APLD

Investing for the Compute Bottleneck

An unstoppable force—rising demand for AI compute—is colliding with immovable objects—physical and potentially social/political limits to rapid data center construction. The result, in our view, will be a persistent imbalance between compute demand and supply, primarily benefiting the owners of energized compute capacity.
Compute is the raw material of machine intelligence. The primary compute scaling challenge is shifting from training to inference—from the building of AI models by developers to the running of AI models by users and their agents.
AI output is digital, but its infrastructure is fundamentally physical. While the infrastructure is expanding rapidly, it is still struggling to keep pace with demand, and there are bottlenecks at each part of the energized compute pipeline—the most stubborn of which may be permitted access to the electrical grid.
Bottlenecks may result in persistently high compute prices. Compute could become a tradable commodity as CME Group and Silicon Data plan to launch two compute futures contracts in October 2026, pending regulatory review.
Investment opportunities in energized compute capacity skew toward large-scale diversified businesses and/or institutional investors. Hyperscalers, for example, control about 70% of compute capacity today. Much of the compute buildout is being financed through investment-grade debt or via private markets.
Listed AI compute specialists offer more targeted, high-growth exposure to the energized compute pipeline in public markets. These firms include data center operators specializing in AI/HPC and AI cloud service providers (neoclouds).
Many of these firms are current or former Bitcoin/crypto miners. Investing in Bitcoin miners offers optionality on the value of grid connections for AI use cases.
Risks to the compute bottleneck thesis include faster-than-expected data center construction, model efficiency gains, and a shift toward edge computing.
Artificial intelligence (AI) has become relevant to almost everyone, whether or not they use the technology. The industry’s growth will affect the stock market, the labor market, and possibly electricity bills and local water tables. What lawmakers should do about it will likely feature heavily in this year’s midterm elections.
We are AI optimists but investing realists. The technology will transform the economy and society, in our view, and demand for compute—the raw material of machine intelligence—will be insatiable. But the supply side can only expand so fast, especially when large segments of the voting population are questioning the tradeoffs.[1] The result will be a prolonged bottleneck: a lasting imbalance between compute demand and compute supply.
The compute supply chain faces shortages of many specific inputs, and at any given time, the producers of certain products will be able to charge markup prices. But we expect the sustained value to accrue to the owners of energized compute capacity: the power producers, data center operators, and cloud service providers that collectively turn energy into computation. In our view, these companies will be able to maintain high utilization and high prices into expanding capacity—in other words, they will capture economic rents.
From Massive Training to Massive Inference
In a highly simplified sense, the AI industry can be thought of as three layers (Exhibit 1):
The compute layer, which turns energy into computation;
The model layer, which turns computation into intelligence; and
The application and agent layer, which turns intelligence into action and value.
The parts are inextricably connected. Users typically interface with the application and agent layer (e.g., ChatGPT, Claude Code), which draws intelligence from the model layer (e.g., GPT-5.6 Sol, Claude Opus 5), which runs on capacity supplied by the compute layer (e.g., IREN, CoreWeave). More demand at the top of the stack (i.e., application use) creates more demand at the bottom of the stack (i.e., for computation inputs). Specific companies may participate in a single layer or multiple layers.
Exhibit 1: The AI Technology Stack

We are in the middle of a phase shift in the nature of the processes running through these layers. Starting around 2010-2012, AI model builders realized that scale is what matters: larger models trained on more data using more compute resulted in more intelligence (Exhibit 2).[2] The first AI scaling challenge was therefore driven by model training: growth in the model layer created more demand for the compute layer.
Exhibit 2: More Intelligent Models Required More Training Compute

About half of all AI compute demand still relates to training, according to McKinsey.[3] But this is set to change, partly due to competition at the model layer. Open-weight models like those produced by China’s DeepSeek now operate near the frontier of the industry. Open-weight models can have various commercial restrictions (i.e., they are not fully open source), but in general they are much cheaper to use. The model layer typically prices its output in dollars per million tokens.[4] Open-weight AI models price their output much lower than the proprietary models, creating competitive pricing pressure for the whole industry (Exhibit 3).
Exhibit 3: Open-Weight Models Like DeepSeek Offer Lower Prices

