In this episode of Breaking Analysis, Dave Vellante, Co-Founder and Co-Chief Executive Officer of SiliconANGLE Media, Inc., along with George Gilbert, Principal Analyst for Data and Artificial Intelligence at SiliconANGLE and theCUBE, examine the details of AWS re:Invent 2025. The event emphasizes Amazon's strategy to embrace practical AI solutions known as Worker Bee AGI, contrasting with the pursuit of advanced AGI models. This analysis explores what AWS's move signifies for the future of AI applications within enterprises.
Vellante and Gilbert lead the discussion, with contributions from theCUBE Research and insights from analysts. Gilbert explains AWS's focus on agentic scaffolding and customizable models, illustrating the company's move toward "service as software." The conversation revolves around the evolution of AI and its practical applications, particularly how AWS differentiates itself from rivals by focusing on enterprise solutions rather than consumer-driven technologies such as frontier large language models.
Key takeaways include the importance of data-driven AI advancements in achieving scalable enterprise applications, as highlighted by Gilbert. According to the analysts, AWS's focus on Worker Bee AGI, driven by practical outcomes, marks a shift away from costly and less sustainable AI approaches. They emphasize the significance of aligning technology models with business needs to exploit the tangible potential of AI within organizational frameworks.
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
AWS re:Invent 2025. If you don’t think you received an email check your
spam folder.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open the link to automatically sign into the site.
Register for AWS re:Invent 2025
Please fill out the information below. You will receive an email with a verification link confirming your registration. Click the link to automatically sign into the site.
You’re almost there!
We just sent you a verification email. Please click the verification button in the email. Once your email address is verified, you will have full access to all event content for AWS re:Invent 2025.
I want my badge and interests to be visible to all attendees.
Checking this box will display your presense on the attendees list, view your profile and allow other attendees to contact you via 1-1 chat. Read the Privacy Policy. At any time, you can choose to disable this preference.
Select your Interests!
add
Upload your photo
Uploading..
OR
Connect via Twitter
Connect via Linkedin
EDIT PASSWORD
Share
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
AWS re:Invent 2025. If you don’t think you received an email check your
spam folder.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open the link to automatically sign into the site.
Sign in to gain access to AWS re:Invent 2025
Please sign in with LinkedIn to continue to AWS re:Invent 2025. Signing in with LinkedIn ensures a professional environment.
Are you sure you want to remove access rights for this user?
Details
Manage Access
email address
Community Invitation
Breaking Analysis | Worker Bee AGI - Why AWS Is Betting on Practical Agents, Not Messiah AGI
In this episode of Breaking Analysis, Dave Vellante, Co-Founder and Co-Chief Executive Officer of SiliconANGLE Media, Inc., along with George Gilbert, Principal Analyst for Data and Artificial Intelligence at SiliconANGLE and theCUBE, examine the details of AWS re:Invent 2025. The event emphasizes Amazon's strategy to embrace practical AI solutions known as Worker Bee AGI, contrasting with the pursuit of advanced AGI models. This analysis explores what AWS's move signifies for the future of AI applications within enterprises.
Vellante and Gilbert lead the discussion, with contributions from theCUBE Research and insights from analysts. Gilbert explains AWS's focus on agentic scaffolding and customizable models, illustrating the company's move toward "service as software." The conversation revolves around the evolution of AI and its practical applications, particularly how AWS differentiates itself from rivals by focusing on enterprise solutions rather than consumer-driven technologies such as frontier large language models.
Key takeaways include the importance of data-driven AI advancements in achieving scalable enterprise applications, as highlighted by Gilbert. According to the analysts, AWS's focus on Worker Bee AGI, driven by practical outcomes, marks a shift away from costly and less sustainable AI approaches. They emphasize the significance of aligning technology models with business needs to exploit the tangible potential of AI within organizational frameworks.
Breaking Analysis | Worker Bee AGI - Why AWS Is Betting on Practical Agents, Not Messiah AGI
Dave Vellante
Co-Founder & Co-CEOSiliconANGLE Media, Inc.
HOST
George Gilbert
Principal Analyst, Data & AISiliconANGLE & theCUBE
search
Dave Vellante
>> At AWS re:Invent 2025, Amazon was tasked with serving millions of its legacy customers and speaking to them while at the same time fighting the perception that they're behind in AI. And what AWS is doing is taking a practical approach, not chasing the holy grail of what we call messiah AGI or even frontier LLMs. Rather, what AWS is doing is they're focusing on foundational agentic scaffolding and customizable large language models or small language models consistent with our view that the tangible opportunity in AI is within the enterprise, i.e. worker bee AGI, not messiah AGI. Now, skeptics might say that worker bee AGI is just a form of RPA, let's call it RPA 2.0 because the data is still siloed. While this is true, we don't see this as paving the cow path, rather we see AWS making a move from siloed agentic automation to what we call service as software. Welcome to this special Breaking Analysis. My name is Dave Vellante. I'm here with George Gilbert, and we're going to break down AWS re:Invent 2025 and share with you our takeaways from re:Invent specifically in the context of the theme that we've been hitting on Breaking Analysis around service as software, which we see as a complete transformation in the operational, the business and technology models of virtually every company and organization and every industry. George, thanks for all the time you put in to making some cool slides. We've got a number of them and we've got a lot to cover. So appreciate you being here.
George Gilbert
>> Always a pleasure.
Dave Vellante
>> Okay. So let's bring up the sort of title slide, the highlights from frontier labs to messiah AGI to enterprise worker bee AGI. I love this, George. You got comrade Altman on the left here. You've got the communist arm band. He's chipping away, trying to get to AGI and get expertise from gig workers. Meanwhile, you've got Matt Garman on the right-hand side, shaking hands with Jamie Dimon. We've written a piece around why Jamie Dimon is Sam Altman's biggest competitor, how he gets there. So take us through these themes. Why did you choose to revisit the comrade Altman? And let's talk about where Amazon fits.
George Gilbert
>> Okay. So basically, the OpenAI business and ChatGPT, the product, they led the first leg of the GenAI revolution. And through GPT-4, maybe even GPT-5, that was all built on extracting free data from the internet. In other words, even though their compute costs were high, the costs of the data that fed the intelligence was essentially zero or very low. And as the high quality free data is running out, they are spending more and more now on curating expert sort of reasoning traces. The expertise from very specialized people. And the key point is that this is not really scalable. It's an entirely different business model. And there's been very credible reporting from the information that even in 2029 or 2030, they expect to spend somewhere between 20% and 25% of their revenue on curating this proprietary data. And the point we're trying to make here is like the picture of Altman is he's digging ever deeper into the seams of the Earth trying to pull some coal out in every smaller amounts.
Dave Vellante
>> Kind of chipping away and it's hard knocks.
George Gilbert
>> Yeah. And it's sort of declining marginal utility. Whereas up on the surface of the Earth, we have Jamie Dimon as a proxy for the leading enterprise. All his resources are renewable. It's his proprietary data. And the more agents he gets up and running, the more models he gets running, the more data he generates. He's on an experience curve. So the more he harmonizes and hooks the data and the models together, he's on this flywheel. And so his marginal costs decline. So Sam Altman get his marginal costs increase and they get more specialized utility. And then the Jamie Dimon as a proxy for all enterprises, they're on this learning curve that's accelerating going down.
Dave Vellante
>> Okay. So why the windmills?
George Gilbert
>> It's renewable. Windmills and solar farms.
Dave Vellante
>> Yeah. Solar panels and windmills.
George Gilbert
>> The point is, in the age of AI, it's about the data. It's not the algorithms. The algorithms are done by the labs. And so how you shape your data to train the AI is everything.
Dave Vellante
>> Now, as you and I have talked about, David Floyer has this belief that OpenAI ultimately ... He already realizes that the opportunity is in the enterprise and they will build software on top of their LLM. He agrees with you. The LLM is not really the issue. It's the software on top that will enable enterprises to actually do some of the hard things that today companies like Amazon and JPMorgan Chase and big companies like we've talked to Dell are doing internally, but they will make it much more simple for mainstream enterprises to do that. And they will go after that opportunity. You believe, correct me if I'm wrong, that that's going to be very difficult for them to do because they don't have access to the data. So my question to you would be, well, couldn't they get access to the data by partnering up with the Jamie Dimons of the world?
George Gilbert
>> Theoretically, I would have two challenges with that. And there's a third issue that David's mentioned that I'll go through. So there's three issues. One, will the corporations be comfortable parting with their most sensitive data with frontier lab that can they set up claim rooms and establish trust? Because this is the crown jewels of-
Dave Vellante
>> They have to do that or else it's a fail. We agree on that.
George Gilbert
>> Right. But the way their models and their customization work, the company has to transfer the data to the model lab because with open models, you can bring the weights and a partially trained model onto your premises or into your VPC and no one else sees it. So that's one issue. So let's say they can establish trust. Then the other is this notion of the scaffolding around the agent that integrates it into the application and the workflow that's relevant to your business. This is the mapping of the new, adaptive, probabilistic model as agent and the older deterministic software that was essentially your application, but now retooled to be able to work agents.
Dave Vellante
>> Okay.
George Gilbert
>> And you look at OpenAI introduced two agents like a GUI low-code version and a programmer first version. And they were very competitive tools, but they acknowledged at their developer conference that it didn't match up to even Databricks' developer tools. And Databricks itself doesn't have ... They have the developer tool, the agent tool scaffolding, but they don't have the data scaffolding, which you and I have talked about and we'll get into the system of intelligence. Databricks aspires to it. The point I'm trying to make is there is so much traditional enterprise software that has to turn into a scaffold to make the agent useful and that these frontier lab companies are not traditional enterprise software companies. And to build that scaffolding is essentially to recreate almost decades of enterprise software. And so I think it's more likely that this is a sustaining innovation technically for the enterprise software incumbents and where it's a disruptive innovation is in the pricing model, where they have to go from seat based pricing and for the infrastructure from consumption based pricing, but especially the seat based pricing, they might lose that and go into something more value based, outcome based, usage based.
Dave Vellante
>> Okay. So let's get into that, because we got a lot of ground to cover. We're just getting through the title slide. You bring up the next slide, Anderson, which is the agenda here. We're going to talk about how the business model, operational model, technology model are changing. And we think it's profound and it has impacts for every company that software like economics now accrue to virtually all companies and it's a winner take most market. We'll get into mapping services software to the specific AWS software stack. So some of the takeaways from the innovations we heard this week at re:Invent and we'll talk a little bit about the whole GPU mania that's going on and some of the industry implications. So if you would go to the next slide, Anderson, let's set this up. So we've talked about this earlier in some of our earlier Breaking Analysis, that everything has to change. Customers have to completely transform their operational model and their business model. George, we saw this when we went from on-prem software to SaaS, but it really only affected the IT departments or-
George Gilbert
>> And the vendors....
Dave Vellante
>> technology companies.
George Gilbert
>> Yes.
Dave Vellante
>> Exactly. It really didn't make that much of a difference to the mainstream organizations out there. The other thing that I'll mention here, and we're going to separate, we're showing here the operational and the business pieces, the technology model we're going to deal with separately. What's interesting here is that normally, practitioners would tell you, "Well, the technology, that's not the hard part. It's the people and process. It's the operational. It's the business model. Those are the really hard things." Not only are they hard here, but the technology is hard here. So fill us in on your thoughts.
