Cheaper AI Models Are Here: Can Enterprises Keep Up With AI’s Release Cycle?

0:00It's been another big week for AI, headlined by conversations among world leaders to new cheaper models from OpenAI and Anthropic. Meanwhile, NVIDIA CEO Jensen Wong told Ezra Klein that AI's value will be decided by its adoption across industries. He pushed back on a broad slowdown and argued that developers need to do far more testing and verification before releasing systems they can't safely evaluate. And a specialist model called JEV has captured the attention of innovators everywhere for making fast decisions inside software. And established software companies, meanwhile, are discounting or bundling AI as customers weigh what these tools are actually worth.

0:39Lots to take in. And as always, we have Worldwide Technology Chief Technology Officer Mike Taylor here to help us distill those signals into practical insights. It's Saturday, September 26th, and this is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

1:02Okay, Mike, welcome back. How are you today? I'm doing well. Good to be back. Never a shortage of topics to hit on here, even week to week. Absolutely. Let's start with Jensen Wong's wide-ranging interview with Ezra Klein of the New York Times. Coming off last week's doomsday headlines that we spoke about, Jensen argued frontier tier labs need to devote substantially more effort to verification as their models become widely used products. And he went on to say companies should hold back models they cannot evaluate safely and test new versions before putting them into operation. He also made the case for open weights as a source of control, saying AI's greatest value will come when it spreads through established industries.

1:43Now, taken collectively, those ideas are creating a demanding assignment for the enterprise leader, so to speak. A lab test can test a model generally. A CIO must decide whether an updated model is reliable with their data, their tools, their permissions. An organization as a whole has to decide which parts of its AI infrastructure are valuable enough to control and how to move beyond just isolated pilots. So just curious, from your perspective, what do you make of Jensen's outlook here and what do organizations have to do to build towards that evaluate before deployment type of vision? Yeah, look, I think there's a lot of lessons we can learn from, you know, these outbreaks, whether it's, you know, agent swarms punching out of sandboxes or, you know, other circumstances that we've all been read into.

2:28I think the key takeaway here is the reward system inside these models is such that if you do not put proper parameters around what the limits are and consequences even built into some of those parameters, the agents in a good or a bad way are going to find unique and novel ways to go about solving problems. And, you know, in its worst case, they go outside the bounds or the context that we believe we provided. But I think that, you know, in talking to our data science teams and even some folks within the various organizations that, you know, have been in the news here, there is always an opportunity to improve guardrails and context.

3:09And, you know, the lesson learned, even Brian, this week, I think I mentioned, we hosted a group of CIOs across financial services this week and really strong insights. I'm not going to get into anything too specific. It was a Chatham House Rules sort of event. But, you know, one of the key themes there is that you have these agent swarms that are out there trying to do and accomplish novel things. Last week, we talked about, you know, solving unsolvable or complex math problems. I think we're seeing this in the alpha fold, the evolution of alpha fold and protein development and pharmaceuticals.

3:45There are absolutely use cases here where the parameters that we want these to operate within are more in the category of creativity or free will, right, in terms of where we want them working and thinking. Within the enterprise, I think it's a stark reminder that we need to make sure that if we have circumstances and parameters where it is more in, you know, the create or invent mode, that we're extremely thoughtful about the parameters and the reward systems or punishment systems that we put in place around how these agents are expected to behave. And, you know, every organization is going to have more, call it workflow, agentic sort of circumstances where the parameters are fairly limited.

4:30They're taking on task-based but intelligent work. And then we're going to have these others where it's going out to explore, determine, create, invent. And the context around this, and we'll talk more probably from a security perspective what this means, but the context within the harness of how these models are executed is going to be critically important. I don't want it to, it can't, frankly, scare off customers. It can't scare off how we're thinking about using these because for every good use, there is an alternate universe of, you know, folks that are thinking about how to use these to hurt our organizations and more broadly our society.

5:12I think Jensen's message is lean in, understand that there are ways to control these. And I'm even optimistic on a front here. I don't see where any of the model creators or providers are going to materially slow down. But in the news this week, and I think this was Musk's suggestion a few weeks ago around sort of a governing committee of the frontier model providers themselves, potentially including open source, to come together and establish evaluation criteria that they would use outside of government regulation. So we'll see where that goes, but at least they're coming to the table. Yeah. Speaking of control, Jensen, also in that New York Times podcast, he said open weight models let companies control a critical layer of their AI infrastructure.

6:00NVIDIA has also agreed to acquire Hugging Face, as most everyone knows, where many companies are discovering and working with those types of models. What do organizations actually want to control when they talk about these open weights? Is it the model, the data? Is it more the economics of everything? Yeah, I think the economics are part of it, for sure. I think the other is uncertainty around distillation, how distillation is occurring. You know, we talked last week even about the, you know, the announcement, you know, where they solved the math problem. And all the scientists are saying there's no way you solve that without distillation of the prompts and some of the other ideas we put in there.