Lower-priced output from the AI model layer means lower input costs for the application and agent layer. As a result, value capture and investment dollars are shifting. Large AI labs like OpenAI and Anthropic appear likely to pivot from mostly building and monetizing models to vertically integrating into higher-value applications, agents, and enterprise services.
Moreover, lower-priced tokens will facilitate the growth in use of AI agents—bots that carry out tasks rather than just respond to queries in a chat. Agents execute multi-step processes, so they require a lot of compute. Compared to an AI chatbot, for example, an AI agent may require 5x-50x as many tokens.[5] Some agents may need to be “always on” (e.g., to scan for weather or health patterns), so they will require continuous compute services.
As a result, the primary scaling challenge for the AI industry is shifting from training to inference—from the building of AI models by developers to the running of AI models by users and their agents. Inference demand is already growing rapidly (Exhibit 4). Goldman Sachs projects that token consumption will rise 24x by 2030, driven largely by AI agents.[6]
Exhibit 4: Inference Demand Already Growing Rapidly

What does all this mean for the compute layer? Recall that more demand at the top of the AI technology stack creates more demand at the bottom of the stack. Competitive forces are shifting value capture on the margin between the model layer and the application/agent layer. However, both rely on the compute layer. What matters for the bottom of the stack is aggregate growth in demand for computation from both training and inference. Both are likely to continue growing rapidly, but especially inference, driven by lower-cost models and exponential growth in AI agents. In short, the compute layer needs to keep scaling.
Compute Infrastructure and Its Constraints
AI output is digital, but its infrastructure is fundamentally physical. Producing compute requires a chain of inputs. Like a conveyor belt, every part needs to be in place for the system to work at all. Therefore, supply constraints at any one point can limit growth in new capacity, and today there are many overlapping constraints.
The AI compute layer is complex. To keep things tractable, it can be helpful to distinguish between (i) the products and services used as inputs (e.g., chips, electrical equipment) and (ii) the ongoing operation of energized compute capacity. In effect, shortages of many inputs limit the rate of growth of new compute capacity, which puts the owners of existing capacity in an advantageous position. These companies are already positioned at the pinch point, ready to deliver a scarce and valuable resource.
The businesses that own and operate energized compute capacity can be grouped into three main functions (Exhibit 5).
Power & delivery: The production of electricity sold to data center operators. The customer pays for electricity typically priced in dollars per megawatt-hour (MWh).
Data centers: The purpose-built facilities that provide power, cooling, and connectivity to run computing equipment (i.e., racks of GPUs). The customer typically pays for capacity priced in megawatts (MW).
AI cloud (neocloud) services: The running of GPUs (owned or leased) to provide computation services for AI models. The customer pays for GPU-hours.
Individual companies may focus on one market segment or operate in all three. As discussed below, the biggest players are large, diversified tech firms (i.e., hyperscalers), but there are also many AI compute specialists.
Exhibit 5: Elements of Energized Compute Capacity

A huge amount of capital is being invested to scale up the compute supply chain. McKinsey estimates that AI-driven compute demand will require more than $5 trillion of investment globally by 2030.[7] But while the supply of this infrastructure is expanding rapidly, it is still struggling to keep pace with demand. And no amount of investment dollars can overcome certain physical shortages over the near term.
There are bottlenecks at each part of the energized compute pipeline:
Power & delivery bottlenecks. Power is the first and often largest constraint: a single AI campus can require as much electricity as a small city. Data centers are projected to account for around 50% of expected growth in U.S. electricity demand through 2030.[8]
Generating enough electricity is only part of the challenge. Long interconnection queues mean that securing land or announcing a project does not guarantee that the new generation required to serve it can be energized (Exhibit 6). Behind-the-meter generation can bypass or supplement grid power, but it introduces new bottlenecks around gas access, turbines, and electrical equipment.
Exhibit 6: Connecting to the Electrical Grid Can Take 5+ Years

Data center bottlenecks. Once power is secured, the site must be converted into an operating, AI-ready facility. US data center supply has expanded rapidly, yet vacancy rates have fallen to record lows as demand continues to absorb new capacity (Exhibit 7). Moreover, announced capacity is not the same as delivered capacity. Every site still needs (i) permits that can take years to obtain, (ii) specialized labor to build out the shell, and (iii) critical electrical and mechanical systems—including substations, transformers, switchgear, and cooling. Lead times for some transformers now exceed three years, allowing one relatively small component to delay an entire campus.
Exhibit 7: Data Center Vacancy Rates Low Despite Rising Capacity

