OpenAI, Meta and Microsoft Race to Put AI to Work Inside Your Organization

0:00This week, the AI industry made a bigger pitch for a place inside your business. OpenAI introduced agents designed to keep working on your behalf. Meta announced a new enterprise platform, and companies across the industry are building ways to connect those agents to business data and control what they do once they get there. The capabilities are expanding, and so are the decisions facing enterprise leaders. What work do you delegate? What does it cost? And how do you know you're still in control? As always, we've got Mike Taylor, WWT's Chief Technology Officer, here to help us make sense of all of it.

0:31It's Saturday, October 3rd, and this is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

0:46Okay, Mike, this week, President Trump and leaders from major AI companies signed a voluntary safety accord at the White House. The commitments include internal controls, oversight, independent external assessments, and board review. The agreement really puts the responsibility on the companies developing these technologies and not necessarily a new enforceable regulatory regime. And so for enterprise customers, the question is what those commitments will produce that they can actually use when assessing a supplier or model. So, you know, I'm just curious from your vantage, what should we make of this always-on stream of news coming out of the White House or really any, you know, government or regulatory leaders that relates to policy and regulation here?

1:27Yeah, I think first I'm very enthusiastic about the speed at which a lot of this is moving. I mean, you look at the leaders who attended that, you think about the conversations they had. And I think while they made some amendments to what was the input and the output, directionally, the things that matter there are you have the right people in the room. I think you have an administration that at least is looking at this from a standpoint of what is the right amount of oversight to be applied to this. It's different. And I think just going through the same motions maybe that regulation has in the past is not the right answer here.

2:05And I appreciate that at least there's some consideration for that as we sit here today. So the speed, the people who are in the room, the attention that all this is getting. I mean, if we took a step back, I can't think of any other – call it kind of circumstance or transition where government speed and private participation and alliance is moving quite as well as it is today. That's a good thing. I think you also – we look at this – in the midst of this, the alliance for critical infrastructure was launched as well with a great deal of private partnership involved in this. So, yes, I think we're being led to a structure maybe from either direct government influence or just natural courses that we think make sense.

2:51But the speed at which all this is happening, I think, is unique enough to call out, as it may have happened in other circumstances, a lot slower. Even when our customers were moving to the cloud, cloud regulation was about two or two and a half years behind not the invention of cloud, but the actual adoption of it. So I'm optimistic on that front. And you have the right leaders. You have leaders who are stepping up, not just Jensen, but in the critical infrastructure piece, you look at the participants in that and the individuals who are leading within the tech and other critical areas. These are big name heavy hitters that carry a lot of not just influence, but I think also call it a balance of common sense as it relates to how these things are going to be deployed.

3:39OpenAI certainly represented in that room. So let's move on to OpenAI's Developer Day, which held the industry's collective attention earlier this week. The company introduced DOTS, persistent agents designed to take on work that continues over time using connected applications and their own cloud computers. It also expanded Codex's cloud development and code review capabilities and added computer use to its managed agents API. Then there was Decisions API that lets developers give a model a defined set of possible answers to classify an input, route a request, or choose an agent's next action.

4:12And OpenAI also announced additional privacy protections under what it calls private intelligence. Some of these capabilities are available now. Others are in limited preview or planned for later this year. Taken together, the pitch looks to be that AI can assume more responsibility for getting work done. Mike, which of these announcements, if any, changes how an enterprise shapes its AI strategy or how it thinks about delegating to AI that's increasingly autonomous. Yeah, this again for me is positive. And I would put Muse in the same sort of category, although far more on the consumer side, you know, probably as we think about the use there.

4:47But for us, we think about this, this is an evolution of what we would call kind of the claw strategies, right? And, you know, you had a year or so ago, you had, you know, open claw and these delegated tasks that were primarily function and operating on your computer. For most folks, that meant you had to have your computer on. You had to either be there or have that continue to run. This cloud use is a very interesting concept, I think, that's going to have to be adapted to enterprise cloud tenants and their environments. But nonetheless, this to me is an evolution from initial enterprise response.

