The New Bottleneck in AI Is Not the Model. It Is the Infrastructure Beneath It

As AI demand pushes beyond software and into hardware, power, and deployment speed, the companies that solve “time to compute” may shape the next phase of the market. Most of the AI conversation lives at the model layer. The benchmarks. The capabilities. The next demo. Below it sits a less visible market, and increasingly a more decisive one. GPUs are oversubscribed. Power contracts crawl. Inventory sits unused on one side of the market while teams line up for capacity on the other. For Dhyay Bhatt, Niraj Yagnik, and Veronika Bhatt, the founding team behind FPX AI, that gap is a market problem hiding inside a technology one. FPX AI is building a marketplace for AI infrastructure, what the team calls a liquidity and intelligence layer. It operates across hardware, colocation, and cloud rentals, with adjacent work in brokerage, consulting, advisory, and research. The goal is not simply to match buyers and sellers. It is to make a fragmented market more transparent and easier to navigate, so that access to compute is determined less by relationships and more by need. “Compute has become one of the biggest dividing lines in AI,” the founders say. “The companies that can get compute can move faster. The ones that cannot are forced to slow down, regardless of how strong their ideas are.” The three founders bring different strengths. Dhyay leads research, product, and infrastructure strategy. Niraj leads software architecture, systems design, and technical execution. Veronika leads go-to-market, partnerships, and customer relationships. Together they turn infrastructure complexity into trust, clarity, and execution in a market defined by opaque supply, shifting prices, and high-stakes deployment decisions. Their interest in the problem began in research. Niraj worked on AI explainability at UC San Diego. Dhyay worked on model merging and fine-tuning at Duke. Both saw how dependent progress had become on reliable compute well before it became a public concern. In 2023, while renting compute through different sources, they hit the friction directly. GPU availability was inconsistent. Pricing was hard to compare. Capacity was relationship-driven. Verification was harder than it should have been. From the outside, AI infrastructure looked like a fast-growing technology market. From the inside, it often behaved like an opaque bazaar. That fragmentation matters because AI development is time-sensitive. A delay in securing infrastructure can slow training, postpone launches, and leave a company less competitive than better-capitalized rivals. For startups and research teams, the problem is acute. Strong ideas. Limited access to the infrastructure needed to test them. In this sense, “time to compute” is becoming a defining metric. FPX AI’s marketplace is built around that idea. By working across hardware suppliers, colocation providers, neoclouds, and end buyers, the company sees signals one part of the stack alone cannot. A hardware provider understands supply. A colocation operator understands power, cooling, permitting, and grid constraints. A cloud provider understands rental demand. AI infrastructure decisions increasingly require all of these to connect. The picture is also widening. GPUs were long treated as the primary bottleneck, and that remains true in many cases, but the next phase is broader. Site readiness. Grid access. CPU availability. Fiber connectivity. Workload-specific deployment needs. Agentic AI may intensify the shift. As more workloads move from experimentation to continuous production, demand becomes more varied. Inference, fine-tuning, retrieval, autonomous workflows, and enterprise deployments each create different requirements. That is why FPX AI focuses not only on GPUs but also on CPU servers, inference infrastructure, and site readiness. The company’s research arm is part of that strategy. FPX AI publishes on GPU supply, data center constraints, power availability, cloud pricing, and emerging bottlenecks. The platform has grown to more than 1,100 subscribers across AI labs, venture funds, hedge funds, infrastructure operators, energy companies, policymakers, and enterprises. For the team, the two sides feed each other. The marketplace gives FPX AI a real-time view into supply and demand. The research turns those signals into analysis the broader market can use. Much of this work is shaped by where the founders began. Dhyay and Niraj grew up in central Mumbai and have known each other since childhood. That long partnership, they say, gives them the kind of trust a fast-moving infrastructure market demands. Veronika adds a complementary operating perspective, translating technical and market complexity into customer trust, partnerships, and execution. In an industry where perfect information is rare, that combination matters. Pricing shifts. Capacity disappears. Deployment windows change. Power availability often decides what is possible more than customer demand does. The stakes extend beyond any single company. If access to compute stays concentrated among the largest players, AI development risks becoming centralized. Smaller companies, independent research teams, and enterprises without hyperscaler-scale purchasing power could find themselves priced out or delayed. In that scenario, infrastructure becomes a gatekeeper for innovation. “AI will not be shaped only by models,” the founders say. “It will also be shaped by the infrastructure available to run those models at scale.” For all the noise around what AI can do, FPX AI’s attention stays on what it takes to do it at scale. Chips. Power. Square footage. The unglamorous parts of the stack that decide who trains, who ships, and who has to wait. Models win the headlines. Compute decides who gets to compete. \ :::tip This story was distributed as a release by Jon Stojan under HackerNoon’s Business Blogging Program. ::: \