Dr. Phil

The H20s were an intentionally downgraded GPU model designed specifically to meet export controls set by the U.S. government for China. NVIDIA ended up making too many, took too many orders, and weren't able to ship them out in time. And that gets even worse next quarter. NVIDIA is saying they expect around $8 billion in losses on the H20.

You would think that perhaps you could just sell these to other companies, and perhaps in some cases that might be true, but these things were nerfed. pretty hard to comply with the U.S. government's desire to not ship cutting-edge chips to China.

You would think that perhaps you could just sell these to other companies, and perhaps in some cases that might be true, but these things were nerfed. pretty hard to comply with the U.S. government's desire to not ship cutting-edge chips to China.

You would think that perhaps you could just sell these to other companies, and perhaps in some cases that might be true, but these things were nerfed. pretty hard to comply with the U.S. government's desire to not ship cutting-edge chips to China.

So it's certainly been a hectic six months for Nvidia's compliance department, who, along with the CEO, have been fairly loud in making the case that perhaps limiting China to low-powered chips will provide the incentive for the country to finally push through and develop its own chips that can rival the highest performing ones that are coming out of the U.S.

So it's certainly been a hectic six months for Nvidia's compliance department, who, along with the CEO, have been fairly loud in making the case that perhaps limiting China to low-powered chips will provide the incentive for the country to finally push through and develop its own chips that can rival the highest performing ones that are coming out of the U.S.

So it's certainly been a hectic six months for Nvidia's compliance department, who, along with the CEO, have been fairly loud in making the case that perhaps limiting China to low-powered chips will provide the incentive for the country to finally push through and develop its own chips that can rival the highest performing ones that are coming out of the U.S.

As they say, necessity is the mother of all invention. And there have been some chirps that perhaps Huawei, the Chinese company that would be most likely to challenge Nvidia, has made more ground in this area than people assume. I mean, I have been there.

As they say, necessity is the mother of all invention. And there have been some chirps that perhaps Huawei, the Chinese company that would be most likely to challenge Nvidia, has made more ground in this area than people assume. I mean, I have been there.

As they say, necessity is the mother of all invention. And there have been some chirps that perhaps Huawei, the Chinese company that would be most likely to challenge Nvidia, has made more ground in this area than people assume. I mean, I have been there.

On the call, Jensen Wang, the CEO, said Microsoft processed about 100 trillion tokens over the last year, which is a 500% increase year-over-year. A token is the base unit of LLM processing. It's slightly more nuanced than this, but for the sake of brevity, one word equals one token. Anytime a word is taken in from a query or output as a response, that's one process token.

On the call, Jensen Wang, the CEO, said Microsoft processed about 100 trillion tokens over the last year, which is a 500% increase year-over-year. A token is the base unit of LLM processing. It's slightly more nuanced than this, but for the sake of brevity, one word equals one token. Anytime a word is taken in from a query or output as a response, that's one process token.

On the call, Jensen Wang, the CEO, said Microsoft processed about 100 trillion tokens over the last year, which is a 500% increase year-over-year. A token is the base unit of LLM processing. It's slightly more nuanced than this, but for the sake of brevity, one word equals one token. Anytime a word is taken in from a query or output as a response, that's one process token.

So that means Microsoft is processing about 200 billion pages of single-spaced text per year. But as Nvidia noted on its blog, the exponential increase in token processing has come from generating tokens, output, essentially generating words. So while we're inputting more and more data to these things, they're spitting out from what Nvidia is saying a far larger asymmetric amount.

So that means Microsoft is processing about 200 billion pages of single-spaced text per year. But as Nvidia noted on its blog, the exponential increase in token processing has come from generating tokens, output, essentially generating words. So while we're inputting more and more data to these things, they're spitting out from what Nvidia is saying a far larger asymmetric amount.

So that means Microsoft is processing about 200 billion pages of single-spaced text per year. But as Nvidia noted on its blog, the exponential increase in token processing has come from generating tokens, output, essentially generating words. So while we're inputting more and more data to these things, they're spitting out from what Nvidia is saying a far larger asymmetric amount.

This jives with what David Solomon, the CEO of Goldman Sachs, said a few months ago, that an LLM can accurately generate 95% of an S-1 statement. An S-1 statement is about a 300- to 400-page document, a super dense financial and legal document that companies file when they're about to IPO. He said it used to take a six-person team two weeks to do that.

This jives with what David Solomon, the CEO of Goldman Sachs, said a few months ago, that an LLM can accurately generate 95% of an S-1 statement. An S-1 statement is about a 300- to 400-page document, a super dense financial and legal document that companies file when they're about to IPO. He said it used to take a six-person team two weeks to do that.

This jives with what David Solomon, the CEO of Goldman Sachs, said a few months ago, that an LLM can accurately generate 95% of an S-1 statement. An S-1 statement is about a 300- to 400-page document, a super dense financial and legal document that companies file when they're about to IPO. He said it used to take a six-person team two weeks to do that.

So a bit speculative here, but that exponential increase in token processing might signal that these things have actually started to see more adoption in business use cases, even if those cases aren't made formal within the company.