Dylan Patel

@dylan522p on X

Writes about AI and Science. 81% of posts open with an observation. A typical post reaches 668 likes.

Remix their style

Posts: 27
Likes: 27K
Typical: 668
Best: 3.9K

What they write about

AI: 74%
Science: 26%

How they open

observation: 81%
number: 15%
story: 4%

Every post of theirs in the library

Dylan Patel@dylan522p

4% of GitHub public commits are being authored by Claude Code right now. At the current trajectory, we believe that Claude Code will be 20%+ of all daily commits by the end of 2026. While you blinked, AI consumed all of software development. Read more 👇 t.co

Opens with an observation

Dylan Patel@dylan522p

Datacenter capex is going to 0 Grok struggling deploying Anthropic extremely low rate limits for paid + Cursor Google not offering 2.5 on API OAI not serving o3, struggling to serve 4o image Azure, still not serving many models including o1, o3, deep research, 4.5

Opens with an observation

Dylan Patel@dylan522p

OpenAI now just needs to sandbag the next model release So the US doesn't export control them So they gain massive market share It's imperative that all the OAI employees don't vauge hype post their next model release as the greatest thing ever Don't know if they are capable tho

Opens with an observation

Dylan Patel@dylan522p

So this is like 2T in loss market cap for a $6M training run (ignoring cost of research, ablations, distilled data from GPT, capex for their various clusters, etc). Imagine if China invests $300m in a training run. It would reduce the world's GDP to $0.

Opens with a number

Dylan Patel@dylan522p

OpenAI Jalapeno is spicy Usually first generation chips aren't competitive, but OpenAI is beating Nvidia Blackwell and even Rubin This is huge news! We got to go into OpenAI's lab to dissect their new chip and dive into the software, architecture, and performance Incredible work!

Opens with an observation

Dylan Patel@dylan522p

Huawei's new AI server is insanely good People need to reset their priors This is why banning H20 without banning tools and sub components is idiotic because Huawei is not far behind H20. The admin needs to act fast to slow down Huawei's ramp or the H20 ban will be useless

Opens with an observation

Dylan Patel@dylan522p

Deepseek has over 50k Hopper GPUs to be clear. People need to stop acting like they only have that 10k A100 cluster. They are omega cracked on ML research and infra management but they aren't doing it with that many fewer GPUs

Opens with a number

Dylan Patel@dylan522p

OpenAI hasn’t even deployed TPUs yet and they’ve already saved ~30% on their entire lab wide NVIDIA fleet. This demonstrates how the perf per TCO advantage of TPUs is so strong that you already get the gains from adopting TPUs even before turning one on. The piece covers a lot

Opens with an observation

Dylan Patel@dylan522p

It was absolutely amazing and surreal to get to grill Satya on his AI strategy. He has some really good responses, but also some areas where we disagree. See the interview below and the report we wrote up and posted today! t.co

Opens with a story

Dylan Patel@dylan522p

Microsoft's initial plan was to leapfrog both Google and Amazon to become the largest tech infrastructure company in the world. Then they got scared. They backed away from OpenAI, backed away from some leases, and even slowed down their construction pace for major datacenters.

Opens with an observation

Dylan Patel@dylan522p

The US + Saudi Arabia / UAE AI deals are complete wins Everything is there to win over the Middle East and accelerate infrastructure spend. The main subtext here is that China is locked out of the Middle East AI infrastructure investments Now we must execute on security

Opens with an observation

Dylan Patel@dylan522p

Our 5-month journey conducting independent analysis & benchmarking of AMD MI300X vs Nvidia H100 + H200 Detailed, open source low-level benchmarks performance vs TCO Comprehensive public recommendations It’s not just immature software, they need to change how they do development

Opens with an observation

Dylan Patel@dylan522p

Intel needs a culture reset Lip-Bu Tan is a strong choice to respin Intel Foundry into a customer-first organization Given he left the board because of frustration with the slow-moving bureaucracy, he likely will give high priority to speeding up + slimming down the company 1/4

Opens with an observation

Dylan Patel@dylan522p

So many people hated on us for the AWS acceleration thesis a couple months ago And we were right of course Only we can do these calls due to our deep knowledge and inights with core research, datacenter, accelerator, and Tokenomics models. Amazon $AMZN Sales@semianalysis.com

