2026 AI Infrastructure Picks: Riding the Compute and Energy S-Curves

The core investment thesis for 2026 hinges on a clear technological S-curve. We are moving from a paradigm where AI's value was captured during the training phase to one where the real economic engine is inference-the act of using a trained model to answer a question or perform a task. This shift is not incremental; it is a fundamental reorientation of compute demand.

By next year, inference workloads are projected to account for roughly two-thirds of all AI compute, a massive leap from just a third in 2023. This isn't a minor trend; it's the next phase of exponential adoption. The market for specialized chips built for this new workload is set to explode, growing to over US$50 billion in 2026. This creates a multi-year infrastructure buildout, but the nature of that buildout is changing.

The vision of inference chips moving to the edge and replacing massive data centers is overly simplistic. The reality is more complex and demands a different kind of infrastructure. While inference will drive demand for specialized, efficient chips, the sheer volume of queries means that a majority of computations will still be performed on cutting-edge, power-hungry AI chips housed in large, expensive data centers. These facilities, valued at hundreds of billions, are not going away. Instead, they are becoming the central nervous system for a new era of AI services.

This sets up a dual infrastructure need. First, there is the race to build inference-optimized silicon at scale. Second, and critically, there is the need for massive, reliable energy supply. Data centers are already significant energy consumers, and their electricity demand is projected to grow at a staggering rate. The shift to inference-heavy workloads means this demand will be sustained for years, not just a fleeting spike. The buildout is not just about chips; it is about the entire stack of power generation, transmission, and management required to keep the lights on for the AI economy. The 2026 inflection is not just a change in software-it is a foundational shift in hardware and energy infrastructure.

Pick 1: Nebius (NBIS) – Renting the Specialized Compute Rails

Nebius represents a pure-play bet on the infrastructure layer for inference. Its business model is straightforward and capital-light: it rents space and cutting-edge GPUs within data centers to clients who need specialized compute power. This is the fundamental rails for the new AI economy. The company's entire value proposition hinges on the shift to inference, where demand for efficient, dedicated chips is surging and high utilization is paramount.

The demand for this capacity is so intense that Nebius has "sold out" all of its available capacity. This isn't a minor backlog; it's a clear signal of a fundamental shortage in the specialized compute stack. The company's growth trajectory underscores the scale of this buildout. It expects to reach a $7 billion to $9 billion annual run rate (ARR) by the end of 2026, a massive leap from its $551 million ARR at the end of Q3. This explosive expansion is a direct function of the inference paradigm, where every query demands a slice of high-performance silicon.

From a strategic angle, Nebius's model captures the marginal cost reduction that generative AI enables. By providing a scalable, on-demand layer of compute, it allows clients to avoid the massive upfront capital expenditure of building and maintaining their own facilities. This aligns perfectly with the infrastructure need we identified earlier. The company itself owns some of its data center locations, giving it a hybrid model that balances control with the flexibility to scale. For an investor, Nebius is a vehicle to ride the exponential adoption curve of inference, capturing value from the essential, high-demand capacity that fuels the next wave of AI services.

Pick 2: Bloom Energy (BE) – Powering the Onsite Data Center Revolution

The exponential growth in AI compute is hitting a hard physical limit: power. As inference workloads surge, the electricity demand from data centers is projected to climb sharply. Servers alone account for around 60% of electricity demand in modern facilities, making them a massive new load center for the global grid. This creates a critical vulnerability. The International Energy Agency estimates that $720 billion in grid upgrades will be required by 2030 to meet this new demand, a monumental task that will take years to complete.

This is where Bloom Energy's business model becomes a strategic play on the infrastructure S-curve. The company provides onsite power through solid oxide fuel cells, offering a direct, always-on alternative to relying on a strained public grid. This is not a niche solution; it is a response to a fundamental friction in the AI buildout. As data centers become the central nervous system for the new economy, their power needs must be met with reliability and speed. Bloom's technology fits that need, allowing operators to sidestep grid bottlenecks and ensure the lights stay on for inference-heavy workloads.

The market has already priced in this paradigm shift. Bloom Energy's stock has skyrocketed roughly 400% over the past year, driven almost entirely by its adoption in powering AI data centers. This surge is a classic sign of exponential demand hitting a supply-constrained infrastructure layer. The company, once a struggling IPO, is now among the priciest energy stocks, trading at 125 times forward earnings. This valuation reflects the market's bet that onsite power solutions will be a critical, high-margin component of the AI infrastructure stack, not a temporary fad.

