AI Workloads: Founders Must Adapt by 2028

The latest projections from Gartner, published in late 2025, confirm what many of us in the industry have already observed: AI workloads are set to double by 2035. This isn’t a speculative trend. It’s a quantifiable trajectory with deep implications for every startup building on or with artificial intelligence. As an investor who has seen countless promising ventures falter not because of product-market fit, but due to an inability to scale their underlying technology, I can tell you this with absolute conviction: your infrastructure strategy today will dictate your survival tomorrow. Many founders, particularly those from a software-first background, mistakenly view infrastructure as a downstream concern, something to be optimized once product-market fit is achieved. This is a fatal flaw in the age of generative AI and complex machine learning models, where computational demands are not linear but exponential.

Consider the typical startup journey: a proof-of-concept built on a single cloud provider, perhaps using readily available GPUs. This works for initial development and early user acquisition. However, as your model matures, your user base expands, and your data volume explodes, that initial setup becomes a straitjacket. I’ve witnessed companies burn through millions in venture capital trying to refactor monolithic architectures, migrate petabytes of data, or untangle vendor-specific integrations that seemed convenient at first. This isn’t just about throwing more money at the problem. It’s about the fundamental design choices made years before that constrain agility and innovation. The notion that you can simply “lift and shift” your way to massive scale is a fantasy. The compounding effect of inefficient resource allocation and technical debt will crush even the most brilliant product ideas.

The Multi-Cloud Imperative for Founders

The single most critical decision founders face regarding their AI infrastructure in the coming decade is embracing a multi-cloud strategy. This isn’t about hedging your bets. It’s about strategic resource optimization and risk mitigation. Relying solely on one major cloud provider (think Amazon Web Services, Microsoft Azure, or Google Cloud Platform) might offer simplicity in the early stages, but it creates deep vendor lock-in that will haunt you as your computational needs diversify. Different AI models, from large language models requiring immense GPU clusters to specialized computer vision tasks benefiting from custom accelerators, perform optimally on different underlying hardware and software stacks. A multi-cloud approach allows you to cherry-pick the best environment for each specific workload, maximizing performance while minimizing cost.

For instance, a startup focused on real-time inference for a personalized recommendation engine might find cost advantages and lower latency by distributing inference across edge locations or specialized regional data centers offered by different providers. Training a foundational model, conversely, might require the sheer scale and specialized hardware available from a hyperscaler with a strong focus on AI research. The flexibility to burst workloads to the most cost-effective or performant cloud at any given moment is invaluable, especially when dealing with unpredictable demand spikes. According to a 2025 report by Gartner, organizations adopting a deliberate multi-cloud strategy for AI reduced their infrastructure spend by an average of 18% over a three-year period compared to single-cloud counterparts. This isn’t just theory. It’s a demonstrable financial advantage that directly impacts your runway and valuation.

Some argue that multi-cloud introduces undue complexity, requiring specialized talent and increased operational overhead. This is a valid concern, but it’s one that can be mitigated with modern infrastructure-as-code tools like Terraform or Ansible, and container orchestration platforms such as Kubernetes. These tools abstract away much of the underlying cloud-specific configurations, allowing teams to define their infrastructure consistently across environments. The initial investment in these tools and the necessary expertise pays dividends by providing a flexible, resilient, and cost-optimized foundation for future growth. The alternative, a single-cloud vendor dictating your pricing and technological options, is far riskier in the long term. You wouldn’t put all your investment capital into a single stock. Why would you do it with your most critical infrastructure?

Modular Design and Data Pipelines: The Unsung Heroes

Beyond the multi-cloud decision, founders must prioritize modular design and strong data pipelines. Your AI infrastructure should not be a monolithic block but a collection of loosely coupled services, each performing a specific function. This architectural philosophy, often associated with microservices, extends directly to how you manage your computational resources, data storage, and model deployment. When your AI workloads double, the ability to independently scale different components of your system becomes paramount. Imagine a scenario where your inference engine experiences a sudden surge in requests, but your data ingestion pipeline is relatively stable. With a modular design, you can dynamically allocate more resources to inference without over-provisioning your entire stack.

This modularity also facilitates experimentation and iteration, which are vital in the fast-paced AI field. You can swap out a specific model serving framework, integrate a new data source, or upgrade a particular hardware accelerator without disrupting the entire system. This agility translates directly into faster development cycles and a quicker response to market demands. I’ve seen too many promising AI products get bogged down by tightly coupled systems where a minor change requires extensive re-testing and redeployment of the entire application. That’s a recipe for stagnation.

Equally important are your data pipelines. AI models are only as good as the data they are trained on, and as workloads double, so too will the volume and velocity of data. Founders need to invest in automated, scalable data ingestion, transformation, and storage solutions from the outset. This means implementing tools for ETL (Extract, Transform, Load) that can handle diverse data sources, ensure data quality, and make data readily accessible to your training and inference systems. Think about data versioning, lineage tracking, and strong monitoring. Without these, your AI efforts will be plagued by “garbage in, garbage out” scenarios, leading to unreliable models and wasted computational resources. A 2024 survey by Pew Research Center highlighted that over 60% of companies struggling with AI implementation cited data quality and pipeline issues as their primary impediment, far outweighing concerns about model complexity itself. This isn’t about having a data lake. It’s about having a well-managed, navigable, and reliable data river feeding your AI systems.

The Human Element: Talent and Governance

While technology forms the backbone, the human element remains indispensable in working through the doubling of AI workloads. Founders must recognize that building and maintaining sophisticated AI infrastructure requires a specialized skill set. The days of a single DevOps engineer handling everything are rapidly fading. You need dedicated talent in areas like MLOps, data engineering, cloud architecture, and even AI ethics. Attracting and retaining this talent in a competitive market is a significant challenge, but it’s a non-negotiable investment. This means fostering a culture of continuous learning, providing access to modern tools, and offering competitive compensation packages. Many startups underestimate the total cost of ownership for AI, often overlooking the significant human capital investment required to keep these complex systems running efficiently and securely.

Beyond technical expertise, governance and ethical considerations are becoming increasingly central to AI infrastructure planning. As AI systems become more pervasive, regulatory scrutiny is intensifying. The European Union’s AI Act, for example, sets stringent requirements for transparency, data quality, and human oversight. Founders need to proactively embed ethical considerations into their infrastructure design. This includes implementing strong auditing capabilities, ensuring data privacy and security (especially with sensitive data), and designing for explainability where possible. Ignoring these aspects is not just a moral failing. It’s a significant business risk that can lead to hefty fines, reputational damage, and loss of user trust. Establishing an internal AI ethics board, even a small advisory group, can help anticipate and mitigate these risks before they become critical issues. This isn’t just about compliance. It’s about building a sustainable, trustworthy AI product that users and regulators can have confidence in.

A common counterargument is that small startups lack the resources for such extensive infrastructure and governance. While understandable, this perspective often misses the opportunity to build foundational elements correctly from the start. It’s far easier and cheaper to implement modular architecture and basic data governance early than to untangle a spaghetti mess years down the line. Many open-source tools and managed services can democratize access to sophisticated infrastructure capabilities, reducing the initial investment barrier. The key is strategic planning and a clear understanding that infrastructure is not a cost center to be minimized, but a strategic asset to be optimized.

The doubling of AI workloads by 2035 is not a distant threat but a present reality that demands immediate and decisive action from founders. Prioritize a multi-cloud strategy, design for modularity, invest in strong data pipelines, and cultivate specialized talent while embedding ethical governance into your core operations. Those who embrace this proactive approach will not just survive. They will define the next generation of AI innovation.