Daniel Newman on AWS, AI Infrastructure, and the Future of Tech Investment | AWS re:Invent 2025

Hey everyone. Welcome back to our live coverage day, one of Amazon or AWS Amazon Web Services, reinvent. I am really happy to be joined by this guy here.

You may not recognize him. He looks 10 years younger. Sand's the goatee, he's lost weight.

Come on, he's in fighting shape. My friends, you know, the Steelers could use some players. Maybe you could help us out here.

You watch that. They would, they would Pitiful. That Was, I was proud of Aaron to come up and basically say no.

Just say, oh, Afterwards, We need to do better. And I'm part of the problem. You know, like that's As there's a leadership message there, right?

I mean, I, I do, I give them credit. It doesn't make it hurt any less. I think there's a lot of parallels.

AWS they're, they're doing that right now. They know that they sort of miss the first wave of this AI pivot. 0.

And they're recognizing the things they need to do, and they're, and I think they're making those pivots this week. They made a really big pivot. I think you're gonna tell me Matt broke his, had bloody notes too, but No, um, that Was good though, right?

Yeah. That was way To get back on track. The way you brought that back in here, it's like, I've been here Before.

Absolutely. So, if you don't know this gentleman, Daniel Newman, CEO Fu group, he's my friend. Um, I'd love working with him and he always has great insights.

Like this little tidbit you just brought us back to AWS So we, I was talking with Mike Ard this morning. Certainly this is all AI all the time this year at Reinvent. Um, what's your take?

I mean, obviously they're making a pivot. They're going hard at ai. Well, I think they recognize, so we are in the era of AI cloud and, you know, future, and we do our signal evaluation.

AWS did not score in top two. Did not. It, it has fallen behind both Google and Microsoft in this era.

Now, AWS has this massive advantage called a humongous customer base. They have a huge customer base from the first cloud era. And by the way, anyone that's worked with CIOs, as long as you have, I mean maybe as long as I have knows that in the enterprise, it's not the same as the consumer, right?

And it's not the same type of pace of sentiment. There isn't a new thing that comes out and everybody just ditches Google for chat. GPT when it comes to cloud.

These companies are deeply integrated in AWS And so while this AI pivot maybe has forced some companies to do more multi-cloud and have seen some workloads go to Google and some, the opportunity is still really in place for AWS. So they had a couple of big things they needed to prove this week. Uh, the first thing they needed to prove is that they really are the place for the enterprise to commit building ai, generative AI applications open up to more developers.

com community knows really well. Sure. Um, that they have infrastructure, that they have an approach both their partnership with Nvidia and making sure the market understood that they have the Nvidia they needed.

That was something that early on, they maybe rotated a little quick to their homegrown TRA chips before it was the right answer for a lot of their customers. And then of course, um, you know, they needed to, to show that they could really open their, um, aperture to the developers of the AI era. You know, and I kind of, I kind of glossed on that.

But like, you know, they, they focused on that with their announcements with Kira, what they're announcing there. Yeah. Um, agent Core with the ability to build agents, they need to be the place where people are building the applications that run their business.

So they had that mission this week. I think Matt Garmin in his keynote, he always gives a good keynote. He's a real product guy.

He's a real product Guy. He really is. He, you could see It, it was a very product kind Of That's the AWS Y.

Yeah. That is the AWS Y. And, and of course, you know, I think the one thing that they, in the last 10 minutes they did like a speed round of every other announcement that they had, you know, uh, Kubernetes and all their regular instances and CPU instances and storage buckets and everything else they're doing.

Mm-hmm. Um, but like you could tell this was all about being a prove it moment that we are a cloud that's ready for the AI era. And I think they did a good job.

Absolutely. Couple of, you know, plays off of what you said. Number one, you're right, this almost wasn't about cloud.

We didn't hear as much about S3 and Lambda and Serverless and all these cool things that AWS pioneered. We did hear a lot of Agen ai, you mentioned DevOps, actually one of the three main agents that they announced, they called the DevOps agent, which we, you know, I thought, you know, tip of the hat to them for that. Um, another thing they announced though, and I don't know, maybe it didn't register on your radar screen, is, um, what's it called?