5:24We could never do OpenClaw. We could never do claw-like capabilities, issuing passwords or secrets to these. And while these announcements don't completely answer that question, it's moving in the right direction for this. And I think the other piece that is just a tenant of technology adoption in general that's important to track here, consumer use and consumer patterns have impacted the enterprise since the advent of the internet and the iPhone and everything else in between. So the more that people begin using or experimenting with these in their personal lives, I think they're going to – whether their environment is highly restricted and highly regulated or they're a little maybe more open and experimental and forward-leaning, the use of these tools is going to be extremely powerful.

6:09I think there's still a bridge between certainly what Muse is doing. And I even think about everybody gets a dot. If you haven't tried out your dot, everybody gets one. But I still get a little piece of the devil and angel on my shoulder of like, how many passwords do I want to persist? And this is from a personal perspective, right, in this. But I think as we get into enterprise use cases, we'll have more controls. It's going to continue to move. And I look forward to the consumer experimentation with these. I think it's going to drive a lot of creativity in terms of what's there and beyond, you know, the technical pieces that you called out there, the model, you know, the model that GPT-6-1 that was released, our teams are thrilled about the performance of this model.

6:55And, you know, both from a capability standpoint and early testing has been in and around developer use cases, but where we had situations or scenarios in the past where developers were somewhere between the highest end frontier models and even using things like, you know, Fable and Astra, they're seeing, you know, Astra-like and I would say at the same time Fable-like performance from a developer infrastructure operator use case standpoint out of 6.1 right now. And it's 20 % the cost. And so a couple market forces we've talked about in the past, yeah, we're going to continue to drive more adoption.

7:34We're going to continue to drive more use. The cost and the sophistication of these is getting better. The harnesses we're building around them make them more efficient. But it was – I still think we're in – I am. I have to remind myself of the timetables that were on here and whether we've reached full recursive learning within the development of these models or not. They're certainly coming faster and the leaps continue to be better. We don't see – I mean we don't see that stopping. That pattern is going to continue out there. So excited about the evolution of these and really kind of as it relates to the CLAW strategy, what it's going to mean to individual users and the relationship to enterprise.

8:14Yeah, OpenAI, I mean, certainly making a broader claim on enterprise work here. And this week, Meta made it clear that it also wants to have a larger role there. You mentioned Muse, but Meta announced a new enterprise platform business to bring its AI models, agents, and coding tools to, sorry, corporate customers. To lead it, interestingly enough, the company hired the chief executive of MongoDB, who previously held senior roles at ServiceNow and Cloudflare. He'll report directly to Mark Zuckerberg, and this move puts an experienced enterprise software executive in charge of turning Meta's AI technology into products that the business can deploy.

8:49For customers, it probably just adds another major contender to an already crowded field. You know, Mike, what's your read on this? Meta knows how to run AI at an enormous scale, obviously, but how much of that experience transfers to the security, support, and governance requirements that enterprise customers actually need here? Yeah, overall, I'm thrilled that the, you know, some of the biggest, most innovative and best capitalized companies in the world are taking on this space that sits between what we would refer to as, you know, one-off sort of use of AI, as primitive as, you know, prompting to, you know, hey, I've developed one agent that does this for me to the idea that to really build scale.

9:33And I think for a question a lot of our customers give us, Capture ROI is you have to build these things into applied AI solutions. And so everything that they announced, whether they figure it out first, NVIDIA is running at these very same opportunities, competition drives success. I think it drives progress in terms of what we're looking at. And the space between individual agentic development and fully applied AI solutions that contemplate agency, that contemplate security, and that lead to kind of end-to-end applied AI solutions. I want as many innovators running at that as possible. To this point, on behalf of internal to worldwide and our customers, we just had to build a lot of that.

