Mapping the AI Infrastructure S-Curve: Where IREN, CRWV, and NBIS Stand

The AI infrastructure market is riding a fundamental exponential growth curve, driven by insatiable demand for compute. To keep pace with projected capacity expansion through 2030, at least $3 trillion in investment is required. This isn't just a growth story; it's a paradigm shift in how that compute is used. The market is transitioning from a primary focus on training massive models to a new era dominated by inference-the process of using those trained models to answer questions, generate content, and perform tasks in real time.

This shift is already underway. In 2026, inference workloads are forecast to account for roughly two-thirds of all AI compute, up from a third in 2023. This isn't a minor trend; it's a structural change that will reshape the entire infrastructure stack. The demand for specialized hardware is a key signal. The market for inference-optimized chips is projected to grow to over US$50 billion in 2026. This explosive growth signals a fundamental change in where and how AI is processed, moving beyond the massive, centralized data centers that currently house the most expensive training chips.

The bottom line is that while the initial scaling law for training may be slowing, the overall demand for compute is accelerating. Inference queries from billions of users will add up, creating a new, persistent workload that requires a different kind of infrastructure. For investors, this means looking past the hype cycle and focusing on the companies building the rails for this next phase-those that can supply the chips, the power, and the data center capacity needed to run the AI models we all depend on.

The Physical Bottleneck: Power and Scale

The exponential growth of AI infrastructure is hitting a physical wall: the power grid. The demand for electricity to run data centers is no longer a future risk; it's a present, accelerating constraint that will determine which companies can scale and which will stall. The latest forecast shows data center power demand is now expected to hit 106 gigawatts by 2035, a 36% increase from just seven months ago. This isn't just a number; it's a signal that the market is recalibrating its view of the bottleneck, with the grid's ability to deliver power becoming the new rate-limiting factor.

The problem is concentrated in the servers themselves. On average, servers account for about 60% of electricity demand in modern data centers. This makes power efficiency and supply the primary cost and operational risk. As AI workloads intensify, the power density within these facilities is rising sharply, pushing against the limits of what existing grid connections can handle. This creates a vicious cycle: more AI demand requires more data centers, which require more power, which strains the grid further.

The scale of new projects underscores the pressure. Of the nearly 150 new US data center projects tracked by BloombergNEF in the last year, nearly a quarter exceed 500 megawatts in size. That's more than double last year's share, signaling a shift toward massive, power-hungry facilities. This boom is colliding with grid realities. In regions like PJM, data center capacity could reach 31 gigawatts by 2030, nearly matching the new generation the Energy Information Administration expects over the same period. The bottom line is that the infrastructure layer for AI is not just about compute chips; it's about securing the massive, reliable power supply needed to run them. Companies that can navigate this power bottleneck-through strategic siting, efficiency gains, or direct power partnerships-will be best positioned to ride the S-curve. Those that cannot will face severe scaling limits.

Company Positioning: IREN, CRWV, and NBIS on the S-Curve

The S-curve for AI infrastructure is now defined by two forces: exponential demand for compute and the physical constraint of power. Within this dynamic, the positioning of individual companies reveals starkly different paths. Some are riding the wave with proven demand, while others are navigating the turbulence of scaling and valuation.

IREN Limited exemplifies the volatility inherent in chasing the growth narrative. The stock surged earlier this year on analyst upgrades and a long-term contract with Microsoft, a clear vote of confidence in its data center build-out. Yet, that momentum quickly reversed. Since its early November peak, the stock has lost approximately one-third of its value. This sharp correction, following a disappointing first-quarter earnings report, highlights a critical inflection point. The market is no longer rewarding potential alone; it is demanding proof of execution and financial discipline as the growth story matures. For IREN, the S-curve is steep, but the path to the plateau requires navigating this new reality of heightened scrutiny.

By contrast, Nebius (NBIS) is operating in a demand-supply vacuum that fuels its explosive growth. The company is a full-stack provider, building and deploying AI data centers with GPUs and software to train and deploy models. Its strategy is to scale power capacity by 4x to 5x this year, a direct response to a projected 9-gigawatt shortfall in U.S. data center supply for 2026. This demand-surplus environment has driven phenomenal results: revenue grew 437% in the first nine months of 2025, and the stock tripled last year. Yet, that success has priced in perfection. Nebius now trades at a steep price-to-sales ratio of 65, a valuation that leaves little room for error. The company's expansion plans are ambitious, but the market is now asking whether its growth trajectory can justify such a premium.

CoreWeave (CRWV) occupies a similar infrastructure layer but faces a different market expectation. It is a key provider, often used by hyperscalers to offload compute workloads. However, the narrative is shifting. As the growth story matures, the market is demanding profitability. This is the natural evolution of any S-curve: early adopters get valuation premiums for growth, but the plateau requires operational efficiency. CoreWeave's challenge is to transition from a pure-play expansion story to a model that demonstrates sustainable margins. Its position is strong, but its future upside is now tied to execution on this profitability pivot, not just capacity additions.

The bottom line is that all three companies are positioned on the steep part of the AI infrastructure S-curve. IREN is testing its footing after a valuation peak, Nebius is scaling into a massive demand gap but at a rich price, and CoreWeave is being asked to prove it can profitably run the rails. Their next moves will determine if they can ride the exponential wave to the plateau or get caught in the turbulence of scaling.

Catalysts, Risks, and What to Watch

The thesis for exponential AI infrastructure growth now faces a critical test. The market is moving from narrative to numbers, and several forward-looking signals will confirm whether this is a sustainable paradigm shift or a bubble about to pop.

The first major validation point is the launch of inference-optimized chip makers. The market for these specialized chips is projected to reach over US$50 billion in 2026. The coming year will be the first to see public revenue figures from these companies. Their ability to scale production and capture market share will be the clearest signal of the inference workload shift. If these numbers miss the mark, it would challenge the entire narrative of a new, cheaper compute layer and could deflate the growth story for the broader infrastructure stack.

Simultaneously, the physical bottleneck of power must be monitored. Grid constraints could become a hard cap on data center expansion. Watch for announcements from utilities and regulatory bodies on new transmission lines, power plant approvals, and capacity auctions. The forecast that data center demand could hit 106 gigawatts by 2035 is a stark warning. If grid upgrades lag, it will create a supply shock for the entire ecosystem, forcing companies to delay builds or pay a premium for power. This is the new rate-limiting factor, and its resolution will dictate the pace of the S-curve.

The most systemic risk, however, is a failure in the financial engine. Moody's has flagged that demonstrating actual revenue generation will become "increasingly important" to silence chatter about an AI bubble. The massive capital spending-projected to peak at $600 billion in 2027-must eventually translate into profits. The warning about circular deals involving companies like OpenAI is a red flag. If revenue growth fails to keep pace with investment, it could trigger a credit risk event, forcing a deleveraging cycle that would slow the entire build-out. The market is now pricing in perfection; any stumble in execution could be punished severely.

The bottom line is that the path forward is defined by these three catalysts: chip maker revenue, grid capacity, and financial sustainability. For investors, the next 12 months will separate the infrastructure rails from the hype.