I finally finished the video, where I looked at various @intel products for inference. Part 1: Vibe coding with VS Code &self-hosted LLMs youtu.be/Gj0znnjqUEY?si=uuF6… via @YouTube

When there's alot of a external noise. Best to tune it out and focus on your own execution. I have faith that Lip-Bu is making the best decisions for long term shareholders. If you don't then sell the stock and move on. Thank you for your attention to this matter.

Eric Bana did not say this! It is now clear this is a lie. What are your thoughts of how @AlboMP is running the country and the impact of the latest Federal Budget on the Australian economy?

SpaceX can't produce anything without someone else's IP They need Intels IP for packaging and logic, or try to be Rapidus 2.0 with IBM If Intel is out, then they need TSMC to run a Fab and they can get some equity in it (So not TeraFab lol). Or try IBM/Rapidus On memory they have to do everything from scratch (7 years probably to a produce) or get a license

Lest take a look at @tculpan reports 21 January 2026: arguing that propping up Intel will not fix semiconductor economics and that complaints about TSMC as a brake misunderstand the industry. 23 January 2026: Intel’s estimated 2026 total capex of about $14 billion equals TSMC’s 2026 additional capex, used to dismiss the idea that Intel is the solution. 2 June 2026, “Exclusive: Intel is Struggling to Supply Laptop Chips Built Around its New 18A Node.” The public framing is a shortage at Intel’s comeback node and client worry about future supply. That is negative on Intel; it does not praise TSMC in the headline. 2 September 2026, “Exclusive: Inside TSMC, Intel and Apple’s Circle of Trust.” Subtitle is Intel’s attempts to win Apple business versus TSMC’s strategy to stay ahead. His own note on it called Intel’s position a self-imposed Catch-22: unwilling to build capacity without orders, unable to get orders without proof of delivery. 2–3 October 2026, the Terafab exclusives (“TSMC Exploring Plans to Work With Elon Musk’s Terafab,” then Musk confirming the talks). The reported contrast is that TSMC is exploring owning and operating a Texas factory for Terafab, while collaboration with Intel is not making progress. This guy is a Taiwan state sponsored asset. Taiwan will do anything to protect $TSM and that usually means attacking its biggest competitor in the U.S which is $INTC

I think theres been some difficulty in the negotiations and perhaps @elonmusk is using $TSM as a threat to gouge $INTC for a better deal on their process / licensing arrangements. I am starting to wonder if Intel should back out of this. It seems very unlikely that TSMCwan would allow a cutting edge node like A14, TSMC's competitor node to 14A, to be based on US soil though, so I dont think we should be fooled into getting a bad deal. Elon should have just bought a stake when we Intel was cheaper. $SPCX

I want to point out that none of the Terafab related engagement for Intel was remotely priced into the stock anyway. As an investor, Intel's story is now. And Terafab was not remotely a reason the stock went up or deserved to go up. Intel is (to the market) a strong CPU story, hence the limited re-rating we did get. But, Foundry is a story more than what Terafab will be for Intel. Don't forget who has engaged with Intel. The likes of Apple, Google, Nvidia, and Amazon. Terafab is an overly hyped up project that won't prove returns for many years. That is not the story. In my honest opinion, I don't think Intel was to do more than be a consultant and simply a licensing partner. That is all. In providing capacity, if Elon wants to leverage TSM, so be it. Then let them do everything. Intel should not (and I don't use this word lightly) WASTE capital on Terafab. Focus on ASICs, focus on margins, focus on delivering your chips built on 18A.

