AI Agents Face Access Barriers and an Infrastructure Test

Hey everybody, and welcome to Monday on the Techstrong gang. And as usual, we got an awesome number of things to talk about with a great number of guests. Before I dive into that, though, I would just point out that if you've been a regular watcher of the show, you've noticed that Alan Shindle's been absent as of late.

And this has been going on for a bit, but trust us, he's just fine and there's no need to put his picture on a milk carton just yet. He'll be back in a few days. All right.

Let me introduce our guest. Steven, good to see you as always, my friend. How are you doing?

It's great to be here. I got to do a little traveling, got to see a little baseball. I'm pretty good.

Excellent. Tom Hollingsworth, are you guys both in the same spot? No.

I think there's rules against us being in the same place at the same time. It's something about the singularity. But don't worry, one of these days we will show up on the same episode together and it'll be lots of fun.

Guy Currier, good to see you again. This is like twice inside of the last five episodes. This is amazing.

Yeah, it's great to be back again. And I think the singularity might extend to me as well, Tom, with respect to you. You and I have never been in the same place.

Well, Yvette and I are both in New York, but Yvette's back. She's been traveling the world as well. Yvette, always a pleasure to see you.

How are things? It's good to be back with the gang. There you go.

Well, let's just jump into this because we're all supposed to be living in this new AI agentic economy, but a funny thing happened along the way. " And then it got a little weirder too, because Amazon said that you couldn't use the Meta agent, and now there's a whole bunch of other websites out there that are saying the same thing. So suddenly, Yvette, what's going on here?

But I thought we were all supposed to be living in this happy world where my AI agent's going to do all my shopping, and there's going to be all these B2B transactions, and did somebody not get the memo? What's happened? Well, the agentic economy has plenty of horsepower, but it's stalling at the door, no pun intended.

If you've seen over the past couple of weeks, months, every chip vendor at OCP, especially last week, built racks for millions of agents. And the companies holding our money and our shopping carts just changed the damn lock. So read the fine print with the banks, folks, because the banks drew a very specific line.

P. Morgan, Citibank, Bank of America, and Schwab are all stopping outside agents from logging in with your credentials and moving your money. So Citi still lets customers grant limited read-only access to some of their banking providers, so agents get a window seat and no steering wheel.

But the reason is security. Security is the headline, and it's super legitimate. Schwab's CEO said that they block 10 to 20 unauthorized agents a week.

And he asked a question, "The whole industry still owes all of us an answer. " Like who? Because if you look at the box, all the risk is on you.

None of the labs have the risk. The banks don't have the risk. P.

Morgan, Bank of America, and Wells Fargo within the reach of agents. That's a lot of freaking damn money, well, in my humble opinion. But as an auditor, here's the thing.

A human in the loop only counts as control when they can actually prove that they did something. What did they approve? " So until someone writes an authorization standard, every bank is going to write their own.

All right. So if that's the case, Tom, is it your impression that any of this is fixable and we just kind of stumbled coming out of the gate, or is this going to be a more complex, permanent kind of problem? It's a fixable problem down the road because all these systems have to do is figure out how to attribute the behavior to an AI agent.

The bigger question is what can you hold an AI agent legally liable for? And that's the biggest problem. The banks don't really care one way or the other.

Money's going to move one way or the other. Yeah, there's some security issues related to this, but ultimately, well, it's like credit card fraud, right? We have a clear standard that says whoever has the least amount of security in the credit card transaction chain is the one who's ultimately liable for any credit card fraud, which ironically enough is how we ended up with chip and PIN finally, or at least chip.

I think the bigger problem is on the flip side of this, where companies like Amazon are blocking agents from doing purchases. And we've even seen this from Amazon's behavior that's changed over the last couple of months, where they no longer send you a summary of what you ordered. " It's because AI agents will go do exactly what they were told.

Right? I want you to go to Amazon and buy this microphone. They can't be swayed by ads.

They can't be swayed by search results. And that's a problem for companies like Amazon and Google because just like all the things that are in the checkout lane at your local big box store, like Walmart or Target, those are designed to get you to make impulse purchases. Bots don't make impulse purchases.

Bots don't look at advertising. " Well, Tom hit the nail on the head because do you all know how much money Amazon made in ads alone? $68 billion last year.