George Gilbert
>> Okay. So even before we get to the technology, let's talk about the operational model is for 60 years, we've had many technology transitions, but fundamentally we were still building silos. We might have replaced the UI, we might have had a new device, we might have had a new operating model for the software, but the operational model was ever more silos. The change is now that we're building, we're going from these craft silos, like craft manufacturing when you had these work cells, we have knowledge work cells now because that's all we know how to automate and we're going to have an end to end built order assembly line for knowledge work. That's a profound process change, because the org chart that we have now, which is designed like to make a machine efficient, doesn't work when you're doing end-to-end outcomes. So that's the process change, the operational model change. The business model change just really quickly is you're now essentially charging like time and materials. It's almost like fee-based and it's cost plus. When you're offering outcomes, what happens is you can charge either by outcome or by value or something that's a proxy to that. And this is the biggest change, not just the pricing model, but we've always seen financial capital supply supporting physical capital that depreciated and you had economies of scale and so you had marginal cost advantage because you could amortize that fixed capital investment over many more units.
Dave Vellante
>> Non-recurring engineering expense upfront. A little bit of maintenance ongoing, but your effective marginal costs went down close to zero. Not with cloud. You had cloud COGS, but as Satya says, back in the days of CD-ROM went down to basically the cost of a CD-ROM, but now it's different.
George Gilbert
>> Now, there's profound difference and I don't think people appreciate this.
Dave Vellante
>> The tokens. You're paying for tokens, right?
George Gilbert
>> That's the cost side, but now your capital is the intelligence and the expertise that's digital and that compounds, and it compounds exponentially. So that the more data you get from experience, the richer your experience base is. So you get an advantage both on cost and on differentiation. And that's why you get winner take most economics.
Dave Vellante
>> Okay. So the software company economics are changing. They got to play not only cloud costs, but they got to pay token costs. The consumers presumably are going to benefit from ongoing improvements in GPU, price performance, et cetera. And the marginal economics that we used to see accrue to confer to software companies will confer advantage to the organizations that get on the learning curve of AI soonest. That's what you're saying.
George Gilbert
>> It's the difference between a depreciating asset, whether it's a physical asset or a process design, or now it's compounding expertise. And so the notion of get big, get fast, it's get volume, get fast, and then no one can catch you.
Dave Vellante
>> Okay. So now, we're going to talk a little bit about the technology model. Actually, before we go there, pricing has to change as well. And you sort of a little alluded to that before. Well, in some cases it was a perpetual license model or a seat based model to a consumption model in SaaS. We still have a lot of seat based models in SaaS, but you saw the pricing models change when we went to SaaS. That has to change again. You referred earlier to outcome based. I was talking about tokens before. You're probably not going to price on tokens. That's not going to be the ideal. You're going to price on outcomes, but your unit economics are going to be determined, at least in part, by your cloud cost, your COGS, and the token costs. So that's going to set the baseline. That's the breaking point.
George Gilbert
>> This gets back to the difference between models and agents. So the model guys are in this token grind where through distillation and advances in the frontier of intelligence, their costs are coming down something like 10x to 30x per year, some obscene number when it's just the raw model and the API. But when it's an agent and there's a learning loop in there, their prices are much more stable. And so that gets back to what we're going to talk about, which is the scaffolding around it to build that agent.
Dave Vellante
>> Right. Okay. Let's go on to the technology model and talk about what the transformations are there. Thank you, Anderson. So you're showing here that we shifted from these applications or we're shifting from these application centric silos to this data centric platform. So this necessitates the breaking of those silos, something that we've been talking about for years and years and years. How do you see that playing out? Why is this technology transformation so much more potentially complicated than say the shift to cloud?
George Gilbert
>> Okay. Because the shift to cloud was just a different operational model for the same silos. We broke things up into microservices, but we were still siloed.
Dave Vellante
>> From a data standpoint?
George Gilbert
>> Yeah, from a data standpoint and actually from the application logic itself. What we're doing now for the first time, we turned data into a corporate asset with the rise of the relational database. Now, where we need agents and humans to have a shared understanding of the business, we need a system of intelligence where the processes, the rules about how the business runs, they become a shared asset. That we've never done. And so it turns on its head, 60 years of investment. And how we do that, you know I've been working on this for months, like a migration path for how we might do this, that's going to be really difficult.
Dave Vellante
>> So why do you have in this slide Tower of Babbel, what are you inferring?
George Gilbert
>> That's the legacy version.
Dave Vellante
>> If you bring that back up, Anderson.
George Gilbert
>> Because every siloed app has its own language and so they don't talk. And if you want a picture of what's going on in the business, now you build a lakehouse, but even within the lakehouse you have silos because you might have-
Dave Vellante
>> Schemas, different things.
George Gilbert
>> Yeah, different schemas for sales, for service, for logistics. And if you want to ask questions across them to say what happened and why, it's very difficult. You need to go stitching bridges together with data engineering.
Dave Vellante
>> Okay, great. So bring up the next slide, you're going to invoke Bill Gates here.
George Gilbert
>> Well, this is the-
Dave Vellante
>> What is this? Kicking whales down the beach.
George Gilbert
>> Dead whales, dead whales. He used to say this actually, when I was at Lotus, about the DOS user interface we had, he was like Lotus and WordPerfect, trying to get to graphical user interface and preserve compatibility is equivalent to kicking dead whales down the beach. You can't do it. I mean, he was kind of taunting us and it was good marketing, but it captured the spirit of the challenge. This is kicking a thousand dead whales down the beach because it's 60 years of investment in silos that we now have to somehow harmonize and no one really has a great solution for that.
Dave Vellante
>> Okay. Let's bring it back the next slide, which is our agenda. We're going to now dive into what this all means from an AWS perspective. We're going to map that. We saw a number of announcements this week around ... A lot of it was making stuff GA, but doubling down on Kiro, a lot of stuff for the software development light cycle, aging core extensions, Nova, a lot of announcements on Nova, Nova Forge, we're going to talk about a lot, and so let's get into it. All right. So here, we're showing a slide, service that software requires customers transform their technology model. This is a slide we've shown before. We're bringing in the Tower of Babbel. We're showing how what we see as the software stack emerging. The green portion of this is the system of intelligence. We see that as the high value piece of real estate. So George, take us through this and then we're going to map AWS on top of that.
George Gilbert
>> Okay. So just really quickly, there's this sense that the data platform is your foundation for building AI based applications, but the data platform, it's not machine or agent readable. And it even has to go through a layer like your BI tool to be human-readable.
Dave Vellante
>> When you say data platform, you're talking about Snowflakes, Databricks, Redshift.
George Gilbert
>> They've done an amazing job of collecting in one platform data from all the different operational applications and the websites and external sources. So you have one data estate, essentially like a logical data estate, but it's not all harmonized.
Dave Vellante
>> Bring that slide back up if you would. So, okay. So the data platform would be the bottom left here.
George Gilbert
>> Yeah, in the orange.
Dave Vellante
>> And you're showing ... Yep. And then we're really good at what happened.
George Gilbert
>> Yes. And it's these snapshots, these two-dimensional snapshots of what happened. And there's really many silos in there, like we were talking about earlier, because there's different schemas for different departments or functions. And so it can answer the question what happened within a fairly narrow scope. And if you do feature engineering and some machine learning, maybe you can say what's likely to happen in a narrow scope.
Dave Vellante
>> Okay.
George Gilbert
>> But what you really want is to build this system of intelligence, the green layer that is 4D map and end to end view across the business. This is what like Palantir is advertising and some of their customers are building it, although it's very expensive because they're sort of doing it by hand. It's not yet repeatable.
Dave Vellante
>> They're forward deployed engineers are making a lot of dough.
George Gilbert
>> 950,000 a year charging them. So you're basically building SAP customer by customer. Each customer is building it again, the equivalent of doing it by-
Dave Vellante
>> Heavy lift, but it's actually high impact.
George Gilbert
>> High impact. Yes, exactly. Now, if they can make that repeatable, that's a different story.
Dave Vellante
>> Okay.
George Gilbert
>> Okay.
Dave Vellante
>> So go ahead. Do you want to finish up on the-
George Gilbert
>> The key point here is the systems of agency. All the agents are not going to be really effective unless they sit on top of that system.
Dave Vellante
>> Sorry, bring that slide back up. The systems agency is the-
George Gilbert
>> Is the yellow.
Dave Vellante
>> Is the yellow portion of this slide. We call it the agent control framework, if you will, which is important, but it doesn't have the intelligence. It gathers its intelligence from the green layer, which is really the centerpiece of this set.
George Gilbert
>> Whether it's an agent or a human, they need to see across the enterprise to know that if I put a stop on this order because I don't have a part, I'm going to annoy a high value customer and that's going to impact a large contract. You need that broader scope and you need to understand it harmonized so that they're all sharing the same meaning.
Dave Vellante
>> Today, humans make that decision.
George Gilbert
>> Right.
Dave Vellante
>> In order for agents to make that decision confidently in a governed manner, you've got to have that 4D map or that digital representation of the enterprise.
George Gilbert
>> And if a human is overseeing the analytics, the human has to be able to go through all the different scenarios and they can't keep mapping all that stuff if they're doing it by hand.
Dave Vellante
>> Right. I mean, you could do what ifs and you could do what ifs very powerfully, but then if you really want to go deep and across the organization, it gets much more challenging.
George Gilbert
>> Right.
Dave Vellante
>> Okay. So today, this week, we're in day four here, we heard a lot from AWS. So let's bring up the map of the AWS into this software stack. I'll just cut to the chase. What's missing is that green layer. It's not obvious. And when you talk to AWS about how they're doing things internally, they essentially, like JPMC, when you talk to them, like Dell, when we've talked to them, they have to build that green layer themselves. It doesn't come out of the box. It's not a solution, but help us map into the AWS into this model.
George Gilbert
>> So where they were very strong, like really strong was ... It looks like Bedrock is becoming more mature. Like last year, I think they were struggling with two and three nines, that's kind of reliability, that sort of thing. That was a sort of red alarm fire to get Bedrock up and running. They needed that abstraction layer above LLMs.
Now, AgentCore is this control framework for building multi-agent systems, and that looks really promising. And we're going to go into that, but even more impressive was that they did these first-party agents like Kiro autonomous agents for coding, and we'll go into this. The DevOps agent, security agent, these look like ... They have to meet customer acceptance, but the story they told around them was very compelling. And again, what was in the middle, the system of intelligence, the context that drives an agent. The agents are very context-sensitive. It's AI. AI is programmed by data, and so you need to shape your data. There's a lot missing there still. And then at the bottom data platform later, there's SageMaker lakehouse, and maybe we'll see Neptune as a-
Dave Vellante
>> So hold on a second. So SageMaker, it's a big emphasis last year, especially too around S3, Iceberg table capabilities. And so S3 is becoming much more than a get put object store.
George Gilbert
>> Yes.
Dave Vellante
>> Yeah. And then Neptune, Neptune is their graph database.
George Gilbert
>> Yes.
Dave Vellante
>> And they don't really talk about their knowledge graph that much. It doesn't get a lot of high visibility in things like that.
George Gilbert
>> This is where like the difference between Microsoft and Amazon is. Microsoft is, they want to think three to five years out where their customers should be, and then they'll announce early versions of products to try and lead them there. And maybe also to, in this case with Fabric IQ, to freeze the Palantir sales momentum. Whereas Amazon is more, "We're just one step ahead of our customers. We're hearing the problems they're working on and we think we can do something a little bit more principled." And so, but you could see right now, they're using Neptune to help customers build these AI workloads, sort of bottom up workload by workload, but you could see a day sometime in the future where Neptune would join S3 buckets, S3 tables as a way to structure the data.
Dave Vellante
>> All right. So bring that up again one more time, Anderson. So what Amazon, our takeaway is Amazon, they're focused on the agent specific tooling. Swami said, "We want to be the best place to build agents." We give them high marks for that. What's missing is that system of intelligence piece. Nobody really has that. You mentioned Palantir, Celonis, but this is the sort of new territory, but bringing together that data and process knowledge. Neptune looks like it potentially could be a linchpin of the future, or they'll do something different, like do something with S3 buckets or, again, maybe follow in Microsoft's footsteps, which is unlikely that Amazon would do something like that. Okay. So now let's dig into some of the pieces, specifically the DevOps agents. Let's bring that up. You've got the software stack of the future and you've got the green arrow pointing to the topology of system intelligence. You got the green box with the arrow. You got the yellow box with the arrow pointing at the DevOps agent. You've got Matt Garman talking about the DevOps agent. Why is this important? How does it fit?