The Forge. Nova Forge. Nova Forge.

So I look at that and say, wow, that gets me excited. And then I think, well, how many organizations do really want to build their own foundation model? Well, and that's a great point.

And, and that was something I probably should have had on my third thing in the first list. But first of all, Amazon is part of their prove it. So I mentioned, you know, the train and the vertical stack of infrastructure.

They talked a lot about that. But the other part of the prove it is I talked to, uh, investment, uh, bankers. I talked to other analysts.

I talked to media and press regularly of, I talked to enterprise customers. A lot of 'em didn't even know Amazon built models, right? So, you know, they had Titan originally they did Titan, and then they really went in with Nova.

But like really opening the market to understanding that they are in that business. 'cause that really is the complete, we have all the compute mm-hmm. And all the infrastructure.

We have all the developer tools and frameworks, and then we have the models, right. And being able to say, like, and then of course Bedrock, where you could basically bring all this, you plug it in and you, and you deploy the, the, the applications. That is the full stack story.

And you know, you saw last week when people started to get the idea that Google and TPU could be a competitive offering to the sort of open GPU era. Mm-hmm. Um, specifically Nvidia, but it could be a MD, it could be anything.

Um, and despite the fact that I would argue that most of what was presented is not quite factual about that, the thing that's made Google so attractive, the reason it's risen to all time high is it's market cap. So you're Saying what was offered by Google, not what was offered by AWS or both? No, No.

What I'm saying, well, first of all, I'm, I'm eventually gonna get to my parallel. It just takes me a while. But like, okay, what Google's doing in the full stack mm-hmm.

Has given the market a lot of appreciation for Google. Yes. Now, Google is unique and they were Rewarded.

It's corpus of data is unique. It's a little different. But Amazon has quite a bit too, from its ads, from its commerce business, a lot of very unique data.

And of course it has a lot of enterprise data, which is where the majority of data still actually sits. But Google's finally getting credit for being full stack. They're getting credit for saying, Hey, you built TPU, you've built the networking, you've built compute chips, you've built, uh, They have the vertical stack.

You've built the agent and applications and Vertex and builders. And of course then the models, Gemini proving to be very good. Amazon wants to follow suit.

They want to say, Hey, we got Traum. Hey, we've got Nova. Hey, we've got Agent Core.

Hey, we've got Kira. Hey, we've got like all the things that you that are required for basically an enterprise to say we can run all our AI in one cloud potentially makes it more valuable. Amazon hasn't gotten a lot of credit for that.

And so this was an important inflection. Now we had to see how much the market digests that. Yeah.

And how much they believe it. Let me ask you a question, though, at the TRA chips and the Google TVs too. Is this really a competitor to Jensen and the NVIDIA people?

Or is Broadcom? Well, Broadcom makes the chips for, for, for Google. So Broadcom is the full end-to-end design.

They do it all. Um, But not for the Amazon tra No, no. TRA is different.

TRA is Marvell. It's all chip. It's actually sourced through a number of different suppliers.

Um, but, but largely Marvell. Um, I think the right question here is are custom AI chips competitive to merchant silicon and specifically the Nvidia ecosystem? And like I said, to a lesser extent, you could argue the A MD ecosystem.

Sure. Well, they, they're the up and calmer. Yeah.

But like, I think the answer is, and I I, I think we've talked about this, we've market modeled it. We do believe the custom chips will actually grow faster towards the end of the decade. And here's the reason why.

There are a small subset of companies that are the largest buyers of infrastructure compute. So there's a benefit. And the reason Google really invested, and now remember, there's seven generations in.

It wasn't like they came out with a new chip and everyone's like, oh, it's gonna replace Nvidia. Seven generations in, I think in their sixth generation, they were able to train a first kind of high performing, large language model, Gemini advanced that was on their own infrastructure, which was a big breakthrough. 'cause obviously before the idea was like, everything has to be trained on Nvidia.