Right? And we all know agents are terrible window shoppers, as Tom indicated. But don't you think that the Thanksgiving dinner between Amazon and Meta is going to be very weird this Thanksgiving?

Right? Because these folks are in business together, apparently. And since Amazon cut off Meta's news on the 20th, and I double-checked and confirmed it's still cut off as of this morning, these two businesses are supposed to be partners.

Meta signed a multi-billion dollar deal to run agentic workloads on Amazon's Graviton chips. So I'm saying, wow, I spent multi-billion dollars on running on your cloud, and you going to block my s**t? Ooh, Thanksgiving's going to be spicy.

So I do think that people need to understand that Perplexity has also had the same challenge, and they were also taken to court. So I do think that this is going to be something that the banks are going to draw a moat around saying... We've all seen it.

If you have Chase, you try to use Zelle, how many warnings do you get? You got to click the box like, yes, I know that person is sending me money. Yes, I'm going to be saying that.

They are trying to make sure that you moving money, you understand who's moving it. And with Perplexity, they've gone to court. So I do think that this is going to fall out.

But in the end of the day, everybody, guess who's in the middle? Us, the consumers. So Steven, let me ask you this because Tom said something that caught my attention.

It's like, so the AI agent just does what it's told. Well, I don't know if I believe that, but secondarily, when I'm shopping, am I going to go say to the AI agent, "Buy me the red stapler from this company," or am I just going to tell it to go buy a stapler, and then whatever it comes back with is fine? I think it's going to be the latter there.

And I think that that's the thing. I take issue with what Tom said about AI agents not being susceptible to suggestive sell. No, the problem is they're super susceptible to suggestive sell.

Those of you who paid attention to AI Field Day last week got to see me do a long presentation about how Meta's Muse is the worst consumer product ever made. I also talked about this on Security Boulevard last week, among other places. And I am here to say the problem, I think, is what Yvette said at the beginning, which is that it's all about liability for these companies.

If agents are able to connect to your bank account, well, they could do some interesting things with your money, and the bank does not want to be left holding the bag when an agent logs in without your knowledge and does stuff. to your account and your money. I think that it's all about that.

And as soon as Anthropic and Meta and Google and the rest of these companies can figure out how to make a deal with the banks, and ultimately with the insurance companies over liability, it's going to be game on. Because everybody in the financial system, I think, should be aware or maybe already is aware, that these agents are incredibly susceptible to suggestion. They're incredibly susceptible to manipulation.

" And the next thing you know, there's one getting delivered at your house, whether you said buy it or not. And let me tell you, the last person who's going to say, "Oh, we shouldn't let this happen," is the credit card companies. They're going to be like, "Hell yeah, bring it on.

More transactions. " It just comes down to liability, and somebody needs to figure it out. And frankly, I'm bullish on this.

I think as soon as we have a clear assignment of liability for AI agents, we are going to have an explosion of activity, but also a glorious explosion of lawsuits and insurance payments and all sorts of things. And maybe that will reign in this craziness. I don't know about that.

I don't know. I have no idea because, Mike, you know what I'm about to say. You got folks today don't even read the terms and conditions of the s**t that they got.

They want to have squirrels flying planes. They're going to download the app, go boop, boop, boop, boop, boop, go to the bottom. So- Squirrels flying planes?

Yes. You think that they're going to, under the guise of, "Hey, if I don't have to spend the five minutes to send my wife flowers, I can have a freaking agent do it," who's connected to their bank. Go send Matilda some peonies, right?

That's what the agent's going to do under saving me the five minutes I don't have to spend time to do that. That says a lot about you and not your agent. But that's the gap that I think that most people don't understand.

At the end of the day, you are going to give an agent control of your bank, your life, all everything, to do s**t for you under the guise of so you don't have to do it to save time. To do what? " Right?

If I didn't walk up to the corner to get the flowers myself, I'm going to tell you, my wife's going to say that doesn't count, but that's a separate issue. Tom, rebut anything Steven said. With all due deference to my colleague, I see where he is coming from, but I think that the problem is that, yes, there is a class of people who just go to Amazon and type something in and buy the first thing.