10:16And with the AI coding assistance, it's never been easier to do. But I, I think, along with our customers, we want as many options as possible. Yeah. I mean, so Meta's move underscores a challenge that really every enterprise AI provider has to solve these days, which is how do you take that technology built to serve millions of people and give it an accurate understanding of a company's business? I mean, this is really kind of where the rubber meets the road. It needs to understand how that company defines a customer or calculates revenue or measures performance, any number of other things. And those definitions can differ between departments.

10:50That's the problem behind an industry effort Microsoft recently joined called Apache OSI. It's developing a common format for sharing business definitions across AI, analytics, and business intelligence tools, all in an effort to make enterprise data more useful to AI. Now, this is notable given Microsoft previously restricted partners from connecting competing data tools to the likes of Power BI. Now it's joining an effort to let customers define business context once and then reuse it across different platforms. For enterprise leaders, that could mean more freedom to choose which tools and models work with their data and less effort rebuilding the same business definitions in every system.

11:29You know, Mike, how much does an enterprise's ability to use AI depend on whether its software providers let other tools work with its own data? Yeah, this is going to be a big area of focus for the SaaS providers. And I think, you know, hats off to Microsoft and what they're signaling and communicating here. I think, you know, the mileage varies based on SaaS provider and our customers are learning that as we go here. I still look back at just the history of companies and the development of growth strategies that – and I don't mean open in the sense of open source. I think we have open source present throughout enterprise environments today, but for the most part, it's a distribution of and managed by someone, like someone owning that.

12:18We're going to need more open strategies here in general, and I think those that lean into that signal early are going to get more looks than those that don't. And no great epiphany there. But these machines and kind of operating at machine speed, they're going to require data variables that we're not even thinking of or accustomed to the processes that we manage today. And we spent some time up in the Northeast this week with a handful of CEOs. And what I was encouraged by listening to them is this should start with the examination of the processes we have today. Why do we do this? Why is it done this way?

13:00And if you removed prior constraints and variables like data or data access or the ability to, you know, call it interrogate data for the purposes of outputs, it changes everything. It changes everything about how you think about things. So I think it's an encouraging sign. I think, you know, in a lot of ways, Microsoft is a big SaaS provider today. They've got a number of different tools. You mentioned a couple of them that are there, but they have a huge amount of enterprise data in the unstructured form, some structured in and around their other systems and technology. But we're going to need more access to that, not just with them, but other SaaS providers.

13:36Yeah, I mean, shared data that can create shared definitions, that can help an agent make a better decision, by extension, the enterprise a better decision. But once it's ready to act, a different question really starts to take over, which is, does it have permission, the agents that is? And can we stop that agent from doing something if it goes wrong? Two developments this week, actually, that address that question from slightly different angles. NVIDIA launched its open agent safety platform. Its open-shell software sets a boundary around how agents execute tasks. A separate watchdog design called Sentry monitors their behavior independently.

14:08NVIDIA says it can quarantine an agent that crosses those boundaries in just milliseconds. Meanwhile, Okta and a group of partners are developing a shared architecture for agent security through the Blueprint Alliance. That effort covers identifying agents, limiting their permissions, tracing who delegate at work to them, and monitoring what they do while they're running. Both announcements point to a growing focus on controls that operate outside of any type of model specifically. What are you hearing from leaders as it relates to things like non-human identities and permissions? What should enterprises be able to enforce while the agent is actually out and working on our behalf?

14:46Yeah. So the first conversation we're having with our customers today, and this is at the CTO to CEO level, is whatever you imagine your workforce to be two years ago. It was 50 ,000 employees, whatever it was. Just assume going forward it's going to be a mix of agents and employees, right? And so the crux of that conversation is really today – and I think the frontier providers are starting to understand this. The frontier model providers are in the business of trust. Whether they realize it or not, I think that's come to fruition here over the last two and a half, three months in particular.