So that's, that was a pretty big inflection. But it, but Google, you think about what Google does, it does AI all day. It's doing massive volumes of infras of, of inference all day long on its own infrastructure.

It needs to think about its margins. So a company like Google that's doing that much scale, of course, might look at, Hey, here's three or four specific workloads. Let's build a custom chip that really works for everything we do for search for recommendation engine.

Mm-hmm. We'll build it at scale. We'll invest big upfront, but our, our, our cogs will get a lot better than when we don't have to pay that 75% margin to Nvidia.

Having said that, though, like the, they're always, at least as far as we see it gonna be probably one if not two generations behind in terms of the most advanced NVIDIA chips. So token economics, inference, uh, efficiency, uh, performance, memory, throughput, all those things that are really critical to training, uh, pre-training to doing large models, um, but also just to scaling tokens in like agentic eras may or may not be as efficient on those, um, on, on the custom chips when you need the flexibility. And so what I think ends up happening is it's really our, our vision of the AI market is it's all hands on deck.

You'll see me say this anytime you see me talk about the bubble. And social is like right now, every single wafer that TSMC can produce a chip on is being sold. So Nvidia has a certain amount of capacity.

Broadcom has a certain amount of capacity. Broadcom can make a certain amount of units, and every one of them is being built these companies. So you gotta expect, by the way, not just Broadcom, it's others, but Amazon, Google, Microsoft, Oracle, OpenAI, um, they're all meta.

They use so much AI that they're gonna use some of their own chips. And by the way, this isn't new. They've been doing this for a while.

We've seen the arm movement with CPUs that move certain workloads off Intel and off mt. Right. This is a business decision, but it's not necessarily because they believe it's the most performant of the best technology.

They're trying to fill gap, hit margin levels, understanding that not every workload needs to be on the most advanced chip. And, and of course in the end, they're looking at delivering EPS value. And Absolutely.

If they're meta and they're doing ads, can we build a lesser priced high performance chip that just focuses on serving ai, AI slop to us all day long? As we, as we, as we run around our meta application, There's, there's, there's a world for AI slop, but you know, it it, but this is not a new strategy. No.

Right? It, it, it was always, it reminds me of when I first got into security 25, 30 years ago, Asics, ASIC based security appliances. Yeah.

This was before we had SaaS or NY of that Stuff. It's just asics too. Right.

And, and that's what, and back basically it's the same stuff over, you know, history repeats itself. You just, that that's what we're dealing with. And you know what, for certain security functions, you know, custom made asics still Penny for penny dollar for dollar gave you the best bang for the buck as long As it was for the right use case map.

It, it's that narrow use case. Yeah. But it's the same thing here.

Metas serving ads is, is a particular one. Um, I'd still look, if I was a betting person, I like NVIDIA's, you know, seat maybe better than some of the other players. Yeah.

But if, if they could maintain just one or two generations behind Daniel, that's a lot of value. Freaking market. They can get a lot of value.

There's a lot of need. There's a lot of, um, you know, sort of deprecated workloads, just like on compute. Like people didn't just throw their last generation, uh, data center server absolutely CPU away.

They, they use 'em for less important workloads and they would, you know, they would prioritize new workloads and they would upgrade and they would replace, and they would add. Right now most of what's going out is, is is new. Right.

Um, and interestingly enough, like, you know, Nvidia, I say it's three to five years minimum before the custom chips could eat meaningful market share, if ever. And I still think it's more of an and than an or. And I think there's a lot of kind of the, the biggest risk to, to NVIDIA over time is as these big buyers, the ones that are buying so much of their technology, are able to do more and more on their own chip and with their own margin structure, is will that create any margin pressure On Nvidia?

On Nvidia? So it's not so much volume. I think if anything, the risk is more sits in the margin.

But then there's other things too, like that Nvidia does. It's just so unique. Like their, their entire backbone being optical.