George Gilbert
>> Okay. Think of a DevOps agent as, at the most basic level, you've got this sprawling estate of infrastructure, middleware, application components, the microservices. It's not like SAP where it's this one, a single vendor created end-to-end framework so they can look inside and they know where to look. It's so much more open-ended. So think of the DevOps agents as a way of saying, "I need to look across my digital operations. And when something goes wrong in all these thousands or tens of thousands of sort of components, I need to find out what's the root cause, what caused that?" And then if I can figure it out with high confidence what caused it, then I can put together a remediation plan potentially autonomously when the confidence is high enough. And so the lesson to take away from this is the DevOps is kind of like training wheels for the system of intelligence because this is your digital operations. And the system of intelligence is not just your digital operations, but it's your business operations, all your business operations. So what you need to do here is you need a map of your digital operations. And that's what they talked about with topology, which is they're trying to learn a map of how the components of your Amazon services and infrastructure fit together. Now, I talked to different vendors. I talked to Datadog and Dynatrace and I think one other.
Dave Vellante
>> Splunk?
George Gilbert
>> Splunk and Elastic.
Dave Vellante
>> Okay.
George Gilbert
>> And they all have their own takes on how to do this and they all have different views. And this was maybe the most surprising lesson for how we might build a system of intelligence more generally, which was each has a rich picture of their own domain. And they also have some insight into the other domains. For instance, Dynatrace, their customers put a probe in each host and that gives them really deep visibility and so they can build this causal graph. They can build the digital twin of your digital operations. And so what happens with the DevOps agent from AWS is when it can't figure out, it figures out a problem and it can't find it in its perimeter, it says, I think it goes to the Dynatrace and says, "I'm having trouble with this, figuring this out, the source of this. Can you run that query for me?" So this might be how we get to a bottom up system of intelligence generalizing where in the future we have smaller number of broader platforms.
Dave Vellante
>> You've argued that ... Well, maybe at least I think you've argued that you've got to have a bottom up or some have argued that you've got to have a bottom up view of the world that top down is not going to solve this problem. Let's dig into that a little bit. I tend to be a top down, bottom up, middle out kind of person. But what do you mean by that? And do you think you need a bottoms up? Because the problem with bottom up sometimes is you end up boiling the ocean, it gets too complicated and then things change before you can get your arms around it. And I feel like AWS is definitely bottom up. I feel like Microsoft is more top down and I feel like Celonis is sort of both.
George Gilbert
>> So let me give you like a very concrete pre-historic example of a problem with going pure bottom up. So in the James Michener novel, it was the one about Israel, The Source. They're trying to create a tunnel to a water source so that if they ever get attacked, they have a tunnel. The water source is outside the town. So they start digging from the water source and they start digging from the town and they don't meet. That's the problem with bottom up, because you don't have the top down guidance. So you kind of need both. And the question is, how much? Now, the way there seems to be some consensus where you say, "I want to take one outcome and I want to gather all the data I need to drive that outcomes improvement." So that would be, I'm going to implement the AWS DevOps agent for let's say a specific set of services and I might need data from the Dynatrace system to make sure I get that end-to-end visibility. So make that work, then extend it to the next service. That might be one way of having the top down part is I need this process to work. The bottom up part is stitching the two together.
Dave Vellante
>> All right. Let's talk about Kiro. Kiro was announced this summer at the NYC Summit. And anybody doesn't think that Amazon is in the AI game. I mean, just look at some of the innovations that they're announcing. In fact, you felt like Matt Garman undersold Kiro, the Kiro autonomous agent in his keynote. I've had feedback from some developers that Kiro is actually something that is increasingly more and more important to them. It's beyond vibe coding. What is Kiro? Why is it important in this conversation?
George Gilbert
>> So we've been talking about harnesses, right? We've been talking about how the models are improving really fast and you need the scaffolding to turn it into an agent or an agentic application. And as our colleague, David Floyer says, there's challenges in a third-party keeping up with the evolution of the model by changing the scaffolding. So let me give you an example. GitHub Copilot kicked off the agentic coding or the assisted coding revolution with code complete. It was like hit the tab and it'll finish your line or finish that section of the code. But then Cursor came along when the models were better with different scaffolding. They customized the IDE so that you could chat and tell the agent what you wanted it to do. So it was different scaffolding and it could take longer time horizon tasks and accomplish more. And then Google just introduced something called Antigravity, which was based on Windsurf that they acquired, but this was still vibe coding, Cursor, Windsurf, these are vibe coding. And vibe coding has taken the world by storm, but it's only been in the last 12 months really. But along this theme where the models are advancing really fast, you might have to rethink your scaffolding entirely. What AWS did that was really compelling that we didn't see from anyone else. And that unfortunately, Matt Garman did not articulate, I think very, very clearly, was this is a categorical break from vibe coding. Vibe coding, the artifact is still the code. You live in the code. It just helps you write a ton of code. What's new here is that you start with a requirements document and the agent actually helps you flesh out the requirements document. It tells you, "Hey, this part isn't specific enough."
Dave Vellante
>> Missing out some pieces.
George Gilbert
>> Yeah, exactly. Exactly.
Dave Vellante
>> Here's where this would actually firm it up a little bit.
George Gilbert
>> Yes.
Dave Vellante
>> So bringing in some other best practice and ideation. Yeah.
George Gilbert
>> It's like a pair programmer, but at the ideation level for the requirements doc. The requirements doc feeds, creates a design doc, and so these are specifications. The code is throwaway so that when you want the software to evolve, you don't try and go into the code and maintain it and figure out what's going to break. You iterate on the requirements and the design doc and you regenerate the code. And what's even more astonishing is that their transform product, which is for migration, they can take a legacy code base and re-engineer or reverse engineer a design doc from that. They think they can also generate a specification doc, a requirements doc. In other words, they can take now ... The big problem has been all these vibe coding things have been for greenfield, but how do you upgrade brownfield, your legacy stuff? It looks like you might be able to take legacy code and turn it into a spec, which then is iteratable and maintainable. But the point is, the new artifact is the spec, not the code. It's like you just generate the component.
Dave Vellante
>> And that's why when they announced this last summer, they said ... They gave kind of, I don't want to say lip service, but they gave a nod to vibe coding because it was such a hot new topic, but they said, "But this is more." And they still have not laid that out the way that you just did, so thank you for that. Okay, let's go to the next slide, which is AgentCore. So this is a big deal. AgentCore, when we talk about scaffolding, this is part of that. I mean, you got Bedrock. Bedrock has been growing in prominence and is the sort of fundamental component. It's that middle layer of their three layer stack, but take us through AgentCore. What is AgentCore? What's new in AgentCore? And how does it fit into this discussion?
George Gilbert
>> So AgentCore is the control infrastructure or control platform that manages the behavior of teams or armies of agents and makes sure they behave. It's your governance model. Here, it's your memory, things like code interpreter. These are utilities that you need, the runtime, the observability. And let me address two things, the observability and policy, which is the governance. We're familiar with data platforms. We've been covering the data platform guys for years, but that one is easy because when you govern that, it's like, who can access what data under what conditions? And the simple version is like row based, you might add column based, you might add tags so you can put a little policy around it, but it's still pretty simple. When it's an agent, the agent is taking actions, multiple actions using potentially multiple tools. So the policy needs to be able to say, "Should this agent be allowed to take this action with these parameters in this context, given what it's trying to accomplish?" That's like exponentially more sophisticated, so you need a different sort of policy framework around it. And this is all part of, if I want to put agents in production, I need some sophisticated scaffolding. And that's why this is the heavy enterprise software, deterministic enterprise software that ... Now, in this case, Amazon's providing it that I don't see-
Dave Vellante
>> The LLM vendors delivering.
George Gilbert
>> Yeah.
Dave Vellante
>> We'll bring back Floyer and have that discussion. Okay. Let's move on to probably my view anyway, the most important announcement that I heard this week was Nova Forge. George and I, we've written extensively about how Jamie Dimon is Sam Altman's biggest competitor, why that is the case, how to get there, how to essentially build, fuse that gap and that data gap and build that system of intelligence. Please bring up the next slide. Nova Forge is that sort of glue, if you will, no pun intended with Amazon Glue, but basically your premise, George, that you've put forth is that customers are going to need, enterprises are going to need open weight models, they're going to be to bring in their own data, and that's how they're going to be able to customize their enterprise for their proprietary advantage. That's what Nova Forge is. It's the first open source, open weight with training data available to customers to leverage and apply to their specific enterprise to build worker B agents.
George Gilbert
>> From a major American vendor.
Dave Vellante
>> Yeah.
George Gilbert
>> We talked about this at-
Dave Vellante
>> I don't think DeepSeek does it. There may be some other-
George Gilbert
>> That have the intermediate checkpoints. It may not be DeepSeek, but yes. So there are a lot of startups building on DeepSeek because it's cheaper, but this is different because they're giving you the training data so you can substitute some of your own.
Dave Vellante
>> Yeah, I don't believe DeepSeek gives you the training data.
George Gilbert
>> Yeah, I think you're right, it's just the weights. But here, what's significant is there's a bunch of themes here, which is if you embed yourself and train this model, let's say you do some pre-training, then you do some reinforcement learning at the end to calibrate and elicits just the right behavior, you are doing ... Remember the old Christensen integrated innovation, you are welded into that model. So when the next best one comes out in six months, you have to go through that process all again. And I don't know how difficult that is. I think maybe it's not that the process is difficult, it's that the behavior may change and so you have instability in your system. But the other key point here is someone like Anthropic, which lives on its API business primarily, they'll say, when I went to their booth, they'll say, "Oh, don't get into the reinforcement learning and you don't want to do the continued pre-training." Because then you're so welded in, they'll say, "You get a little less integration if you just do supervised fine-tuning with us, but you ride along the frontier of the frontier LLM because as soon as the next model comes out, you're ready to take advantage of it."
Dave Vellante
>> Okay.
George Gilbert
>> And that's Floyer's point of view.
Dave Vellante
>> Well, and he feels like that the open AIs of the world and Anthropic will be the next great software companies that will build software on top of the LLMs and make it easier for organizations to integrate their data because it's too complicated to deal with open weights and open weight models and all that complexity.
George Gilbert
>> So I would just push back one thing and say, he made me write for mainstream enterprises, but if ISVs are building applications for the mainstream enterprises and now we go beyond RPA 2, agents as RPA version 2 and they're now microservice version 2, you're probably not going to want just a few frontier models because you're going to want a lot of specialization.
Dave Vellante
>> So if you're not going to want a few frontier models, what are you going to want? You're going to want specialized small language models?
George Gilbert
>> Yes, yes. You will want the frontier models. They'll be the orchestrators. They'll be for the most sophisticated planning, but you're going to want to be able to specialize models.
Dave Vellante
>> Okay. Let's go back to the agenda if we can and take a look at the next section. We're going to look at this power shift from GPU rich to infrastructure smart. All right, let's dig into that. Go to the next slide, if you would. We're on slide 14, powering through. Okay, so let's get into it. So you're talking here about ... Well, it's interesting. So Matt Garman, he's showing us the section of his keynote where he said NVIDIA is the best place to run ... Or AWS is the best place to run NVIDIA GPUs, which is, again, ironic because people think, "Oh, AWS, they don't have their allocation and just all kinds of politics going around there. And they've got their own Trainium. You hear about Trainium, you hear about TPUs. Microsoft has its own." CNBC is all up in arms because they're saying, "Oh, there's all this competition to NVIDIA." What's the point of all this? We know GPUs are expensive. My take, Floyer's kind of educated me on this is that NVIDIA's got the volume. They're going to have the best performance per watt. Volume drives learning curve and they're going to be in the driver's seat unless it's theirs to lose. What's your take on all this?