AI cloud bottlenecks. A powered data center is still only an empty shell until it contains a complete computing system. GPUs must arrive alongside high-bandwidth memory (HBM), advanced packaging, and high-speed networking. SK hynix expects a shortage of high-performance memory supply to persist potentially through 2030 (Exhibit 8).[9] Hyperscaler-designed chips broaden accelerator supply and reduce dependence on third-party chipmakers like Nvidia, but many leading designs still depend on the same capacity-constrained inputs like HBM and advanced-packaging.
Exhibit 8: Bottlenecks Across the Compute Layer

Political approval bottlenecks. Even when the physical inputs are available, public approval can restrict supply. Concerns over electricity bills, grid reliability, water use, and land are increasingly translating into political opposition to new data center development. Community resistance blocked or delayed an estimated $98 billion of US data center projects during one three-month period.[10] New York has imposed the nation’s first statewide temporary moratorium on new hyperscale data center permits. Political action can slow development even after the physical bottlenecks have been overcome.
The implication: The compute bottleneck will not be solved anytime soon, and demand growth will likely outstrip supply growth for at least several more years.
Compute as a Tradable Commodity
Tight compute supply will likely mean relatively high compute prices. Therefore, the asset most directly tied to the bottleneck may be compute futures. Like oil or natural gas futures, compute futures will provide a hedging and/or speculative instrument for the price of compute itself. CME Group and Silicon Data plan to launch two compute futures contracts in October 2026, pending regulatory review.
The on-demand services provided by the AI compute layer are typically priced in GPU-hours—that is, the price for using one graphics processing unit (GPU), the standard chip for AI cloud services, for one hour. There are many different types of GPUs, and each AI cloud service provider offers unique features that may affect model performance. Therefore, each GPU and each provider has a unique price. Compute futures will be based on standardized benchmark indexes. The CME plans to list two contract types, one for the Nvidia H100 chip and one for the higher-performance Nvidia Blackwell B200 chip (Exhibit 9).[11]
Exhibit 9: Tight Supply May Keep Compute Prices Elevated

Financing the Buildout
For buy-and-hold investors, the bigger opportunity, in our view, will be investing in the businesses that already own the energized compute capacity. Compute futures provide direct price exposure to the scarce asset, but equity and credit markets provide a way to capture capacity growth and/or fund the industry’s capital needs. Although there are many avenues for compute-related investment, they are often geared toward large-scale diversified businesses and/or institutional investors.
The largest players in the compute market are hyperscalers—Google, Microsoft, Amazon, Meta, and Oracle. Collectively, they operate roughly 70% of global compute capacity (Exhibit 10).[12]
Exhibit 10: Hyperscalers Control Largest Share of Compute Capacity

The financing requirement for data center buildout will be enormous, and hyperscalers can likely fund much of this internally. Collectively, the largest hyperscalers are on pace to earn over $830 billion in EBITDA and spend approximately $730 billion on capital expenditures in 2026 (Exhibit 11). Their scale, recurring cash flows, and investment-grade balance sheets make them natural anchors of the AI buildout. However, hyperscaler equities provide only diluted exposure to AI infrastructure because these companies operate mature businesses spanning advertising, enterprise software, e-commerce, and consumer technology.
Exhibit 11: Hyperscaler Earnings Funding Compute Buildout

Debt markets offer a more direct claim on cash flows supporting the AI investment cycle. Morgan Stanley expects global AI-related debt issuance to exceed $570 billion in 2026, spanning corporate bonds, leveraged loans, asset-backed securities, and other structured instruments.[13] The largest single category is investment-grade bonds issued by hyperscalers. These instruments generally exchange upside for contractual income and greater seniority, potentially backed by data centers and other assets.
Private capital is also playing an increasingly important role. Technology exposure across the broader leveraged credit market—including syndicated loans and private credit—likely already exceeds $500 billion, with software and services representing approximately 15%[14] of the Morningstar LSTA US Leveraged Loan Index and an estimated 20%–30%[15] of private-credit portfolios (Exhibit 12). Private lending may increasingly finance the data centers, compute providers, and power infrastructure as financing requirements outgrow traditional balance sheets and projects require customized debt structures.
Exhibit 12: Tech Companies Tapping Private Debt Markets

