The AI Factory - Data Center of the Future | Sid Nag, Tekonyx & Corey Quinn, Duckbill

In this interview from theCUBE + NYSE Wired: AI Factories - Data Centers of the Future, Sid Nag, chief research officer of Tekonyx, joins Corey Quinn, chief cloud economist of Duckbill, to talk with theCUBE's Dave Vellante about the gap between AI factory hype and the economic and operational realities enterprises must navigate to build scalable intelligence. Nag reframes the AI factory concept not as a $500 billion data center build but as a logical entity combining curated data, multimodal LLMs, orchestration engines and compute fabric into what he calls "scalable intelligence" — actionable outcomes a CIO can deliver across business functions. He also introduces the "hybrid processing unit," arguing that AI workloads require a mix of CPUs, GPUs, NPUs and TPUs rather than any single dominant chip, reflecting a fundamentally more distributed and heterogeneous compute reality. The conversation also explores the precarious position of neo clouds, with Quinn warning that pure GPU rental businesses face commoditization the moment hyperscaler capacity constraints ease — unless they move up the stack with platform services, much as AWS did layering value atop EC2. Nag draws a parallel to exchange businesses that commoditized rapidly and urges neo clouds to elevate their value proposition beyond compute. The panel then weighs AWS's ambitions with Amazon Q, which both guests find credible as an emerging enterprise work surface — but critically undermined by internal fragmentation, with Agent Core, Connect and other services still operating as disconnected products. Looking ahead, Quinn emphasizes that companies are finally demanding proof of where AI actually generates value rather than spending without accountability, while Nag flags a widening chasm between vendor silicon investment and enterprise readiness to pay. From the commoditization of inference and the survival pressures facing neo clouds to the internal fragmentation haunting even the most ambitious hyperscaler platforms, the discussion provides a practical lens for separating durable AI infrastructure value from supercycle noise.

Sid Nag, Tekonyx & Corey Quinn, Duckbill

In this interview from theCUBE + NYSE Wired: AI Factories - Data Centers of the Future, Sid Nag, chief research officer of Tekonyx, joins Corey Quinn, chief cloud economist of Duckbill, to talk with theCUBE's Dave Vellante about the gap between AI factory hype and the economic and operational realities enterprises must navigate to build scalable intelligence. Nag reframes the AI factory concept not as a $500 billion data center build but as a logical entity combining curated data, multimodal LLMs, orchestration engines and compute fabric into what he calls "scalable intelligence" — actionable outcomes a CIO can deliver across business functions. He also introduces the "hybrid processing unit," arguing that AI workloads require a mix of CPUs, GPUs, NPUs and TPUs rather than any single dominant chip, reflecting a fundamentally more distributed and heterogeneous compute reality. The conversation also explores the precarious position of neo clouds, with Quinn warning that pure GPU rental businesses face commoditization the moment hyperscaler capacity constraints ease — unless they move up the stack with platform services, much as AWS did layering value atop EC2. Nag draws a parallel to exchange businesses that commoditized rapidly and urges neo clouds to elevate their value proposition beyond compute. The panel then weighs AWS's ambitions with Amazon Q, which both guests find credible as an emerging enterprise work surface — but critically undermined by internal fragmentation, with Agent Core, Connect and other services still operating as disconnected products. Looking ahead, Quinn emphasizes that companies are finally demanding proof of where AI actually generates value rather than spending without accountability, while Nag flags a widening chasm between vendor silicon investment and enterprise readiness to pay. From the commoditization of inference and the survival pressures facing neo clouds to the internal fragmentation haunting even the most ambitious hyperscaler platforms, the discussion provides a practical lens for separating durable AI infrastructure value from supercycle noise.

Share this session

Clips
More from the series

Sid Nag

Founder, CEO & Chief Research Officer

Tekonyx

Corey Quinn

Chief Cloud Economist

Duckbill

play_circle_outline Network-as-Computer: Orchestrating Data, Models and Hybrid CPUs/GPUs/NPUs/TPUs in the AI Factory

play_circle_outline On‑prem vs cloud inference: tokens, frontier models, economics, capacity constraints

play_circle_outline Neoclouds versus hyperscalers: CoreWeave and commoditization, reliability, value-added services

play_circle_outline Agentic AI Disrupts ISVs: Probabilistic Agents, Rapid Adoption, and the Demise of Traditional App Models

play_circle_outline Enterprise economics and FinOps: tokenomics, cost discipline, ROI and CFO scrutiny

play_circle_outline Edge and small language models: on-device/custom SLMs for enterprise-specific workloads

play_circle_outline Market and governance: model vendor IPOs, vendor-enterprise chasm, government/public sector role

Episode Info
Transcript

Sid Nag, Tekonyx & Corey Quinn, Duckbill

June 18, 2026

Sid Nag

Founder, CEO & Chief Research Officer Tekonyx

Corey Quinn

Chief Cloud Economist Duckbill

This discussion explores artificial intelligence factories, hybrid compute and the economics of inference. Sid Nag of Tekonyx and Corey Quinn of The Duckbill Group join theCUBE Research hosts John Furrier of theCUBE Research and Dave Vellante of theCUBE Research for a recorded conversation at the AWS NYC Summit and NYSE Wired.