George Gilbert
>> Well, I would add to that on that perspective, they also have locked up the leading process nodes at TSMC, the most capacity. So it's like Apple was 15 years ago. No phone maker could compete with them because they locked up all the hardware for that generation for each year.
Dave Vellante
>> And I am too, happy that Intel is now at least on a path to maybe sustain profitability, but just the volume's coming out of their fab and the ability to support third-party customers in the fab is still to be proven.
George Gilbert
>> Yeah. So there's another perspective on this. Well, let's deal first with why AWS was not the volume customer for NVIDIA, even if they wanted to be originally.
Dave Vellante
>> Okay. Why is that?
George Gilbert
>> We went first to Microsoft and neoclouds and maybe Google and Meta because AWS was always in the infrastructure smart. They designed their data centers with their Nitro sort of networking and data processing unit a lot to build a composable data center, more instance types with more storage and networking flexibility and crucially the software, the hypervisor to be able to work with all that. And we covered that.
Dave Vellante
>> I mean, that was a secret weapon in the last five, seven years.
George Gilbert
>> Yes, exactly. But when NVIDIA came along and when there was this new version of the WinTel moment when ChatGPT came along and it ran on NVIDIA. Jensen only wanted to sell the full data center kit. He didn't want to sell just a rack.
Dave Vellante
>> Yeah, he had the hardware, he had the software, he had the tooling, all the networking, everything.
George Gilbert
>> So the customers who were not infrastructure smart, who got caught flat-footed, they had to buy everything from NVIDIA, so they got bigger allocations. Now there was a wrinkle.
Dave Vellante
>> Like the neoclouds.
George Gilbert
>> The neoclouds, Microsoft. Oracle did it actually not because they were behind so much as they never re-architected the database, so they needed a scale up architecture to get scale out of their database. The bigger issue was something no one talked about for the last year. It was like this dirty little secret, which was-
Dave Vellante
>> What's that? We got a CUBE exclusive here.
George Gilbert
>> The volume GPU for the last 12 months has been the GB200, the Grace Blackwell 200. So it's got the Grace CPU and the Blackwell. And the failure rates on these have been 50%.
Dave Vellante
>> Five-zero?
George Gilbert
>> Five-zero. I got that from two sources.
Dave Vellante
>> Okay. And nobody's talking about it because?
George Gilbert
>> They're terrified of making Jensen look bad and then getting their allocations cut.
Dave Vellante
>> Okay. So everybody talks about the energy problem and the lack of data center, space, power, water, cooling. There's another piece that nobody's talking about, which is the failure rates are very high.
George Gilbert
>> It has big implications on the business side, which was, this was the year CapEx just went through the roof and Wall Street was like, "Where's the money?"
Dave Vellante
>> Where's the return?
George Gilbert
>> Right. Because 50% of them weren't working. So that was the problem.
Dave Vellante
>> So they're underutilized for reasons that most people didn't understand?
George Gilbert
>> Okay. Two implications. One, the GB300s passed the GB200s in volume in October.
Dave Vellante
>> So bring that slide back up, the earlier one, the GPU slide. Thank you.
George Gilbert
>> The new generation, they're on a one-year refresh cycle. Supposedly the new generation, it appears to install quickly and have a much higher quality.
Dave Vellante
>> We were talking to Lambda about the GB300s, they were thrilled.
George Gilbert
>> They were.
Dave Vellante
>> Yes.
George Gilbert
>> Okay.
Dave Vellante
>> Saying they were running great. And so again, Lambda, the neocloud, they probably get some allocation because they're willing to buy the whole stack. I don't know that for a fact, but that would make some sense. And they're GPU specialists. Okay.
George Gilbert
>> Because there was this big theme from the guys on the Latent Space podcast who distilled everything, that there was the GPU rich and the GPU poor and the GPU rich were able to obviously do things that the GPU poor scrounging around with no capacity, couldn't even dream about. But now the changes to the infrastructure smart, the guys who really did build their own accelerators and a composable data center architecture because like there's enough Trainium out there in the Amazon environment that you can run a frontier model like Anthropic on Trainium for inference. Now, I'm not saying it's the best place to train Anthropic. Both sides have been saying, "Oh, they're building us a half million chip training facility." But Dario let it slip once in an interview, I think with The Economist that, well, it would have taken only a fourth as many NVIDIA-
Dave Vellante
>> Amazon made a big investment in Anthropic and one of the conditions was you're going to train on Trainium because they're going to help each other and they're going to integrate more tightly. So that's understandable.
George Gilbert
>> But then he went and signed an even bigger deal with Google for training on the TPUs. And now, Google is selling TPUs apparently to Meta and possibly others.
Dave Vellante
>> Again, but Google's volume is never going to be where NVIDIA is unless NVIDIA trips. I mean, NVIDIA is selling to Google's competitors, it's selling to the neoclouds. Are the neoclouds going to buy from TPUs and risk losing their allocation?
George Gilbert
>> Not the neoclouds.
Dave Vellante
>> Not likely. Is Amazon going to buy TPUs? Is Microsoft going to buy TPUs?
George Gilbert
>> No.
Dave Vellante
>> No. And so again, volume is everything in semiconductors. And so I think this whole TPU thing is just way overblown.
George Gilbert
>> But by having a credible competitor, it may create pricing pressure because the buyer can say, "I have an alternative."
Dave Vellante
>> Well, 70% margins will probably be under pressure. Okay. But there's still the software, the tooling, the libraries, the ecosystem, I mean, all that. Let's move on, because we're out of time here, but let's go to the next slide. This is the scaffolding. So you have the messiahs, him on the left chasing AGI. On the right side, you have all this wonderful scaffolding. Explain this and put it into context.
George Gilbert
>> Okay. So this is just me having a little fun with the frontier. Well, particularly Sam, because Dario doesn't think he's Messiah and doesn't think he's coming down the mountain with the tablets and everything. I mean, I'm mixing metaphors. I think that was Moses, not Jesus.
Dave Vellante
>> Okay. Yeah.
George Gilbert
>> But the point is-
Dave Vellante
>> Both Jews.
George Gilbert
>> Okay. The bigger thing is that on the right side, it's let's just build practical stuff. And ironically, that's what China is saying. "Look, let's let them chase AGI." The practical stuff. And that's all that we've been saying, which is the ISV community, the hyperscaler community, they're building the hard enterprise software scaffolding at the agent layer and at the data layer. And let me bring those two together. We talked about AgentCore because that's the agent control layer and the agent governance layer. But the other scaffolding is when you bring the data and the action space together. That sounds abstract. The actions are the tools when you turn your processes into workflows that are callable tools and then your data is your analytic data, that's the context that tells the agent what to do. When they come together, they come together in a knowledge graph. That's when it's a system of intelligence. And so that's the other scaffolding. So when you have both in place, that is hard enterprise software. And these LLM guys, these are not DBMS guys.
Dave Vellante
>> And when the companies that I talk to that are ... Amazon itself, not necessarily Amazon Web Services, but Amazon.com, companies like big banks, JPMorgan, other big banks, technology companies like Dell, when you dig in and ask them about that, what we call the system of intelligence or the knowledge graph, they are having to build that themselves because it doesn't exist out of the box today. I go back to one of our first pieces on these topics was Uber for All. We did this stuff with Madrona. We did some work with Muglia where the whole notion of people, places, things, activities, Uber, drivers, riders, activities, prices, locations, et cetera, all come together in a digital representation of an enterprise, Uber for all being a solution that comes essentially out of the box, an easier to deploy solution that gives me that 4D map of my enterprise. That's what ultimately has to get built out and why we say this agentic era is going to take the better part of a decade to really gestate and unfold and have impact beyond. It was sort of having that conversation with Mike Gannon at Snowflake. He didn't say it's here today, but he's like, "No, no, we're doing a lot today, but it's within the analytics sphere. It's not across the entire operational state and processes of the enterprise. When we say it's going to take the better part of a decade, that's what we mean to really have that real time digital representation of the enterprise, that digital twin, as you like to call it."
So, okay, let's go to the next slide. This one cracks me up. We got the camel with fleas on it and you got all these logos, which I presume you got more of them than fleas and a camel's back. Explain this one, George. I'm having more fun here
George Gilbert
>> All right, this is the point that everyone ... There was this scene in the Pixar movie Up when the dogs are in their ... The plane is chasing the house, I think, trying to shoot it down, and all they had to do was yell squirrel, and the dogs just went like crazy. All anyone had to do was yell agent and VCs came running.
Dave Vellante
>> Yeah, I was going to say, Larry Ellison would have a field day with this one.
George Gilbert
>> Yeah.
Dave Vellante
>> Remember his rap on cloud?
George Gilbert
>> Yes. It's water vapor. The point here is, in the agent, there is the model and there is the scaffolding and there's how much scaffolding. It ranges. Like Harvey has knowledge of the legal workflows, but most of these are very, very thin scaffolding compared to what we were just talking about with the agent control layer and the system of intelligence. And the point is, a lot of these agents, there's so many of them because they're not doing the hard work of building that scaffolding. The big software companies are building the scaffolding, and I think that's where value's going to lie. And that's why I say this, you don't differentiate your agent development tool by some little wizzy feature. It's integration with the data in the action space and the agent control apparatus. And that's why I say the rest of these guys, there's more of them than their fleas on the average camel.
Dave Vellante
>> All right. So let's close with the next slide because this brings it all home and shows us what this future looks like and where some of these players fit. As we say, that high value piece of real estate is in that center, that platform, that's with a system of intelligence. Companies like Celonis, like Palantir, probably two of the leaders in this space taking different approaches. Palantir obviously very service is heavy, but doing some amazing work with software. Our AI, other graph database companies, clearly ServiceNow has designs there, SAP, Salesforce. Virtually any major SaaS company with process logic and business knowledge is going after that space.
George Gilbert
>> They need to build that data platform, but there's also that one other thing that the business model problem, which is that as they put more of the process logic with the data and you can do more with agents, then they lose the seat subscription. And that's the business model complex.
Dave Vellante
>> They've got an innovator's dilemma there. The guys on the left, we've got a $50 billion business that's been built just to sort of present dashboards, which is getting upended and disrupted. Now, they're of course going to try to bring talk to your data into the equation. You've got the data platform guys down there, Snowflake and Databricks. And then beneath that, you've got SageMaker and lakehouse. You got Bedrock up here, which is that system of agency, that agent control framework. Governance is throughout. I mean, all these guys are doing some form of governance. And then, of course, you've got a lot of activity. Last year we heard tons about Q. Today, it's Kiro. You've got the whole software development life cycle changing from one that's sort of linear in the pre-GenAI world to one that's non-linear and interactive. And that's a whole different workflow. So just as every part of the hardware stack is changing, compute storage, networking, every part of the software stack is changing as well. Let's wrap, George. Give us your final thoughts on re:Invent 2025 and this whole move to service as software. It feels like we're taking baby steps to get there.
George Gilbert
>> So big picture, going back to the notion of the messiah AGI, and the consumer agent, and then the enterprise worker bee. I would say I think on the consumer side, we're going to see some of the air come out of it in the sense that the natural ... It's a sustaining innovation, it looks like more and more, in that the Google ecosystem, including Android, are natural surfaces for Google's Now industry leading model. And then as screwed up as Apple's initial response was, now that they're punting and using Google, they have great services to essentially make Siri what the dream always was for Siri.
Dave Vellante
>> With minimal CapEx.
