I am not the best at anything.
What I am is sneaky-dangerous across a bunch of things simultaneously.
That's the decathlete. 🏃♂️
That's what I spent 30yrs figuring out.
That's why I stopped trying to pick a lane.
👂Listen. 🖊️ Write. ⌨️ Type. 🤖 Listen.
A short history of pretending to pay attention.
(Drag through 22,000 years of the same meeting via interactive version in comments.)
1/
$NVDA “bought” Groq on Xmas Eve.
Five months later its investors are putting another $650M in (per @axios) — not for chips, but for the inference cloud everyone wrote off.
That’s the story. And it’s bigger than Groq.
🧵
5/
There’s a real bear case — no lock-in, rich multiples, sovereign concentration. But it’s about who keeps the margin, not whether the category survives.
And nobody can hand you a clean “inference market share” number. The demand isn’t in question. The measurement is.
4/
Best part: inference isn’t one job.
Prefill (reading the prompt) is compute-bound — old GPUs crush it.
Decode (writing tokens) is memory-bound — GPUs idle.
So old GPUs aren’t landfill. And decode chips (@GroqInc, @cerebras) complement GPUs rather than replace them.
3/
The popular read on Groq: “they sold everything, the team left, nothing’s there.”
Not quite. The $NVDA license was non-exclusive. Groq kept the LPU IP, the deployed fleet, the customers, and a real speed edge — and GroqCloud was carved out and kept running.
2/
The independent inference cloud — selling tokens, not training runs — quietly became a fundable category.
• $CBRS IPO’d past $50B.
• Together raising at $7.5B.
• Fireworks at $15B.
• Baseten at $11B.
• Groq at $650M.
Six companies. One category.
On Thursday I got the best colonoscopy report I’ve had in years. 💪
This winter my 10yr old daughter had her first flare-free winter since being diagnosed with Crohn’s disease at 6. 🙌
Neither happened by accident.
Both happened because of decades of research, better treatments, and incredible doctors.
🏥🩺⚕️🔬🧪👨🔬🥼👩🔬
We are making one final push for the Crohn’s & Colitis Foundation before Take Steps on June 6. Huge thanks to all who have participated already. If you haven’t, please consider contributing if you’re able. 🙏 Thank you ❤️
Full story (and donation link) in comments.
This is why “waiting for the numbers” on enterprise AI adoption is a pretty stupid way to invest in AI. (1) By the time the numbers reflect the impact it’s (way) too late and (2) even when the impact is happening (ie now) it’s not reflecting the numbers. Have had a handful of investors tell me there is no real AI revenue / impact. Ok. Sure. Carry on. Nothing to see here.
Debt capital markets is the hottest role at AI infra companies now. (It’s not. But it’s showing up a lot now. For a reason.) The latest one: @fluidstack
@andrewdfeldman Timely. (And very much agreed tho it took me three decades to figure out.) Posted this about being a “decathlete” yesterday. https://t.co/jrgbqaUbkp
I am not the best at anything.
What I am is sneaky-dangerous across a bunch of things simultaneously.
That's the decathlete. 🏃♂️
That's what I spent 30yrs figuring out.
That's why I stopped trying to pick a lane.