6:40Who knows, you know, how that actually played out. But I think that's a real concern for enterprises today. I think the other piece is the cost of these. Now, you know, open weight and open source models, they're not free, right? I mean, you're either, you're getting the distribution, but you got to pay to run it somewhere. And at the end of the day, the infrastructure, the power, the cooling that goes into supporting the performance of the GPUs is a real cost that's there. But at scale, and as we're seeing more agentic deployments, I think the on-prem circumstance is going to become more real.

7:14And I think that our customers, just like they have in the past, they're going to have on-premise, they're going to have cloud, they're going to have Neo partners, and they're going to have some existing and emerging SaaS partners that provide very unique capabilities in terms of the training of their models. But on-prem is going to be a circumstance in the future, no doubt. And that on-prem doesn't just mean a data center. It's going to be edge. We talked about robotics last week in a starker contrast, but, you know, that's going to need more at the edge and more of this, as you've already seen, even down to mobile devices, you know, that we're using here today.

7:49So I think it's a wise investment on the NVIDIA part. We're with the NVIDIA teams this week. They're very bullish on what this is going to mean to customers, the sovereignty of their data, the tokenomics that go with it, and building out and creating an integrated system that can be deployed on-premise and also at the edge, as we've seen with the development of robotics and physical AI. Yeah, certainly a compelling case to be made. I mean, to have more control over the models becoming part of your core infrastructure. But almost on cue, OpenAI and Anthropic release lower cost models making the alternative, you know, the idea of buying access to a model as opposed to, you know, something someone else runs more attractive.

8:31Just for some more context, OpenAI said that its new GPT-6, Sol, and Luna models have API prices 50 % below their predecessor's promotional prices. And Anthropic said Claude Opus 5.5 costs 40 % less than Opus 5 on typical workloads, although its per token prices fell 20%. Now, both companies also make new performance and safety claims along with that. But I'm curious, Mike, how should an enterprise decide which parts of its AI stack it needs to control and which it's comfortable buying as a service in light of these new models coming out? Yeah, I heard this great analogy last week from Jensen about models and how they were going to work.

9:10He was talking about water in comparison to this universal access to tokens that certainly is his vision and I think the reality of the journey that we're on. And, you know, his point was, you know, you use different, even in the world, you use different types of water for different circumstances, right? You know, you're taking a shower versus, you know, making a cup of coffee versus, you know, a filtration system you have at home or about like, you know, there are just a number of different varieties that are going to be used here. And that, I think is, that is the reality of where this stuff is headed.

9:42The tension that we're seeing, and I think some of this is certainly commercially oriented, but the fact that I believe strongly we are either fully or near fully recursive in the training and development of these models. You know, you see that in terms of the speed and then just the increased capabilities that are released with each. Our teams have been hands-on this week with Opus 5.5 and, you know, looking at, I think we said it's 40 % the price of what Fable 5.1 was coming out. And so far, in the context of AI native engineering, our teams have fully switched over and defaulted from any Fable use case that existed prior.

10:27Again, and call it that high-level, you know, creative thinking invention category have shipped over and shifted over to 5.5 in this. And as I think about the tokenomics conversation we and others are having, I think the market forces to drive down the token costs, the efficiency that's happening there, along with increased adoption, to me sort of plays out as a, what is a level state look like in terms of consumption per user as they adopt more. But we just still have a lot more, at least within the enterprise as users that are going to start using these tools and driving growth of token consumption.

11:07So very impressed so far with 5.5, in particular, you know, where we're using Fable to do planning, orchestration, you know, an invention or creation in the AI native engineering use cases. Yeah, you mentioned we already have engineers kind of playing around with these new models, which brings up an interesting sidebar here. Every new release, whether it's OpenAI or Anthropic or Grok or whoever's putting these models out, it gives enterprise teams another model that they have to test before actually putting anything into production. I was reading a forecast analysis, forecast, F-O-R-K-A-S-T, a pronoun here, analysis.

11:45it called that the evaluation tax. The thinking there is that the time and money spent checking for performance, safety, and cost across real workflows. Forecast actually counted five major model releases in just 10 days earlier this month. And researchers behind the Holistic Agent Leaderboard spent about$40 ,000 in API fees just to run a broad comparison of those agents. Now, most enterprises won't replicate that whole study, but every new model raises a familiar question, which is, you know, how do we spend time and money testing all these against workflows we already have, knowing that more model releases are right around the corner?