George Gilbert
>> Yes. Exactly. Okay. So then that leaves OpenAI as trying to somehow squeeze in there on the desktop, maybe as a third-party app on the phones. Now, the point of this is that I think that means more focus will now look on the enterprise and distinguish it between two categories. One is this worker bee stuff we've been talking about, but the other is agent enabling both the user interface and the backend logic of every software development tool or every software tool that's out there. So like Kiro that we talked about or security or DevOps, but you do that for every service that every vendor has, and that's just started. So I don't think that the bubble is bursting. I think you're going to see some deflation on the OpenAI side, but I think you're going to see still a furious amount of activity to catch up on those two categories of the enterprise. It's going to be difficult to do the enterprise worker bee AGI beyond the little RPA 2 to start with and microservice 2, but there's going to be a huge productivity increase in all software development tools.
Dave Vellante
>> Okay. And then ultimately, that migrates into virtually all aspects, all processes within the organization and confers that marginal economics advantage. Speaking of bubble, I don't know if you heard Dalio the other day, it was a couple weeks ago, a week ago, I can't remember, on TV, crashed the market talking about where 80% of the way into the bubble needs to be a prick. I think he was probably just trying to help his gold trade, but I don't know. It feels kind of bubblicious, but he's right in that you need to prick the bubble in order for it to burst and we're not there yet.
George Gilbert
>> Well, what could be the prick, like historically like a geopolitical event or tighter money? Yeah.
Dave Vellante
>> Well, and you recall during the dot-com, we saw many 10% pullbacks and I don't know how many, but it was more than zero. And so we just kind of had one, not quite 10%, but some disruptions. All right, George. Great work. Thank you. Amazing. Actually, you got here on Tuesday?
George Gilbert
>> Yeah.
Dave Vellante
>> You got here Tuesday. You did a lot of research in a day and a half, so thanks for taking all the time, putting together these great slides. Really appreciate it.
George Gilbert
>> All right. Thanks, Dave.
Dave Vellante
>> Thank you for watching. This wraps up re:Invent 2025, and thanks for watching this Breaking Analysis. We'll see you next time.
Breaking Analysis | Worker Bee AGI - Why AWS Is Betting on Practical Agents, Not Messiah AGI
search
Dave Vellante
>> At AWS re:Invent 2025, Amazon was tasked with serving millions of its legacy customers and speaking to them while at the same time fighting the perception that they're behind in AI. And what AWS is doing is taking a practical approach, not chasing the holy grail of what we call messiah AGI or even frontier LLMs. Rather, what AWS is doing is they're focusing on foundational agentic scaffolding and customizable large language models or small language models consistent with our view that the tangible opportunity in AI is within the enterprise, i.e. worker bee AGI, not messiah AGI. Now, skeptics might say that worker bee AGI is just a form of RPA, let's call it RPA 2.0 because the data is still siloed. While this is true, we don't see this as paving the cow path, rather we see AWS making a move from siloed agentic automation to what we call service as software. Welcome to this special Breaking Analysis. My name is Dave Vellante. I'm here with George Gilbert, and we're going to break down AWS re:Invent 2025 and share with you our takeaways from re:Invent specifically in the context of the theme that we've been hitting on Breaking Analysis around service as software, which we see as a complete transformation in the operational, the business and technology models of virtually every company and organization and every industry. George, thanks for all the time you put in to making some cool slides. We've got a number of them and we've got a lot to cover. So appreciate you being here.
George Gilbert
>> Always a pleasure.
Dave Vellante
>> Okay. So let's bring up the sort of title slide, the highlights from frontier labs to messiah AGI to enterprise worker bee AGI. I love this, George. You got comrade Altman on the left here. You've got the communist arm band. He's chipping away, trying to get to AGI and get expertise from gig workers. Meanwhile, you've got Matt Garman on the right-hand side, shaking hands with Jamie Dimon. We've written a piece around why Jamie Dimon is Sam Altman's biggest competitor, how he gets there. So take us through these themes. Why did you choose to revisit the comrade Altman? And let's talk about where Amazon fits.
George Gilbert
>> Okay. So basically, the OpenAI business and ChatGPT, the product, they led the first leg of the GenAI revolution. And through GPT-4, maybe even GPT-5, that was all built on extracting free data from the internet. In other words, even though their compute costs were high, the costs of the data that fed the intelligence was essentially zero or very low. And as the high quality free data is running out, they are spending more and more now on curating expert sort of reasoning traces. The expertise from very specialized people. And the key point is that this is not really scalable. It's an entirely different business model. And there's been very credible reporting from the information that even in 2029 or 2030, they expect to spend somewhere between 20% and 25% of their revenue on curating this proprietary data. And the point we're trying to make here is like the picture of Altman is he's digging ever deeper into the seams of the Earth trying to pull some coal out in every smaller amounts.
Dave Vellante
>> Kind of chipping away and it's hard knocks.
George Gilbert
>> Yeah. And it's sort of declining marginal utility. Whereas up on the surface of the Earth, we have Jamie Dimon as a proxy for the leading enterprise. All his resources are renewable. It's his proprietary data. And the more agents he gets up and running, the more models he gets running, the more data he generates. He's on an experience curve. So the more he harmonizes and hooks the data and the models together, he's on this flywheel. And so his marginal costs decline. So Sam Altman get his marginal costs increase and they get more specialized utility. And then the Jamie Dimon as a proxy for all enterprises, they're on this learning curve that's accelerating going down.
Dave Vellante
>> Okay. So why the windmills?
George Gilbert
>> It's renewable. Windmills and solar farms.
Dave Vellante
>> Yeah. Solar panels and windmills.
George Gilbert
>> The point is, in the age of AI, it's about the data. It's not the algorithms. The algorithms are done by the labs. And so how you shape your data to train the AI is everything.
Dave Vellante
>> Now, as you and I have talked about, David Floyer has this belief that OpenAI ultimately ... He already realizes that the opportunity is in the enterprise and they will build software on top of their LLM. He agrees with you. The LLM is not really the issue. It's the software on top that will enable enterprises to actually do some of the hard things that today companies like Amazon and JPMorgan Chase and big companies like we've talked to Dell are doing internally, but they will make it much more simple for mainstream enterprises to do that. And they will go after that opportunity. You believe, correct me if I'm wrong, that that's going to be very difficult for them to do because they don't have access to the data. So my question to you would be, well, couldn't they get access to the data by partnering up with the Jamie Dimons of the world?
George Gilbert
>> Theoretically, I would have two challenges with that. And there's a third issue that David's mentioned that I'll go through. So there's three issues. One, will the corporations be comfortable parting with their most sensitive data with frontier lab that can they set up claim rooms and establish trust? Because this is the crown jewels of-
Dave Vellante
>> They have to do that or else it's a fail. We agree on that.
George Gilbert
>> Right. But the way their models and their customization work, the company has to transfer the data to the model lab because with open models, you can bring the weights and a partially trained model onto your premises or into your VPC and no one else sees it. So that's one issue. So let's say they can establish trust. Then the other is this notion of the scaffolding around the agent that integrates it into the application and the workflow that's relevant to your business. This is the mapping of the new, adaptive, probabilistic model as agent and the older deterministic software that was essentially your application, but now retooled to be able to work agents.
Dave Vellante
>> Okay.
George Gilbert
>> And you look at OpenAI introduced two agents like a GUI low-code version and a programmer first version. And they were very competitive tools, but they acknowledged at their developer conference that it didn't match up to even Databricks' developer tools. And Databricks itself doesn't have ... They have the developer tool, the agent tool scaffolding, but they don't have the data scaffolding, which you and I have talked about and we'll get into the system of intelligence. Databricks aspires to it. The point I'm trying to make is there is so much traditional enterprise software that has to turn into a scaffold to make the agent useful and that these frontier lab companies are not traditional enterprise software companies. And to build that scaffolding is essentially to recreate almost decades of enterprise software. And so I think it's more likely that this is a sustaining innovation technically for the enterprise software incumbents and where it's a disruptive innovation is in the pricing model, where they have to go from seat based pricing and for the infrastructure from consumption based pricing, but especially the seat based pricing, they might lose that and go into something more value based, outcome based, usage based.
Dave Vellante
>> Okay. So let's get into that, because we got a lot of ground to cover. We're just getting through the title slide. You bring up the next slide, Anderson, which is the agenda here. We're going to talk about how the business model, operational model, technology model are changing. And we think it's profound and it has impacts for every company that software like economics now accrue to virtually all companies and it's a winner take most market. We'll get into mapping services software to the specific AWS software stack. So some of the takeaways from the innovations we heard this week at re:Invent and we'll talk a little bit about the whole GPU mania that's going on and some of the industry implications. So if you would go to the next slide, Anderson, let's set this up. So we've talked about this earlier in some of our earlier Breaking Analysis, that everything has to change. Customers have to completely transform their operational model and their business model. George, we saw this when we went from on-prem software to SaaS, but it really only affected the IT departments or-
George Gilbert
>> And the vendors....
Dave Vellante
>> technology companies.
George Gilbert
>> Yes.
Dave Vellante
>> Exactly. It really didn't make that much of a difference to the mainstream organizations out there. The other thing that I'll mention here, and we're going to separate, we're showing here the operational and the business pieces, the technology model we're going to deal with separately. What's interesting here is that normally, practitioners would tell you, "Well, the technology, that's not the hard part. It's the people and process. It's the operational. It's the business model. Those are the really hard things." Not only are they hard here, but the technology is hard here. So fill us in on your thoughts.
George Gilbert
>> Okay. So even before we get to the technology, let's talk about the operational model is for 60 years, we've had many technology transitions, but fundamentally we were still building silos. We might have replaced the UI, we might have had a new device, we might have had a new operating model for the software, but the operational model was ever more silos. The change is now that we're building, we're going from these craft silos, like craft manufacturing when you had these work cells, we have knowledge work cells now because that's all we know how to automate and we're going to have an end to end built order assembly line for knowledge work. That's a profound process change, because the org chart that we have now, which is designed like to make a machine efficient, doesn't work when you're doing end-to-end outcomes. So that's the process change, the operational model change. The business model change just really quickly is you're now essentially charging like time and materials. It's almost like fee-based and it's cost plus. When you're offering outcomes, what happens is you can charge either by outcome or by value or something that's a proxy to that. And this is the biggest change, not just the pricing model, but we've always seen financial capital supply supporting physical capital that depreciated and you had economies of scale and so you had marginal cost advantage because you could amortize that fixed capital investment over many more units.
Dave Vellante
>> Non-recurring engineering expense upfront. A little bit of maintenance ongoing, but your effective marginal costs went down close to zero. Not with cloud. You had cloud COGS, but as Satya says, back in the days of CD-ROM went down to basically the cost of a CD-ROM, but now it's different.
George Gilbert
>> Now, there's profound difference and I don't think people appreciate this.
Dave Vellante
>> The tokens. You're paying for tokens, right?
George Gilbert
>> That's the cost side, but now your capital is the intelligence and the expertise that's digital and that compounds, and it compounds exponentially. So that the more data you get from experience, the richer your experience base is. So you get an advantage both on cost and on differentiation. And that's why you get winner take most economics.
Dave Vellante
>> Okay. So the software company economics are changing. They got to play not only cloud costs, but they got to pay token costs. The consumers presumably are going to benefit from ongoing improvements in GPU, price performance, et cetera. And the marginal economics that we used to see accrue to confer to software companies will confer advantage to the organizations that get on the learning curve of AI soonest. That's what you're saying.
George Gilbert
>> It's the difference between a depreciating asset, whether it's a physical asset or a process design, or now it's compounding expertise. And so the notion of get big, get fast, it's get volume, get fast, and then no one can catch you.