I am in the other camp, which is I have done 45 months of research on whatever I'm going to buy, and I can write a bot that will go out to Camel Camel Camel and find out if it's actually the lowest price that it has ever been, if this is really a deal and, Yvette, to your point, Black Friday is going to suck for a lot of these companies whenever I can go, "Oh, well, you just jacked the price up three weeks ago, so you could discount it by 45% to make it look like a sale. But in fact, you guys historically put this on sale in March, and that's when I'm going to buy it. " It's almost like I'm treating physical goods like stocks.

" And that's when the lawsuits are going to start, because as it turns out, letting computers do all the stock trading is bad. Oh, crap, we already have that. Never mind.

Never mind. Guy, you want to be the voice of reason here? How do you see this playing out?

Sure. I'll be the voice of reason because I'm really depressed now. And I got to say, this is a fine kettle of fish that we are caught in here.

And Yvette, I think you framed it up at the beginning, and we haven't gotten back to it, when you're talking about the one-point whatever trillion dollars. And then move on to the deal between Meta and Amazon that became a no deal and everything because, folks, oh my God. I feel like the subtext to a lot of this discussion is we're waiting for these tech titans and big companies and everything to work.

Listen, we know where the liability is. The liability is in the designer of the agent or the creator of the model or what have you. But once there's six different companies involved, it gets really complicated, and us poor users, however expert, Tom, I do think that there's going to be a maturity in usage just across the board as people get used to it.

But geez, if they had a podcast, a Techstrong gang back in 1892, I think we would all be on it talking about the railroads and the oil barons and the this and that in practically the same terms. I feel like- You know, the- ... it's just another case of these grand, huge billionaire, big companies, blah, blah, blah, just control...

Not controlling, because it's out of their control, but just their moves and decisions deciding our fate. So there. That's my- Alan's not here, so I have to pitch the book.

You mentioned railroads, so you know. Exactly. "In This Victory Trap" by Alan Jim.

" I'm going to say what y'all are not saying, but I'm just going to say it because that's what they bring me on this show for. It's like this: We got these six to seven companies that have created a vacuum on human connectivity. They're sopping it up.

I remember the days when you used to go to the bar to meet people and date. All that is online. And if you look at these same people who have vacuumed up all these interactions that used to be physical and brought them down to the digital realm from human connectivity and communication, they are not the best people that have human communication skills.

We are turning over all of this stuff to people who can't manage their way out of a paper bag or not lose their stuff on camera. Really, literally. They own the whole thing.

So if you want to turn over to people who morally have challenges, are probably emotionally bankrupt, to determine our safety, our risk, and where your money's going, we are in a world of hurt. I'm just going to say it. We-- Pack it up.

There's going to be the four horsemen coming across the sky. That's what I'm saying. I said it, and I'm going to keep to it.

All right. Thank you for saying this. I'm going to end this conversation here because we can go on forever.

But there is one part of this thing I don't get, going to Tom's point. So my AI agent will go buy something, and a seller will have an AI agent that wants to sell something. Somewhere, these AI agents are going to negotiate with each other, but they have different competing missions, and eventually they'll cancel each other out.

Am I understanding this correctly? Oh, Guy, go ahead. You just got there me.

Listen, this is what happened with Priceline. Everybody, it was Priceline, right? Yep, Priceline.

The one we did for mayonnaise or whatever- Yep ... was that it turned out that what people were paying was more, not less, because of how they made that market. And what you're talking about there, Mike, is market-making by agents.

Yep. And go back to- Curse you capitalists ... but no, but Tom actually put it right.

Because he instructed his agent what to do. Do you really think 90% of the people using freaking agents know how to prompt right? They can't even balance a freaking checkbook.

No, 95. And now they want to give an agent instructions on how to buy s**t, do s**t, and come on, man. Y'all, come on.

Not everybody's a Tom. You actually balance your checkbook? I'm impressed.

All right. Here we go. We do need more Toms.

I am going to leave this here, though, and shift the gear because we had a great event last week with the folks at Tech Field Day talking about AI. Steven, I know you were the host for this stuff. There was a lot of vendors represented, a lot of delegates.

Give us the highlights. What's going on here, my friend? Sure.

Yeah. This was a fun one because obviously, I don't know if you've heard about this AI thing, but it seems to be making a lot of difference in the world. It's kind of a big deal.

So yeah. So this was AI Field Day. This was notable, I think, in the just absolutely incredible depth that we got on the real world of AI I think that was the most incredible thing.