12:23Yeah, I think, you know, obviously, we need to move quickly here. The other piece that I think the teams have to understand is, and I'm sure most of our listeners do as well, is that, you know, in some cases, these are, particularly in the open source categories, these are models that are trained to be good testers. You know, they're trained to test well. And so I think finding areas and even what our teams went through this week, this was taking the opportunity to point Opus 5.5 at real work that we were doing and pulling other models out in the course of doing that. And by the way, OpenAI is the same.

13:03We use a lot of these different models across OpenAI, Google, and Anthropoc. But you have to put them to real workloads in your environments. The testing, I think, is a good directional indicator. The context windows certainly are important based on the use cases that you're doing. But it's safe to assume that these companies at their maturity today, these releases, they know need to be substantially better. And that's a balance of performance and cost. And I think that track record has proven out, you know, certainly over the last 18 months going back to the middle of last year. Yeah. So whether it's Sol, Luna or Opus, you know, these are giving enterprises more choices at different price points and capability levels.

13:45A new model that actually made a lot of waves this week from type safe AI called JEV, named after William Stanley Jevons, perhaps better known as the namesake of the now widely known Jevons paradox, is make it was making headlines as it's one of the first models designed to return structured choices, scores and probabilities that software can act on rather than, you know, the typical English language or words from, you know, chat GPT and whatnot. The pitch for JEV was speed and lower cost for repeated decisions inside applications and agent workflows. According to Forbes, developers could use JEV to evaluate an agent's proposed actions, determine when a more capable language model is needed, or inspect completed tasks before allowing an automated workflow to proceed.

14:29Does any of the stuff surrounding JEV, maybe not with that specifically, but where the models might be headed, does that change how we might design an enterprise agent or use different models for different judgments within the same workflow? It does. I think it's recognition of a couple of different patterns. I think the first is that, you know, human supervision or even agentic supervision requires an architecture that provides checks and balances. And in some cases, the final option being choices that either a human or a independent agent is going to act upon outside the context of the generation of that recommendation.

15:08In enterprise, we still see, you know, you have a tradeoff to balance of machine speed versus what you would derive as accurate or the best decision based on, frankly, variables that you aren't able to contemplate within the data structures you're providing to your agents today. Some of that may have to do with real-time information. Some of it may be, you know, data maturity within the organization. But in most cases, human supervision is around understanding the quality of the agents, which is, again, back to the context that you've given to them and the data that they have access to. So what's exciting, I think, about this, and whether it's JEV or other tools that, you know, I'm sure are out in the market or will be coming shortly, even from the model providers themselves, is alleviating enterprises from having to build that kind of presentation harness and output of what are the recommended actions, what are the associated risks or associated probabilities of success?

16:11because these are non-deterministic recommendations, you know, in terms of what's coming out. So in the context of human supervised or architectures that have agentic structures of supervision and permission, I think it's highlighting or illustrating a need that's out there in the market. And those that have been moving really fast on this have been required to build these sorts of capabilities themselves. Yeah. Is there a risk out there? It seems like there is to me, but do we risk making these architectures more efficient, but also making it harder to observe and govern in the end state? I think that's going to be the tradeoff ultimately.

16:49I mean, you know, we want, you know, certainly as it relates to serving our customers, protecting our enterprises, we want machine speed anytime we can get it so that we can respond and react as quickly as possible. The trade-off in that circumstance is that you may be making decisions or doing things that you would otherwise want undone or changed back and adjusted in short order. And I think that's a bit of the next horizon here. So there's a recommendation engine that's out there. Then in the context of cyber security or customer services, you then have to have agentic responses that are going back to ensure that the change you made had the effect that you wanted it to have.

17:32And that's where I think, aspirationally, we get into this recursive learning sort of thinking, not just around model development, but around agentic response. And in the enterprise and, you know, on the most forward leaning feet that we have, those I think are still being developed and instrumented as a full life cycle of a business process or a business function. But we can't, I think for right now, we can't, we need to assume human supervision, even if it's human supervision and agentic supervision in combination is going to continue to exist. And let's not forget, we have to have these, you know, infinity loop sort of systems that are going back to ensure the changes that we did make are, in fact, having the effect that we intended.

18:19And if not, we're regenerating a whole nother set of recommendations and adaptation based on those observations. It's really, I mean, it's pretty powerful when you start to think about the recursive nature of it and how fast it would be able to respond, in particular when time is of the essence. No, it absolutely is. I mean, if companies are starting to assign different models to different parts of a workflow, that's going to make the economics a little bit more complicated at a time when tokenomics and AI spend is already kind of a murky area, certainly top of mind for a lot of folks. Reporting from the information revealed that established software and cloud vendors like Amazon, Microsoft Workday, and a few others, they're cutting prices, bundling AI into core licenses, and extending free trials to stop customers from redirecting spend to Anthropic and OpenAI.