Dave Vellante
>> Okay. So now, we're going to talk a little bit about the technology model. Actually, before we go there, pricing has to change as well. And you sort of a little alluded to that before. Well, in some cases it was a perpetual license model or a seat based model to a consumption model in SaaS. We still have a lot of seat based models in SaaS, but you saw the pricing models change when we went to SaaS. That has to change again. You referred earlier to outcome based. I was talking about tokens before. You're probably not going to price on tokens. That's not going to be the ideal. You're going to price on outcomes, but your unit economics are going to be determined, at least in part, by your cloud cost, your COGS, and the token costs. So that's going to set the baseline. That's the breaking point.
George Gilbert
>> This gets back to the difference between models and agents. So the model guys are in this token grind where through distillation and advances in the frontier of intelligence, their costs are coming down something like 10x to 30x per year, some obscene number when it's just the raw model and the API. But when it's an agent and there's a learning loop in there, their prices are much more stable. And so that gets back to what we're going to talk about, which is the scaffolding around it to build that agent.
Dave Vellante
>> Right. Okay. Let's go on to the technology model and talk about what the transformations are there. Thank you, Anderson. So you're showing here that we shifted from these applications or we're shifting from these application centric silos to this data centric platform. So this necessitates the breaking of those silos, something that we've been talking about for years and years and years. How do you see that playing out? Why is this technology transformation so much more potentially complicated than say the shift to cloud?
George Gilbert
>> Okay. Because the shift to cloud was just a different operational model for the same silos. We broke things up into microservices, but we were still siloed.
Dave Vellante
>> From a data standpoint?
George Gilbert
>> Yeah, from a data standpoint and actually from the application logic itself. What we're doing now for the first time, we turned data into a corporate asset with the rise of the relational database. Now, where we need agents and humans to have a shared understanding of the business, we need a system of intelligence where the processes, the rules about how the business runs, they become a shared asset. That we've never done. And so it turns on its head, 60 years of investment. And how we do that, you know I've been working on this for months, like a migration path for how we might do this, that's going to be really difficult.
Dave Vellante
>> So why do you have in this slide Tower of Babbel, what are you inferring?
George Gilbert
>> That's the legacy version.
Dave Vellante
>> If you bring that back up, Anderson.
George Gilbert
>> Because every siloed app has its own language and so they don't talk. And if you want a picture of what's going on in the business, now you build a lakehouse, but even within the lakehouse you have silos because you might have-
Dave Vellante
>> Schemas, different things.
George Gilbert
>> Yeah, different schemas for sales, for service, for logistics. And if you want to ask questions across them to say what happened and why, it's very difficult. You need to go stitching bridges together with data engineering.
Dave Vellante
>> Okay, great. So bring up the next slide, you're going to invoke Bill Gates here.
George Gilbert
>> Well, this is the-
Dave Vellante
>> What is this? Kicking whales down the beach.
George Gilbert
>> Dead whales, dead whales. He used to say this actually, when I was at Lotus, about the DOS user interface we had, he was like Lotus and WordPerfect, trying to get to graphical user interface and preserve compatibility is equivalent to kicking dead whales down the beach. You can't do it. I mean, he was kind of taunting us and it was good marketing, but it captured the spirit of the challenge. This is kicking a thousand dead whales down the beach because it's 60 years of investment in silos that we now have to somehow harmonize and no one really has a great solution for that.
Dave Vellante
>> Okay. Let's bring it back the next slide, which is our agenda. We're going to now dive into what this all means from an AWS perspective. We're going to map that. We saw a number of announcements this week around ... A lot of it was making stuff GA, but doubling down on Kiro, a lot of stuff for the software development light cycle, aging core extensions, Nova, a lot of announcements on Nova, Nova Forge, we're going to talk about a lot, and so let's get into it. All right. So here, we're showing a slide, service that software requires customers transform their technology model. This is a slide we've shown before. We're bringing in the Tower of Babbel. We're showing how what we see as the software stack emerging. The green portion of this is the system of intelligence. We see that as the high value piece of real estate. So George, take us through this and then we're going to map AWS on top of that.
George Gilbert
>> Okay. So just really quickly, there's this sense that the data platform is your foundation for building AI based applications, but the data platform, it's not machine or agent readable. And it even has to go through a layer like your BI tool to be human-readable.
Dave Vellante
>> When you say data platform, you're talking about Snowflakes, Databricks, Redshift.
George Gilbert
>> They've done an amazing job of collecting in one platform data from all the different operational applications and the websites and external sources. So you have one data estate, essentially like a logical data estate, but it's not all harmonized.
Dave Vellante
>> Bring that slide back up if you would. So, okay. So the data platform would be the bottom left here.
George Gilbert
>> Yeah, in the orange.
Dave Vellante
>> And you're showing ... Yep. And then we're really good at what happened.
George Gilbert
>> Yes. And it's these snapshots, these two-dimensional snapshots of what happened. And there's really many silos in there, like we were talking about earlier, because there's different schemas for different departments or functions. And so it can answer the question what happened within a fairly narrow scope. And if you do feature engineering and some machine learning, maybe you can say what's likely to happen in a narrow scope.
Dave Vellante
>> Okay.
George Gilbert
>> But what you really want is to build this system of intelligence, the green layer that is 4D map and end to end view across the business. This is what like Palantir is advertising and some of their customers are building it, although it's very expensive because they're sort of doing it by hand. It's not yet repeatable.
Dave Vellante
>> They're forward deployed engineers are making a lot of dough.
George Gilbert
>> 950,000 a year charging them. So you're basically building SAP customer by customer. Each customer is building it again, the equivalent of doing it by-
Dave Vellante
>> Heavy lift, but it's actually high impact.
George Gilbert
>> High impact. Yes, exactly. Now, if they can make that repeatable, that's a different story.
Dave Vellante
>> Okay.
George Gilbert
>> Okay.
Dave Vellante
>> So go ahead. Do you want to finish up on the-
George Gilbert
>> The key point here is the systems of agency. All the agents are not going to be really effective unless they sit on top of that system.
Dave Vellante
>> Sorry, bring that slide back up. The systems agency is the-
George Gilbert
>> Is the yellow.
Dave Vellante
>> Is the yellow portion of this slide. We call it the agent control framework, if you will, which is important, but it doesn't have the intelligence. It gathers its intelligence from the green layer, which is really the centerpiece of this set.
George Gilbert
>> Whether it's an agent or a human, they need to see across the enterprise to know that if I put a stop on this order because I don't have a part, I'm going to annoy a high value customer and that's going to impact a large contract. You need that broader scope and you need to understand it harmonized so that they're all sharing the same meaning.
Dave Vellante
>> Today, humans make that decision.
George Gilbert
>> Right.
Dave Vellante
>> In order for agents to make that decision confidently in a governed manner, you've got to have that 4D map or that digital representation of the enterprise.
George Gilbert
>> And if a human is overseeing the analytics, the human has to be able to go through all the different scenarios and they can't keep mapping all that stuff if they're doing it by hand.
Dave Vellante
>> Right. I mean, you could do what ifs and you could do what ifs very powerfully, but then if you really want to go deep and across the organization, it gets much more challenging.
George Gilbert
>> Right.
Dave Vellante
>> Okay. So today, this week, we're in day four here, we heard a lot from AWS. So let's bring up the map of the AWS into this software stack. I'll just cut to the chase. What's missing is that green layer. It's not obvious. And when you talk to AWS about how they're doing things internally, they essentially, like JPMC, when you talk to them, like Dell, when we've talked to them, they have to build that green layer themselves. It doesn't come out of the box. It's not a solution, but help us map into the AWS into this model.
George Gilbert
>> So where they were very strong, like really strong was ... It looks like Bedrock is becoming more mature. Like last year, I think they were struggling with two and three nines, that's kind of reliability, that sort of thing. That was a sort of red alarm fire to get Bedrock up and running. They needed that abstraction layer above LLMs.
Now, AgentCore is this control framework for building multi-agent systems, and that looks really promising. And we're going to go into that, but even more impressive was that they did these first-party agents like Kiro autonomous agents for coding, and we'll go into this. The DevOps agent, security agent, these look like ... They have to meet customer acceptance, but the story they told around them was very compelling. And again, what was in the middle, the system of intelligence, the context that drives an agent. The agents are very context-sensitive. It's AI. AI is programmed by data, and so you need to shape your data. There's a lot missing there still. And then at the bottom data platform later, there's SageMaker lakehouse, and maybe we'll see Neptune as a-
Dave Vellante
>> So hold on a second. So SageMaker, it's a big emphasis last year, especially too around S3, Iceberg table capabilities. And so S3 is becoming much more than a get put object store.
George Gilbert
>> Yes.
Dave Vellante
>> Yeah. And then Neptune, Neptune is their graph database.
George Gilbert
>> Yes.
Dave Vellante
>> And they don't really talk about their knowledge graph that much. It doesn't get a lot of high visibility in things like that.
George Gilbert
>> This is where like the difference between Microsoft and Amazon is. Microsoft is, they want to think three to five years out where their customers should be, and then they'll announce early versions of products to try and lead them there. And maybe also to, in this case with Fabric IQ, to freeze the Palantir sales momentum. Whereas Amazon is more, "We're just one step ahead of our customers. We're hearing the problems they're working on and we think we can do something a little bit more principled." And so, but you could see right now, they're using Neptune to help customers build these AI workloads, sort of bottom up workload by workload, but you could see a day sometime in the future where Neptune would join S3 buckets, S3 tables as a way to structure the data.
Dave Vellante
>> All right. So bring that up again one more time, Anderson. So what Amazon, our takeaway is Amazon, they're focused on the agent specific tooling. Swami said, "We want to be the best place to build agents." We give them high marks for that. What's missing is that system of intelligence piece. Nobody really has that. You mentioned Palantir, Celonis, but this is the sort of new territory, but bringing together that data and process knowledge. Neptune looks like it potentially could be a linchpin of the future, or they'll do something different, like do something with S3 buckets or, again, maybe follow in Microsoft's footsteps, which is unlikely that Amazon would do something like that. Okay. So now let's dig into some of the pieces, specifically the DevOps agents. Let's bring that up. You've got the software stack of the future and you've got the green arrow pointing to the topology of system intelligence. You got the green box with the arrow. You got the yellow box with the arrow pointing at the DevOps agent. You've got Matt Garman talking about the DevOps agent. Why is this important? How does it fit?
George Gilbert
>> Okay. Think of a DevOps agent as, at the most basic level, you've got this sprawling estate of infrastructure, middleware, application components, the microservices. It's not like SAP where it's this one, a single vendor created end-to-end framework so they can look inside and they know where to look. It's so much more open-ended. So think of the DevOps agents as a way of saying, "I need to look across my digital operations. And when something goes wrong in all these thousands or tens of thousands of sort of components, I need to find out what's the root cause, what caused that?" And then if I can figure it out with high confidence what caused it, then I can put together a remediation plan potentially autonomously when the confidence is high enough. And so the lesson to take away from this is the DevOps is kind of like training wheels for the system of intelligence because this is your digital operations. And the system of intelligence is not just your digital operations, but it's your business operations, all your business operations. So what you need to do here is you need a map of your digital operations. And that's what they talked about with topology, which is they're trying to learn a map of how the components of your Amazon services and infrastructure fit together. Now, I talked to different vendors. I talked to Datadog and Dynatrace and I think one other.
Dave Vellante
>> Splunk?
George Gilbert
>> Splunk and Elastic.
Dave Vellante
>> Okay.
George Gilbert
>> And they all have their own takes on how to do this and they all have different views. And this was maybe the most surprising lesson for how we might build a system of intelligence more generally, which was each has a rich picture of their own domain. And they also have some insight into the other domains. For instance, Dynatrace, their customers put a probe in each host and that gives them really deep visibility and so they can build this causal graph. They can build the digital twin of your digital operations. And so what happens with the DevOps agent from AWS is when it can't figure out, it figures out a problem and it can't find it in its perimeter, it says, I think it goes to the Dynatrace and says, "I'm having trouble with this, figuring this out, the source of this. Can you run that query for me?" So this might be how we get to a bottom up system of intelligence generalizing where in the future we have smaller number of broader platforms.