Nag explains hybrid processing units and network-first architectures and they argue the network becomes central as computing moves to hybrid processing. Quinn emphasizes inference portability, token costs and neocloud economics a... Read more

explore Keep Exploring

How is computing changing with AI, and what do concepts like the network-as-computer, hybrid processing units, an “AI factory,” and “scalable intelligence” mean for enterprise IT? add

How do you expect customers will choose between running large language models on-premises versus using cloud APIs for token generation, and what factors will influence that choice? add

Can specialized "neocloud" providers (e.g., CoreWeave) deliver better AI infrastructure than hyperscale clouds, and what must they do to remain viable long-term once supply and demand re-balance? add

What will happen to general-purpose computing and the large installed base of x86 infrastructure and software—will it continue to be innovated or be replaced, and how do you see its future? add

How will the move from on‑premises to SaaS — and now to AI/agentic, token‑based models — change enterprise technology, operating, and pricing models, IT spending (CapEx/OpEx), and the prospects for incumbent enterprise software vendors like Salesforce, Workday, and SAP? add

How will inference-optimized chips and local/small language models change the economics and deployment of enterprise, on‑premises, and edge AI—particularly for company-specific, domain-tailored models? add

What surprised you about recent developments in the AI industry (pace of commercialization, vendor strategies, and adoption of agentic systems)? add

bolt Powered by CUBE AI

Sid Nag, Tekonyx & Corey Quinn, Duckbill June 18, 2026

search

John Furrier

>> Palo Alto studio connecting Silicon Valley and Wall Street. I'm John Furrier, the host of theCUBE, here with Dave Vellante, my co-host.

Dave Vellante

>> Everybody welcome to theCUBE's NYSE Wired studio here at the Buttonwood Podium high above the New York Stock Exchange Options Exchange. My name is Dave Vellante and this is our AI factory series. We were this week at the AWS NYC Summit. I had the pleasure of cajoling a couple of top analysts here to join me to talk about AI factories, to talk about what we learned in the hyperscale market generally and AWS specifically this week. Sid Nag is back. He's the chief research officer at Tekonyx. Sid, good to see you. It's quite the title. You got promoted.

Sid Nag

>> Glad to be back. Exactly. Life's been good.

Dave Vellante

>> Of course, theCUBE alum, Corey Quinn is here as the chief cloud economist at Duckbill. It used to be the Duckbill Group. You changed name now to just Duckbill.

Corey Quinn

>> We did. We raised a round. We're building software now. And when you say group, it sounds consulty and folksy and not at all fundable. So here we are.

Dave Vellante

>> No kidding. You raised a round.

Corey Quinn

>> Yeah.

Dave Vellante

>> Oh, congratulations.

Corey Quinn

>> Thank you.

Dave Vellante

>> I missed that news. That's fantastic. But hopefully you're still in control and-

Corey Quinn

>> Oh, yes. Oh, yes.

Sid Nag

>> Not fundable, yet fungible.

Dave Vellante

>> Yeah. Well, obviously it's fungible.

Corey Quinn

>> We'll say that. Exactly.

Dave Vellante

>> All right. So AI factories. What is it? Is it real? Is it a buzzword? Is it just marketing or is there substance underneath?

Corey Quinn

>> Depends entirely on who says it and what they're trying to sell today. At some point the idea of AI building AI and software that just automatically builds more software to do what all software currently does, which is sell more AI.

Dave Vellante

>> Okay. Well, so-

Corey Quinn

>> It's an ouroboros, but a very expensive-

Dave Vellante

>> Ouroboros. All right. But Sid, I mean, clearly compute is evolving. The network obviously evolved. And last year, last March at GTC, Nvidia went deeper down at the stack with STX, which is sort of they get storage into the game now. Of course, building agents on top. Is this a new architecture and what does that mean for the future of computing?

Sid Nag

>> Well, I think network first of all is becoming sexy again, so that's the new phrase. So network is the computer in many ways.

Dave Vellante

>> Scott McNealy said-

Sid Nag

>> The whole Scott McNealy thing, right?

Dave Vellante

>> He said, "I should have called it cloud."

Sid Nag

>> Yeah, "I should have called it cloud."

Dave Vellante

>> Put the dot in dotcom.

Sid Nag

>> But honestly, computers evolved so much with the world of AI, like you just talking about before the show started. There's so many different processing units.

Dave Vellante

>> What did you call it? You said hybrid processing units?

