“I don't know how anyone can read these numbers as anything but damning.”
Walk into Nestlé’s headquarters in Vevey, Switzerland, and the first thing you notice is the calm: Maybe it’s the light coming through the striking modernist building off Lake Geneva; maybe it’s the fact that everyone The Stack talks to in passing seems to have been working at the company for decades, maybe it’s just because it’s a Friday and half the team at HQ are working from home, but the mood is distinctly sedate.
Scratch the surface though and there is, inevitably, a great deal happening. It would be strange if it weren’t. Nestlé, after all, is the world’s largest food and beverage company, selling coffee (Nescafé, Nespresso, Starbucks Coffee at Home), cereals (Cheerios, Shredded Wheat, Nesquik), and petfood (Purina, Friskies, Tidy Cats) among other well-known logos – thirty of which are “billionaire” brands in their own right.
It sells around one billion products a day; has a footprint in 185 countries, 270,000+ staff and 335 factories in 75 countries. At any given moment its staff are interacting with farmers, with freight providers, with scientists, with a complex distribution network and of with a vast network of retailers and customers globally.
The digital systems that underpin its complex supply chains, factories, and enterprise IT are also hard at work: In 2022 Nestlé made one of the biggest enterprise ERP cloud migrations to Azure in history, spanning 200 SAP ECC instances, 10,000-plus servers, and 1.2 petabytes of data. It is now migrating to S/4HANA; streamlining its “data foundations”; testing agentic workflows; and busy building digital twins of its plants.
A major drive is underway to further optimise its data systems with the aim of driving more innovation, optimising production, distribution and overall efficiency.
As Nestlé CEO Philipp Navratil put in the company’s annual report, the firm is “simplifying our structure, clarifying roles and reducing duplication” across the organisation in a bid to further “digitalize its operations, automate its processes, embed AI and other new technologies, and fully leverage its shared services.”
(These will, he noted, “result in some headcount reductions.”)
The man in charge of much of this digital and data layer is Nestlé’s Group Chief Information & Digital Officer (CIDO) Chris Wright – a Britain-born, Switzerland-based physicist by training, who joined Nestlé in 2004 as a graduate. Wright, who became CIDO in January 2026 (having led IT as CIO since 2021), oversees an IT estate that comprises some 7,000 applications, of which roughly 1,500 are global, and a large cloud footprint. It rolls very little of its own IT infrastructure, bar a data center supporting R&D. His broad IT function employs some 5,000 people at Nestlé.
Most enterprises have learned a hard lesson about infrastructure economics over the past decade: on-demand pricing is an expensive way to run a predictable workload.
Whether that baseline sits in their own data centres or in the public cloud, mature organisations now largely handle it the same way. They right-size it, commit to it through reserved instances, savings plans or owned hardware, and keep the expensive pay-as-you-go bursts for the spikes and the genuinely unpredictable.
The same discipline is now starting to arrive in enterprise AI, and the economics are, if anything, starker. Infrastructure, whether on-premises GPUs, reserved cloud capacity or a neocloud, can be clearly budgeted for. Tokens bought from frontier models typically cannot. An emerging pattern is to run AI workloads on committed infrastructure and use price-per-token frontier models only when the task demands.
Frontier models are the on-demand tier: powerful, elastic and priced by the unit.
“When you pay for infrastructure, it is much more predictable than paying per token” - Nutanix CTO
As AI increasingly moves into production, that unit, the token, is becoming one of the most unpredictable lines in the IT budget; or at least, one of the most predictably expensive ones – something that has given rise to a burgeoning “tokenomics” industry and a recognition that, from model choices to harness optimisation, getting AI right is going to take some refinement and a lot more robust oversight.
(Not least because developers will want to use the most high-powered models and the most expensive. To borrow a metaphor from Nutanix’s Dan Ciruli: “If you tell the developer you can drive the Rolls Royce all the time, they will.” Yet sometimes, of course, you don’t need to be driving the latest V12 “Phantom” to pick up the milk.)
“Earlier AI was an experiment. Now it’s going from experiment to scale,” as Nutanix CTO Manosiz “Mano” Bhattacharyya sums it up for The Stack. And without controls, bills can run into tens or even hundreds of millions of dollars, and they swing wildly.
An application that starts calling a reasoning model, or ingests badly cleaned data, can multiply token consumption overnight. Nutanix, which lets users manage VM-centric or containerized application infrastructure across on-prem, private, or public clouds, is now doing the same for AI workloads. The company recently pivoted its own developer AI workloads to open-weight models running on local GPUs. In doing so, it has cut costs by 80% and the GPUs will pay for themselves in 12 months, according to the company.