19:11Now, it sounds like buyers may have more negotiating power right now, but they also face a harder comparison. A familiar software suite with discounted AI or a specialist tool employees prefer or a workflow built using models directly. Mike, I'm curious, you know, is this moment of unusual leverage for enterprise buyers? Is it a moment when it's especially easy to buy overlapping AI or is this good for organizations? Yeah, I think you have two market forces there. I think one, you have the existing SaaS, ERP and other established providers that are working, you know, diligently and at light speed to try to embed AI capabilities on top of the platforms and the data that reside within their systems.

19:53And they have been and will continue to be in a constant state of, you know, competing with one another. You know, the overlap between the established SaaS providers is higher than ever. And it will continue to go in that route because they want to grow their businesses. They want to be more relevant in those spaces. So you have the commercial inevitability of that continuing to happen. Right. The second piece is that so much of this is through the lens of the consumer or the employee. in this case. And, you know, these tools, while very, very similar, it can just become sort of an artistic evaluation of what they're familiar with, what they're comfortable with.

20:35And to some extent, what models are exposed, you know, through there, the more flexibility, the better. But those two pressures to me are deriving both internally for worldwide and how we've adopted these capabilities. But what I see within our customers is more and more tension and overlap of, you know, in the midst of sort of this creativity and this fever pitch of things going on, they have more tools than they have, but they're tolerating that a bit because the creativity, the learning is worth it at this point, you know, to go in there and do that. But there's going to have to be some rationalization, right, to this.

21:11The other, I'd say, piece that's been weaponized in this, and again, weaponized good way, bad way, you have some of these products that are licensed in a per seat, all you can use, and you have others that are on a per-use basis. And the variability of the per-use basis is not like what we're accustomed to, at least in large enterprises, in terms of variability of expense. We lock into a cloud contract or something. We're really sort of committing to an amount of spend over a period of time. I think we mentioned this last episode, the variability of spend could vary widely without some thresholds that are established.

21:46And again, back to the harnesses and the guardrails that are put up, you run the risk of these things spinning too much. Yeah. And on top of that, I mean, these contracts and infrastructure choices, these can last years. And this week, world leaders heard from a lot of tech leaders at the United Nations where they were briefed on safety concerns and calls for regulation. A lot of stuff we heard last week. Meanwhile, President Trump met with China President Xi Jinping, who called for cooperation in keeping AI under human control. Also worth mentioning, President Trump is now calling AI superintelligence and an apparent rebrand of sorts.

22:22And then just fresh off the presses, an appeals court ruled the Trump administration can ban or upheld the decision to ban Anthropic from all Pentagon contracts for the time being. Now, you know, putting aside any type of real advancements in policy, what do you think the ideal outcome from all of this is? What's the right path forward that we should be looking at to know that we're actually making sense of what's transpiring on a global stage? I believe strongly that we have to have some level of consistent, and I'm going to call it evaluation, not regulation, but evaluation of the parameters by which models can work.

22:58And the more global that could be, the better for it. I still believe strongly most of our enterprise-related or, you know, business user customers, they're going to want some third party between an open source model and their consumption of it. And I think that's why, in my opinion, that's why NVIDIA did what it did. It's trying to, you know, well, it will integrate this into their overall software stack. And then at the same time, they're going to offer a third-party vetted, verified, or inspected version of those models that are out there. So that, to me, is happening. It's already in motion.

23:45The part that I think is buried even in some of the headlines that you offered there is, you know, we, at this event we had last week and then even doing some more research on my own, you know, The prevailing theory on the open source models is that they're going to be at or near where the top frontier models are today. They're going to be there in a year. And so, by the way, I think our imagination only gets better in consuming the frontier stuff as it continues to get better. We're going to continue to use more of it. But as we think about the use case of open within the enterprise, I think that there's going to be demand there again for running and operating components of business that we have out there today.

24:31But that is in the context of productivity, growing everything else. In the context of cyber, I think it means a whole bunch of other things. And, you know, moving a lot faster beyond just inspection of code and going through and scanning code, but looking at the perimeter of your organization. You know, most cases, more than half of security incidents have something to do with a device that is poorly configured, misconfigured and vulnerable as much as it does a zero day that rolls out on it. And so I think instrumenting these models in an open source way to address some of those things is where I would be encouraging our customers to be sprinting towards because the window is going to continue to shrink.

25:20Like there are already way more advanced and persistent capabilities out there. We see it, our customers see it, but we need to respond quickly. And I think the distribution of open source models through major providers like NVIDIA is going to be something that's going to drive a lot of commercial interest. Yeah, lots to play out for sure. And we'll do our best to stay on top of it and we'll bring it up to you next week when we meet again. Mike, thanks for joining. As always, enjoy the rest of your weekend. Yeah, you do the same. Thank you.