Sid Nag

>> It's a logical thing. It's a hybrid processing unit. It's not a physical thing you touch like a GPU or a CPU or a LPU or NPU or a TPU. It's a logical thing. It's a hybrid processing environment because as we heard at the Amazon conference this week, but also something that I've been talking about for a while, is that an AI workload has many different facets and each facet is serviced best by the right processing unit. It could be CPU for certain things, could be a GPU for other things, and so on and so forth. I mean, your iPhone has an NPU built in there, right? It has all sorts of agentic functionality that's being built in, AI functionality with Siri and things of that nature. So we're going to be experiencing a very interesting time where computers becoming very hybrid in that nature. And so the AI factory to me is a logical entity that helps end users. I don't care what all the vendors hype about building $500 billion data centers. The end users have to find a way to use AI in a pedestrian fashion and that is essentially how do I take curated data? How do I take all these different multimodal and multi-model LLMs? How do I take all the orchestration engines? How do I take all the compute fabric and all the other software tools that spits out what you call systems of intelligence, right? Is that term you use?

Dave Vellante

>> Yeah. It's not my term, it's Jeffrey Moore's.

Sid Nag

>> Okay. So I call it scalable intelligence, right? And that is something a CIO can take and action that individual constituents, be it business, IT, supply chain, procurement, R&D, marketing, research, whatever. To me, that's a successful-

Dave Vellante

>> So I want to go up the stack in this conversation, but Corey, I'll ask you. So you built a great company and a career on helping customers-

Corey Quinn

>> Understand their cloud builds.

Dave Vellante

>> Understand.

Corey Quinn

>> Yeah. And now we're doing a lot of it with inference too. Same story. It's the most portable cloud workload we've ever seen is inference. It is effectively the same thing from multiple providers. So it comes down to performance, latency, and better commercials.

Dave Vellante

>> And token costs, right?

Corey Quinn

>> That's where the better commercials come into play.

Dave Vellante

>> Okay. So I want to ask you, so there's this debate going on in the industry is, oh, you have to do this on prem because you can generate your own tokens, you don't want to buy them through an API, but of course you can be a token generator in the cloud.

Corey Quinn

>> You can.

Dave Vellante

>> How do you see that all sort of playing out in the customer base today?

Corey Quinn

>> I think that it's becoming rapidly commoditized. The question then becomes one of economics. Do you have the wherewithal real estate capacity these days, ability to get Nvidia GPUs in the door, power these things, run them responsibly? Then it's a pure mathematics question, and it also depends, of course, what model you're going to run locally. Frontier models are not open weight. So you're going to have to make your peace with whatever it is we're doing inference on. It's going to be this specific kind of model. Whereas if you're chasing the frontier, they drop a new thing, it's available through Bedrock within hours and okay, I guess we're going to go use it there now. But that'll slow.

Dave Vellante

>> I want to stay on this idea. Okay. So take the neoclouds. Somebody put out a tweet the other day that you shared with me that they were basically, they likened it to some kind of dot-com...

Corey Quinn

>> Bubblish.

Dave Vellante

>> The market is definitely bubblicious. But the premise of the neoclouds, like a CoreWeave is we can do AI better than the hyperscalers because they have to, quote, unquote, bolt on AI onto their general purpose clouds. What do you say to that?

Corey Quinn

>> I think you can build a data center extremely quickly if you think about it being only single purpose. Well, I don't need to worry about redundant power. It's not going to store anything long-term. So the data stuff we can cheap out on. We can skip the fire insurance payments and skimp on the cooling. But there's going to be reliability issues, there's durability issues. And in the fullness of time, these things are not going to be constrained forever. Companies would not be dealing with the neoclouds if their existing provider, whoever that may be, had sufficient capacity for them. So the question is what happens? Do the neoclouds grow up into something that is more robust, more sustainable in different ways before the capacity constraints ease or do they not and suddenly everyone just starts to bring all of their inference back into whatever provider they're on, maybe on prem, and they die a quiet death of not being needed anymore?

Dave Vellante

>> And the likelihood is probably a combination of that, Sid. I mean, to Corey's point, what happens when supply and demand come in balance? What does a neocloud or an AI cloud have to do to survive?

Sid Nag

>> Well, I think what Corey said is spot on, but there's another element to neocloud that one is missing, I think, right? The whole value proposition of neoclouds is 90% compute capacity utilization at a commoditized rate. And if the hyperscaler can't meet my demand, then I go to the neoclouds. But I'm reminded of the days where there were companies that were doing exchanges, connecting different environments through a central exchange like an Equinix or whoever, and that business got commoditized really fast. So I still remember our advice to them was to uplevel the game, provide value-added services on top of that commodity business. So my question with the advent of neoclouds is what is CoreWeave and their peers, what are they going to do to elevate their value proposition? Because if they're just selling rental compute capacity or GPUs, that's not a sustainable business.

Corey Quinn

>> And then commoditize becomes a price war and good luck outbidding Amazon.

Sid Nag

>> That's right. Exactly. Yeah.

Dave Vellante

>> But people used to say that about Amazon, that the infrastructure was going to get commoditized. This is a company that has what, 35, 38% operating margin.

