fodkin @fodkin

the real cost of AI, one GPU at a time Fab 18, Tainan Joined June 2023

Tweets

8
Followers

50
Following

10
Likes

55

fodkin @fodkin

2 weeks ago

Jensen Huang Built a $4 Trillion NVIDIA Empire on One Insane Math Curve, Its AI Chips Double in Power Every 1.43 Years and MIT Just Measured the Machine Behind It MIT FutureTech tracked every NVIDIA datacenter GPU from the mid 2000s to 2025 and timed the growth AI math on these chips doubles every 1.43 years, that's about 127x more power every decade Memory only doubles every 3.3 years, so compute runs away from memory about 16x per decade Prices double every 5 years and power every 15, so every dollar buys way more AI each year Jensen Huang rode that curve while one H100 sold for an estimated $25,000 to $40,000 US export rules cap what China can buy, leaving a 23.6x performance gap on the best chips Proposed changes could cut that gap to 3.54x and NVIDIA already took a $4.5 billion H20 charge AMD and others are closing in, but MIT says not fast enough to shake NVIDIA's lead

fodkin @fodkin

2 weeks ago

1 0 4 219 4

0 0 4 131 3

View Details

fodkin @fodkin

2 weeks ago

Nvidia sells a chip that costs about $3,300 to make for up to $30,000 and that 9x markup turned Jensen Huang's company into a $120 billion a year profit machine while AMD, Google and Amazon race to break it one cost model puts the actual brain of the chip at around $300 and the memory at $1,350 an Epoch AI analysis found memory now eats 63% of what AI chips cost to build and the logic just 13% that's roughly 89% gross margin on every chip Nvidia ships Nvidia kept about $120 billion in profit on $216 billion of revenue in fiscal 2026 the secret weapon is CUDA, software Nvidia shipped in 2007 years before anyone cared about AI Microsoft and Oracle signed up for AMD's chips and Anthropic agreed to use up to a million Google TPUs 5 minutes and you'll read a chip company's earnings like a pro, just watch the margin line ↓

starmex @starmexxx

4 months ago

x.com/i/article/2070…

33 35 364 1.4M 1K

1 0 4 219 4

View Details

fodkin @fodkin

2 weeks ago

Einstein's 1905 full paper: famaf.unc.edu.ar/~dpaz/EM2/Eint…

0 0 0 16 0

View Details

fodkin @fodkin

2 weeks ago

Jensen Huang packed 2 miles of copper into one Nvidia rack because optics would burn 20 kilowatts while a $300 million trading cable backed by Netscape's Jim Barksdale lost to radio towers and sold for $127 million and the speed of light explains both Einstein's 1905 paper on special relativity made the speed of light the limit no signal can beat Grace Hopper handed out 30 cm wires and called each one a nanosecond because that's how far light goes in fiber a signal needs about 5 nanoseconds per meter, so crossing a data center costs 2000 GPU clock ticks Michael Lewis wrote about that cable in Flash Boys, it was built to cut Chicago to New Jersey to 13 ms microwave towers did it in about 8.5 ms because light in air barely slows down Nvidia crams 72 Blackwell GPUs into one liquid cooled NVL72 rack moving 130 TB/s to keep the wires short 5 minutes and you'll budget latency like a chip architect↓

starmex @starmexxx

4 months ago

x.com/i/article/2070…

33 35 364 1.4M 1K

1 0 1 154 0

View Details

fodkin @fodkin

2 weeks ago

Nvidia's gigawatt AI factories would need 200,000 kg of copper just for busbars so Jensen Huang is reviving the power war Edison lost to Tesla and Westinghouse Nvidia, Infineon, Vertiv and Texas Instruments are moving AI power to 800 volts DC by 2027 Edison's company merged into GE after he lost in the 1890s, and 130 years later Nvidia is betting on his DC the napkin math is P = V × I and loss = I²R, the same megawatt with 15x less current and 219x less heat Microsoft, Meta and Google wrote their own DC rack spec while AI data centers head toward more power than all of Japan uses the same formula works on your solar panels and the 800 volt Porsche Taycan charging at 270 kW, bookmark it

starmex @starmexxx

4 months ago

x.com/i/article/2070…

33 35 364 1.4M 1K

0 0 0 102 0

View Details

fodkin @fodkin

2 weeks ago

Nvidia’s best GPU is capped at roughly the size of a postage stamp, Cerebras built one chip the size of a dinner plate with 4 trillion transistors, and OpenAI signed a deal worth more than $10 billion to run on it the reason is one formula, yield equals e to the minus defects times area, double a chip’s size and the share of working dies drops from about 44% to 20%, Jensen Huang even admitted a design flaw in Blackwell crushed yields and called it 100% Nvidia’s fault, and this video shows how Cerebras beat that math with spare cores every AI chip lives or dies by this one line, 5 minutes, bookmark it