So from the delegates who are literally out there doing the work and seeing the market change, to the companies that presented to the other-- One of the big highlights of this for me was not just the fact that we had CoreWeave and DDN, who are perhaps-- They make up probably 20, 30% of the enterprise AI industry at this point between the two of them. It's pretty incredible to have them come in and present, and they both absolutely rocked it, I got to say. But to have this group of delegates coming in and to have invited guests coming in, really bringing the goods.

So we had, for example, ETR coming in and showing the latest data on AI adoption in the enterprise and sort of the subjective of AI right now, which was just something you just don't get because they have this network of literally hundreds of people that are making these decisions in the enterprise space come, and they brought that to the table. We also had Signal65 come in and present Pinnacle, which is probably the coolest objective data about models these days. We had a marvelous moment.

So Google released a new Gemini model literally the evening before the presentation, and we get on there the next morning and everybody's like, "Hey, I wonder what that Google model's good for. " They throw the slides up and they've already tested it. They've already ranked it on the Pinnacle evaluations and shown just how good that model is.

And that was one of those really like, whoa, it hasn't even been 24 hours and you've already done the entire full evaluation suite against it. So that was amazing. We had the delegates come and present their own stuff, their own work.

Everything from John Willis, who is friends to the gang here, talking about quantum computing, which is a must-watch, to Dave Graham talking about building a digital twin of Leonardo da Vinci and feeding him information over time to see what he would do if he lived for five, six centuries. Just really wild stuff. And I've griped and on about news.

So I can't even summarize it simply because it was just this incredibly rich under the ground vein of goodness. I don't know. Tom, in the enterprise, I keep hearing the following.

Everybody's okay with training in the cloud for the most part, unless the data's super sensitive. But when it comes to inference, they want to put that closer to the point where their data is, and all that stuff is either in a local data center or maybe even at the edge. But nobody has the infrastructure to run that stuff successfully and the amount of data that's involved with all that.

So we seem to be kind of stuck somewhere where the cow is half over the fence. It's like we all kind of like the idea, but we're not quite sure how to operationalize it. Terrible.

Well, the problem that we've run into before, because this is not a new concept. It even predates cloud when we were starting to talk about big data. People don't want to have this stuff just sitting around doing nothing, right?

Think of "The Incredibles," right? " That's what they want. They want it to see-- They want to see it in action, right?

And that's why everybody and their kid runs out and buys the latest Mac Mini, Mac Studio whenever it comes out with 32 gigs of RAM or whatever spark that once they sell a car and a kidney to afford it, because they want to do things on their desktop with AI. They want these agents to vibe code the GTA 6 demo or whatever, but they don't care about it when it's working, right? Because when it's working, it's boring.

It's like watching a hard disk defrag or watching a compiler operate. And the reason that companies want it on-site is because then that gives them the control over people, how they use it, right? Think about all of that token maxing crap that we dealt with back in, what was it?

April. Yes, just six months ago, all we could talk about was how much people wanted to burn all of these tokens to do all these amazing things, and then when it turns out that you actually have to pay for that because you're renting those tokens from somebody else, now it's a problem. So the trap is they think that by moving inferencing on-site, that they're going to make that all cheaper somehow.

I don't know. Yvette, what's your take? What are you hearing about folks in the enterprise about where they really are when it comes to AI?

Because I'm on the journalism side of this thing, so everybody comes to me with the horror stories, but is there a balance or what's your take? No, I don't think there's a balance because the labs are hooked up with the enterprise consultants who are selling the lab stuff, which are selling the fear of missing out of not using the new thing. So it's like Whac-A-Mole.

I'm seeing Whac-A-Mole. " Give it a job. Not every job, like a job, b job.

They cannot articulate what that is, and that is why most enterprises are like they're on the struggle buses because they're throwing it at everything and they don't know how to train it, control it, secure it, put data around it, but they want to make sure that they're using it in some type of way. That's my read. So all this is like it's all great.

Until they answer the question of what are we trying to do? What problem are we trying to solve? With the right partner to tell them the stuff that they need to hear, not they want to hear, it's just going to be Whac-A-Mole across the enterprise like it is.

And that's where we're going to get these reports coming out of Gartner and MIT saying that the ROI, they're still looking for it. Yeah. And I said this before, the only way if you want to catch a squirrel, use bright shining lights.

