Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

SoftBank sells its entire stake in Nvidia

cnbc.com

Post reply on HN
Print view View original

191–200 of 235 posts

Re: SoftBank sells its entire stake in Nvidia

#191

Post by jasonwatkinspdx »

lukeschlather wrote: post #130
skylurk wrote: post #107

Earlier quoted context omitted.

Sorta? If that was the full story, TPU would not be a thing.

I'm not an expert in chip design by any means but I think it's fair to say that TPU is a marketing term and it's not substantially different from a GPU like an H100. H100's cores are also called "Tensor Cores."

You are entirely mistaken. The TPU and GPU are organized very differently particularly with how the memory subsystem works.

In the big picture, TPUs are systolic arrays. They don't have threading, divergence, or similar.

GPUs in the big picture are SIMT, a hybrid of SIMD and multithreading where individual data streams in SIMD are relaxed to allow them to diverge somewhat.

Memory Wise the TPU can keep the partial products right in the array. Parameters and weights are held in large on die scratch memories, and backing those are streams coming from the HBM. The TPU acts as a single giant CISC coprocessor, and has much more predictable memory and communication patterns vs a GPU, which its design exploits for higher efficiency at inference vs GPUs.

So even if they use the word "tensor" and both have HBM based memory systems, how those are actually architected is very different.

Re: SoftBank sells its entire stake in Nvidia

#192

Post by rchaud »

timonoko wrote: post #101

I sold third at 180. I have seen it happen so many times in 50 years. You sit on your pile of gold and just watch it slowly disappear. But also selling ALL is a bad move, says one who sold all of Nokia at 1000 markka in 1996.

You're thinking like a lowly retail investor who can only spend his own money. Softbank spends other people's money. They have to show some profits now and then to make their financials look good and thereby attract more funding.

Re: SoftBank sells its entire stake in Nvidia

#193

Post by rchaud »

sixeyes wrote: post #95
reconnecting wrote: post #7

Earlier quoted context omitted.

$5.83B is too large to be a round trip. This is clearly an exit.

Why would they exit nvidia while still keeping upwards of $10B or whatever it is in OpenAI? Genuine question.

Nvidia is investing $100bn in OpenAI: https://nvidianews.nvidia.com/news/openai-and-nvidia-announc...

By purchasing OAI shares now, Softbank will probably be able to sell them to Nvidia at inflated prices later.

Re: SoftBank sells its entire stake in Nvidia

#194

Post by downrightmike »

creer wrote: post #166
lifthrasiir wrote: post #13

Note the first line: > SoftBank said Tuesday it has sold its entire stake in U.S. chipmaker Nvidia for $5.83 billion as the Japanese giant looks to capitalize its “all in” bet on ChatGPT maker OpenAI. They are switching gears, not exiting, folks.

As usual also "sold its entire stake" can be written as "investors eagerly bought SoftBank's entire stake of nvidia".

Bag holders gonna buy the bag

Re: SoftBank sells its entire stake in Nvidia

#195

Post by roncesvalles »

yesco wrote: post #112
hobofan wrote: post #49

Earlier quoted context omitted.

In many ways OpenAI is transitioning towards an end-user facing product business. They have by far the strongest brand among consumers and are positioning themselves to take on Google/Meta in the ad business. By proxy, having the strongest frontier model becomes less and less necessary for them and instead building a strong product by properly layering medium-strong models in a cost-efficient way is the priority.

I'd argue their brand might be too strong, ChatGPT has already begun to enter the same semantic space as "Velcro". Everyone I know seems to have tried it yet quickly you begin to realize that for most people ChatGPT == LLM, it seems everyone is using "ChatGPT" on completely different platforms. In the end, regardless of technical understanding, people will always shop around on price if the feature set is similar eno…

The thing is, laypeople aren't using anything other than Google Search even for LLM answers.

If I want an LLM answer to "is erythritol bad for you", I'm not firing up ChatGPT. I'm just typing it into Google, and the LLM answer it spits out is pretty good.

ChatGPT needs to be significantly more compelling for most people to use it for one-shot LLM answers over Google Search. And the minute Google removes the one-shottedness of its search answers, it's over for ChatGPT.

Imo ChatGPT is just "a feature not a product", in the search engine space, as the adage goes.