Sid Nag

>> I think Amazon's a great example. They took the whole EC2, S3, all their basic IaaS business and up-leveled it to platform as a service capabilities and other capabilities on top of that. And now with the advent of AI, they're driving the pull through of IaaS through all these value-added services. So that's what the neoclouds have to do.

Dave Vellante

>> How much of Amazon's revenue do you think is EC2? Less than half. Let's-

Corey Quinn

>> AWS's revenue?

Dave Vellante

>> Yeah, AWS, AWS.

Corey Quinn

>> Last time I ran the numbers was right about half. So then there's a question of, all right, well, other things run on top of EC2. Is that qualified as EC2 revenue? Probably not. They don't break out numbers for this and they call everything these days an AI service. So who knows?

Dave Vellante

>> But my point is it's a contributor, I believe anyway, it's a contributor to profit.

Corey Quinn

>> It is. Yeah. Absolutely.

Dave Vellante

>> I don't think they're giving away EC2 in order to make margins up the stack.

Corey Quinn

>> No.

Sid Nag

>> I don't think that's what we're saying. I'm saying that Amazon's not selling EC2 as a standalone product anymore. There's not a pull through coming of EC2 through other value-added services they built.

Dave Vellante

>> Like Snowflake.

Sid Nag

>> Whatever.

Corey Quinn

>> So their Kubernetes offering, it brings them along for the ride.

Sid Nag

>> Exactly. All the microservices, all that. So what is neocloud's answer to that? I don't see an answer yet.

Dave Vellante

>> Well, presumably they have to build out beyond compute, storage, networking, other services, and help ISVs build software on top of it.

Corey Quinn

>> Remember the early days of the cloud and people were highly skeptical of it because we have our own data centers and even though they tend to lose power three times a week and the security guard leaves the door open when he has a little too much to drink in the mornings. Great. That's okay. You own the base, rent the peak for the peaks and valleys from cloud. And over time it became, why are you running your own data centers again? And companies still do, but someone launching today likely is not going to go down that path the way that it once was the only option. These days, born in the cloud is a very common story of successful companies.

Dave Vellante

>> Well, the Dell Tech World recently talked about it has many thousands, I want to say the numbers was north of 3,000, what they call AI factory customers in the enterprise. Now most of that or much of that is going to guys like CoreWeave. Most of the stuff that's going in the enterprises is air cooled, right? They're not doing the latest and greatest Grace Blackwell and they won't be doing Vera Rubin in mainstream enterprises, at least not for a while. There is an argument that says that accelerated computing is a new form of computing, CUDA is the new sort of developer stack, and that the entire general purpose stack in the fullness of time, to use a phrase that we like to use, will be absorbed into the AI factory model, compute, storage, networking, offloads, and that ISVs are going to ultimately have to write to this new accelerated computing stack or they'll become irrelevant. What do you guys think about that?

Corey Quinn

>> I think that's an awfully lofty term to describe the grim reality of today where a stupendous amount of the software we have is more or less just gluing APIs together maybe with some data enrichment as well. There's no such thing as an agentic CPU, as Tobias Mann said, and I echo the sentiment, where there's so much of what we're doing on a computer that is not inference based. I mean, theoretically, could you make everything inference based? Yeah. All right, every day at five o'clock, I want to rotate the log files. Well, okay, instead of doing that, now I want you to think about rotating the log files first and 90% of the time you will and 10% of the time maybe you'll decide not to. That starts to get very non-deterministic, very weird, but also kind of wasteful. It's not necessary in those same ways.

Dave Vellante

>> So you would agree that... So what happens to the general purpose computing, the whole trillions of dollars of X86 infrastructure and software that's out there? Just continues to get innovated, doesn't get innovated away? What do you see happening there?

Sid Nag

>> Well, I think you mentioned ISVs, right? It's going to be interesting what the role of the ISV is going to be going forward. I think that's what we should be talking about.

Dave Vellante

>> Let's talk about it.

Sid Nag

>> Because when you think of... We just came out of the Amazon conference. When you look at things like AgentCore or even Quick, which has capabilities to spin up agents and applications and APIs in real time, that's going to be the new world of app development and that's going to be something that ISV should be seriously worried about, right? Now that's of course going to drive a ton of compute power that is going to be necessary. And again, that compute power is going to be highly distributed, right? It's not going to be centralized.

Dave Vellante

>> Okay. And I do want to go up the stack, I promise I will, but I want to stay in this for a minute. So when we went from on prem software to SaaS, everything changed. Technology model, the operating model, the pricing model, everything changed largely within IT, and the same thing seems to be happening again. The pricing model is going to change, the whole deceit based model.

Corey Quinn

>> Several times in a month, yes.

