We shared some thoughts last year on the enterprise AI value chain and why it matters. Of course, tech moves pretty fast these days, so it’s time to take a look at how the concept of the AI value stack has evolved. Read on to discover where and how AI creates, captures, and loses value in the current market – and what this value stack means in relation to your API ecosystem.
The shift from tech stack to value chain
The first rush of excitement about what emerging AI can do and how enterprises can integrate it has given way to a more strategic focus on the value that it delivers. Enterprise-grade AI solutions need to prove their mettle when it comes to making businesses more efficient, innovative, automated, and scalable. They also need to show that the investment required to train, enable, and optimize AI delivers sufficient returns.
As such, we’re seeing the emphasis shift away from the capabilities of the tech stack to the AI value chain
Why it matters for business strategy
Any investment in hardware or software needs to align with an organization’s strategic goals. This is why enterprise architecture plays such a foundational role in business strategy and efficiency. A framework for embracing AI is no different – it needs to support the company’s ability to progress and succeed. Embracing the technology for the sake of it isn’t enough; doing so can end up being a costly distraction.
So, how do you explore the potential of AI in a way that aligns with your strategic direction? This is where the AI value chain comes in. Before we look at where that value comes in, it’s important to understand the layered infrastructure that is making the rapid expansion of AI possible.
Layer #1: The foundation – infrastructure and silicon
Underpinning everything is silicon. This metalloid chemical element is used as a semiconductor in everything from computer chips to solar cells, thanks to its ability to control electricity. Global appetite for semiconductors is so intense that it is driving rapid innovation in manufacturing, with one Welsh company successfully trialing manufacturing them in a furnace in space in 2025.
Compute and specialized hardware
The rapid growth of AI has increased demand for the silicon that powers its compute ability. From custom-designed chips and manufactured components to the foundational hardware within AI servers and systems, silicon is at the heart of the technology.
The dominance of GPUs and custom chips
We can see this clearly when we look at leading tech companies. Businesses that design and build silicon-based chips and integrated circuits are the bedrock of the modern economy. An example is NVIDIA, whose H100 and A100 GPUs are silicon chips fabricated on advanced silicon wafers. These are mostly manufactured by TSMC (Taiwan Semiconductor Manufacturing Company), which is the world’s largest and most advanced semiconductor foundry.
AMD’s Instinct MI300 accelerators are another example. Its silicon chips use advanced packaging (chiplets, HBM), making AMD a key AI accelerator industry player.
Other examples of the dominance of silicon in AI infrastructure include Apple’s A-series and M-series systems-on-a-chip (SoCs), including the Neural Engine, which are custom silicon designs, and Google’s tensor processing units (TPUs), which are custom silicon application-specific integrated circuits that are purpose-built for AI workloads. AI chip startups are also bringing innovation and unconventional approaches to the market (Cerebras’ wafer-scale engine springs to mind).
Driving this dynamic sector is the role of compute as the most profitable layer of AI infrastructure. The processing power used to run calculations, compute is crucial to the matrix operations that train and run AI models, typically provided by CPUs, GPUs, or AI accelerators. Capital-intensive, scarce, and tightly coupled to performance, compute power is enabling companies to capture outsized value in the current market. Why? Because while algorithms and data are needed to scale, AI progress ultimately stalls without massive compute.
Physical realities
The rise of AI has done more than level up demand for silicon. The energy sector has also had to step up to meet the needs of this power-hungry new technology, as have data center providers, who are racing to build ever-larger homes for the extensive array of servers and other equipment that AI requires.
Data centers and energy
AI at scale comes with a critical need for massive amounts of energy, cooling, and hyperscale data centers. Examples range from Microsoft’s Fairwater (covering 315 acres in Wisconsin) and Meta’s Hyperion (a four million sq ft campus in Louisiana), to the 10 GW, 5M GPUs-strong Nvidia-OpenAI partnership.
The level of investment needed to create such enormous facilities – and fund the cost of the power it takes to run them – is another factor driving an increased focus on the value that AI delivers.
