TSMC Execs Dismiss OpenAI Chief's $7 Trillion Chip Plan as 'Podcasting Bro' Vision

Taiwan Semiconductor Manufacturing Co (TSMC) executives have dismissed OpenAI CEO Sam Altman's ambitious chip-making proposal as unrealistic, according to The New York Times. Altman, seeking to boost AI computing power, pitched a $7 trillion plan to build 36 semiconductor plants over several years during a visit to TSMC's Taiwan headquarters. TSMC leaders reportedly found Altman's proposal so far-fetched that they privately referred to him as a "podcasting bro," reflecting skepticism about his grasp of the semiconductor industry's complexities. The world's largest contract chipmaker, already grappling with multi-billion dollar expansion projects, viewed Altman's scheme as overly risky given the massive capital requirements and market uncertainties.

The problem with tech bros is

[Rosco P. Coltrane]

However hare-brained their "visions" are, they're still dangerous and nefarious. Particularly Sam Altman.

good for them!

[LazarusQLong]

that 'vision' sounds entirely too optimistic, and doesn't appear to take into account anything in the real world... But that is just how it appears to me. I mean, what is the hard data that proves enough ROI to make that massive investment justifiable?

Re:

[LazarusQLong]

I mean, a vision doesn't necessarily need to have hard data behind it, but this sounds more like he was trying to make a pitch to this company to invest billions, hence the need for actual ROI projections based upon some level of reality.

Re:

[UnknowingFool]

a pitch to this company to invest billions, hence the need for actual ROI projections based upon some level of reality.

Pitching the idea to build 36 chip fabs with $7T in the next several years is like pitching an idea to increase the birth rates in some countries by flying in a lot of women to those countries. Only on the very surface level does it make any sort of sense. The part Altman got right is the money required. The practicality and details are ignored.

Let's start with building 36 factories has so many logistical issues to overcome. Zoning, land ownership, facilities resourcing (these plants use a lot of water and

Re:

[cstacy]

I think a quadrillion dollars should be invested in chipmaking. Altman doesn't have real vision. Every man woman and child should be making chips.

Well AFAIK they all do.
But we don't call them "chips" where I live.
(And people do make trillions shoveling them. Like Sam Altman!)

Those who know...

[Targon]

Intel put a LOT of money into trying to get their chip making back on track, and so far have failed miserably. It's not just about spending money, it's about having the right talent, even when you have the right equipment. The people at TSMC understand this all too well.

Re:

[sg_oneill]

Right. If it was just a case of throwing money at it, the Chinese would have had the whole market sewn up long ago and you better believe Apple wouldn't be using TSMC fabs.

But even if it could be done. Nobody is going to be giving Sam Altman 7 trillion to make chatgpt chips. AI might be important, but its not THAT important. Not yet, anyway.

Re:

[m00sh]

Right. If it was just a case of throwing money at it, the Chinese would have had the whole market sewn up long ago and you better believe Apple wouldn't be using TSMC fabs.

But even if it could be done. Nobody is going to be giving Sam Altman 7 trillion to make chatgpt chips. AI might be important, but its not THAT important. Not yet, anyway.

China doesn't have access to the equipment now. They were on track and were strategically blocked.

They don't have access to ASML equipment. They can't make their chips on TSMC either. However, they are welcome to buy Apple products made from these technologies but not the nVidia chips that could do AI.

Re:

[bill_mcgonigle]

When you have only bad choices go for the least damage. 4 years max or 8 years max is the hidden variable.

I know, everybody thinks the wrong one will spell the end of civilization, just like the previous six times.

Re:

[quantaman]

When you have only bad choices go for the least damage. 4 years max or 8 years max is the hidden variable.

I know, everybody thinks the wrong one will spell the end of civilization, just like the previous six times.

The difference is only one of the two choices has actually taken fairly concrete steps to destroying the US's tradition of performing willing and peaceful transfers of power.

So I need to have dinner with Bill, or a tiger that will eat my face. The problem is I've always found Bill really disagreeable and called him a 'face eating tiger'... so I guess the choice really is a toss-up.

Re:

[Targon]

Harris obviously is smart enough to outsmart the "stable genius" during the debate. She said it early what Trump was going to do, and she beat Trump easily. The only people who say that she isn't intelligent are those who listen to Fox News and the pseudo-conservative talk show radio.

Re:

[bigfinger76]

The debate was a wash.

No it wasn't.

Re:

[DarkOx]

Harris only one the debate because should the moderators were on her side, doing everything they could play gotcha with Trump while giving here a pass. They even admitted they only planned to fact check Trump.

Harris for her part managed say EXACTLY NOTHING during the entire debate. She clarified no positions on anything and articulated no substantive policy on any subject. She won sure, but only because a massive handicap, given to her. In any other venue it would have been an embarrassing performance.

Process?

[bill_mcgonigle]

This is a bad time to start building AI fabs.

Some bench prototypes use 1000x less power using different techniques.

AI problems, at least some classes, are robust against small error rates.

Traditional calculations need traditional chips with perfect error-corrected outputs.

Engineering trade-offs are obvious.

But OpenAI would love a $7T investment. In SV the biggest failures can be rewarded.

Altman and TSMC guy are optimizing for different optima.

Re:

[ndverdo]

it's now almost 2 years after the release of ChatGPT. There are many vendors pitching x10 and more from ASIC to FPGA to Wafer. Hard to understand that in 2 years matrix-vector optimized for transformer training and inference is not all over the place in lieu of the absurd pricing for CUDA devices.

Submitting a more than 100k context right now - just first step initially processing it is probably more than 0.5 kW.

Money is not capital

[RobinH]

This is what happens when you either think money can buy anything, or that money is capital. We use money to measure capital, but capital is actually the physical assets and human capital (available labour) that we can use to accomplish a project. When you're talking about the chip industry, which operates right on the edge of what's possible, the capital available to expand it by leaps and bounds simply doesn't exist, no matter how much money you throw at it. If you try, you'll just cause massive inflat