Dave Vellante

>> Right. Several times in a month. Tokens now you think... Okay. And so more CapEx. I would think IT spending, tech spending is going to increase dramatically, unless this AI thing is just BS. On average, 4% of companies' revenues are spent on IT, on technology. That's going to explode and it's going to be through tokens and presumably driving productivity. If you buy that, then-

Sid Nag

>> I don't buy it. Go ahead.

Dave Vellante

>> Okay, but that is the premise and that as a result, all the software stack is going to change. You say no.

Sid Nag

>> Because I don't think enterprises have unlimited budgets, first of all. I mean, we're already talking about FinOps or tokenomics in the world of AI.

Dave Vellante

>> So Salesforce would be in great shape then. Workday, Salesforce, SAP, ServiceNow, they're good. Market's got it wrong in your opinion, right?

Sid Nag

>> Yeah, exactly. Why? Because these are... So if you think about the enterprise, the enterprise systems are based on applications that are deterministic, like the ones you rattled off, right?

Dave Vellante

>> Yeah, yeah.

Sid Nag

>> Now here comes agentic or the world of agents, that's highly probabilistic, right? To marry the two together and make it work is not going to be that simple. And there are elements associated with, which I'm sure Corey can add to this, the whole economic prediction, right? People are focused on agents doing tasks correctly, but they're not focused on agents doing tasks correctly at what cost. You think that's going to not be a worry of the CFO tomorrow? Of course they're going to be worried.

Dave Vellante

>> Depends on the value that they're generating.

Corey Quinn

>> Depends on the continuing rise of this as well. And in the fullness of time, because all right, I used to have to go out to the mainframe because my local dumb terminal had nothing on it in terms of capacity whatsoever. At some point they're going to start making chips specifically tuned for inference. Local models are going to get good enough. Maybe it's not going to re-architect your entire application, but a local call to basically write a fancy for loop and rename all these files for me is well within the grasp of-

Dave Vellante

>> You're saying at the edge.

Corey Quinn

>> Exactly.

Dave Vellante

>> I agree with that.

Corey Quinn

>> And that becomes free except for the cost of power to run your laptop.

Sid Nag

>> Not just the edge, but also on premises, right? If I'm Bank of America or if I'm Goldman Sachs, I want to build a small language model that's applicable to my domain, right? So that could be a derivative or could be curated from a frontier model, but then I create an SLM and I ingest my company specific data that a frontier model can never scrape because sitting on legacy CMDBs and database records and all that. And now I've got a very customized model that I don't need a H100 to run on as a price of a Honda Civic, right? It's going to be a very different economic model that's going to emerge with the advent of these company-specific, enterprise-specific small language models. That's where the industry's going.

Dave Vellante

>> Okay, but H100-

Corey Quinn

>> H100 is for coding and they looked into it, but you don't hear about that anymore where we take our general purpose coding assistant and then we train it on your company's code base. Which is a way of saying, we're going to take this smart robot and then give it a traumatic brain injury. And surprise, surprise, it didn't go so well for the companies that tried that and now the general purpose coding models are superior than those custom trained on your code based things with just a bit of steering.

Dave Vellante

>> So I want to go deeper in the stack before we go up again. I want to talk about silicon. So things like Nitro, Graviton, secret weapon for AWS for a number of years. They really had a great lead on the rest of the market. But they were competing essentially with a flattening out X86 curve and Intel was shooting itself in the foot. Now the game has changed. Nvidia is setting the cadence of the industry. They're on a, let's call it a one-year extreme co-design, 10X every year. Maybe you stretch that out a little bit. But nonetheless, they're breaking historical boundaries in terms of price performance.

Corey Quinn

>> It's breaking the laws of physics and supply chains.

Dave Vellante

>> Yeah. Which, right, they've got to manage that and they're locking up supply chains, doing what they have to do to achieve that. My question is, how do you guys see, how would you stack rank the hyperscalers in terms of their ability? Let's start with AWS because you had Nitro, Graviton, ARM-based processors, home run.

Sid Nag

>> Trainium.

Dave Vellante

>> Trainium, Inferentia, which no longer has an R&D roadmap but it's being supported. But you wrote about this the other day. I said, "Inferentia's dead." You said, "Well, not quite dead." But what's the status of the silicon efforts inside of AWS? How would you stack rank that?

Corey Quinn

>> Of the three majors, I would say that you have Amazon who is right now the front-runner on this. Google is doing amazing things with their TPUs, but they would be a much more serious contender if they weren't continually getting distracted by whatever shiny thing moves at the moment. That has always been the Google problem. They don't stay the course on anything. Microsoft is, they say interesting things, I enjoyed my time at Build, they're doing some neat stuff, but they are in absolute shambles. It was reported today that GitHub is now moving to Amazon, moving to AWS for some stuff because Azure keeps falling over. They said nicer things about it, in the service of AI. Terrific. Great. If GitHub goes down anymore, they're going to write another copy of the business book, 4-Hour Workweek, to describe it. They can't keep the service up.

Dave Vellante

>> So you would put-

Corey Quinn

>> Under crushing demand, to be fair.