Opens with an observation

Dylan Patel@dylan522p

LLAMA 3 8B was amazing but will be overshadowed Phi-3 mini 4b, small 7b, medium 14b this week, and the benchmarks are fucking insane Synthetic data pipelines are massive improvements over internet data Flywheel only continues with big models too when these techniques are applied

Opens with a number

Dylan Patel@dylan522p

We are going super deep into robotics from models to supply chains Part 1 is a banger start to the series We are covering from the basics of hardware like batteries, motors, actuators, magnets, materials We also cover the big historic robotics firms and China's rise

Opens with an observation

Dylan Patel@dylan522p

Scaling Laws O1 Pro Architecture Reasoning Infrastructure Orion and Claude 3.5 Opus “Failures” AI Lab Synthetic Data Infrastructure Inference Tokenomics of Test Time Compute The Data Wall Evaluation's are Broken RLAIF Inference Time Search Scale is Needed t.co

Opens with an observation

Dylan Patel@dylan522p

AMD software is improving rapidly Cuda isn't a moat forever, but Nvidia is building new ones with the Python DSL, Dynamo, and more Meanwhile Nvidia hardware advantage is huge this year, but perf/TCO of 355X has attracted some customers MI450X is actually competitive with Rubin

Opens with an observation

Dylan Patel@dylan522p

Intel on the Brink of Death Board Short Circuits, Cultural Rot, Product Focus Is Flawed x86 Has No Moat Foundry Must Survive Nvidia's PC CPU Push - 3 chip tegra strategy Sierra Forest, Granite Rapid, Clearwater Forest, Diamond Rapids, Rogue River Forest t.co

Opens with an observation

Dylan Patel@dylan522p

Databricks DBRX model is AMAZING, generally great but CRUSHES code. 132B parameters, 12T token, 16 experts, 4 per forward, 36B active. ~2.6e24 HumanEval5, 0-Shot DBRX - 70.1% GPT-4 - 67% Gemini 1.5 Pro - 71.9% Mixtral - 54.8% Grok - 63.2% LLAMA 2 - 32.2% t.co

Opens with an observation

Dylan Patel@dylan522p

Only NVIDIA, AWS, Google have successfully deployed rack scale architecture Trn3 is the 2nd after Nvidia with switched scale up topology, which is better for frontier mixture of experts models AWS Trainium3 course corrects on software Switch design choices are... interesting

Opens with an observation

Dylan Patel@dylan522p

Debunking the false narratives around AI Datacenters. First it was that water usage is high, but it's not. This time it's about power prices. The narrative around powe pricing is silly and politically charged with no actually backing Manipulated statistics by people with agendas

Opens with an observation

Dylan Patel@dylan522p

This is a public cry to not waste the silicon and have better AI infra Rivos, DSF, Grand Teton, Ariel and the cut-down MI450X were defensible on some narrow metric that a group inside Meta Infrastructure was optimizing for, but each was a poor decision for the company as a whole.

Opens with an observation

Dylan Patel@dylan522p

Open models show 2.5x faster, 6x more expensive Lower batch size, speculative decoding harder Pareto optimal curve for Deepseek at t.co shows this Claude Opus 4.6 is 100 Tok/s/user Deepseek at 100 is 6k Tok/s/GPU At 250 tok/s/user it's closer to 1k

Opens with a number

Dylan Patel@dylan522p

The tariffs are much more complicated than they seem Industry and the market are misunderstanding the ramifications We cover everything from fabs to packaging to servers to cables to optical modules to datacenter equipment The impacts differ greatly We also found a huge loophole

Opens with an observation

Dylan Patel@dylan522p

AI reveals all frauds Supermicro accounting, stuffing channels, having servers shipped to China $SMCI Bank research is actually SemiAnalysis reports in bullet points Half the bullet points are ctrl + c / ctrl + v They didn't even rephrase with chatGPT 😭 t.co

Opens with an observation

Dylan Patel@dylan522p

We expect significant restructuring of the slow and ineffective middle management at Intel. Although his recent time was spent in management, Lip-Bu has technical bona fides and should approach the role with a technical eye. 2/4

Opens with an observation