Yeah. And you know, like, you know, nv, they don't, and Spectrum acts like, you know, but they're, they, the transport that between the GPUs is so efficient and so fast that like the latency issue is, is really, when you talk about clusters of this size, it's, it's, it's really meaningful. Like until we get co packaged optics on, on all the backends of all these, of, of all these spines and everything, it It, look in my day it was the buses.

Yeah. Right. On the, on the motherboards and forth, still so forth.

Well, it's still my day. I've had a long day so hard on yourself. It's been a long day, but it, it was the buses and that, you know, that latency in there.

But what I think the other thing is, you know, necessity's, the mother of invention, it keeps Nvidia on its toes to be constantly innovating, to be constantly staying that one or two generations. And they've been great at that talking about, you know, Blackwell three hundreds are, are now ramping shipping and volume. Uh, we already know Ruben's coming and we know Feynman's gonna come after that.

They're coming out two generations a year. And by the way, that's very hard to do with custom. It's very hard to do.

Like I said, there's a few companies in the world that can do it. And then the thing is, is like it's all run in their cloud. Like I will really, like, I've heard Foxconn's ramping up building servers with the Google TPU chip that could be sold outside.

But like, um, I think there's, you know, like NVIDIA's a merchant silicon company, meaning that enterprises can buy it. Neo clouds can buy it. Like, you know, like, is Amazon gonna really buy a volume of Google chips like ever?

You know, of course not. They're probably gonna build their own. I mean, Meta's unique, Meta's not competing with the hyperscalers, though.

Meta's slightly different. Well, they have a different, they're not three hyperscalers, They're not the, they're not the selling cloud services In terms of a cloud, but they're a hyperscaler in and of themselves. But like, like the net of it is, is like, I just think it's, it's all hands on deck.

I I think that, you know, we're hearing by the way, it was $1 trillion of expected AI infrastructure by like 20, 29, 20 30. You got a fancy Game going. We're getting, we're getting a thing going on live here.

I don't know, it's ESPN though. So, but, uh, But um, now that number's been risen, possibly two to 3 trillion. So what I'm saying is like, look, we do market sizing, we look at the market and it's kind of simple.

There's market size and overall tam and then of course there's margin in the TAM risk. And so on the, on the market size, it's like the market just keeps getting bigger. We're building up 65 gigawatts right now of capacity, um, up to 80 here in the us.

And I mean, you look at, the numbers are anywhere from 20 to $50 billion of spend per gigawatt on infrastructure. It's massive. And so, like, they can't build enough TPUs, they can't ramp the supply chain fast enough.

And that's another thing is like, we know Intel's on the way up and Intel's going to win, uh, more and more foundry deals. Some percentage Of has To work. No, it literally has to because there's no other Way.

TSM Can't build enough fabs fast enough. So, but what I mean is every wafer is gonna be sold. And so right now the point is, is like every wafer and every chip that that Nvidia can build is gonna be sold their backlog where they have half a trillion for next year, half a trillion order of visibility to next year, not including their open AI deals, not including their open AI deals.

So a lot of people are like, well, it's open AI risk. Well this isn't, it's $500 billion of potential sales between now and 13 months from now. It's incredible.

Not every vendor has that cushion though. No, Not everyone does. But, but my point is, is like if, if real realistically A TPU or, or a train was a risk to nvidia, you would start to see it show up.

They, I don't think NVIDIA truly has a real risk on the horizon. No, But they're fighting that battle in the market every day. Absolutely.

The market is trying to create the FUD that they do. And what I'm saying is they don't, but it, again, it's not zero sum. There's so much zero sum thinking.

It's not at the cost. Crazy. Alright.

A SEC Yep. Let's clean that up, um, at the cost of Nvidia. And that's my big problem is like, look, maybe over three or four or five years you'll see some maybe Nvidia shed a couple points of market share.

Maybe the TPUs and the XPS will gain a few points. Maybe a MD will get a few points of market share. I expect all these things to happen.

Mm-hmm. I mean, I think Qualcomm and others that are eventually, but the market's huge. There's enough, right?

