AI Tech Stack: Build vs Buy
Executive Summary
The marketplace of AI products will offer benefits to enterprises, but the defensible value creation of AI will come from enterprises creating differentiated AI solutions
Every enterprise has an opportunity to leverage AI to create new, valuable solutions by focusing on its unique data, workflows, and institutional expertise
There is no single correct AI architecture; the decision of what to build in-house versus what to leverage from partners is a crucial strategic choice that must align with an enterprise's ambition, scale, and desire for control
The goliaths are likely to flex their scale and take a full stack approach, trading off agility for long-term defensibility
Ultimately, the value is captured at the application layer, where a company’s ability to define valuable use cases and execute determines success
The Defensible Value of AI
The recent breakthroughs in artificial intelligence have led to a fundamental redefinition of what is possible with technology. Yesterday’s limitations have been shattered, and the combination of rapidly advancing capabilities and incredible cost reductions opens a greenfield for new, valuable goods and services.
This is why we’ve seen a cambrian explosion of new AI products. The 2025 Stanford AI Index Report shows that the number of newly funded AI startups has increased from 31 in 2019, to 214 in 20241. This wave of creation isn't limited to startups; tech incumbents are also embedding AI across their entire portfolios. Salesforce, for example, is enabling its customers to build their own AI 'agents' that can autonomously perform complex tasks like analyzing sales data and updating customer relationship management records2.
Thanks for reading! Subscribe for free to receive new posts and support my work.
The marketplace of AI products will transform functions common across industries, from customer support to software engineering. The standardization of these tasks make them ripe for automation and commoditization. It’s important to note that even commoditized products can still be transformational. For example, products for AI for software development can provide immense value to any company with developers.
Enterprises that successfully adopt these products will reap significant benefits in efficiency and productivity. But because these products are easy to adopt, they will become the standard, and not a source of enduring advantage. The true, defensible value of AI will not be found in buying off-the-shelf products, but by leveraging an enterprise’s most unique assets: its proprietary data, workflows, and deep-seated expertise, to create differentiated solutions.
We are already seeing the early innings of this in financial services. Morgan Stanley, in partnership with OpenAI, trained a custom AI on its own vast library of intellectual property, such as market analysis, investment strategies, and advisor notes3. The result is a tool that gives its 16,000 financial advisors instant, curated access to this proprietary knowledge. Bank of America developed its AI-powered virtual assistant, Erica, which taps into the proprietary data of millions of customers to provide personalized financial guidance and proactively monitor spending4. Both are prime examples of building a competitive moat with proprietary AI.
Crucially, this opportunity is not limited to the goliaths. Their scale is certainly an advantage, but it can be overcome with awareness and focus. This is the classic story of how disrupters outcompete incumbents. Every enterprise has an opportunity to leverage AI to create new, valuable solutions.
To do so requires a clear technology strategy. The modern AI stack can be broken down into four key layers: Applications, Platforms, Models, and Infrastructure.
The critical decision every business leader faces is this: what layer do I need to own, and what layer can I buy?
Outsourcing the Hardware Battle
AI infrastructure significantly differs from traditional workloads, utilizing specialized GPUs rather than standard CPU-based servers. Managing GPU hardware introduces considerable complexity distinct from traditional servers, prompting enterprises to either maintain their own GPU clusters or utilize public cloud providers like AWS, Google Cloud, or Azure.
For most enterprises, the cloud is the optimal choice, offering rapid experimentation, value realization, and avoiding the capital-intensive complexity of maintaining GPU clusters. Cloud providers leverage their immense scale to manage hardware competition and negotiate advantageous pricing, insulating enterprises from direct hardware supply challenges.
However, large financial institutions, such as JPMorgan Chase, increasingly adopt a hybrid approach: building in-house GPU capabilities alongside cloud usage5. This strategy enables rapid experimentation in the cloud while maintaining control over mission-critical applications on-premises, providing a strategic hedge against cloud constraints and enabling cost optimization at scale.
Thus, infrastructure strategy aligns closely with enterprise scale and strategic goals:
Cloud First: Ideal for most enterprises due to rapid innovation capabilities and reduced upfront investment
Hybrid: Suitable for large enterprises like JPMorgan Chase, offering strategic flexibility and protection against cloud dependency
On-Premise Only: Rarely adopted except by organizations with extreme security or unique operational needs capable of competing directly against hyperscalers
The Brain of the Operation
The model is the brain of any AI application. The choice of model determines an application's capabilities, its cost, and its underlying risks. For any enterprise, the decision at this layer boils down to three distinct paths: leveraging proprietary models, hosting (and customizing) open-source models, or attempting to build a model from scratch.
