00:00And look, this is a market when it comes to compute as something you've described
00:04as a market structure that's broken. Anj, exactly what is broken and what is AMP trying to do about
00:10it? Yes, great to be here. Thank you for having me, Danny. What's broken is what was broken,
00:16the same thing that was broken, you know, 100 years ago when electricity showed up,
00:21right? Electricity was a general purpose technology and it changed everything. And
00:25when it became clear and legible to the world that you could use electricity to make useful
00:29products like steam engines and cars, people started hoarding. And the reason they started
00:35hoarding that electricity is because it was so valuable to have access to that input at a time
00:41when it wasn't clear to people how it would be used, when it would be required, because at the time
00:48people were doing a lot of research with electricity. There was Bell Labs, you know, Edison was working
00:51on figuring out how do you electrify the world. And so when you have a scarce input like electricity
00:57or coal and the output is really valuable, people who have the largest balance sheet start hoarding it.
01:04What that resulted is in is about like 30 to 40 years of just absolute chaos and pain in both
01:12America and other Western nations until there was a breakthrough in standardization. What it required
01:20was two solutions, two important sort of alignment mechanisms to make sure that that electricity was
01:26distributed in even an accessible way to the people who needed it the most. And then that there was
01:31coordination to enforce the standards. The first was the standardization of electricity on ACDC,
01:37right? Until then, it was really hard to actually convert those two. So, you know, megawatts were not
01:42always just megawatts. And then the second was a coordination mechanism called the grid. And so what we
01:48figured out as a society is that when you have a scarce resource, what you need to do is to
01:54share,
01:54you need to pool capacity and needs across everybody. And to do that, you want to standardize
01:58that resource and then enforce through mechanisms like utility companies and grids,
02:03the ability for different types of customers like, you know, residential homes and factories and
02:09universities to all have the peace of mind that their baseload is taken care of, but that they can
02:16spike up and down. I see where you're going with this to create this again in AI data centers and
02:20compute. But how do you compete with these big hyperscalers that can throw so much money behind
02:26hoarding all of the power and all of the AI compute? Yeah, that's a great question. So back to the
02:32future, we have created a grid. My engineering co-founder created this grid at Google at scale.
02:38It's called a Borg export GQM scheduler. I won't bore you with the details. But not only have we created
02:43a
02:43grid that pools capacity across a bunch of different clouds and providers as well as AI labs. Today, I'm really
02:49excited to announce that we're announcing an unveiling $100 billion frontier grid security
02:55program. Over the next five years, we plan to raise $100 billion for grid security in America and our
02:59allied nations. We've raised a meaningful amount of that already. And we've also announced that we've
03:06deployed a meaningful amount of that in role models who are helping land, energy, power, you know, all
03:11parts of the supply chain in America, all the way to the model air and beyond to unveil, you know,
03:17what the
03:18good examples are for good behavior. One of them is Anthropic. We have deployed a meaningful amount
03:22in Anthropic this year. And we expect to do more. We're inviting partners. Of that $100 billion, 20%
03:28is committed from our own capital. And the remaining we're inviting partners to. In the last, you know,
03:32two weeks, we've received over $6 billion of interest. And so, you know, how does it, how do you make
03:38a dent
03:38in the universe? Well, you build products and solutions. And then you get people around the table and
03:41say, this is one of the most valuable infrastructure investment opportunities of all time. You know, by your own
03:46admission, about $5 to $7 trillion worth of AI infrastructure spend is expected over the next
03:51five years. And we think you should take 2% of that and put that in grid security. We think
03:54it's
03:54a pretty fair trade. And thank you for bringing us that announcement. That's super interesting. I mean,
03:58it's a market is kind of you're describing as one that's, and many people have, is chronically
04:02undersupplied. What then do you look at for potential signs of overbuilding? Are we anywhere near
04:09that? Is there any froth you think in the system? I wish. Look, because we are in the
04:15pre-standardization era of compute, what's happening is the spread between base rates and on-demand rates
04:20right now from the data we're looking at is roughly 2 to 5x. So, for example, if you look at
04:24a chip like an H100,
04:25which is a three-year-old chip, if you can even find it on any of the hyperscaler marketplaces, which
04:30honestly, you can't for this year, the on-demand rates are 5x. For GB300s, it's somewhere between 2.5 and
04:383x. And I think
04:39this is going to get worse. My hope is that by standardizing compute, we can have flops flow like
04:43megawatts. But for this year, from the data I'm seeing, it's not looking good. Not for independent
04:49teams who are not inside of hyperscalers.
04:52Before we let you go, I have to ask about your course at Stanford, which has been dubbed AI Coachella,
04:57which is awesome. And please let me just sit in the back at some point. And I want to ask
05:02you,
05:02because there was recently an op-ed from a Stanford senior basically describing almost an apocalyptic
05:08attitude among the students there, that like all the generational wealth has been had, that if you haven't
05:14already been early at Anthropic, which we should say you were one of the founding investors at, that you're
05:18kind of done for, and you should have already left Stanford and graduated and founded your own AI startup. I
05:24wonder if you
05:24get a sense from some of your students, from this idea that perhaps everybody is already far behind in AI
05:31race,
05:32and the winners have already been made.
05:35Oh, these kids. Look, I find the, you know, the gen alphas or whatever they are, they're just anxious,
05:41you know, and we need to remember that as stewards, as adults, as role models, we've got to remind them
05:46that they're, they're like sitting on the cusp of history. It's the best time to be alive and be
05:52graduating. The, you know, the, the senior project for this class is this one person frontier lab. And what we
05:57tell them is with the right tools, you can scale yourself endlessly. And I genuinely mean that.
06:02In fact, I expect to invest millions of dollars into alums from this class, you know, a former TA,
06:07we've taught this class for four years in a row. This one, this year just happened to be really
06:10popular. It's 500 students. It's the largest CS class on campus, as far as I'm aware. But I think
06:16of our, you know, we have to remind these kids that there's opportunities there. As long as they put
06:20in the work, you can't outsource your thinking to AI, or rather, sorry, you can outsource your
06:26thinking to AI, but you can't outsource your understanding. And when I spend time with the
06:30students, their brains are so incredible, because they've grown up on so much information. And the
06:36ones who actually are able to sit down and channel their focus are able to, and understand the topic,
06:42whether it's, you know, energy, uranium enrichment, or machine learning. I mean, this generation is,
06:46the opportunities available to them, as long as they're willing to not outsource their
06:51understanding and put in the work to create new kinds of products and services, I think it's going
06:55to be a generational cohort. I mean, I've never seen anything.
Comments