Do that. And that's what most folks are doing when it comes to AI. Just like, "Oh, let's do this.

" You can't manage your enterprise, and you won't be able to get return on investment on that. Does that include squirrels flying airplanes? Squirrels flying airplanes and raccoons, yeah.

I literally have-- So my digital twin is a badass raccoon. Yes. Mm-hmm.

With lots of attitude. That's my digital twin, yeah. I don't know.

Stephen, what's your take? Because what Yvette just described was a whole lot of AI theater. Show me the money.

Yep. Mm-hmm. To provide additional dollars to sell for a dime.

The problem is, where are we going to go with this? And there's a lot of concern. I will just say, like I said, I don't want to specifically attribute this to anybody, but we were in San Jose during OCP.

We were downtown, we were talking to people at these companies. There's a realization among the companies, the AI companies themselves, along with the consultants and the analysts and the users, that we have to figure out a way out of this weird AI industrial complex that's losing money at record pace. And yet there is a bright spot, and the bright spot is we are seeing some useful productive applications for the technology.

We are starting to see companies turning the corner. I think one of the big takeaways, for example, from the CoreWeave session is this was not the let's build a lot of data centers and finance a lot of GPUs CoreWeave. This was the other part of the company that they've started investing in, which is let's hire a bunch of enterprise IT folks from Silicon Valley who actually know how this stuff works, and have them build us a platform that is useful and allows companies to get value from AI.

So for example, their platform allows you to take a model and do things like fine-tuning and reinforcement learning in a very approachable visual way to come out with a model that's a lot more practical and a lot more usable to you. So essentially As an enterprise, if you want a model that you can run locally profitably, you need to do some fine-tuning on that model. Instead of building your own training cluster, which is ludicrously expensive and basically impossible now, you can use CoreWeave's Forge to do this in a very friendly graphical way that manages the whole process for you, and then take that model and run it locally.

That's a thing you could do, and it's actually a really useful thing to do with CoreWeave's incredibly high-dollar data center full of AI hardware. And so, we were hearing people speaking appreciatively of this kind of idea of using the NeoClouds for the job that they're good for, which is basically these huge training runs and so on, and then figuring out where to run it afterward, whether it's a NeoCloud or a conventional hyperscaler or on-prem or a service provider. And I think that shows the way where the industry can start making profit on AI instead of just pouring money on the fire.

But that being said, of course, the other big thing that everybody's talking about is just this incredible lack of hardware. Everybody was just rending their garments over the fact that you just can't get hardware and how expensive it is. These Spark systems have doubled in price since NVIDIA released them.

People are all like, "Oh my God, I wish I had bought the Big Mac Mini when it was available. " Just stuff like that is a common refrain in Silicon Valley now. All right.

We're going to have to leave this here, but I would just point out one thing. " And then the first wave of people play with it, and then they quickly figure out that they need a small army of consultants to implement it. In the age of AI, we're calling those forward-deployed engineers, but it's the exact same people.

And that just tells you something about, well, just whether or not this thing is ready for prime time or not, because the more consultants required to implement it, the more you're probably locked into those consultants, and around and round we go. All right. Hardware.

Stephen brought this up. Guy went to the OCP conference and had a walkabout, and basically he's going to give us his impressions. But I think during the conference, things that we covered on Techstrong IT, I think AMD had some new racks that they were touting and- I got all that for you, yeah.

All right. NVIDIA was talking about stuff, and of course Intel shared a white paper, so there you go. A very significant one.

So yeah, thanks, Mike. So OCP is the Open Compute Project or Open Compute Project Foundation. They had their global summit last year.

It was the biggest one yet. They sold out about two weeks before, which was a first for them. Roughly 15,000 was their capacity.

So, that's a pretty big conference. Now, it used to be that OCP was mostly for the hyperscalers. It was founded by Meta, after all, back when it was called Facebook.

And for all the good standards work that they've done over the years, I think it's fair to say they're mostly focused not just on data centers, but on the hyperscale core, huge build-outs. For the rest of us, I've enjoyed attending for years because it's really cool whiz-bang stuff for all of us to look at, but I did feel a little bit like I was pressing my nose against the glass, like not everybody could implement it. And so I think one of the nice trends that we're seeing is how that is starting to break out into...