Most enterprises leverage proprietary models from leading labs such as OpenAI (GPT series), Google (Gemini), or Anthropic (Claude). These models represent the pareto frontier of capability relative to cost. Typically accessed via cloud platforms (e.g., OpenAI through Azure, Anthropic via AWS or Google Cloud), these providers wrap models with enterprise-level security, compliance, and data privacy controls.
However, leveraging proprietary models involves trade-offs, primarily around data control and strategic dependency. While cloud providers guarantee API data privacy and baseline security, enterprises sacrifice direct observability and granular data oversight for model access. Strategic dependency risks remain, but have significantly decreased as cloud platforms offer multiple substitutable proprietary models through common APIs, reducing vendor lock-in.
Enterprises seeking deeper control often consider open-source alternatives like Meta’s Llama series6. These models enable extensive customization and fine-tuning on proprietary data. Yet, open-source models typically lag behind proprietary counterparts in performance, a gap that has expanded in the last 6 months. Additionally, many US enterprises restrict non-US-based open-source models due to security and compliance concerns, narrowing viable options and emphasizing operational burdens.
Building a language model from scratch, as Bloomberg did with BloombergGPT in 2023, offers deep domain specialization but entails competing directly with rapidly evolving proprietary models from heavily funded AI labs7. BloombergGPT initially excelled at specific financial tasks, but was soon matched or surpassed by more versatile proprietary models.
Ultimately, enterprises must strategically balance the performance and convenience of proprietary models against the control and customization advantages of open source. Most will benefit from strategic flexibility, leveraging customized open-source models but remaining agile enough to adopt proprietary solutions if performance gaps widen significantly.
The AI Factory Floor
If models are the engines of AI, the platform is the factory floor—connecting AI models to value-creating applications. It provides tools, services, and governance for securely building, deploying, and managing AI solutions at scale. Enterprises must strategically choose between building their own platform or leveraging an external one.
Most enterprises will benefit from leveraging external platforms, primarily provided by major cloud providers like Google's Vertex AI8, Microsoft's Azure AI Platform9, and AWS's AI suite10. These platforms compete intensely to become enterprises' default AI infrastructure, offering seamless access to models and leveraging the significant "data gravity" already existing in enterprise environments.
Independent software companies such as Dataiku11, C3 AI12, and Palantir13 offer compelling alternatives, particularly valued for cross-cloud functionality and reduced vendor lock-in risk. However, these cross-cloud platforms may introduce additional complexity, including increased integration challenges and potential latency, compared to more integrated cloud-provider platforms.
Building an in-house platform offers significant strategic advantages: it avoids vendor lock-in, customizes security and governance layers, and creates a standardized development environment accelerating AI application development organization-wide. However, this path involves substantial cost, ongoing maintenance, significant talent demands, and the risk of lagging behind market innovations.
Ultimately, the decision depends on enterprise scale and strategic ambition:
Most enterprises should leverage external platforms, choosing between deepening their relationship with a primary cloud provider or adopting independent multi-cloud platforms
Building in-house is suitable for organizations with significant scale, deep engineering resources, and a long-term vision where AI is central to competitive differentiation, making the platform itself a strategic asset
Where Value Is Captured
The preceding layers of infrastructure, models, and platforms are the foundation, but the Application layer is where business value is truly realized. It is here that technology is forged into differentiated solutions that create competitive advantage. Without a successful application strategy, any AI tech stack is merely a cost center.
Success at this layer hinges on clearly selecting the right use cases and executing effectively. The most promising applications directly leverage an enterprise's unique assets: proprietary data, specialized workflows, and deep institutional expertise. Equally important is understanding the current capabilities and limitations of AI, focusing on solving high-value business problems that competitors cannot easily replicate.
Large enterprises who invest heavily in custom-built stacks and AI infrastructure will inherently spread their resources across a broad portfolio of use cases. While this diversified approach generates extensive learnings, and mitigates risk by not overly relying on any single initiative, it also inevitably dilutes their focus. Each individual use case is one of many use cases competing for the attention of the small pool of AI talent. Without disciplined prioritization, the large enterprises risk overweighing breadth of use cases, over depth.
Conversely, smaller firms and new market entrants must operate under resource constraints that compel intense strategic discipline. Without the luxury of spreading their bets broadly, these companies must meticulously identify the highest-value use cases and channel their investments deeply into them. The resulting laser-like focus not only drives a deeper commitment to each initiative, but also encourages bold experimentation, creating an opportunity for these agile players to disrupt established processes and gain rapid competitive advantage.