The AI Compute Specialists
Although smaller and higher risk than the tech giants, AI compute specialists can provide more direct investment exposure to the compute industry. A variety of firms focus mostly or entirely on owning and operating energized capacity for AI compute. Their future earnings potential is closely tied to growth in the compute supply/demand imbalance (Exhibit 13).
Exhibit 13: Compute Specialists Can Provide Targeted Exposure

The table below lists 11 firms that can be considered AI compute specialists.[16] They all have essentially the same business model: acquire the right to draw a large amount of electricity at a specific site, build a facility capable of running computing equipment, and sell the use of it under long-term contracts. The key scarce asset is the grid connection. As discussed above, interconnection queues are long, so firms that can deliver grid connections faster can potentially secure attractive contracted rates.
Exhibit 14: AI Compute Specialists Are Growing Rapidly but Remain Capital Intensive

A few of the firms are involved in multiple parts of the process, but essentially there are two categories: (i) data center businesses that function as landlords that own power access and lease the space for approximately 10-15 years[17] and (ii) AI cloud operators (neoclouds) that buy and install the GPUs and sell access under contracts lasting about one to three years.[18]
These are high-growth, high-capex, loss-making businesses (Exhibit 15). For the most part, they have high planned capacity but low delivered capacity. They are financing the construction of compute capacity that will be delivered over the next several years. The investment thesis centers on whether they can deliver that capacity on time and on budget, whether prices will remain high when contracts are executed or renewed, and to some degree on the credit quality of the compute customer (often a hyperscaler).
Exhibit 15: AI Compute Specialists Are High Growth but Currently Loss Making

Bitcoin Miner Optionality
Of the 11 US-listed AI compute specialists, nine are current or former Bitcoin miners, and CoreWeave is a former Ethereum miner.[19] Only Nebius has no crypto mining heritage.
Bitcoin mining and AI compute have a lot in common (Exhibit 16). Bitcoin mining uses power to run ASICs to produce hashes, which are ultimately converted into the digital asset Bitcoin. AI compute uses power to run GPUs to produce GPU-hours, which are ultimately converted into tokens, another type of digital asset.
Exhibit 16: Parallels Between Bitcoin Mining and AI Compute

Because Bitcoin mining is power-hungry, miners found themselves in the enviable position of having a grid connection when the AI boom came along. AI or high-performance computing (HPC) applications earn roughly two times more than Bitcoin mining per megawatt-hour (MWh), on average.[20] As a result, virtually all publicly traded Bitcoin miners have converted or announced plans to convert a portion of their energy access to AI/HPC use cases. Of the 11 AI compute specialists listed above, four are former miners that have pivoted entirely to AI/HPC (Cipher, Core Scientific, Applied Digital, and Galaxy).
Investing in Bitcoin miners beyond the fully converted AI compute specialists can be considered a way to capture the option value of their grid connections. There is usually a reason that the remaining facilities have not converted to AI/HPC already—for example, their power source may not be ideally suited for producing compute. But if the compute scarcity problem remains acute, more Bitcoin miners may be able to secure construction financing and lock in long-term compute contracts, possibly benefiting their share prices (Exhibit 17).[21]
Exhibit 17: Bitcoin Miners Offer Optionality on Value of Grid Connections