Layer #2: The cognitive bridge – models and software
On top of the foundational infrastructure layer sits the cognitive bridge – the models and software that turn all that compute power into something businesses can use. This encompasses the algorithms, training techniques, and software abstractions that create business-ready, enterprise-grade AI, creating value through performance, adaptability, and ease of integration.
Foundation models and algorithms
As foundation models and algorithms evolve, we’re seeing a shift away from task-specific systems to reusable, multi-domain intelligence. These large-scale models are trained on massive datasets using extensive compute, maximizing the value extracted from each and every algorithm. They encode general-purpose capabilities (in relation to language, vision, and reasoning) that can be adapted downstream and deliver performance that scales predictably in line with data, parameters, and compute investment.
This shift provides enterprises with greater control over how they grow their AI usage and how much they spend on doing so. An intelligent, interoperable model that scales predictably is easier to integrate, monetize, and optimize, meaning businesses can more accurately analyze the value their AI solutions deliver.
Proprietary vs. open source models
AI, like so many other parts of the tech space, has given rise to both proprietary and open source models. Both can serve as the intelligence engine of the AI value stack, with enterprises choosing their approach according to their use cases and preferences around vendor lock-in (as a quick aside, we run through some of the benefits of open source when talking about Tyk Open Source API Gateway here).
Closed, proprietary ecosystems (such as GPT-4 and Gemini) are usually monetized through APIs and platform lock-in, with providers optimizing them for reliability, scale, and enterprise use cases.
Open-weight AI models (such as Llama) offer greater transparency and customization, enabling experimentation, cost control, and on-prem deployment. This shifts value capture towards fine-tuning, hosting, and integration.
The innovation layer (software)
Software is what converts AI model capabilities into systems that you can deploy and maintain. This is critical for moving AI from experimentation to production in a way that’s achievable, scalable, and valuable.
Frameworks and tooling
Serving as the glue between compute, model, and application, core machine learning (ML) frameworks and tooling abstract hardware complexity, while MLOps tools enable you to manage training, deployment, monitoring, and scaling.
These tools form part of the AI control stack that makes AI usable. Examples include TensorFlow and PyTorch, along with Tyk AI Studio, which enables you to take control of your AI and maximize its value.
Layer #3: The interface – applications and services
Sitting on top of models and software are the interfaces we use to interact with AI – applications and services such as chatbots, copilots, and AI agents. You can use these highly modular elements to turn your AI strategy into tangible, enterprise-level success. They translate model intelligence into workflows and decisions, enabling you to automate and innovate using a data-driven approach. Their modular nature supports rapid iteration and integration across use cases.
The new “front door” of the internet
Users increasing access information and services via an AI interface. This shift from browsing and searching to conversing and instructing directly captures user intent, reducing friction and improving relevance.
Interfaces as the new operating system
This role of AI interfaces as the new front door of the internet comes with a need to orchestrate tools, data sources, and applications behind the scenes. Chat and agent interfaces provide a unified interaction layer on top of this, controlling not just information retrieval but task execution. This positions AI as the primary gatekeeper of user intent, delivering greater value than traditional search engines and browsers.
AI-as-a-service (MaaS)
Using AI-as-a-service means you can access AI capabilities on demand, rather than building them in-house. This lowers barriers to entry for advanced AI adoption, enabling businesses to focus on differentiation rather than infrastructure. You can read more here about choosing the right path to build AI capabilities in your enterprise.
Note that AI-as-a-service is usually abbreviated to model-as-a-service, or MaaS. This not only emphasises the fact that what you’re consuming is a model via an API but also avoids any unfortunate pronunciation of the abbreviation.
Cloud integration and APIs
The ability to consume AI through APIs (like OpenAI or Anthropic), rather than building from scratch, democratizes access to supercomputing power like never before. Added to this, the Arazzo specification provides the ability to refine OpenAI workflows for complex tasks.
Using APIs and cloud-based solutions to consume AI means your enterprise can remain agile and flexible in its approach, powering innovation.
Strategic control points: Where is the value?
As we mentioned above, value tends to concentrate at points of scarcity and control. This means that not all layers capture equal economic returns. Instead, power accrues where there are those who own bottlenecks and/or user relationships.