Dave Vellante

>> And Amazon was the first to really, I think it was the first to develop its own silicon. I know TPUs were in there too. I mean, to Google's credit, they've got probably more H100 capacity than anybody, but that's fine. That's maybe a nice advantage. But I would agree, I think Amazon's got some experience there and they can leverage that. Can they use this Graviton playbook vis-a-vis Nvidia? That's kind of the big question because they're competing.

Sid Nag

>> I don't think it's a matter of stack ranking because everyone's doing it differently. Amazon hasn't showed any interest in taking Inferentia and Trainium and making it commercially available for others to use.

Dave Vellante

>> I don't buy that though. I just don't see Google succeeding at that. I mean, they're going to use it for their own internal things.

Sid Nag

>> Well, Google is working Meta to . Amazon's not doing that.

Dave Vellante

>> But to what end? Are they going to become a merchant silicon vendor really?

Sid Nag

>> We don't know. We don't know.

Dave Vellante

>> I'd be surprised if that's the case. Maybe. Maybe they see they need to get the volume and Broadcom has volume.

Sid Nag

>> I mean, I think compute is going to be so multifaceted that you can't make a comment like that today. You don't know what's going to be-

Dave Vellante

>> Because there's going to be so much demands.

Sid Nag

>> Yeah, in different .

Dave Vellante

>> In a variety of workloads.

Corey Quinn

>> Let's not conflate ARM with GPUs. GPUs have their own unique architectures on a per manufacturer basis by and large. Graviton is an ARM chip. It is built on the ARM standard. You can compile the same software and run the binary on a Graviton instance as a Raspberry Pi sitting in my sewing. Why I have one of those when I don't sew, we don't know. But it is the same thing. It is you can get that architecture anywhere. If you build for an Nvidia GPU, that is basically what you're doing without a fair bit of reconfiguration.

Sid Nag

>> That's precisely right. I mean, an inferencer is not doing large matrix multiplication as a Nvidia GPU's doing. They're very different needs, right? So you can really stack-

Dave Vellante

>> Right. And Trainium, Trainium is a great inference chip. Right?

Sid Nag

>> Yeah.

Corey Quinn

>> Surprise, Amazon was bad at naming something.

Dave Vellante

>> Well, they had a training chip and-

Sid Nag

>> There are technology issues, there are commercial issues, there are positioning issues. So stack ranking is not an apples to apples thing.

Dave Vellante

>> That's something we do in theCUBE. But Microsoft well behind. Okay.

Sid Nag

>> Microsoft is my... We don't know what Microsoft's doing Maia. They haven't told much about it .

Dave Vellante

>> Maybe use it for internal stuff. .

Corey Quinn

>> Thing is they've been dead before many times and somehow they're still here in a front-runner.

Dave Vellante

>> Well, they got good cash machine. Okay. Let's move up the stack a little bit because the thing that I learned this week at AWS is they're serious about a work surface with Quick. Looks like they've got that right. And they've got a backend, what we call system of intelligence. They're working on that. That's AWS Context, I think they call it. And there's other pieces in there. AgentCore has capabilities, there's governance. Their governance seems to still be distributed and they seem comfortable with that. We'll let the ecosystem do whatever governance and they will keep it in Glue and SageMaker or whatever tooling we have there. They don't seem to be wanting to pull that together into an abstraction, which I thought-

Corey Quinn

>> Or they're incapable of doing it effectively. The stack has always been perilous for Amazon.

Dave Vellante

>> Maybe Quick can help. Well, they would argue Connect is a-

Corey Quinn

>> Is the single breakout success that they've seen.

Dave Vellante

>> It proves the rule, I guess. But nonetheless, you have this interesting battle going on. You've got the hyperscalers. AWS now in the mix. Clearly Google had shown its hand before. Microsoft is there. You went to Build. They showed some pretty interesting things at Build.

Corey Quinn

>> They really did.

Dave Vellante

>> You've got the SaaS vendors that you say are fine because they got determinism. And it's true. Salesforce with their data cloud, their cloud, Data 360, I think they call it now.

Sid Nag

>> Can't keep up with the names.

Dave Vellante

>> Right. Palantir has got the best ontology. ServiceNow has designs. They want to participate in that. No doubt SAP's got great data models. So they're in the mix.

Sid Nag

>> And don't forget the model builders.

Dave Vellante

>> And then exactly, that's the third. You got the model builders. You also have Snowflake and Databricks now participating both in that end user surface for developers and for business users.

Sid Nag

>> And don't forget the Wall Street bros, like the-

Dave Vellante

>> What do you mean?

Corey Quinn

>> Careful, they can hear us.

Sid Nag

>> Like Blackstone. The Blackstones and Apollos getting into the AI business.

Dave Vellante

>> The PE companies. Okay. So are we destined for silos of intelligence, I'll call it, more fragmentation in the tech business?