Re: SoftBank sells its entire stake in Nvidia

#196

Post by roncesvalles »

consumer451 wrote: post #181
yugioh3 wrote: post #129

Earlier quoted context omitted.

yeah it made me wonder what all this fuss is about. perhaps im numb to big numbers now. but $6b is a miniscule fraction of NVDA.

My main thought was "how bad is Softbank at their job to have only been in on NVDA for such a paltry sum?" Really living up their reputation? Or, am I way off-base here? I am just an outside observer and I would love to be educated. Did it makes sense for Softbank to be in so little on the company "selling shovels" to all their "gold mining" investments?

Or, how down bad is Softbank that it needs to sell its NVDA stake to raise $6B?

Re: SoftBank sells its entire stake in Nvidia

#197

Post by lenerdenator »

toomuchtodo wrote: post #147
lenerdenator wrote: post #142

Earlier quoted context omitted.

> It's just baffling they are still getting billions to spend. When you realize that most investors at that level are degenerate gamblers, and that they're even worse about thinking "the Generals were due!" than your buddy who can't look away from the DraftKings app for more than five minutes, it's not so baffling anymore. The only difference is that when you have that much money, you own the casino.

It is hard not to love the degeneracy of Son Masayoshi as a gambler. The trick is to not take it seriously. https://news.ycombinator.com/item?id=21799821 Bill Hwang of Archegos Capital Management is a close second (lost $20B in two days, and basically brushed off margin calls while his highly overlevered bets were collapsing, incurring $10B+ of losses at the investment banks he worked with). https://www.bloomberg.com…

> It is hard not to love the degeneracy of Son Masayoshi as a gambler. The trick is to not take it seriously.

Therein lies the problem. People do take it seriously, particularly on the idea that he's now due, and keep feeding him money to make these bets with, instead of tossing him out of the boardroom, like he should have been with his bet on WeWork.

Re: SoftBank sells its entire stake in Nvidia

#198

Post by boredtofears »

Topfi wrote: post #175
boredtofears wrote: post #146

Earlier quoted context omitted.

> Both DuckDuckGo and Kagi are, purely focused on result quality in 2025 far better than Google currently I wish this was true

You don't have to take my word for it. Open the three next to each other, type in a search query. Only one will provide part of the results as a wildly inaccurate "overview" that somehow manages to link to sources telling you the opposite or contradicts itself inside the text box, which can also not be deactivated. Those are the results you get from Google and that alone makes quality worse, pure and simple. More sub…

I had DDG set as my default for years and found myself resorting back to google so often I had to finally admit to myself their results were inferior even to the diminished google result. I just didn’t seem to get the result you’re claiming on the everyday type of queries I make.

Re: SoftBank sells its entire stake in Nvidia

#199

Post by adastra22 »

jasonwatkinspdx wrote: post #190
adastra22 wrote: post #85

Earlier quoted context omitted.

The fundamentals are different. Bitcoin mining is not intrinsically suited to acceleration on a GPU. It is a not-very-wide serial integer operation. AI inference on the other hand is basically just very large floating point tensor matrix multiplication. What does an ASIC for matmul look like? A GPU.

No, GPUs are not a particularly ideal architecture for AI inference, it's just inference needs way more memory bandwidth than a general purpose CPU's memory hierarchy can handle. > What does an ASIC for matmul look like? A systolic array, and ultimately quite different than a GPU. This is why TPUs et all are a thing. In general with a systolic array you get a quadratic speedup. For example with a 256x256 array, it ta…

The tensor cores in your NVIDIA chip are systolic arrays.

Re: SoftBank sells its entire stake in Nvidia

#200

Post by tim333 »

Topfi wrote: post #102
ForHackernews wrote: post #96

Earlier quoted context omitted.

There's no viable alternative for either of those yet. Massive network effects around Microsoft office, and Google's only real competition is the equally-bad Bing.

Both DuckDuckGo and Kagi are, purely focused on result quality in 2025 far better than Google currently, but that does not matter if one has no need beyond the most basic results. And much of the MS Office suite is actually a worse copy of another product, merely copied and forced via network effect. There are more than enough far better alternatives to e.g. Teams and OneDrive. The latter even makes Excel far less re…

>Windows is easily replaced for the truly average person who spends their entire computer use inside a Chromium browser nowadays

Maybe I'm out of touch but I don't know anyone for whom that's true. They all use either games or Office or Photoshop type stuff.

I mean you can replace it but all the options - mac, linux or chromebooks have friction.

Post reply on HN