Steven Martin

@usr_bin_roygbiv Not such a bargain then. So Much for that idea

@Colosteve2000 not right now under any circumstances the cost to rent for 1 year is 80% of the cost to buy new

Hot take: They're not subsidized their margins are insane. They are just absolutely raping api customers. Anyone who has used deepseek or hosted anything and done the math on hardware/power costs knows this

@demishassabis @bodonoghue85 Now imagine this 4x faster using Cerebra’s technology and even more crazy add @TileRT_AI. Obviously you aren’t likely to need both but imagine the speed!

Awesome to see this innovation in text diffusion. DiffusionGemma is lightning fast, 4x faster than other Gemma 4 models! Congrats to @bodonoghue85 and the team who worked so hard on this - excited to see what people build with it!

@LikesFriedRice @usr_bin_roygbiv Well if you try to use DeepSeek the same way as GPT 5.5 yeah you could have problems. Every model has its own idiosyncrasies.

@usr_bin_roygbiv Isn’t DeepSeek terrible at coding. I’ve tried v4 twice and regretted it both times. Went back to Codex 5.5

@albustime @usr_bin_roygbiv @albustime I ask this with the utmost respect. If you are privy to the actual costs please share that knowledge with us. Even if it’s a generalization please consider it.

@usr_bin_roygbiv LOL anyone who has maintained thousands of H100's at the slurm level knows that your 3090-based numbers are not related to reality. The fact is rate limits and surge-routing make subscription cheap, and this API comp is done by people without knowledge of these dynamics

@treasureh8nter Lip-Bu Tan has stated that he isn’t going to build until he has the orders. Has this weeks news changed that? So far the people that know that answer aren’t telling.

SemiAnalysis says $INTC currently does not have enough money for major CapEx spending. They need to raise capital through dilution 👀 a 4% to 5% dilution would raise ~$25B They need to build new fabs and they simply don’t have the funds right now!!! 😱😱😱

@Im_IrushiK I think it depends are each individual. Using your vision to build something. Or punching clock

Unpopular opinion - 9-5 with a high salary is better than owning your own startups. Give me a single reason why you disagree with me

@tunguz Crap I voted for deepmimd as well thinking it was DeepSeek. Open source wins every time for me

Which AI lab do you trust the most to act in users’ best interests?

Per Commercial Times TSMC 3nm wafer supply remains constrained even though $TSMC just increased production to 175,00 wafers a month. Note to #JPMORGAN the customers that don’t get selected have to go somewhere! There are 2 choices Samsung Electronics or $INTC or do without! There are zero other choices right now

@IanCutress With Intels 18A filling up. There aren’t many non legacy fabs left. Samsung’s going to start getting more deals for their top fabs soon

People making noise about Samsung manufacturing IO dies for TPUs as a big shock. Samsung been making a good number of Intel chipsets for generations. 🤷

@intelfabs If the numbers are even halfway close $INTC will be at the trillion dollar cap by then if not sooner.

Bank of America just DOUBLE upgrade Intel to buy!!! Yes, DOUBLE UPGRADE $INTC BofA is modeling Intel as a future AI infrastructure company where foundry + packaging contribute nearly half of the growth story. BofA is no longer valuing Intel as a turnaround. They're valuing it as a future AI infrastructure platform. The foundry numbers are far larger than most people realize (very important) Apple wafers = $10.6B MediaTek TPU = $12.4B Terafab = $8B Other ARM server CPUs = $10B Packaging may actually be more important than wafers (that is ultra pro max important The Cadence 14A mention is actually a bigger signal than many investors think The CPU thesis is much more aggressive than people realize Ownership may be the most important near-term catalyst BofA says Intel is only: 16% institutionally owned, despite being the fifth-largest semiconductor/AI infrastructure company by market cap.

@IanCutress Even older than that I remember having to type in games out of video game magazines as i didn't even have a tape recorder for my computer. I still remember Zork and the Infocom text adventures. It changed my life!

You had a mouse? If my games weren't text, they came on tape and used a joystick.

@IanCutress @awscloud I presume its to keep costs under control and availability at the scale Amazon wants.

@Colosteve2000 @awscloud One could ask why they're not supporting MRDIMM...

It's here: @awscloud Graviton5. ➡️ 192 Arm V3 cores, 1 MB L3 per core ➡️ TSMC 3nm ➡️ 12ch DDR5-8800, 800+ GB/sec ➡️ *96x PCIe Gen6 ➡️ 420GB/s die-to-die bandwidth ➡️ Up to +25% performance vs G4 per core It'll be one of the biggest Arm CPU deployments in the datacenter.

@imnotharsh I think he is tired of being wrong. The puts and calls suggest a price as high as $151 by December. The catalysts keep coming in day after day.

$INTC Intel double-upgraded by Bank of America from Underperform to Buy and raised its price target to $135 from $96. Key Details from BofA’s Note, led by Vivek Arya: - Rationale: Increased confidence in Intel’s server CPU opportunities and external foundry business. - Earnings Outlook: BofA now projects earnings power of more than $6 per share by 2030 (specifically $6.24 EPS estimate), up significantly from the prior $3–4 range. - Valuation Method: They applied a 25x multiple to the 2030 EPS estimate, then discounted it back two years to reach the $135 target. The old sum-of-parts approach (based on 2028 estimates) was seen as under-representing longer-term CPU and foundry potential. Analyst Vivek Arya has repeatedly issued an Underperform (or Sell-equivalent) rating on Intel shares, often with price targets significantly below prevailing market levels. Recent targets have included levels around $40, with reiterations of Underperform even amid positive company developments such as Q1 results or design wins.

@MarkLundeberg I work in logistics. I can confirm tankers can easily implode

So, the 'condensate' name in BEC really is analogous to condensing steam: if we fix the external pressure (instead of volume), it does starts to shrink, though you still need to remove the latent heat. I guess though the container could collapse like this: :-)

Bose Einstein condensation is often depicted as a sudden collapse into a 0-entropy state like the animation below, but that's not true with fixed volume. In fact it's a gradual process as can be seen by the entropy approaching 0 smoothly.

@lousy38623 @MarkLundeberg @IanCutress you asked for citations :) Here you are. embeddedrelated.com/showarticle/14…

@lousy38623 @MarkLundeberg @IanCutress Most automobile chips are legacy nodes. 🚗 Engine Control Units (ECUs) → 40nm–130nm 🚗 Body Control Modules → 40nm–180nm 🚗 Airbag Controllers → 90nm–180nm 🚗 Power Management ICs → 90nm–350nm 🚗 Microcontrollers (MCUs) → 28nm–90nm 🚗 Automotive Radar → 28nm–40nm

A reminder that 90%+ of all fab litho equipment that has ever been made is still being used. ASML claims 95%. For other tooling it's probably around 80-90% too - I regularly get emails from a GF equipment reseller and they seem to cycle through tools.

@TheAIShrink @zerohedge If Oracle drops that much Berkshire Hathaway could just buy the entire company. Thus getting all those data center’s all in one swoop. Don’t tell me Buffet wouldn’t want one last big deal

@zerohedge Oracle's negative FCF = capex sprint, like every big tech company. Buffett looked at the margin timeline and decided: not my show at 94. The rest? Still all-in.

ORCL Free cash flow was negative $23.7 billion for fiscal year 2026,