Sid Nag

>> I think so. I think-

Dave Vellante

>> History would suggest that, wouldn't it?

Sid Nag

>> Yeah. I mean, I do think, however, that all of these things are going to converge at some point, right? Because these things... At the end of the day, where are people going to make money out of all of this? From the end users. And end users don't want a ton of technology being thrown at them. That's where things like AI factories matter. That's where things like systems of intelligence matter. So show me a pathway where whatever Amazon is doing with Quick, I think Quick's a amazing tool. I didn't realize how powerful it was. But I see for the first time Amazon has done something interesting in this industry in bringing all their different capabilities together into a single work surface. Some people call it the enterprise operating system. Are you going to wake up in the morning and instead of seeing a Windows screen, are you going to see a Quick screen? That day may arrive.

Dave Vellante

>> Well, that's the intent, right?

Sid Nag

>> That day may arrive. And others are going to follow. I mean, Google's not going to sit back and they're going to try to do something similar with Gemini Enterprise and so on and so forth. I think that's really where the battle is going to be fought in my opinion.

Dave Vellante

>> So there's a school of thought, some say the LLM vendors are going to own that because they don't have to make money yet and they've got access to compute. Others say, no, they're never going to be able to write that deterministic software.

Sid Nag

>> Well, the LLM vendors can build a work surface that-

Dave Vellante

>> Well, they are doing that.

Sid Nag

>> ... is available on your laptop. Then of course they can do it. They haven't shown any propensity to do that yet.

Dave Vellante

>> Well, what's Claude Cowork and Codex from OpenAI? That's a work service.

Sid Nag

>> That's a very atomic function. It's not an all encompassing function. Quick and Build, I mean, that paradigm can build pathways into a Salesforce or a Amazon.

Dave Vellante

>> But what-

Corey Quinn

>> It is if you all these things and violate your company security policy. The nice thing about Quick is that presumably it works with the way that your data -

Sid Nag

>> Yeah, exactly. It's a window to the world essentially, right?

Dave Vellante

>> And to quote-

Sid Nag

>> And I'm not saying Amazon's the only company that's going to do it. Everyone's going to follow.

Dave Vellante

>> But to quote George Gilbert, there are more user work surfaces agents than there are fleas on a camel's back. And so they're not all going to survive. What impressed me about what I saw at AWS is they've got that backend system of intelligence that is emerging and it looks like it's a combination of AWS Context is what they call it.

Sid Nag

>> I think that's what they call it, yeah.

Corey Quinn

>> I want it to work, but I am so highly skeptical given that it seems that Amazonians I've talked to for a decade now seem incapable of talking to other Amazonians. The left hand doesn't know what the right hand's doing and it never has.

Dave Vellante

>> They have to in order to make that work. You're right.

Corey Quinn

>> If they can't solve that, then the rest of it's for naught.

Dave Vellante

>> Well, I can't remember.

Sid Nag

>> I think he makes a good point because I was impressed with Quick, then I saw the AgentCore demo, I was impressed with that, but that's like divergent paths.

Corey Quinn

>> It's like you two should talk.

Sid Nag

>> Wow, what the hell? And then I go and see a contact center demos called Connect. Connect has no connected tissue, sorry for the bad pun, with Quick, right? So why is a contact center agent not using Quick? Why is he using Connect? Right?

Dave Vellante

>> Well, and then you've got-

Sid Nag

>> These are the problems that have to be solved at some point.

Dave Vellante

>> You got S3 tables.

Sid Nag

>> And they're all based on agentic AI.

Dave Vellante

>> You got S3 tables, that has some value there. You've got, as you mentioned, AgentCore. There's the SageMaker sort of piece, which is in this Glue, that catalog. You've got now Quick. You've got, not called Elasticsearch anymore, it's Open Search, right?

Corey Quinn

>> This is fantastic if I want an ETL pipeline that I have to go build from parts at Home Depot. That's the problem. They haven't .

Dave Vellante

>> But the tech is all there.

Corey Quinn

>> Yeah. Maybe Claude Code can put it together for them.

Dave Vellante

>> And they've got the knowledge graph in Context, right? AWS Context.

Sid Nag

>> Yeah. And Quick too, and Quick. It's amazing. It's powerful. Yeah. The knowledge is there.

Dave Vellante

>> Is it the same knowledge graph?

Sid Nag

>> No, it's different. That's the thing. It's a different-

Dave Vellante

>> In Quick than Context? Are they different knowledge graphs? Are you sure?

Corey Quinn

>> Of course it is.

Sid Nag

>> Yes, it is.

Dave Vellante

>> Okay. Well, they got to get that all together.

Sid Nag

>> I went to the key. I didn't sit through a slide show. I went to do the demo. I went into the weeds. They're all different. That's a problem.

Dave Vellante

>> Well, so that's the-

Corey Quinn

>> This is their fourth and fifth bite at the apple. They've been saying for three or four years now, they're great at AI. This is the first time that actually is credible with what they've built in Quick desktop. But it's like, yeah, Lucy has yanked that football away enough times that I'm a little skeptical at taking a kick at it. Other people are too.