Identifying the bottlenecks
With AI, bottlenecks most commonly occur at the infrastructure and interface layers of the stack. They shift over time as technology commoditizes, with these bottlenecks frequently impacting both pricing power and margin capture.
The battle for compute vs. customer
There are two main power centers when it comes to bottlenecks: owning the infrastructure (e.g. NVIDIA) vs. owning the user relationship (e.g. ChatGPT or Perplexity).
Compute owners control scarce, capital-intensive infrastructures, capturing value through a combination of performance and comply constraints. Customer owners, meanwhile, control user intent and distribution, capturing value through engagement, data, and switching costs.
The risk of the middle squeeze
These two key advantages – owning the infrastructure and owning the user relationship – mean that companies that simply wrap models risk being crushed between high compute costs and commoditized model access. Their limited differentiation leads to margin compression and, without control of data, users, or infrastructure, their value is at risk of leaking out.
Building defensibility in the AI economy
If your enterprise is seeking long-term success in the AI economy, it’s important to build defensibility into your strategy. That means focusing on sustainable differentiation and ensuring your products offer something better – better performance, better value, better experiences, and so on. Not only that, but they must compound that advantage over time, delivering longevity that promotes ongoing customer loyalty to your brand.
You need to engineer this kind of defensibility and support it to flourish, not assume it will occur naturally. That means focusing on value and scalability as you design and build each platform, process, and product.
The “four Ds” of AI product strategy
This intentional approach to designing and building is crucial to long-term product strategy success. You can use the “four Ds” as you developer your framework, aligning technical capabilities with business value in relation to data, decisioning, design, and deployment. Each dimension reinforces the others, layering strength and resilience into your products – but bear in mind that weakness in one also has the ability to underline the whole system).
Data, decision, design, deployment
When it comes to AI products, focus on:
- Data: Unique, high-quality inputs that are hard for your competitors to replicate.
- Decision: Clear logic for how AI outputs drive actions.
- Design: Intuitive interfaces that embed AI into workflows with user-centric design.
- Deployment: Continuous learning, feedback loops, and improvement in production.
Creating moats and flywheels
You can grow defensibility by compounding effects. Like a flywheel, where gains compound over time, you can create a cycle where compounding creates durability and longevity, making improvement self-reinforcing.
Over time, these compounded effects harden into moats, creating barriers that your competitors can’t easily cross, as they lack the same data, scale, or feedback loops your organization has created.
Data feedback loops
The concept of building robust data feedback loops to create a unique data asset is a simple one. Your user interactions generate proprietary data, which improves model performance and personalization. That drives more usage, which over time becomes hard for competitors to copy.
Vertical integration
Another way to build in value and secure your margins is to use vertical integration. This is where companies aim to reduce depending on external providers and own multiple layers of the stack. Doing so can not only protect margins, it can also improve performance optimization.
Future outlook: The evolving stack
The AI value stack is still in flux and likely to remain so for quite some time. New layers and control points continue to emerge (look out for Tyk’s 2026 LEAP event for more on the AI control stack – details will be on our events page in due course). As capabilities mature, so will advantages and opportunities to capture value, making this a dynamic space to be in over the next few months and years.
Emerging trends
At the time of writing, emerging trends in relation to the AI value stack include intelligence moving closer to users, AI systems becoming more autonomous, and new economic models forming around trust and control.
Of course, AI doesn’t operate in a vacuum, so trends across the tech sector also come into play, from distributed cloud deployment approaches to increased workflow and CI/CD pipeline automation.
Edge AI and autonomous agents
Two other areas to watch in terms of emerging AI value stack trends are edge AI and autonomous agents. Why? Because edge AI is moving inference to local devices, thus reducing latency, cost, and privacy risk, and agents are shifting AI from passive assistance to autonomous task execution. This is reframing interfaces around outcomes instead of prompts, making it a dynamic area of evolution for AI deployments.
Get more from your AI
We’ve pointed out previously that AI readiness starts at the API layer. If your API house isn’t in order, in terms of security and governance, it doesn’t bode well for any AI implementations.
To ensure you lay solid foundations and achieve value as you adopt and scale AI, download our free strategic blueprint for enterprise: AI, APIs, and AI readiness.