15:26We've had other topics we've discussed weekly. You also have OpenAI holding back on their latest Astro release for concerns that they have. Within the workforce in that context of an agent, today they're treated – we're looking at them as they're barely qualified and untrusted. But, you know, and so you think about what you do as a manager or leader or any of you listening. What do you do as a manager or leader to an employee you have that you don't trust? Well, you micromanage the heck out of them probably. Put them on a development plan. I mean, you know, there's a lot you could do there. But fundamentally, we're starting from a basis of no trust.

16:04So what do you need? You have to monitor, right? You have to have signal technology because these things are moving at machine speed for what they're doing. So back to kind of that analogy of what's between a prompt and an enterprise solution. There's a lot of room in there. And a big part of it is how do you evaluate agency and how do you begin establishing trust based on performance? Well, these things are performing potentially thousands, if not millions of tasks and functions a day. The logging of what they're talking to, the success of the actions that they've taken. And then what's still sort of out there, I think, floating around is who's doing the evaluation of kind of pre-agent, post-agent performance within a business function, right, you know, that we had there.

16:50So I look at all that space between this is an important one to fill. How do you log? How do you determine who it's interacting with? I do think back to maybe why Astra 6.1 was held back and certainly what you'd read. How do you also determine what's just regular activity and bad activity? And I go back to you have two circumstances to manage within that. You have the variable of what you know and how you want to manage those things being shaped and managed. And then fortunately in many circumstances, but unfortunately in others, these agents and swarms certainly have the ability to invent new methods and techniques that while we may say something like, you know, don't invent new methods or techniques to go in and do this, or we may say do not access this or do not approach that, how do you ensure the context of that instruction exists and persists within these agents more holistically?

17:47holistically. And that's what these folks are running at. That's what NVIDIA is running at with what they're building out. I think Cisco's taking a look at this from a network layer perspective. How do you look at the bits and bytes of traffic? I think they have a point of view there too in their AI threat defense platform. And you see this with every one of our key suppliers today. But it's going to take a lot of telemetry and the collection of that. And then some ways using AI to evaluate at machine speed how those variables come in and can be evaluated. Yeah, our good friend G2 Patel on a prior episode likened it to a teenager.

18:20So now we have swarms of teenagers running around. That seems scary. We're going to need these controls. And, you know, obviously those controls are going to matter as companies introduce more models into their workflows. Increasingly, they're doing that because the economics demand a closer look. And so along those lines, the Financial Times had an interesting report where more companies are turning to open weight models as they look for ways to manage their AI costs. Specifically in the report, it found that AT &T says it runs about 40 % of its AI workloads on open models with a goal of reaching 70 % within a year.

18:52Now, Tinder is routing some queries to open weight models to help control spending. Didn't think we'd end up talking about Tinder on this podcast, but here we are. No, we're here. And digital realty is using a mix of open and proprietary systems based on sensitivity and complexity of the task. Mike, last week you noted that open models, they're not free, but that the economics become more compelling as a gentic workload scale. A lot of organizations are dealing with this exact scenario. So what would you tell an enterprise that a workload is worth moving or how would you think about that from commercial API to a self-hosted open model once you include the cost of infrastructure, staffing, ongoing operations?

19:27I think I would – some of this is a parallel to cloud and kind of just workloads in general. I think the barrier to deploying – and I think we have to categorize these in two ways. One is are you using open-weight models and then where are you using them? Are they running in a Neo? Are they running in a public cloud? Are you running them on-prem? But if we leave that alone for right now in the interest of time, I would say it is a certainty from my perspective that within an enterprise, you are going to be running anywhere from 10 to 30 different models today. And maybe that even, you know, that list could grow.

20:07And, you know, Jensen had a really interesting comment. He used the analogy, I think it was last week or the week prior to, you know, water. You know, the water I drink may not always be the same water I take a shower with. And he went through this. He was on a roll. It was great, actually. And I think hopefully it's something that, you know, even people outside of technology can relate to in a sense that the open weight – and really when we talk about this, it's, you know, how many parameters do you really need to have that, you know, to make some of these decisions? And then in some senses, how much do you want to either train or coerce those models into preferences that you're going to have, right?

