Dylan Patel

@modestproposal1 @hashemim86 I mean we track over 5,000 DCs, their permitting, regulatory filings, physical construction timelines, etc. Most people will be surprised at level of investments being made but fact is they are doing it

@magicsilicon @dwarkesh_sp @asianometry Good post

People keep asking about the high purity quartz from Spruce Pine, NC. There are some articles online saying it's then only source for a specific part of the semi supply chain. They are being a bit over exaggerated, and the impact is not nearly as large as being fear mongered.

I hate soft tech. They don't even know how their shit runs. Buncha weak ass pansies. Deep and hard tech all day.

Hold up what? We use Pledius Neo which is 30-cm resolution. All I had to do is get interviewed and send them my US passport and can task the satellite for $5k to $6k to track datacenters. Private sector surveillance is better than the US Navy?!?! https://t.co/aGi0a7lhG4

People still don't get that TrendForce spot pricing means nothing. All AI Accelerators (HBM), most phones (PoP LPDDR), and increasingly many PCs (Apple, Qualcomm, Intel Lunar Lake) have on package dram purchased direct from manufacturer. Old school channel checks don't work $MU

Kind of sad for ML Twitterverse that this has not already been made before https://t.co/a7xkgw9QsY

SRAM scaling from N5 to N3E to N2 to A16 is very complicated. Headline bitcell size itself doesn't shrink much, but periphery seems to scale a lot through different DTCO. I'm trying to wrap my head around the methods of how/why Anyone of my followers wanna chat or DM about it? https://t.co/8la5GQ4GWR @Darth_Goldsmith ?

Jokes on y'all, you each owe me a $150 million ASML EUV tool https://t.co/nwZeS0ZL6y

@CommyHipster Yes let's lay off 30,000 US semiconductor design engineers in cost cutting exercises. Sounds p great for China

Qualcomm buying Intel is dumb, but some positives Creates a powerhouse in consumer chips, mobile to PC, but results in a lot of duplicate IP, custom CPU, GPU, ISP, Al, etc So, there would be significant cost cutting Qualcomm has no ability to turn around datacenter or foundry /1 Keep both x86 and Arm both reduce number different cores, focus way more on x86 though Kill Qualcomm iGPU keep Intel's Kill Intel VPU keep Qualcomm Hexagon Kill Intel Wifi keep Qualcomm Kill Intel ISP keep Qualcomm

@one_clever_boi @fchollet This is a meme about keras people on twitter lmfao

Proof that Pytorch is falling from its place at the top. Keras is winning. Higher performance. @fchollet is a true mastermind. https://t.co/WAeRJneu3b

All these AI grifter Twitter influencers are confusing LLM with MLM

Mfers taking Ozempic to get to the optimal weight They needa be taking Adderall to get the optimum weights

Hock Tan has forsaken you because you have sinned You dared to defy, and now in his wrath, he rains down M&A upon you Your bandwidth will be throttled, servers tremble, you will beg for mercy But there will be no mercy For you have sinned Hock's judgment is swift and unrelenting https://t.co/uneWibSQqe

@hyhieu226 @tri_dao Flex attention on PT is p fire

@highyieldYT @realmemes6 Arrow lake 20a was cancelled

@nosunkcosts @xai They are running all the GPUs with generators + 50MW grid power.

@Ar_Douillard Thanks for the recc Arthur, I'll check it out

Multi-Datacenter Training OpenAI's Ambitious Plan To Beat Google's Infrastructure Gigawatt Clusters, Telecom Networking, Long Haul Fiber, Hierarchical & Asynchronous SGD, Distributed Infrastructure Winners, Silent Data Corruption, Stragglers https://t.co/unluurzsnd

The Brazilian Supreme Court now has to get rid of all of their Inspiron laptops.

@swyx @gkossakowski @GroqInc @sundeep Mine accounts for it