Why These Two: The Complementary S-Curves

The case for Nebius and Bloom Energy rests on a simple, powerful truth: the AI infrastructure buildout is a two-pronged problem. It requires not just specialized compute, but also a massive, reliable power supply. Together, these two companies represent the first-principles solution to the two core bottlenecks of the new paradigm.

Nebius captures the compute S-curve. Its entire model is built for the shift to inference, where demand for efficient, dedicated chips is surging. The company's explosive growth is a direct readout of this exponential adoption. It has "sold out" all of its available capacity, a clear signal of a fundamental shortage in the specialized compute stack. This isn't speculative growth; it's the market validating the need for a scalable, on-demand layer of silicon. Nebius's projected run rate of $7 billion to $9 billion by year-end is a multi-year ramp-up, capturing the value of the foundational compute rails.

Bloom Energy captures the energy S-curve. As inference workloads drive electricity demand from data centers, the physical limit of the grid becomes a critical vulnerability. Bloom's onsite fuel cells provide a direct, always-on alternative, sidestepping the need for a decade-long grid upgrade. The market has already priced in this paradigm shift, with the stock skyrocketing roughly 400% over the past year due to its adoption in powering AI data centers. This isn't a bubble; it's the market betting that reliable, scalable power will be a high-margin, essential component of the AI infrastructure stack.

Together, they form the fundamental infrastructure layer. Nebius provides the specialized chips and the rental model that reduces marginal costs, while Bloom Energy provides the power that keeps those chips running. They are complementary solutions to the dual constraints of compute and energy. For an investor, this pairing offers a clean way to ride the early stages of exponential adoption in the two most critical, supply-constrained layers of the AI economy.

Catalysts and Risks: The Path to Exponential Growth

The thesis for both Nebius and Bloom Energy is built on exponential adoption curves. The forward path, however, is not guaranteed. Investors must watch for specific catalysts that validate the infrastructure buildout and be aware of headwinds that could disrupt the S-curve.

The primary catalyst for Nebius is the pace of inference chip adoption and the resulting utilization of its capacity. The company has already "sold out" all of its available capacity, a strong early signal. The next major milestone will be its ability to scale production and deployment to meet this demand, translating its projected $7 billion to $9 billion annual run rate into consistent, high-margin revenue. First major customer wins for new infrastructure providers like Nebius will be a key validation point. The market is watching for evidence that clients are moving from proof-of-concept to large-scale, committed usage, which would confirm the shift to inference as a sustained, high-demand workload.

For Bloom Energy, the catalyst is the acceleration of onsite power adoption by data center operators. The stock's 400% surge over the past year reflects market anticipation. The next phase is execution: can Bloom scale its fuel cell manufacturing and installation to match the explosive growth in AI data center power needs? The company's ability to secure long-term contracts with major cloud providers or hyperscalers will be a critical signal of market penetration and pricing power.

The most significant risk to the thesis is a faster-than-expected shift to edge inference. The narrative suggests inference chips could be deployed on edge devices, reducing the need for massive centralized data centers. While Deloitte predicts a majority of computations will still be performed in large, expensive data centers, a rapid edge adoption could fragment the demand Nebius is built to serve. This would challenge the centralization thesis and compress the addressable market for large-scale compute rental providers.

A parallel risk is the pace of grid upgrade funding versus AI load growth. Bloom Energy's model assumes grid bottlenecks will persist, making its onsite solution essential. If governments and utilities accelerate funding for grid modernization-potentially through initiatives like the $720 billion in upgrades needed by 2030-the urgency for alternative power sources like fuel cells could diminish. The company must demonstrate that its solution offers a faster, more reliable path than waiting for a decade-long grid overhaul.

The primary metric to monitor for Nebius is its utilization rate and pricing power. High utilization confirms demand is real and not speculative. Pricing power, reflected in the ability to maintain or increase rates as capacity remains sold out, signals a supply-constrained infrastructure layer. For Bloom Energy, the key metric is the rate of new data center power contracts signed, which will show whether its fuel cells are becoming the default choice for AI facilities.

The bottom line is that both companies are riding powerful, foundational trends. Their success depends on the speed of adoption and their ability to scale. The path to exponential growth is clear, but it is paved with execution risks and technological shifts that could alter the infrastructure landscape.