Other Ways to Invest in the Buildout
Beyond AI compute specialists and converted Bitcoin miners, investors can gain targeted exposure to the compute scarcity thesis through (i) legacy infrastructure players with recurring revenue, (ii) emerging high-growth companies, and (iii) decentralized alternatives.
Legacy players include established power producers, data center operators, and suppliers of critical electrical and cooling equipment. Established data center operators like Digital Realty serve a diversified mix of workloads and generate relatively steadier recurring revenue through long-term leases and service agreements but offer less direct exposure to AI than converted miners. Power providers such as Bloom Energy can supply onsite generation that can help data centers bypass grid delays, while companies such as Vertiv provide the electrical and cooling systems needed to operate AI facilities.
Private markets offer high-growth alternatives focusing on addressing bottlenecks or enhancing efficiencies throughout the energized compute stack, including:
Inference-optimized chips: Developing hardware to run AI models faster and more efficiently (Etched, raised $700 million at $21 billion valuation).[22]
Alternative-energy data centers: Pairing compute with new power sources including ocean waves (Panthalassa, raised $140 million at nearly $1 billion valuation).[23]
Superconducting power delivery: Carrying more electricity through constrained transmission corridors (Veir, raised $75 million at a $170 million valuation).[24]
Advanced cooling: Using dew-point cooling to reduce power and water use, leaving more grid capacity available for GPUs (YC-Backed Madrone).[25]
Small modular nuclear reactors: Developing reactors that could provide carbon-free onsite power (Valar Atomics, raised $1 billion at a $6 billion valuation).[26]
Certain crypto assets also provide additional exposure to this theme. Bittensor, a decentralized network for AI development, enables participants worldwide to contribute compute and other AI services in exchange for token rewards. Other avenues include decentralized GPU marketplaces such as Akash, tokenized rights to inference capacity such as Venice’s DIEM, and onchain GPU-backed credit through protocols such as USD.AI.
Risks to the Compute Bottleneck Thesis
The largest risk is that compute supply catches up faster than expected. Accelerating GPU production, new data center capacity, model-efficiency gains, custom chips, and a shift toward edge computing could reduce scarcity, weighing on utilization, pricing, and margins. Hardware obsolescence poses an additional industry-wide risk, although exposure will vary by provider. Individual compute providers also face company-specific risks, including customer concentration and reliance on debt-funded expansion. Data center operators face construction delays, cost overruns, and the risk that planned capacity cannot be energized or leased, while power providers face uncertainty around regulation, grid-upgrade costs, and the location of future demand.
In a bear case for owners and developers of energized compute capacity, the largest potential losers would include highly leveraged AI cloud providers, developers with large unenergized pipelines, and companies whose valuations assume that capacity will be delivered on time and contracted at attractive rates. Bitcoin miners valued for their AI-conversion potential may also be vulnerable if their sites lack the reliability, fiber connectivity, or cooling infrastructure required for AI workloads. Hyperscalers are generally better positioned to absorb these pressures, although overbuilding could still weigh on free cash flow and returns on capital.
Conclusion
AI may be digital, but the production of intelligence remains fundamentally physical. As the industry shifts from massive training to massive inference, meeting growing demand must still pass through the same constrained chain of power, data centers, chips, and cloud infrastructure. Rather than trying to predict which model or application will win, we believe the more durable opportunity lies with the owners of energized compute capacity—the scarce infrastructure on which many successful models and applications will depend.
[1] Pew Research polling suggests that ~50% of American adults are “more concerned than excited” about AI, while only ~10% are “more excited than concerned.”[2] “Scaling Laws for Neural Language Models,” Jared Kaplan and coauthors, January 2020.[3] Source: McKinsey. [4] Tokens are small pieces of text—such as whole words, parts of words, or punctuation—that an AI language model reads and generates. [5] Source: Gartner, Goldman Sachs. [6] Source: Goldman Sachs. [7] Source: McKinsey [8] Source: IEA report on energy and AI. [9] Source: Reuters. [10] Source: Data Center Watch. [11] Source: CME. [12] Source: Epoch.ai. [13] Source: Reuters. [14] Source: Morningstar. [15] Source: BIS. [16] Constituents determined by Grayscale Investments based on company filings. Includes US-listed companies that own at least 100 MW of electricity capacity at identified sites (firms only leasing capacity are excluded), sell to outside customers (i.e., not for own use like hyperscalers), and derive at least half of their contracted capacity or forward revenue from AI/HPC. Regulated utilities and equipment and construction suppliers are excluded. Constituents as of September 2026. [17] Source: CBRE. [18] Source: Compute.exchange. [19] Including hosted mining; 9 of the 11 also hold Bitcoin on balance sheet. Source: Bitcointreasuries.net. [20] Source: TheEnergyMag. [21] Constituents determined by Grayscale Investments based on company filings. Includes US-listed companies that earn at least half of their revenue from Bitcoin mining (including hosting). Excludes companies already included in the AI compute specialist basket. Minimum market cap of $250 million and minimum ADV of $5 million. Basket is equally weighted. Constituents as of September 2026. [22] Business Insider [23] Business Wire and Financial Times [24] Business Wire and Private Market View [25] Y Combinator [26] Valar Atomics and New York Post