100%
1/1 predictions confirmed
4 pending · 0 expired
13
companies · 42 data points
And what's wild to me is that probably on the margin, Anthropic is losing share to OpenAI,
You know, SpaceX was a magical moment. And there's only one SpaceX, but to quote Bill Belichick, you know, it's on to Cincinnati. Got it. And there's a lot I think I had for SpaceX, so the SpaceX story isn't over.
the one thing OpenAI is like, they're a very good product, just really, really good. And I just always go back to the statement from Eric Vistria, you know, who's a venture capitalist at Benchmark, brilliant man. And he just, he said it was on X. You know, it might be that if OpenAI never releases another model, they're still a really valuable company because they have so many users and they're so good at product and they do have a lot of distribution.
this is the first time that OpenAI has not been able to clearly beat a competitor on all metrics, you know, a month after a new frontier model was kind of released from one of their chief rivals.
the companions from Grok have been interesting to play with, and they certainly make it more engaging.
Is XAI and Apple are natural partners. There's been a lot of news about Apple thinking about buying perplexity or mixed draw, but that's just a bandaid. Those companies don't get Apple what they need. And Apple's had an incredible partnership for many years now with Google that's generating tens of billions of dollars for both companies. And I think there is solid industrial logic for partnership. You could have Apple grok, safe grok, whatever you want to call it, you know, so that Apple feels c
particularly to crock for being the best product is the best product. Doesn't always win in technology.
And so from an industrial logic perspective, something that I think makes a lot of sense, particularly since open AI bought Johnny Ives hardware startup, which will place them into competition with Apple at some point. Is XAI and Apple are natural partners. There's been a lot of news about Apple thinking about buying perplexity or mixed draw, but that's just a bandaid. Those companies don't get Apple what they need.
And sports, you know, they say defense wins championships and on the internet distribution wins championships. And grok for has formidable competitors with lots of distribution and Google and. Meta. Meta. If they get their act together. Microsoft. Microsoft.
the companions from Grok have been interesting to play with, and they certainly make it more engaging.
XAI and Apple are natural partners. I think there is solid industrial logic for partnership. You could have Apple grok, safe grok, whatever you want to call it, you know, so that Apple feels comfortable. And that probably also helps grok in the enterprise.
particularly to crock for being the best product is the best product. Doesn't always win in technology. And grok for has formidable competitors with lots of distribution and Google and. Meta. Meta. If they get their act together. Microsoft. Microsoft.
And I think it's worth mentioning, this model was trained on Hopper. So this model is probably about as far as you can take the last generation of NVIDIA GPUs. The next models we see, you know, Grok 5, you know, 05 from OpenAI, whatever they're going to call the next Gemini, they will be trained on Blackwell, and I think that will be a really big step function.
And I think the benchmarks that matter are RKGI 2 and Humanity's Last Exam, where it did, in the case of RKGI 2, it's called semi-private, which is important because it means that some of the questions were held back because the criticism of these models is they memorize the answer. And so RKGI 2, it has not seen the questions before, and it did roughly twice as good as, you know, the state-of-the-art Google OpenAI and anthropic models. And then on Humanity's Last Exam, you know, we can call it
XAI and Apple are natural partners. There's been a lot of news about Apple thinking about buying perplexity or mixed draw, but that's just a bandaid. Those companies don't get Apple what they need. And Apple's had an incredible partnership for many years now with Google that's generating tens of billions of dollars for both companies. And I think there is solid industrial logic for partnership. You could have Apple grok, safe grok, whatever you want to call it, you know, so that Apple feels comf
And Apple's had an incredible partnership for many years now with Google that's generating tens of billions of dollars for both companies. And I think there is solid industrial logic for partnership. ... And then it also helps both Apple and Google with this DOJ trial.
And grok for has formidable competitors with lots of distribution and Google and. Meta. Meta. If they get their act together. Microsoft. Microsoft.
I think it will become a standard and it makes it really easy for an LLM like Stripe can just integrate with MCP.
Anthropic developed something called the model context protocol that OpenAI just adopted. I think it will become a standard and it makes it really easy for an LLM like Stripe can just integrate with MCP. And then any LLM that uses MCP can, you know, interact with Stripe. And this is solving a big, big problem for agents in terms of just making them, you know, much easier to use, much more standardized.
core weave bought weights and biases, which I think is a really interesting acquisition for them that kind of, kind of further differentiates them and maybe decommodifies them a little bit.
