work out Building a datacenter-class AI supercomputer from scratch is one of the most capital- and talent-intensive bets in tech. The engineers capable of pulling this off are rare — and easily poached.

💡Dojo’s Story Announced at AI Day in 2021, entered production in 2023. Over ~4 years, Tesla poured several billion dollars into it. They did some impressive stuff along the way. Most notably — innovating in packaging, cooling and SW.

It was a genuine moonshot — and not all moonshots work. But I really admire Tesla for taking it on. I think they underestimated just how complex it is to build a supercomputer from scratch.

New episode is just dropped - don't miss this one😎

Week 3 at @Stanford No time to post, barely time to breathe XD Today we had a pitch practice with a communication coach who works with top @Google execs — and one thing really stuck with me: Silence is power. We’re often quick to react/fill in pauses, but let the silence sit 😌

What important truth do very few people agree with you on?

7. This shift won’t happen overnight — but the direction is clear. Follow @AnastasiInTech for more

6. Companies to watch: 🔹 Startups: Lightmatter, Ayar Labs, Q.ANT, LightIntelligence, Black Semiconductor 🔹 Public: Broadcom, Marvell, Lumentum (part of NVIDIA's photonics push), Cisco, POET Technologies

5. Add a DAC — and now the same device can perform analog multiplication — multiplying input light with an analog weight. That’s the basis of matrix multiply–accumulate (MAC), the core operation behind many AI workloads.

4. How photonics works — in a nutshell: At the heart of it is an optical modulator, like a Mach–Zehnder Interferometer (MZI). It encodes digital data (1s and 0s) into the intensity of light.

3. And here’s the bottleneck: Traditional copper interconnects are reaching their physical limits in: ❌ Bandwidth ❌ Energy efficiency ❌ Physical reach That’s where photonic interconnects come in.

2. To keep up, modern AI workloads now need entire data centers to function like one giant GPU. But that puts enormous strain on one thing most people overlook: The interconnect.

1. Each new generation of LLMs demands significantly more compute. → AI compute demand is growing by 4.6× per year → Hardware performance improves by just 1.3× per year This mismatch creates huge pressure on infrastructure.

🚨The biggest transition happening in next-gen computing right now. A quick thread from my recent keynote — on why photonics is rising and the top companies to watch in this space:

New episode on the channel - very proud of this one 😎

Obsession is my superpower

Hi friends! Super excited to be attending NVIDIA #GTC25 in person this year. 🚀 Don’t miss out—register now for the free virtual event using this link https://t.co/YEQjMpBbsG and get a chance to win NVIDIA RTX 6000 Ada GPU. #GTC2025, #NVIDIA #AI, #JensenHuang, #GTC25