You think if $INTC had not made a deal with @elonmusk, $TSM would even entertain him right now? LOL. Remember, they’ve shooed him away not too long ago. This has little to do with Taiwan and Elon suddenly making a deal, and everything to do with Taiwan suppressing IFS. If there’s one thing we’ve learned about Taiwan, it’s that they are ruthless when it comes to self-preservation and $TSM competition. There are no red lines they’re not willing to cross. Fabricating rumors, releasing “negative news” pieces right before earnings, risking their IP for the chance of keeping IFS from having a real shot… We shall see if anything actually comes out of this “news” other than another attempt to knock Intel down…I doubt it. Because Taiwan is proving one thing: it is no friend of the U.S. And allowing its IP to be used by the one man who could potentially beat them at their own game seems like a disastrous move for Taiwan. I don’t believe they would ever seriously go for it. My guess is this is mostly about knocking Intel down and buying $TSM some time while they regain their footing on capacity issues. $INTC

“When you go mountain climbing, the first thing you are told is not to look at the peak, but to keep your eyes on the ground as you climb. You just keep climbing patiently one step at a time.” Akira Kurosawa

Yes, of course if that “trend line holds” then AI capex will be $64T in 2032, and $128T in 2033, effectively doubling global GDP. Where’s Anthony Odlyzko when you need him…?! #GetTheNet

Deutsche sees Muse reaching ~8% of $META revenue by 2030 with agent commerce fees driving ~75% of upside case. What makes that really interesting is the model assumes only ~0.3% of daily users ever pay and gives no credit for Muse completing purchases through $SHOP and Stripe while helping businesses create and spend more on Meta ads. So Muse can monetize the transaction itself while also feeding more dollars back into Meta’s core ad business,which makes the upside MUCH bigger.

Yann LeCun's (@ylecun ) latest talk at ETH Zürich Scaling LLMs to reach AGI is "impossible" A large language model is trained on about 30 trillion tokens, which is roughly 10^14 bytes of text and would take a person about 400,000 years to read. A 4-year-old child receives about the same amount of data, 10^14 bytes, through vision alone in about 1 year and 10 months. In his view, intelligence is the ability to learn new tasks quickly or perform them without prior training, as a teenager learns to drive in about 20 hours. Scaling increases stored knowledge, but it does not produce this ability to adapt. ---- From "Perfology Clips" YouTube channel, (link in comment)

前彭博的人今天寫了一句，台積電在看德州那座 Terafab。 你第一眼會以為，喔，要合作了。先別急。沒有合約，沒有金額。台積電還是那句，市場傳聞，不評論。 馬斯克三月就講過這座廠。特斯拉、SpaceX、xAI 要的是一年 1 太瓦的運算。他說得很白，台積電如果做得出這個量，這座廠根本不用蓋。八月地點定了，休士頓外的 Grimes County，第一期 168 億美元。更早的申報更誇張，第一期 550 億，做滿上看 1190 億。邏輯、記憶體、封裝、測試，想塞在同一個園區。四月英特爾先跳進去，拿 18A、14A。魏哲家在法說上沒有接話。他只說，客戶還是客戶。蓋一座廠要兩三年，量產再一两年。沒有捷徑。 那「台積電來蓋、馬斯克來包產能」是誰講的？不是公司。是網路上有人把空白補上了。台積電擁有、台積電營運，SpaceX 出錢或先把產能買斷，不拿貨也要付錢。彭博前一天剛好寫，台積電在看德州第二個園區，達拉斯附近，不只一座，一座就要 200 億美元起跳。亞利桑那已經押了 2650 億。你再蓋，總得有個人大到可以把樓撐住。 像餐廳要加蓋一層樓。圖還沒畫完，先得有人把位子整桌訂走。馬斯克胃口很大，可是他沒有一座已經在跑的先進廠。台積電有手藝，缺的是一個理由，值得在德州再押幾百億。英特爾先搶到一張製程的新聞。三星在泰勒也已經黏著特斯拉。真要是台積電自己在德州把廠跑起來，這兩家的位子都會被擠掉一點。 魏哲家不太可能把配方交出去。那東西不是一張圖，是無塵室裡天天在調的參數。廠掛在別人名字下面，配方會走路。所以如果最後真的發生，比較像是台積電的廠，馬斯克來下單。不是馬斯克的廠，台積電來當顧問。