There's enough to lose a Few points. So, so, you know, if you're modeling, and one of my models are all saying is that NVIDIA's gonna be bigger than we thought it was a year ago. It's gonna be bigger than we thought it was six months ago.

And even while all this other stuff is starting to take some market share, there's still growing Tam. It's a problem that, you know, we wish we had in the research business, you know, that there was that much demand, but like God's ears, you know, the bottom line is, is that I just don't worry about it. The bottom line is Nvidia, the $10 trillion company could be, There's no reason it, I think it'll be six next year though.

All right. You heard it here, Daniel. I want to pivot and switch gears a little bit.

I want to talk few terms signal. Okay. Uh, we've been talking about it here on Tech Trunk TV and in our Tech trunk stuff.

And you know, look, you and I both know, right? A good percentage of our market is all about up and to the right. It's just, you know, the, the way we were brainwashed, I guess coming through.

Talk signal to our audience a little bit, what makes it so unique and so special and why they, and, and guys, you don't have to buy a subscription for today. It's available to you right now. You don't need to, you know, all of the other stuff, but Talk.

Yeah, I mean, we, we believe there was a fundamental problem with the way technology's evaluated. You see, uh, models being introduced weekly, monthly with meaningful improvements, scaling laws mostly still intact. But like the improvements you're seeing between GPT-3 and four and five, or Gemini one and two and three, or even just, uh, the, we just talked about Nvidia a whole bunch.

Like twice a year, they're able to roll out a major release, upgrade cycle or release of their stuff. But yet, right now, if you want to evaluate those technologies, you work with an analyst firm, you, you do a year of interviews and meetings and you fill out forms and then you, you get a report that comes out and it's six months outdated, maybe a year by the time it gets published. And largely by people who can't touch the stuff regularly.

So you've got all kinds of just mismatching now in terms of technology and enterprise decision making with technology. So the simple question is, do we eat our own dog food? Do we drink our own champagne?

Are we in the business so we thought we could solve the problem differently? And so the the reality is, can you build a pervasive, autonomous, uh, evaluation platform that really offers true market intelligence, competitive intelligence that can look at a, in an important, say, agent platform, say Neo Cloud, say, uh, CRM platforms, it can look at it up to the minute, can actually really distinguish, uh, it can look at the voice of the customer. It can look at the market models and growth data.

It can look at CIO decision data and, and it buyer decision data. It can look at the analyst perspective and it can really create this thing on the fly to completely revolutionize how buyers are able to be matched with the right vendors and have the right supporting evidence to make a buying decision. And so we built it and we built it in a way where, look, this thing is completely real time.

This thing has the ability to take every piece of information, hence signal that comes. It could be a S one filing from a acquisition or from any sort of market filing that the company makes a, a report that they put out. It could be earnings, it could be voice of customer data.

We did an exclusive partnership with G two. So every, you know, they have millions and millions of reviews. It could be all the analysts, uh, you know, the FU matters.

Yeah. The briefings, everyone at futurum. It could be, uh, conversations had with Techstrong, uh, it could be Tech Field days where we bring experts in and, and, uh, advisors in to talk about different products.

It could be the briefings that we take. It could be events and press releases that come out of these. Why can't we use all this signal in real time to help evaluate the technology?

And by the way, I got feedback today. It was great from, you know, from an enterprise buyer. And they basically said it was the, they looked at our age agentic report.

They said it was the best report they'd ever seen in terms of truly comparing all the age agentic offerings. And then they asked me, they go, how much of that was written by people and how much was ai? And I said, it was a hundred percent ai, a hundred percent ai, big prompt, all ai, big Prompt.

I mean, there's a human in the loop. There's no, there's a ton of work that went into this, but the actual content was written by ai. But using all those inputs, that massive corpus of data, and by the way, this is just for this kind of evaluation, but the beauty of this is this can be for everything we do.

It could be for economic validations, total cost of ownership reports. It could be for insights and, and reports and research. And like, the model is slow.

