Dominic Jainy is an IT professional whose career has been defined by the rapid convergence of machine learning, blockchain, and high-performance infrastructure. As the industry moves from experimental models to massive, trillion-dollar “AI factories,” Dominic has been at the forefront of analyzing how hardware limitations dictate the speed of global innovation. In this conversation, we explore the explosive growth of compute demand, the shift toward agentic AI systems that operate without human prompts, and the emerging “vibe coding” movement that is opening software development to a new class of non-technical creators.
Summarizing the current landscape, the discussion centers on the unprecedented one-million-fold increase in compute demand, the technical superiority of next-generation platforms like Vera Rubin, and the strategic pivot from training models to continuous inference. We also examine the tension between dominant unified platforms and the rise of custom silicon developed by hyperscalers.
Compute demand has surged by an estimated one million times over the last two years. How is this unprecedented scale reshaping the way enterprises plan their infrastructure, and what specific metrics are you using to determine if current hardware investments can keep pace with this trajectory?
The sheer magnitude of a one-million-fold increase in demand has completely shattered traditional procurement cycles, forcing enterprises to abandon predictable semiconductor schedules in favor of aggressive, continuous scaling. We are seeing a move away from viewing servers as modular units and instead treating the entire data center as a single, massive computer designed to handle a trillion-dollar demand for next-generation systems. To measure success, we no longer just look at clock speeds; we focus on whether compute capacity is acting as a hard ceiling on market growth, as any shortage now directly limits a company’s ability to capture new revenue. It is a high-stakes environment where the $500 billion estimates of yesterday have been doubled to one trillion dollars practically overnight, creating a palpable sense of urgency in every boardroom. This “off the charts” growth means that if you aren’t planning for orders of magnitude more capacity than you currently use, you are essentially planning for obsolescence.
The upcoming Vera Rubin platform aims for five times better inference and significantly stronger training performance. What specific technical bottlenecks does this architecture solve for large-scale “AI factories,” and how will these improvements change the return on investment for companies operating always-on compute environments?
The Vera Rubin architecture is a masterclass in solving the “data movement” problem, which has long been the silent killer of efficiency in large-scale AI factories. By offering five times better inference and 3.5 times stronger training performance, this platform eliminates the traditional lag that occurs when GPUs are forced to wait for data from storage or networking. These systems are designed as unified platforms that integrate CPUs, GPUs, and specialized storage to maximize token throughput and context memory, which are the lifeblood of modern AI. For a company running an always-on environment, this translates to a massive leap in ROI because you can process significantly more productive work without a linear increase in power or physical footprint. We are moving toward a world where “AI factories” function like high-efficiency power plants, churning out intelligence at a scale that was physically impossible just eighteen months ago.
AI is shifting from simple prompts to persistent, agentic systems that read, think, and execute tasks independently. How does this “inference inflection” change the design requirements for modern data centers, and what are the primary challenges in managing the power needed for these continuous workloads?
The arrival of the “inference inflection” marks the moment AI stops being a tool you call upon and starts being a persistent worker that reads, thinks, and executes tasks in the background. This shift requires data centers to move away from burst-heavy configurations to architectures that support continuous, real-time AI processing that never sleeps. We are seeing the introduction of new rack-level systems and CPUs specifically engineered for these agentic workflows, focusing on reducing latency so that the AI can make decisions as fast as a human would. The primary challenge is the relentless power draw of a system that is constantly “on,” requiring cooling and energy management strategies that can handle a workload that never dips. It is a sensory shift in the data center—the hum of the fans is no longer intermittent but a constant, heavy roar that signifies a machine that is perpetually thinking and doing.
Breakthroughs in AI are enabling non-technical users to build applications via “vibe coding.” What are the long-term implications for the traditional software development market, and how should organizations restructure their internal teams to capitalize on this influx of non-technical application builders?
“Vibe coding” represents a dramatic expansion of the market, where the barrier to entry for building complex applications is being dismantled for business analysts and non-technical staff. We expect the total number of users in the application market to expand exponentially as these individuals begin creating software solutions based on intent rather than syntax. Organizations should restructure by moving away from silos where “IT builds” and “business asks,” and instead create cross-functional teams where analysts use AI to prototype and deploy tools in real-time. This doesn’t replace the software engineer; rather, it frees them to focus on the high-level architecture and the complex integration of these AI-generated tools. The traditional development cycle is being compressed from months of coding to hours of “vibing,” which allows for a level of institutional agility we have never seen before.
Hyperscalers are increasingly developing custom silicon and seeking alternative hardware partnerships to run their models. How do these emerging proprietary chips impact the broader ecosystem, and what strategic trade-offs must companies weigh when choosing between unified platforms versus specialized, custom-built architectures?
The rise of custom silicon, such as Google’s collaboration with Marvell and their internal Tensor Processing Units, introduces a fascinating tension between specialized efficiency and general-purpose power. Companies must now weigh the “unified platform” advantage—where everything from the chip to the software stack is optimized for compatibility—against the potential cost savings of proprietary hardware designed for specific models. Choosing a specialized architecture can lead to incredible performance in a narrow niche, but it often carries the risk of vendor lock-in or a lack of flexibility as AI models evolve. In contrast, the unified approach provides a safety net of versatility, ensuring that as AI research shifts, the hardware remains relevant. It is a strategic tug-of-war between the raw speed of a custom-built engine and the reliable, massive scale of a standardized platform.
Compute supply remains a significant bottleneck, acting as a direct limit on market expansion rather than just a resource. What practical steps can firms take to optimize their existing hardware during shortages, and how should they prioritize specific AI workloads in a supply-constrained environment?
When compute is the limiting factor for your entire business, you have to treat every GPU cycle like a precious commodity. Practically, firms are looking at optimizing their software stacks to squeeze more throughput out of existing silicon, using techniques like quantization or more efficient storage architectures to remove bottlenecks. Prioritization is key; companies must distinguish between “exploratory” training, which can be deferred, and “revenue-generating” inference, which must be protected at all costs. During these shortages, we see a move toward more disciplined AI roadmaps where only the most productive, agentic workflows are given priority on the premium hardware. It is a sobering reality for many firms to realize that their growth isn’t limited by their imagination or their capital, but by the physical availability of a chip.
What is your forecast for the future of AI-driven computing?
I forecast that the current one-trillion-dollar demand for AI infrastructure is actually a conservative floor rather than a ceiling. As we move toward 2027, the industry will stop seeing AI as an add-on service and start viewing it as the primary operating system for the global economy, leading to even more frequent upward revisions of infrastructure needs. We will see the “inference inflection” result in a world where billions of autonomous agents are continuously interacting, creating a digital ecosystem that requires a level of compute power we can currently only imagine. The companies that successfully secure and optimize their access to this compute will not just be market leaders; they will be the architects of a new industrial era defined by synthetic intelligence. Ultimately, the future of computing is no longer about the machine itself, but about the sheer volume of “productive work” that the machine can autonomously generate for society.