20:42And we used RAG to do some of that stuff. I think at scale now there's a lot of other techniques to do it. when we get into that agentic workflow where it's, you know, steps that culminate into a business process, I'm of the belief that you're not – you don't today need Frontier to do that, right? And by the way, the gap between the open weight, you know, and the Frontier models, that was, you know, a year ago, that was a two-year gap. That's now down to a six-month and you could, you know, you have and others, I'm sure on the pod here, you've read. It's going to – It's not going to be zero, but through some – through techniques and learning and maybe some bad distillation stuff that's going on that we've read about, that window is going to get shorter.

21:29And whether it's a week or a month, I don't know, but it's good enough and it's going to continue to get better. So the reality everybody should have in this is you ought to expect that you're going to have for now maybe five to ten models that you're going to use. Open source is going to be a part of it. And then you're going to get to a scale like an AT &T. Others, you know, you had Eli Lilly on the pod a few weeks ago as well. They've made a big leap into an on-premise because they've got scale. They've got scale to do it. And I think every Fortune 500 to Fortune 2000, as they continue to grow, they're going to have some level of scale there that's going to instigate a conversation on, do we want to continue to consume in the cloud?

22:09Do we want something on-prem? And then obviously between the EU, the US, and other considerations, you're going to have some sovereignty in there as well. Yeah, I mean certainly customers looking for better AI economics to catch up to some value here. At the same time, the companies supplying these systems are facing a more expensive path, a more complex kind of economic landscape. Let's just take Anthropic here. According to IPO documents that were reviewed by Reuters, its revenue grew 12-fold last year to nearly$4.6 billion, but it still recorded an operating loss of more than$8 billion while disclosing that roughly$518 billion in future cloud computing and infrastructure obligations.

22:50That's a substantial bet on future demand here, and the picture really extends beyond anthropic. Axios reported that two Stanford economists estimate nearly a trillion-dollar gap between hyperscaler spending and AI revenue since 2024, billing it, of course, as investment made in anticipation of future demand. If customers are being offered cheaper models while providers are making enormous infrastructure commitments, when does that tension become an operating concern for any enterprise leader and what should they be watching to understand what their exposure might be? Yeah. It's a complicated thing to keep track of as you look at even investments that these organizations are making in one another and some of these buildouts and commitments that they have.

23:37The piece that I look at is we are in the first inning to first half of an inning in terms of demand generation. And I say that from a standpoint of consumer adoption, enterprise adoption, and then horizons or frontier horizons like physical AI, which are just sort of the beginning of all of this. And so I don't have a concern. We are with customers, partners all the time talking about, you know, is there a bubble? Do we see a bubble? I definitely think there'll be companies on the fringes of this that are not going to work out. It's a consequence of any market that's moving at this pace. But the core demand of these things is only going to continue to increase.

24:23And I think some of this will be certainly driven by even the DOT and Muse that we talked about earlier. But enterprises are getting ready. Many of them are already there building these things out. And even those that have large on-premise and sophisticated, they're still doing it in certain areas. It's not ubiquitous. So I think there is a lot of runway here for continued adoption of this. And if I just took 30 seconds and went back five years ago, the concern back then was, oh my gosh, NVIDIA in particular, but include AMD and now Intel and some of this too. They're innovating so fast. The useful life of these things is only going to be 12 months.

25:11And how do you depreciate an asset at this scale? And it's like, well, no, what we found is you've got training, you've got high-end, you've got inference. Most of what's happening today is inference. And those inference compute stacks and data centers, those have a much longer life than people were prognosticating at 8 or 12 months. And so the economics have proven out, at least in the depreciation schedule of these. And the demand curve to me is it's the major lever and there's so much more that is going to be brought into these architectures. Yeah, a lot has changed certainly over the years.

25:48Lots will change over the next coming weeks, months, and years. And that's why we love having you here in the podcast to help keep us informed. Mike, thank you so much for your time as always. Enjoy the rest of your weekend. Yeah, thanks. Take care. All right. Cool.

26:07you