Dave Vellante

>> Well, but in Agentcore, we heard the growth of AgentCore. Swami gave some stats today, so it's obviously getting absorbed. I talk to customers who love it. But this has always been Amazon's greatest strength and its greatest weakness. Jassy years ago said, "Look, we like to be at the primitive level because when the market changes, we can change with it." After a point in time, it gets too complicated. And Microsoft is the other end of the spectrum. Like somebody said, it's a black box that doesn't always work.

Corey Quinn

>> What's the box made out of? SKUs.

Dave Vellante

>> Yeah.

Corey Quinn

>> All different. 75 or more now called Copilot.

Dave Vellante

>> Okay. So we're at the halfway mark of 2026. What surprised you this year and what are you looking for in the second half?

Corey Quinn

>> I would say that I did not expect the model labs to go for IPO as quickly as they have. And I would have thought that by the time you are basically hovering around a trillion dollars in valuation, you would have gotten a little bit better at managing your comms and your messaging and not the story being about what you're building, not about the drama that swirls around you like a miasma everywhere that you go. That has surprised me. I'm looking forward to increased capabilities across the board from the AI space. The things we're talking about now, if we play this back a year from now, it's going to sound hopelessly out of date, sophomoric because the world has changed. Heck, between now and the time we publish it, that might be the case. That'll be several hours potentially. And it is constantly changing. So what I am excited about is that there is value here. Companies are finally caring about what it costs rather than saying, "We care about costs," while their engineering team has a money incineration tournament leaderboard going on. Those two cannot coexist. So we're starting to see, okay, where is the value coming from here? And we're getting answers to that because it's clearly not nothing. And it's also clear that we're not going to summon God via JSON anytime soon. The answer is somewhere in between and figuring out the exact location of that in a per company and a per workload basis is really what the second half of this year's about.

Dave Vellante

>> That's great analysis. I mean, this is going to take the better part of a decade to play out and to show that AI can be self-funding and live up to some of its promises. What are your thoughts on what surprised you?

Sid Nag

>> What surprised me is the rapid adoption of agentic. I didn't expect it to be that pervasive and that soon. So that's definitely a surprise. The other thing that surprised me is a lack of understanding by the AI vendors in terms of the ability for people to pay for what they're building. I still think they haven't figured it out. They think they can just spend billions and trillions of dollars building GPU and TPU and silicon farms and somebody's going to come and have enough applications and enough workloads to consume compute power out of those farms. I still think there's this massive divide and the chasm that has to be crossed between what the vendors are doing and what the enterprises are looking for. And that is just mind-boggling. I don't get it. Aren't they talking to each other?

Dave Vellante

>> I mean the continued-

Corey Quinn

>> A developer's not going to drop 20 grand a month on inference tokens.

Sid Nag

>> Exactly.

Corey Quinn

>> That's sort of looking like what's needed to achieve some of these valuations.

Sid Nag

>> What is missing? Does it take an analyst to go and bridge that gap?

Dave Vellante

>> You've got Sundar, you've got Andy Jassy, you've got Satya, right? Well, Nvidia is making money off of all this. But these are smart people. They've seen companies get crushed in the past by just making really bad decisions. And this CapEx, I mean, if they're wrong, this is a disaster for these guys.

Sid Nag

>> Well, there's a dark horse in this conversation.

Dave Vellante

>> Oracle's the other one, right.

Sid Nag

>> Which is maybe they're thinking that enterprise is never going to consume that amount of AI ecosystem things, whether it's workloads, applications, and maybe the focus should be on the government sector and that's where all the money's going to be spent.

Dave Vellante

>> Well, a lot of things surprised me about that. I have to stop there.

Sid Nag

>> That's a topic for another day.

Dave Vellante

>> Mythos was a bit of a surprise, the government's active role.

Sid Nag

>> Yes. I mean, look at Palantir. Great example. Where are they making most of their money?

Dave Vellante

>> Yeah, government.

Sid Nag

>> Public sector.

Dave Vellante

>> Yeah. Government and forward deployed engineers. But obviously there's some powerful technology there. So what I'm looking for in the second half is Palantir for all. That actually is you can roll out on a horizontal basis without a million dollar a year forward deployed engineer in my shop. That's what I would hope to see. Guys-

Corey Quinn

>> We have forward deployed engineers at home.

Dave Vellante

>> Great. Exactly. That's right. Well, Corey and Sid, thank you so much for spending some time coming down to the Exchange and-

Sid Nag

>> Pleasure.

Dave Vellante

>> ... hanging out with theCUBE. It's great to have you.

Sid Nag

>> Thanks for having us.

Corey Quinn

>> Thanks, Dave.

Dave Vellante

>> All right. Thank you for watching. This is Dave Vellante for theCUBE NYSE Wired. We'll see you next time.