Dave Vellante
>> You've argued that ... Well, maybe at least I think you've argued that you've got to have a bottom up or some have argued that you've got to have a bottom up view of the world that top down is not going to solve this problem. Let's dig into that a little bit. I tend to be a top down, bottom up, middle out kind of person. But what do you mean by that? And do you think you need a bottoms up? Because the problem with bottom up sometimes is you end up boiling the ocean, it gets too complicated and then things change before you can get your arms around it. And I feel like AWS is definitely bottom up. I feel like Microsoft is more top down and I feel like Celonis is sort of both.
George Gilbert
>> So let me give you like a very concrete pre-historic example of a problem with going pure bottom up. So in the James Michener novel, it was the one about Israel, The Source. They're trying to create a tunnel to a water source so that if they ever get attacked, they have a tunnel. The water source is outside the town. So they start digging from the water source and they start digging from the town and they don't meet. That's the problem with bottom up, because you don't have the top down guidance. So you kind of need both. And the question is, how much? Now, the way there seems to be some consensus where you say, "I want to take one outcome and I want to gather all the data I need to drive that outcomes improvement." So that would be, I'm going to implement the AWS DevOps agent for let's say a specific set of services and I might need data from the Dynatrace system to make sure I get that end-to-end visibility. So make that work, then extend it to the next service. That might be one way of having the top down part is I need this process to work. The bottom up part is stitching the two together.
Dave Vellante
>> All right. Let's talk about Kiro. Kiro was announced this summer at the NYC Summit. And anybody doesn't think that Amazon is in the AI game. I mean, just look at some of the innovations that they're announcing. In fact, you felt like Matt Garman undersold Kiro, the Kiro autonomous agent in his keynote. I've had feedback from some developers that Kiro is actually something that is increasingly more and more important to them. It's beyond vibe coding. What is Kiro? Why is it important in this conversation?
George Gilbert
>> So we've been talking about harnesses, right? We've been talking about how the models are improving really fast and you need the scaffolding to turn it into an agent or an agentic application. And as our colleague, David Floyer says, there's challenges in a third-party keeping up with the evolution of the model by changing the scaffolding. So let me give you an example. GitHub Copilot kicked off the agentic coding or the assisted coding revolution with code complete. It was like hit the tab and it'll finish your line or finish that section of the code. But then Cursor came along when the models were better with different scaffolding. They customized the IDE so that you could chat and tell the agent what you wanted it to do. So it was different scaffolding and it could take longer time horizon tasks and accomplish more. And then Google just introduced something called Antigravity, which was based on Windsurf that they acquired, but this was still vibe coding, Cursor, Windsurf, these are vibe coding. And vibe coding has taken the world by storm, but it's only been in the last 12 months really. But along this theme where the models are advancing really fast, you might have to rethink your scaffolding entirely. What AWS did that was really compelling that we didn't see from anyone else. And that unfortunately, Matt Garman did not articulate, I think very, very clearly, was this is a categorical break from vibe coding. Vibe coding, the artifact is still the code. You live in the code. It just helps you write a ton of code. What's new here is that you start with a requirements document and the agent actually helps you flesh out the requirements document. It tells you, "Hey, this part isn't specific enough."
Dave Vellante
>> Missing out some pieces.
George Gilbert
>> Yeah, exactly. Exactly.
Dave Vellante
>> Here's where this would actually firm it up a little bit.
George Gilbert
>> Yes.
Dave Vellante
>> So bringing in some other best practice and ideation. Yeah.
George Gilbert
>> It's like a pair programmer, but at the ideation level for the requirements doc. The requirements doc feeds, creates a design doc, and so these are specifications. The code is throwaway so that when you want the software to evolve, you don't try and go into the code and maintain it and figure out what's going to break. You iterate on the requirements and the design doc and you regenerate the code. And what's even more astonishing is that their transform product, which is for migration, they can take a legacy code base and re-engineer or reverse engineer a design doc from that. They think they can also generate a specification doc, a requirements doc. In other words, they can take now ... The big problem has been all these vibe coding things have been for greenfield, but how do you upgrade brownfield, your legacy stuff? It looks like you might be able to take legacy code and turn it into a spec, which then is iteratable and maintainable. But the point is, the new artifact is the spec, not the code. It's like you just generate the component.
Dave Vellante
>> And that's why when they announced this last summer, they said ... They gave kind of, I don't want to say lip service, but they gave a nod to vibe coding because it was such a hot new topic, but they said, "But this is more." And they still have not laid that out the way that you just did, so thank you for that. Okay, let's go to the next slide, which is AgentCore. So this is a big deal. AgentCore, when we talk about scaffolding, this is part of that. I mean, you got Bedrock. Bedrock has been growing in prominence and is the sort of fundamental component. It's that middle layer of their three layer stack, but take us through AgentCore. What is AgentCore? What's new in AgentCore? And how does it fit into this discussion?
George Gilbert
>> So AgentCore is the control infrastructure or control platform that manages the behavior of teams or armies of agents and makes sure they behave. It's your governance model. Here, it's your memory, things like code interpreter. These are utilities that you need, the runtime, the observability. And let me address two things, the observability and policy, which is the governance. We're familiar with data platforms. We've been covering the data platform guys for years, but that one is easy because when you govern that, it's like, who can access what data under what conditions? And the simple version is like row based, you might add column based, you might add tags so you can put a little policy around it, but it's still pretty simple. When it's an agent, the agent is taking actions, multiple actions using potentially multiple tools. So the policy needs to be able to say, "Should this agent be allowed to take this action with these parameters in this context, given what it's trying to accomplish?" That's like exponentially more sophisticated, so you need a different sort of policy framework around it. And this is all part of, if I want to put agents in production, I need some sophisticated scaffolding. And that's why this is the heavy enterprise software, deterministic enterprise software that ... Now, in this case, Amazon's providing it that I don't see-
Dave Vellante
>> The LLM vendors delivering.
George Gilbert
>> Yeah.
Dave Vellante
>> We'll bring back Floyer and have that discussion. Okay. Let's move on to probably my view anyway, the most important announcement that I heard this week was Nova Forge. George and I, we've written extensively about how Jamie Dimon is Sam Altman's biggest competitor, why that is the case, how to get there, how to essentially build, fuse that gap and that data gap and build that system of intelligence. Please bring up the next slide. Nova Forge is that sort of glue, if you will, no pun intended with Amazon Glue, but basically your premise, George, that you've put forth is that customers are going to need, enterprises are going to need open weight models, they're going to be to bring in their own data, and that's how they're going to be able to customize their enterprise for their proprietary advantage. That's what Nova Forge is. It's the first open source, open weight with training data available to customers to leverage and apply to their specific enterprise to build worker B agents.
George Gilbert
>> From a major American vendor.
Dave Vellante
>> Yeah.
George Gilbert
>> We talked about this at-
Dave Vellante
>> I don't think DeepSeek does it. There may be some other-
George Gilbert
>> That have the intermediate checkpoints. It may not be DeepSeek, but yes. So there are a lot of startups building on DeepSeek because it's cheaper, but this is different because they're giving you the training data so you can substitute some of your own.
Dave Vellante
>> Yeah, I don't believe DeepSeek gives you the training data.
George Gilbert
>> Yeah, I think you're right, it's just the weights. But here, what's significant is there's a bunch of themes here, which is if you embed yourself and train this model, let's say you do some pre-training, then you do some reinforcement learning at the end to calibrate and elicits just the right behavior, you are doing ... Remember the old Christensen integrated innovation, you are welded into that model. So when the next best one comes out in six months, you have to go through that process all again. And I don't know how difficult that is. I think maybe it's not that the process is difficult, it's that the behavior may change and so you have instability in your system. But the other key point here is someone like Anthropic, which lives on its API business primarily, they'll say, when I went to their booth, they'll say, "Oh, don't get into the reinforcement learning and you don't want to do the continued pre-training." Because then you're so welded in, they'll say, "You get a little less integration if you just do supervised fine-tuning with us, but you ride along the frontier of the frontier LLM because as soon as the next model comes out, you're ready to take advantage of it."
Dave Vellante
>> Okay.
George Gilbert
>> And that's Floyer's point of view.
Dave Vellante
>> Well, and he feels like that the open AIs of the world and Anthropic will be the next great software companies that will build software on top of the LLMs and make it easier for organizations to integrate their data because it's too complicated to deal with open weights and open weight models and all that complexity.
George Gilbert
>> So I would just push back one thing and say, he made me write for mainstream enterprises, but if ISVs are building applications for the mainstream enterprises and now we go beyond RPA 2, agents as RPA version 2 and they're now microservice version 2, you're probably not going to want just a few frontier models because you're going to want a lot of specialization.
Dave Vellante
>> So if you're not going to want a few frontier models, what are you going to want? You're going to want specialized small language models?
George Gilbert
>> Yes, yes. You will want the frontier models. They'll be the orchestrators. They'll be for the most sophisticated planning, but you're going to want to be able to specialize models.
Dave Vellante
>> Okay. Let's go back to the agenda if we can and take a look at the next section. We're going to look at this power shift from GPU rich to infrastructure smart. All right, let's dig into that. Go to the next slide, if you would. We're on slide 14, powering through. Okay, so let's get into it. So you're talking here about ... Well, it's interesting. So Matt Garman, he's showing us the section of his keynote where he said NVIDIA is the best place to run ... Or AWS is the best place to run NVIDIA GPUs, which is, again, ironic because people think, "Oh, AWS, they don't have their allocation and just all kinds of politics going around there. And they've got their own Trainium. You hear about Trainium, you hear about TPUs. Microsoft has its own." CNBC is all up in arms because they're saying, "Oh, there's all this competition to NVIDIA." What's the point of all this? We know GPUs are expensive. My take, Floyer's kind of educated me on this is that NVIDIA's got the volume. They're going to have the best performance per watt. Volume drives learning curve and they're going to be in the driver's seat unless it's theirs to lose. What's your take on all this?
George Gilbert
>> Well, I would add to that on that perspective, they also have locked up the leading process nodes at TSMC, the most capacity. So it's like Apple was 15 years ago. No phone maker could compete with them because they locked up all the hardware for that generation for each year.
Dave Vellante
>> And I am too, happy that Intel is now at least on a path to maybe sustain profitability, but just the volume's coming out of their fab and the ability to support third-party customers in the fab is still to be proven.
George Gilbert
>> Yeah. So there's another perspective on this. Well, let's deal first with why AWS was not the volume customer for NVIDIA, even if they wanted to be originally.
Dave Vellante
>> Okay. Why is that?
George Gilbert
>> We went first to Microsoft and neoclouds and maybe Google and Meta because AWS was always in the infrastructure smart. They designed their data centers with their Nitro sort of networking and data processing unit a lot to build a composable data center, more instance types with more storage and networking flexibility and crucially the software, the hypervisor to be able to work with all that. And we covered that.
Dave Vellante
>> I mean, that was a secret weapon in the last five, seven years.
George Gilbert
>> Yes, exactly. But when NVIDIA came along and when there was this new version of the WinTel moment when ChatGPT came along and it ran on NVIDIA. Jensen only wanted to sell the full data center kit. He didn't want to sell just a rack.
Dave Vellante
>> Yeah, he had the hardware, he had the software, he had the tooling, all the networking, everything.
George Gilbert
>> So the customers who were not infrastructure smart, who got caught flat-footed, they had to buy everything from NVIDIA, so they got bigger allocations. Now there was a wrinkle.
Dave Vellante
>> Like the neoclouds.