FinOps-style cost discipline works only because enterprises built controls around it: visibility into usage, policies on what runs where, and budgets that hold. AI bursting needs the same scaffolding, and it needs it before usage takes off, not after.
That’s the gap the Nutanix Agent Gateway is designed to fill. Now generally available in Nutanix Enterprise AI 2.7, it sits between an organization’s agents and the models and tools they call. Developers get a single API across hosted frontier models and self-hosted private ones, so changing where a request runs doesn’t mean rewriting the application. Platform teams get centralised visibility into token usage across model vendors, cost attribution by team and agent, and token-based rate limits and quotas enforced in one place. For agents that reach into business systems, governance for Model Context Protocol (MCP) servers adds tool-level access controls and a full audit trail.
Governance matters as much as cost. Bhattacharyya has watched the same pattern repeat with each new architecture; shadow IT became “shadow cloud-native” and now “shadow AI,” he warns wryly. A central AI gateway gives developers sanctioned, low-friction access, which removes much of the incentive to route around IT.
Can most AI work run on committed infrastructure? It absolutely can, says Nutanix. Not least due to how agents work in practice. Bhattacharyya explains that an agentic loop typically starts with a planning step, in which the LLM works out what to do, followed by a chain of actions. Planning benefits from frontier-grade reasoning. Actions, which make up roughly 80% of token consumption, generally don’t.
“When it comes to actions, much simpler models can do a very good job.”
And as base model capabilities converge, Bhattacharyya argues, bringing AI closer to the data enterprise applications depend upon is another lever enterprises can pull in terms of latency and cost – whether they’re containerised or running on VMs.
Nutanix’s broader pitch is running VMs and Kubernetes side by side via Nutanix Cloud Clusters (NC2), with storage, networking and policy defined once across cloud or on-prem. “You don’t want separate operational dashboards,” says its CTO.
The same principle applies for generative AI workloads: Get oversight, governance, management in one place via its gateway – whether you’re running open-weight models on your own GPUs, making API calls to a frontier model on a hyperscaler, or a healthy mix of the two. Ultimately, enterprises that learned to stop paying on-demand rates for predictable workloads did so because they could see and control what they were running. AI deserves the same discipline: a predictable baseline on committed infrastructure, frontier models on demand, and a gateway in the middle.
Delivered in partnership with Nutanix.
Ken Beinert is Chief Technology & Operations Officer at London Stock Exchange Group (LSEG), where he leads the development of its technical and operational architecture. The Goldman Sachs veteran joined the global markets infrastructure and data provider in summer 2025 – and is responsible for its AI transformation strategy.
The big focus? “Enabling sustainable growth and stability of our customers and their communities.” Among his tactical challenges: delivering that in an exceptionally fast-moving technology landscape – one where data discipline, innovation, and fiscal precision are all paramount. Beinert joined The Tokenomics Foundation’s Executive Director J.R. Storment to talk AI agents, economics, governance, and more.
The below is from the transcript of their conversation, lightly edited for brevity.
Q: Tell us about your AI strategy
“We think about our AI transformation strategy really in three pillars.
“The first is AI-ready data, where we're focused on making sure that our customers have access to our trusted high quality data in APIs and formats that are extremely accessible to AI. The second is transformative products where we embed AI directly into our existing products and evolve those products so that they're truly AI-native, and the third, of course, is what we call intelligent enterprise, where we're focused on enabling our internal colleagues with the right tools and practices.”
Q: When we think about AI economics, [The Tokenomics Foundation] finds that it's about the total cost of AI, because models and tokens are often a quarter of the spend. When we're talking about AI, what's included in that equation for you?”
15:31 When we think about the investment for an AI solution, I think really about it in two different phases. The first is [to…] focus on the problem that we're trying to solve, secondly on the success metrics that we'll use, the spend, the quality, the user experience are all part of defining the problem and defining the right solution.
We're not just planning for token spend. We're also planning for other infrastructure.. We're planning for data cleansing and preparation to make sure that the data is ready for AI solutions, and then of course we're also planning ahead for real-time guardrails, recurring evals, and comprehensive observability and telemetry, so that we can manage both the cost and the quality of our AI solutions over time.
Q: The State of Tokenomics survey showed that companies are increasingly investing in different model options [and building] governance layers in. How do you ensure that [your teams are] using the right tool for the job?