It's broken, it's not immediate. Like we have to be as fast as the news cycle every week. Like two weeks ago, the TPU wasn't a thing this week, it was a thing.

The analyst has to be able to look at what does all this noise mean? And if you're a buyer, what should you be considering right now, the current way this is done, you won't have anything meaningful from the market. 12, three months, six months, 12 months.

12 months, three months to get a paper. Yeah. 12 months to get an evaluation out.

And so we're like, with Signal, what we're really doing is we're saying, Hey, how do we do this faster? How do we make it more accessible? And what we're doing, by the way, is you, you've only seen the beginning.

I mean, you've heard my story in the background, but I truly believe that this is all we're doing right now is the Netflix version of shipping DVDs. Right? We've created, we've shown the market that we can do this much faster.

It can be high quality and it can be a better experience than going into the store and running a video. But realistically, this can be streamed, this can be real time, this can be inter continuous, interactive. This is continuous.

It's Continuous. This can be comparative, this can be so many more things. And why in the world would we want to just continue to settle and why would we wanna allow this industry to continue to evolve at such a slow pace when the stuff that we're evaluating is evolving at such a fast pace?

Is is lightning speed? One last subject. I'm gonna let you go, please.

No, Please let me go if you like, if you like what Dan, and you know, Daniel knows a thing or two about the market, about, about the CPU market and, uh, the tech markets in general. You recently launched Futurum equities. Yeah.

Uh, I guess it's been six months already, hasn't it? Yeah. For people who, and there are a lot of you out there who follow the markets who are investing, how, what's the best way to follow you on, on future equities?

Yeah, Look, I the real quickly, you know, the, the stocks of the tech companies and the technology itself are inextricably linked. Like there's this belief that like industry versus equities is like two different things. It's not, um, the data that feeds decision makers and hedge funds or, or, or, you know, investment banks to buy stocks, uh, is the same data that CIOs are using to decide whether or not to use technology.

I mean, they, they absorb it differently. They read it differently. Maybe the modeling is a little bit more spreadsheet versus a little bit more practitioner.

But like, we basically realize that we have so much insights, so much knowledge, so much signal that we built future home equities to basically do the same thing we do for industry, but just put a little bit of a different lens on it. Mm-hmm. So we built a great team.

Uh, you know, we've expanded the, the corpus of data. We've taken a lot of the market sizing and modeling that we're doing now to help us sort of assess which companies are, are good investments. Now, again, we're not advisors, we're not making recommendations.

We're providing Signal or Alpha like they like to call it to, to investors. And, you know, we're using that platform and some, you know, platforms like, like Substack and Reddit and, and X where investors sit really heavily and we're kind of taking all the great work we're doing in Signal and in our lab, uh, and in, you know, across our, our, our analysis. And we're creating content that's really designed for, for that audience.

And, and within the next few months we'll actually be launching True Sell side research that'll help, um, you know, audiences make, you know, we'll have some, some our opinions, uh, not advice, opinions on whether things are are a good buy, what things should be based on models, what the price is. And by the way, we're doing this with just like Signal. We're building a completely autonomously, we can create these reports based on all the insights and data, everything we have that's publicly available, very clearly firewall.

Mm-hmm. Put them in different S3 buckets by the way. Okay.

Uh, totally making sure that we're inized, That's our AWS reinvent kind of hook in there. There you Go. Um, and basically we believe that the, the markets, the media and the, uh, enterprises all are symbiotic and we're gonna address all of them.

I love it. Daniel, I know you were busy and we pulled you up here. Nothing to do.

I appreciate you always, man. Always. Alright, thank you.

Good to be with you. Futurum Equities, Futurum Signal. We're gonna be, I think this is gonna wrap up our Day One reinvent coverage.

We've got a full day tomorrow. A lot of good stuff. Our friends at Cuse will be back on with us tomorrow too.

We've got a lot to cover with them as well as AWS themselves. But for today, hey man, this is Alan Hummel, fur Techstrong tv. We're out.

Have a great day.