The Four AI Tech Stack Strategies
Navigating the AI stack, from infrastructure and models to platforms, reveals there is no single correct choice, but rather a spectrum of strategic trade-offs. The goal is creating differentiated, defensible AI-driven advantages, informed by enterprise ambition, scale, and desired control:
Full-Stack Owner: Pursues maximum control, developing proprietary platforms, customizing open-source models, and managing hybrid infrastructure. Suitable for large enterprises like JPMorgan Chase, this approach offers robust defensibility but demands extensive resources and sustained investment.
Cloud-Native Builder: Builds custom platforms and applications, fully leveraging cloud-based infrastructure. Ideal for tech-forward enterprises aiming to control their AI capabilities with internal talent without managing hardware complexity directly.
Application Innovator: The most common path, leveraging external cloud platforms and proprietary models while focusing internal resources solely on building unique, customer-centric applications. This strategy maximizes agility by using an in-house engineering team that cultivates long-term AI capability and ensures deep alignment with the business context. Building these internal capabilities is essential for making differentiated AI a core pillar of a long-term strategy.
Strategic Outsourcer: For many organizations, partnering with external contractors or specialized AI firms is a pragmatic and effective entry point. This approach provides immediate access to expert talent and can significantly accelerate development timelines, allowing a business to secure quick wins and build crucial momentum without a massive upfront investment in hiring. While this path may involve higher long-term costs and less immediate knowledge retention, it should be viewed as a strategic and foundational step rather than a permanent state. The experience gained is invaluable for providing the insights to gradually build internal capabilities and evolve, using the momentum to transition toward owning the talent and vision for the most critical AI products.
Ultimately, sustained AI success hinges on an enterprise’s ability to identify valuable, defensible use cases and effectively execute. As AI becomes deeply embedded in business operations, developing internal capabilities at the application layer transforms AI from a purchased commodity into a core strategic asset.
References
Stanford University, Human-Centered AI Institute. Artificial Intelligence Index Report 2025. (2025). https://hai.stanford.edu/ai-index/2025-ai-index-report
Salesforce. (2024, October 29). Salesforce’s Agentforce is here: Trusted, autonomous AI agents to tackle specific business challenges. https://www.salesforce.com/news/press-releases/2024/10/29/agentforce-general-availability-announcement/
Morgan Stanley. (2024). AI @ Morgan Stanley Debrief launch press release. https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch
Bank of America. (2024). Erica® – Virtual Financial Assistant Overview. https://info.bankofamerica.com/en/digital-banking/erica
Alves, D. (2025, February 12). How we’re strategically planning infrastructure years ahead to support AI [LinkedIn post]. https://www.linkedin.com/posts/jpmorganchase_our-cio-of-infrastructure-platforms-darrin-activity-7292913658236469248-JXoL
Meta AI. (2024, April 18). Introducing Meta Llama 3: The most capable openly available LLM series. https://ai.meta.com/blog/meta-llama-3/
Wu, S., Irsoy, O., Lu, S., & Mann, G. (2023). BloombergGPT: A large language model for finance (Tech. Rep.). arXiv:2303.17564. https://arxiv.org/abs/2303.17564
Google Cloud. (n.d.). Introduction to Vertex AI. Retrieved June 17, 2025, from https://cloud.google.com/vertex-ai/docs/start/introduction-unified-platform
Microsoft. (n.d.). Azure AI Platform—Cloud AI Platform. Retrieved June 17, 2025, from https://azure.microsoft.com/solutions/ai
Amazon Web Services. (n.d.). Generative AI on AWS. Retrieved June 17, 2025, from https://aws.amazon.com/ai/generative-ai/
Dataiku. (2024). AI ecosystem with Dataiku: Cross-cloud capabilities. https://www.dataiku.com/product/key-capabilities/ai-ecosystem/
C3 AI. (n.d.). Platform independence: Multi-cloud and polyglot deployment. Retrieved June 17, 2025, from https://c3.ai/what-is-enterprise-ai/platform-independence-multi-cloud-and-polyglot-cloud-deployment/
Palantir. (2022). Palantir Edge AI whitepaper: Enabling Interoperability and Preventing Lock-In with Foundry. https://www.palantir.com/assets/xrfr7uokpv1b/7BxLPkTqJU9QhLTQCjJMo6/eed1457949dc2d1cd6b6e71936c0aa9c/Enabling_Interoperability_and_Embracing_Openness_with_Foundry.pdf