“One of the factors that I think is most key to model selection is doing comprehensive evals of a model in situ for each solution before we launch it, and then of course continuing to monitor those evals over time, and determining as new models become available, as new solutions become available, whether or not we get better outcomes, whether it is user experience, latency, or cost optimization…
Q: We've got frontier models that everybody's using. We've got open weight models. We've got SLMs. Are you using a mix of those? What does that look like?
We are using a mix of them. You can imagine quite a lot of our customer-facing solutions and our more advanced solutions do leverage frontier models, but increasingly we're finding that many of our solutions can be solved with non-frontier models, smaller models, in some cases custom models. We're beginning to see open weight models that really compete successfully with closed weight models, and that's both at the frontier level of model as well as a smaller level of model…
Q: Let's dig in a little bit about how the variability of AI models is cascading into potentially pricing model changes for you, or maybe just margin management. In the historic capex-heavy world, there was more straight predictability on how much you were going to spend. And with AI usage, both for developer productivity as well as the mission-critical AI, there's a lot less predictability. So, how are you seeing that variability cascade into how you're having to price for your customers?
I think the most important part of pricing AI-based solutions is to make sure we understand the business case upfront, and that we do very thorough evaling for quality and cost ahead of time, so we do understand the unit economics.
The second most important part of that is to bake in the right assumptions into those unit economics, so that we have the opportunity to continue to make them successful over time. Like the rest of the industry, certainly what we're seeing is that prices change very rapidly in multiple dimensions for potential AI solutions, but at the same time, what we're often seeing is that unit economics for an existing solution can improve over time, especially as newer models become available or as we tune our solutions based on our observations through our own telemetry and ongoing evals.
Overall cost, on the other hand, certainly can increase as we see customers and colleagues using those same products even more.
Q: You mentioned unit economics a few times. Is that a granular unit economic, like our cost to serve this customer, or is it a bigger concept?
The unit economics really can vary based on a specific use case. But for me, the example that comes most immediately, most accessibly to mind, is cost per task, where what we really want to understand are the typical tasks that a user uses a given AI solution for, so that we can develop the right evals for each of those tasks to measure both accuracy, quality, latency, feel and experience as well as cost.
Q: How do you enable those trade-offs between cost, value, speed, latency, and are there certain jobs or use cases where you heavily index toward one versus another?
Within LSEG, we have a number of use cases where we use AI for an input to our products, but we [also] have data preparation, data cleansing work that runs perhaps overnight and essentially is not latency sensitive, not part of the direct user experience. There, clearly we have an even greater range of models that we can select from, as opposed to use cases where a customer is engaging directly in an experience with a model. If we're talking again, 12 months from now, one problem that I'd love to see us as an industry solve is the challenge of running agents safely at scale. Today, when we select an agent platform, we're selecting solutions for identity management, entitlements, containerization, runtime infrastructure, sandboxes, and of course observability and telemetry. And each of the solutions that we could use has its own differences, its own potential gaps, which can make it quite hard to switch between platforms and between models and between providers. If we solve this problem, it will really help the industry as a whole to deliver better AI solutions.
Q: How are you driving accountability for AI spend in the firm?
“Well, first and foremost, we believe that the person who owns the P&L for the product also owns the AI spend. That aligns incentives and responsibility.
And secondly, we believe in providing central platforms and controls around AI, both for evals and cost controls, as a joint venture between enterprise AI and finance. And the reason for that is that while the accountability needs to remain with the P&L owner, centrally we have the opportunity to stay abreast of best practices and transformative AI solutions that can be rolled out across these use cases.”
Q: How does Tokenomics differ from FinOps for you?
I see AI economics really, in some ways, as an evolution of cloud FinOps, clearly it is more complex. The pace of change of model selection is much greater, with changes in new types of compute, storage, network infrastructure, etc. There's much more variability in cost, and there's just so many different dimensions of cost. For example, caching and fine tuning… These are things that FinOps didn't have to deal with!
Look, I think that the practices that we have in place for FinOps today, we're going to continue to have all those same practices. But we're going to raise the standard for how we do financial management so that it is more real-time… add an increased element of managing for user experience, because certainly what we find in AI, unlike compute, unlike storage, is that there's a much greater element of variable user
The thing that I'm most excited about is understanding the real potential of AI and putting it in place within our organization, it seems like every day there's new insights into practices that we can apply. Almost every day we're seeing new models that have capabilities in advance of what we expected, but entirely new model paradigms that enable use cases that literally did not seem feasible even a week ago.
Delivered in partnership with the Tokenomics Foundation. To learn more read the 2026 State of Tokenomics report.