So for example, the headline news is, I think, at least one that was really evident center stage is what you might call the revenge of the CPU. So Lisa Su of AMD, she said on stage it was her first ever OCP, which is interesting. And one big introduction at the show that she featured was the introduction of Helion.

Now you might have heard of Helios, which is their multi-rack scale infrastructure mostly for AI training or large-scale inference. Helion is a rack system. One version is CPU only.

Now there's 33,000 cores on that rack, so that's more than your average number of cores. Those are AMD EPYC data center CPUs. Another version has up to 16 of the AMD Instinct GPU accelerators, along with also still a ton of AMD EPYC.

So that was AMD rack scale. That's starting to seem a little bit more, especially in the age of AI, like human scale as opposed to hyperscale. NVIDIA, that you mentioned, Mike, they have an OCP compliant rack and compute tray.

So now we're at the rack scale again. So they're part of the CPU party, NVIDIA, because they're splitting their AI factory in two with Vera Rubin, which is CPU and GPU, pretty classic for training and inference, but also a CPU-only Vera rack for agentic. Intel, yeah, they've been on the CPUs for AI argument years ago, I guess for obvious reasons, which I won't name here.

I think the headline for them was not so much products or announcements, just with the message that CPUs are the coordinators of agentic work, which is more and more important, and this nice slogan of put every prompt in its place, which is a decent summary of the white paper. Although you have to admit, that's pretty good. Put every prompt in its place.

They're also still your standard hyperscaler type whiz-bang stuff, Tom, for you. NVIDIA and Cisco announced a 400G per lane optical module. That pops my eyes anyway.

Maybe you have one in your backyard at the moment, I don't know. Don't @ me. There was a cross-vendor hardware security model from Microsoft, Marvell, and Oracle.

All this kind of cross-vendor stuff is typical in a standards environment. My own favorite news, I have to say, because I was in this so-called AI Compute Continuum project from late last year myself, is the extension of a data center standard called DC-MHS to apply to AI at the edge with an actual product, a server from HPE and Intel. So, this idea of this breakout from hyperscale of standards to address in a unified way a lot of these problems, not just in infrastructure generally, but in AI, it ties right back to what you guys were just talking about a minute ago, which is getting AI and AI-based infrastructure all over the place.

And I think that can create real complexities if there aren't standards to help unify them. And the fact of a standard that's viable both in the hyperscale side and all the way out at the edge, that at least to me is pretty encouraging. Of course, Mike, you're one of the most skeptical of the optimists that I know.

So my naive question to you is, if it's a market leader like an NVIDIA or somebody contributing a rack design that's an open spec, is it just making its own design the market? I think open beats closed every time, but do we know whether a spec like that is open in practice until somebody other than the authors- Yeah. What- ...

responds to it? What do we say about standards? It's great that we have so many of them.

So what happens is, is that there's a standard for AMD that it shows to Cisco and all its partners, and they go implement it, and then there's a variation of a standard from Intel and another one from NVIDIA, and then you wind up with all these different flavors of different racks. But Stephen, I've seen this movie before. " And is that where we are here, and is that what this is all about at the end of the day?

Well, it is interesting that one thing that Guy was talking about there absolutely rhymed with what we were seeing at AI Field Day. And as I mentioned, among the delegates were some of the top consultants out there helping companies build AI infrastructure, and they clearly said that they're seeing exactly the same ... transformation happening in enterprise.

Essentially, yes, enterprises have bought huge training clusters or massive amounts of training hardware, whether they're using it or not, I don't know. But now they are very much interested in adding CPU to the mix because they're seeing that inferencing is a very different application. And my metaphor was essentially that the training cluster is R&D, but the inferencing cluster, now that is the factory.

That's where the product is made, and that's where the money comes out. And so companies are very much interested in trying to figure out how they can build efficient inferencing platforms to the point that we were hearing, for example, that the smaller AMD and NVIDIA GPUs are very much in demand because essentially you get one with enough RAM to run your model, and you don't have to have this massive cluster of hardware. The form factor of these systems is very, very different.

They're talking about pizza boxes once again. They're talking about maybe 4U servers with a few independent GPUs in them, each one running workloads, with Kubernetes, that sort of thing. It looks very much like a cloud infrastructure and not very much like this massive training supercomputer that we've been hearing about.