Eric Schmidt explains why a mutual pause of the AI race is impossible: > US incentives will not stop, a local lab or China can cheat, and a UN check cannot read code that computers wrote or outrun US and Chinese labs. > it cannot be verified. Inspectors would have to enter secret laboratories, including military laboratories, and review the full code base. That code is written by computers, so no human team can read it line by line. ---- From "The Economist" YouTube channel, (link in comment)

The biggest bottleneck as we know as industry is: ✨The cost of data movement in processor-centric architectures 
@SKhynix latest AI Ecosystem piece is worth a read which is written by Dr. Onur Mutlu @_onurmutlu_ Professor at @ETH, one of PIM (processing-in-memory's) most prominent advocates. His core argument chimes with our view that AI performance now depends on how little data moves, not how fast compute runs. 
▪️One DRAM access can cost 150–2,000x the energy of a simple arithmetic operation. ▪️ Memory access and data movement can account for more than 90% of system energy for large ML models according to some studies. ▪️ This is where architecture matters and high-speed interconnect approaches such as CXL & NVLink also matters ▪️ CXL and NVLink/NVSwitch are solving different problems: ▫️CXL = capacity. It expands, pools and shares memory across hosts.
 ▫️NVLink = bandwidth. It moves weights, activations and KV cache between GPUs. ▪️Today, GPU memory expansion runs mostly through Nvidia's own paths, such as NVLink-C2C to Grace's LPDDR5X, not through CXL. CXL pooling at scale is still more roadmap than deployment. ▪️The KV cache is what's really driving this. Longer contexts and agentic reasoning make the KV cache grow faster than HBM capacity can keep up. ▪️ That's where Dr. Mutlu argues for disaggregation: moving the KV cache and attention work into separate memory pools or near-memory compute. e.g PIM could be ultimate goal but maybe not for every workload. 
▪️ So, every architectural approach raises memory's share of system value. That includes HBM next to the GPU, CXL pools, near-memory accelerators and PIM. This all points to we need more and more memory in different forms. And @SKhynix is not shy of emphasizing it 😅 $SKHY $MRVL

I don't understand how with the N-2 law, and TSMC's reluctance for cutting-edge chip production offshore, TeraFab would benefit from this partnership. The only advantage they get is having a reputable partner onboard, and by reputable I mean with a proven track record, but then, they compromise on getting access to the most newest chip technology. I don't know whether Intel-SK JV has stirred off TeraFab ambitions, but last I heard, Elon was pretty happy with Team Blue.

Just discussions, but something may come of it

His @greens senator mate @NickMcKim did a video calling me ‘a rich landlord prick’ who only cares about my negative gearing then said ‘Hughesy, you can go fuck your self’ It led to people sending me DMs wishing me dead. I then told him I have four properties, didn’t negative gear any of them, and rented them at 50 percent market rate to help low income people, and gifted one of them to special needs adults. He didn’t take down his video. I then got told he had four properties. I asked him how much they rent for and does he negative gear them. He didn’t respond. The video is still up. I still get people hating on me because of it. The @greens are a complete fraud. The worst type.

Diamond Rapids matters because Intel is no longer designing Xeon as one giant CPU. Intel is designing a chiplet computing system that happens to fit inside a CPU socket. And that may be a much better framework for understanding the next decade of server semiconductor competition. tspasemiconductor.substack.c…

Guy is Taiwanese with an agenda. Rarely writes anything positive about Intel If Elon takes an equity stake in a TSMC fab its not 'TeraFab'. Would be a poor sign that SpaceX revenue won't ramp if he can't afford WFE

I think the story is bs Couple things are funny. Most of the people hired for TeraFab so far are Intel folks. So you have to believe that Intel's management have lost interest in the project. Not that former coworkers can't work together lol. But, both Lip-Bu and Pushkar have talked about their relationship as strong with Elon in last couple months. So this has to have happened recently. Taiwan regards their process technology as a national secret. Culpium says that TSMC would run TeraFab and SpaceX would get an equity stake which is worse than anything Intel was offering. Because TSMC can't license leading edge recipes. This would be a TSMC fab, not a SpaceX one. Intel was offering a process license.