Enterprise comments I get do not now, nor they ever have, validated the notion that cloud-hosted AI of the type that’s the foundation of AI today will transform their business. The great majority point out that to accomplish that goal, they’d need AI agents to operate on core business data that is subject to governance policies that have kept the data out of the cloud, in any form. If this view is correct, it still represents an AI opportunity for model and chip providers, one based on enterprise self-hosting, but not so for the companies investing in AI cloud services—like Microsoft, Meta, Google, and Amazon.
Are these guys then facing an inevitable collapse of their AI dreams? Is there a pathway to having the cloud-hosted model of AI do something transformational? Let’s look at the question first, to assess what would have to happen, and then look at what would be needed to make it happen.
There are three broad technical paths to validate cloud AI. First, use AI to improve ad targeting and raise ad revenues. Second, to somehow make cloud AI play a role in the analysis of core business data despite governance barriers. Third and finally, make cloud AI play a role in the consumer Internet experience. Do any of these paths lead to a credible strategy?
I’m doubtful about the ad sponsorship angle. Yes, it is possible that AI could improve ad targeting, but the four giants I’ve mentioned have invested billions in AI, and that sure looks like an expensive new step in an arms race for revenue that’s limited by the limitations on global ad spending. In any event, it doesn’t seem likely that improvements to targeting would need the level of investment already seen, much less what’s forecast.
The business-focused path would seem to offer some potential. Would it be possible to define a distributed AI model that pushed access to governed data out to enterprise facilities? If the governed component could be trusted, it would seem that this could work, but the architecture would have to be such that any return of data to the cloud would be protected. That suggests that the approach needed would be one where “knowledge” in the cloud would formulate what was essentially a data query that would be transferred to an AI model on premises, and that model would do the analysis. Further work, steps of analysis toward a solution, could be provided along the way to the final results.
What this may involve is a kind of “world model” or “digital twin” that defines a process and acts as a kind of application template. Since the model would describe application elements and information flows, it could be used as the basis for a governance audit, and compliance standards enforced by referencing the element relationships and hosting. A model-builder, which could be any trustworthy organization, would publish models, and enterprises would then shop cloud providers and agent providers for compliant elements, then decide what to use to host them.
The final pathway is the most complex, but we could view it as an application of the same distributed-model approach. A model in this case would be a blueprint for an experience, one that’s created or modified in form by AI. In effect, what is today “content” could become a more personalized/customized experience. You could envision this experience as being entirely created in the cloud (self-generated video or audio), generated with some customization, or perhaps even combining some combination of cloud and customized with real-time video or audio. Sing along with your favorite pop star? Possible.
This sort of thing, the concept of a model as a template for an application/experience, could be extended in a number of ways. First, given that enterprises already see agents as a kind of special case of a software component, you could mix AI with software in general, using the template as a guide. Second, since you need to be able to shop the template fulfillment to a degree, you’d have to assume that the model included the specifications needed for each element, and for the connections. These would not only include requirements but characteristics, which in flow terms would mean QoS. You could, from proper model information, decide where to run something, based on price/performance. If you were a model-builder, you could obviously build models that preferenced your own services, even that required them. The AI giants, of course, have both the skill needed to realize the model approach, and the money.
Some have suggested that this model-driven thing could reshape not only the software market but also the Internet. They see SaaS being translated to a kind of token-as-a-service, which to me means something more like my world-model-as-a-service. They see websites and content being replaced by a delivered AI capability that offers the deal or answers the question. Is that possible?
Maybe. What a distributable world model of applications or experiences demands is a combination of distributed, perhaps even competitive, resource options, or it’s nothing more than a backhand way of letting cloud giants dominate us via AI when they failed to do that with traditional software. Some VCs, in particular, are happy to promote this vision; “services as software over software as a service”. You could achieve some of this perhaps if the model approach gave rise to edge computing, and perhaps some if we assumed that more and more AI was built into premises devices. The IEEE 802.11 initiative to add a kind of AI sharing/hosting to WiFi networks is an example of this. Neither edge nor WiFi/premises advances are a given, though, and even if we had model support for applications that used this combination of resources, there’s still a need for the applications to make a business case.
Anything as transformative as this would demand a pretty significant investment, meaning a significant opportunity for return. The barrier to that is lowered if we assume that AI chips get a lot cheaper, so they can take on missions that can’t support the cost of the top-end GPUs. There are some indications that might come along, but the pace of advance is still an open question. It’s actually easier to envision edge-enhancing developments because most would be driven by business applications that generate a more easily determined ROI.
There are some interesting questions that arise from this all. It’s interesting to wonder how this sort of model might relate to the real-time process model or digital twin. Are the two totally different, or is this model a simple-case example of a more static relationship? Also interesting is the question of how a model that represented a business AI application involving governed data could be made aware of the data. Is that also something that could be modeled? Is a generic model equipped with a template of generic expected data variables, which are then mapped to enterprise databases? Clearly there’s a lot to this that we’ll have to explore over time, if we’re to realize it.