George Gilbert
>> The neoclouds, Microsoft. Oracle did it actually not because they were behind so much as they never re-architected the database, so they needed a scale up architecture to get scale out of their database. The bigger issue was something no one talked about for the last year. It was like this dirty little secret, which was-
Dave Vellante
>> What's that? We got a CUBE exclusive here.
George Gilbert
>> The volume GPU for the last 12 months has been the GB200, the Grace Blackwell 200. So it's got the Grace CPU and the Blackwell. And the failure rates on these have been 50%.
Dave Vellante
>> Five-zero?
George Gilbert
>> Five-zero. I got that from two sources.
Dave Vellante
>> Okay. And nobody's talking about it because?
George Gilbert
>> They're terrified of making Jensen look bad and then getting their allocations cut.
Dave Vellante
>> Okay. So everybody talks about the energy problem and the lack of data center, space, power, water, cooling. There's another piece that nobody's talking about, which is the failure rates are very high.
George Gilbert
>> It has big implications on the business side, which was, this was the year CapEx just went through the roof and Wall Street was like, "Where's the money?"
Dave Vellante
>> Where's the return?
George Gilbert
>> Right. Because 50% of them weren't working. So that was the problem.
Dave Vellante
>> So they're underutilized for reasons that most people didn't understand?
George Gilbert
>> Okay. Two implications. One, the GB300s passed the GB200s in volume in October.
Dave Vellante
>> So bring that slide back up, the earlier one, the GPU slide. Thank you.
George Gilbert
>> The new generation, they're on a one-year refresh cycle. Supposedly the new generation, it appears to install quickly and have a much higher quality.
Dave Vellante
>> We were talking to Lambda about the GB300s, they were thrilled.
George Gilbert
>> They were.
Dave Vellante
>> Yes.
George Gilbert
>> Okay.
Dave Vellante
>> Saying they were running great. And so again, Lambda, the neocloud, they probably get some allocation because they're willing to buy the whole stack. I don't know that for a fact, but that would make some sense. And they're GPU specialists. Okay.
George Gilbert
>> Because there was this big theme from the guys on the Latent Space podcast who distilled everything, that there was the GPU rich and the GPU poor and the GPU rich were able to obviously do things that the GPU poor scrounging around with no capacity, couldn't even dream about. But now the changes to the infrastructure smart, the guys who really did build their own accelerators and a composable data center architecture because like there's enough Trainium out there in the Amazon environment that you can run a frontier model like Anthropic on Trainium for inference. Now, I'm not saying it's the best place to train Anthropic. Both sides have been saying, "Oh, they're building us a half million chip training facility." But Dario let it slip once in an interview, I think with The Economist that, well, it would have taken only a fourth as many NVIDIA-
Dave Vellante
>> Amazon made a big investment in Anthropic and one of the conditions was you're going to train on Trainium because they're going to help each other and they're going to integrate more tightly. So that's understandable.
George Gilbert
>> But then he went and signed an even bigger deal with Google for training on the TPUs. And now, Google is selling TPUs apparently to Meta and possibly others.
Dave Vellante
>> Again, but Google's volume is never going to be where NVIDIA is unless NVIDIA trips. I mean, NVIDIA is selling to Google's competitors, it's selling to the neoclouds. Are the neoclouds going to buy from TPUs and risk losing their allocation?
George Gilbert
>> Not the neoclouds.
Dave Vellante
>> Not likely. Is Amazon going to buy TPUs? Is Microsoft going to buy TPUs?
George Gilbert
>> No.
Dave Vellante
>> No. And so again, volume is everything in semiconductors. And so I think this whole TPU thing is just way overblown.
George Gilbert
>> But by having a credible competitor, it may create pricing pressure because the buyer can say, "I have an alternative."
Dave Vellante
>> Well, 70% margins will probably be under pressure. Okay. But there's still the software, the tooling, the libraries, the ecosystem, I mean, all that. Let's move on, because we're out of time here, but let's go to the next slide. This is the scaffolding. So you have the messiahs, him on the left chasing AGI. On the right side, you have all this wonderful scaffolding. Explain this and put it into context.
George Gilbert
>> Okay. So this is just me having a little fun with the frontier. Well, particularly Sam, because Dario doesn't think he's Messiah and doesn't think he's coming down the mountain with the tablets and everything. I mean, I'm mixing metaphors. I think that was Moses, not Jesus.
Dave Vellante
>> Okay. Yeah.
George Gilbert
>> But the point is-
Dave Vellante
>> Both Jews.
George Gilbert
>> Okay. The bigger thing is that on the right side, it's let's just build practical stuff. And ironically, that's what China is saying. "Look, let's let them chase AGI." The practical stuff. And that's all that we've been saying, which is the ISV community, the hyperscaler community, they're building the hard enterprise software scaffolding at the agent layer and at the data layer. And let me bring those two together. We talked about AgentCore because that's the agent control layer and the agent governance layer. But the other scaffolding is when you bring the data and the action space together. That sounds abstract. The actions are the tools when you turn your processes into workflows that are callable tools and then your data is your analytic data, that's the context that tells the agent what to do. When they come together, they come together in a knowledge graph. That's when it's a system of intelligence. And so that's the other scaffolding. So when you have both in place, that is hard enterprise software. And these LLM guys, these are not DBMS guys.
Dave Vellante
>> And when the companies that I talk to that are ... Amazon itself, not necessarily Amazon Web Services, but Amazon.com, companies like big banks, JPMorgan, other big banks, technology companies like Dell, when you dig in and ask them about that, what we call the system of intelligence or the knowledge graph, they are having to build that themselves because it doesn't exist out of the box today. I go back to one of our first pieces on these topics was Uber for All. We did this stuff with Madrona. We did some work with Muglia where the whole notion of people, places, things, activities, Uber, drivers, riders, activities, prices, locations, et cetera, all come together in a digital representation of an enterprise, Uber for all being a solution that comes essentially out of the box, an easier to deploy solution that gives me that 4D map of my enterprise. That's what ultimately has to get built out and why we say this agentic era is going to take the better part of a decade to really gestate and unfold and have impact beyond. It was sort of having that conversation with Mike Gannon at Snowflake. He didn't say it's here today, but he's like, "No, no, we're doing a lot today, but it's within the analytics sphere. It's not across the entire operational state and processes of the enterprise. When we say it's going to take the better part of a decade, that's what we mean to really have that real time digital representation of the enterprise, that digital twin, as you like to call it."
So, okay, let's go to the next slide. This one cracks me up. We got the camel with fleas on it and you got all these logos, which I presume you got more of them than fleas and a camel's back. Explain this one, George. I'm having more fun here
George Gilbert
>> All right, this is the point that everyone ... There was this scene in the Pixar movie Up when the dogs are in their ... The plane is chasing the house, I think, trying to shoot it down, and all they had to do was yell squirrel, and the dogs just went like crazy. All anyone had to do was yell agent and VCs came running.
Dave Vellante
>> Yeah, I was going to say, Larry Ellison would have a field day with this one.
George Gilbert
>> Yeah.
Dave Vellante
>> Remember his rap on cloud?
George Gilbert
>> Yes. It's water vapor. The point here is, in the agent, there is the model and there is the scaffolding and there's how much scaffolding. It ranges. Like Harvey has knowledge of the legal workflows, but most of these are very, very thin scaffolding compared to what we were just talking about with the agent control layer and the system of intelligence. And the point is, a lot of these agents, there's so many of them because they're not doing the hard work of building that scaffolding. The big software companies are building the scaffolding, and I think that's where value's going to lie. And that's why I say this, you don't differentiate your agent development tool by some little wizzy feature. It's integration with the data in the action space and the agent control apparatus. And that's why I say the rest of these guys, there's more of them than their fleas on the average camel.
Dave Vellante
>> All right. So let's close with the next slide because this brings it all home and shows us what this future looks like and where some of these players fit. As we say, that high value piece of real estate is in that center, that platform, that's with a system of intelligence. Companies like Celonis, like Palantir, probably two of the leaders in this space taking different approaches. Palantir obviously very service is heavy, but doing some amazing work with software. Our AI, other graph database companies, clearly ServiceNow has designs there, SAP, Salesforce. Virtually any major SaaS company with process logic and business knowledge is going after that space.
George Gilbert
>> They need to build that data platform, but there's also that one other thing that the business model problem, which is that as they put more of the process logic with the data and you can do more with agents, then they lose the seat subscription. And that's the business model complex.
Dave Vellante
>> They've got an innovator's dilemma there. The guys on the left, we've got a $50 billion business that's been built just to sort of present dashboards, which is getting upended and disrupted. Now, they're of course going to try to bring talk to your data into the equation. You've got the data platform guys down there, Snowflake and Databricks. And then beneath that, you've got SageMaker and lakehouse. You got Bedrock up here, which is that system of agency, that agent control framework. Governance is throughout. I mean, all these guys are doing some form of governance. And then, of course, you've got a lot of activity. Last year we heard tons about Q. Today, it's Kiro. You've got the whole software development life cycle changing from one that's sort of linear in the pre-GenAI world to one that's non-linear and interactive. And that's a whole different workflow. So just as every part of the hardware stack is changing, compute storage, networking, every part of the software stack is changing as well. Let's wrap, George. Give us your final thoughts on re:Invent 2025 and this whole move to service as software. It feels like we're taking baby steps to get there.
George Gilbert
>> So big picture, going back to the notion of the messiah AGI, and the consumer agent, and then the enterprise worker bee. I would say I think on the consumer side, we're going to see some of the air come out of it in the sense that the natural ... It's a sustaining innovation, it looks like more and more, in that the Google ecosystem, including Android, are natural surfaces for Google's Now industry leading model. And then as screwed up as Apple's initial response was, now that they're punting and using Google, they have great services to essentially make Siri what the dream always was for Siri.
Dave Vellante
>> With minimal CapEx.
George Gilbert
>> Yes. Exactly. Okay. So then that leaves OpenAI as trying to somehow squeeze in there on the desktop, maybe as a third-party app on the phones. Now, the point of this is that I think that means more focus will now look on the enterprise and distinguish it between two categories. One is this worker bee stuff we've been talking about, but the other is agent enabling both the user interface and the backend logic of every software development tool or every software tool that's out there. So like Kiro that we talked about or security or DevOps, but you do that for every service that every vendor has, and that's just started. So I don't think that the bubble is bursting. I think you're going to see some deflation on the OpenAI side, but I think you're going to see still a furious amount of activity to catch up on those two categories of the enterprise. It's going to be difficult to do the enterprise worker bee AGI beyond the little RPA 2 to start with and microservice 2, but there's going to be a huge productivity increase in all software development tools.
Dave Vellante
>> Okay. And then ultimately, that migrates into virtually all aspects, all processes within the organization and confers that marginal economics advantage. Speaking of bubble, I don't know if you heard Dalio the other day, it was a couple weeks ago, a week ago, I can't remember, on TV, crashed the market talking about where 80% of the way into the bubble needs to be a prick. I think he was probably just trying to help his gold trade, but I don't know. It feels kind of bubblicious, but he's right in that you need to prick the bubble in order for it to burst and we're not there yet.
George Gilbert
>> Well, what could be the prick, like historically like a geopolitical event or tighter money? Yeah.
Dave Vellante
>> Well, and you recall during the dot-com, we saw many 10% pullbacks and I don't know how many, but it was more than zero. And so we just kind of had one, not quite 10%, but some disruptions. All right, George. Great work. Thank you. Amazing. Actually, you got here on Tuesday?
George Gilbert
>> Yeah.
Dave Vellante
>> You got here Tuesday. You did a lot of research in a day and a half, so thanks for taking all the time, putting together these great slides. Really appreciate it.
George Gilbert
>> All right. Thanks, Dave.
Dave Vellante
>> Thank you for watching. This wraps up re:Invent 2025, and thanks for watching this Breaking Analysis. We'll see you next time.