So I do agree with what Guy is saying. The pendulum is very much swinging back towards CPU, toward GPU, toward NPU, and away from this monolithic thought of we have to build the biggest GPU clusters we could possibly have, and who cares about the CPU? This is mainframes versus distributed computing all over again, right?

Yeah. Everything is the same all over again, Mike. I mean, really.

No, I'm not even joking. You guys were talking earlier about the productivity trap in AI. This is not different from the PC in the '80s and '90s, Alan Greenspan famously saying that the productivity gains have not been seen, right?

We might not see productivity gains from AI for a really long time. Just saying. I'm literally feeling like I'm back in Amazon days, where you're talking about two pizza teams, you know what I mean?

" That's all I heard. That's all I heard. And when every vendor is agreeing, y'all need to check with who's selling what.

Every number that Guy kind of rattled off is a spec sheet or a design target. AMD. Who's published real-world deployments showing all this s**t, right?

That's my question. Who's going to do that? And I think the smartest, even though it's typical, they came out with a white paper, is Intel.

" Just kids. That's what this needs. Now we need someone without a quota to measure it because you got these same people building the s**t, buying the s**t, selling the s**t, and then telling you how it's going to work.

Because you know what? The auditor can't be the vendor, and neither can it be the benchmark. There you go.

Hey, Tom. As Guy was talking, I couldn't help but think of every year, at least I try to go to the New York Auto Show. It's a great experience.

I get to see all these cars that are really awesome looking, and I can't afford them. And then when I find out later there's only six of them available. Is that kind of what Guy just described for hardware?

" It's funny because as soon as Guy started talking about it, that's exactly what I thought it was, the auto show where you've got these concept cars. And I remember, what was it, the Pontiac, was it the Transporter, which was basically like a mini RV minivan that no one would ever buy because it's not practical. It's almost like sticking jet engines on the bottom of a sports car and running it up a Hot Wheels loop.

Oh, wait, we did that, too. Nobody would do that. Come on.

Nobody would do that. But one thing I do want to talk about that Guy mentioned, and I think it's kind of funny, is this idea that NVIDIA and Cisco got together to create a new spec for optics, USPF, which is not OSPF, which is not OPSF, which is a bunch of different things. The deal with USPF is that it is not water-cooled, which directly puts it in competition with Arista XPO, which is a water-cooled optic.

I think it's funny that Arista or that NVIDIA is backing both. And people who watch the gang a lot may remember a couple weeks ago, I made a reference to "The Good, the Bad and the Ugly" with all the AI venture model folks all having the standoff with each other. This reminds me of "A Fistful of Dollars," the other great Clint Eastwood Western, which one of the reasons why I love it so much is NVIDIA is basically playing Clint Eastwood now.

They have an agenda to get everybody else fighting each other while they have their own thing that they're trying to do. Because NVIDIA is still trying to embrace co-packaged optics, which is going to be a completely different line of thought from these pluggable modules. And I think one of the biggest things that people need to understand is NVIDIA does hold all the cards here because they are the largest manufacturer of these types of devices.

Yeah, I get it. Cisco is the 800-pound gorilla in networking. They're followed closely by HPE and Arista and so many other companies.

But if NVIDIA comes out and says, "This is the new standard, kids," do you think that people aren't going to fall all over each other to back that standard because they know that that's $11 billion trillion kajillion worth of stock that gets pushed out the door every quarter to people who are buying this stuff sight unseen as fast as they can, not because they need it, but because they got to keep it out of their competitors' hands. That's how you make standards, folks, whether you like it or not. All right, Stephen, last word.

We definitely also need to keep an eye on the old, speaking of old battles, the old ARM versus x86 battle, because as CPU becomes more and more important, we are seeing very clearly the CPU companies once again lining up their wares. We've got Intel talking Xeon. We've got AMD talking EPYC.

We've got Broadcom and Qualcomm and NVIDIA talking ARM. We've got ARM talking ARM. I mean- Yeah, ARM was very much in effect at the show last week.

Yeah. Exactly. That was a huge takeaway from the show, and certainly something that we're going to have to keep an eye on.

Which of these CPU architectures is ultimately going to prevail? All right. " And then we will go from there.

But I want to thank our guests for sharing their knowledge and insights, as always. tv lineup. And until then, we'll see you all tomorrow.