I do think core weave runs these big GPU clusters as well as anyone. And there aren't that many people on planet earth who can run them well. And, you know, it's kind of easy, you can just look at the NPV of their contracts and the asset value.
NVIDIA, by giving kind of like a standard reference design, which someone like CoreWeave bought to CoreWeave, you got GPUs in the market a lot faster.
In the next five years, I think zero. I think it's really, really hard. But over the 10 years, who knows? And if you're the CCP, 10 years isn't that long. If you're America, 10 years is an eternity.
I think we're putting enough friction into the system that it does theoretically give America an advantage at the cost of creating tremendous incentives for China to develop their own semiconductor ecosystem and, you know, pressure to kind of, you know, necessity is the mother of innovation. And these export controls are creating an immense incentive for China to be really algorithmically innovative. And you saw that with DeepSeq where there were some real algorithmic innovations.
if America cannot keep illegal drugs out of, you know, out of our own country, these GPUs are arguably dramatically more valuable per unit. You know, like a Blackwell is like, you know, the size of this. And it's just hard, particularly like, you know, America, we're trying to keep drugs out. Here, China is trying to bring the GPUs in and it's kind of significantly, you know, I think harder to prevent than, than, you know, smuggling of illegal drugs. Now everybody does their best, but yeah, the
Blackwell is going to be a very successful product cycle. There's a little bit of a prisoner's dilemma where you're going to spend on Blackwell almost no matter what, because if you don't, you're afraid that you cede a big advantage to a competitor.
sending NVIDIA GPUs to people who are adopting NVIDIA's reference architecture, which they believe is the best architecture for those GPUs. Whereas each hyperscaler has their own slight tweaks to that server, that hyperscaler does it so it looks more like their other servers, but maybe it doesn't run quite as well as the standard NVIDIA reference architecture
If that trend continues past the July quarter, then I would say, hey, that's reason for concern. But this is a very understandable time for that to happen, from my perspective.
I think the only reason NVIDIA was able to grow through this product transition is because of reasoning models like DeepSeq, which are just so compute hungry.
And so the big three kind of cloud computing, you know, hyperscalers, their incremental share of revenue went down. And that was nothing but good for NVIDIA because they now have kind of a more fragmented base of buyers who have less power over them.
I really don't think like if, if NVIDIA had not, not put any money into core, we've not put any money into kind of other of these Neo clouds, I don't think it would have impacted their revenues at all. At all. They would have sold. More to Meta and more to Tesla.
we have some very exciting stuff coming out later this year
if you're only able to get a third of the theoretical maximum the amount of math your chip can do you can only serve a third of that number of customers but if you get 90 you can serve three times more customers
it'll be the same way of the number of headline flops that we claim uh you will find you will get more than 90 utilization of those flops because the whole thing is built around getting to that really high utilization number
we price 10x cheaper than nvidia in terms of dollars per token
for an nvidia 8x h100 box you can expect to pay around 300 000 for that system which is eight of those chips inside of one cage inside of one server you got it that will draw around 10 kilowatts and the rule of thumb is that a walk costs around a dollar a year okay so that would be ten thousand dollars per year to run that eight h100 cluster you got it what this means is that if say the lifetime is five years you're spending three hundred thousand dollars in capex and you're spending ten thousan
i think the in a lot of ways what i'd say broadly is i think for a while it's going to be big businesses are winners big businesses that use ai thoughtfully and the reason is is that with what is happening what what o3 showed us openai's new model the combination of reasoning and test time compute like i think you're if you're a big business and you can pay a million dollars to let an ai think for six weeks about the most important question for your business that's going to be a profound advanta
i think fsd works today and it's gonna yeah it's it's gonna cross into mainstream adoption where i already notice particularly if i'm taking an uber late at night i really really prefer to have a tesla just sometimes you get an uber driver who's who's tired and i just feel a lot safer if they have the fsd running and then obviously using it for yourself is amazing but i think it's going to continue compounding at an accelerating rate
I think Facebook had a sound argument that they could have done that cheaper.
I think there's a lot of reasons to think that in that NVIDIA may finally begin to lose some share for the first time. But I mean, the market is exploding.
you are lifting more than five times the mass to orbit. And based on other statements, that 100 metric tons is a very conservative estimate. And you're doing it at call it 10 to 15% of the cost. So this is a, you know, we can all do the math, but we can envelope it, you know, roughly a 50 X huge change. And this, you know, massively changes unit economics for Starlink for sending anything into orbit.
once you do that, they should be able to send over 100 metric tons to orbit at a variable cost of under $2 million per public data.