The Atoms Over Bits Trade: Investing in the Physical Layer of AI Infrastructure

Executive Summary

Wall Street spent 2024 and 2025 bidding up AI software companies on the assumption that intelligence would be the scarce resource. The market got the scarcity wrong. Intelligence is scaling rapidly and becoming commoditized — frontier model performance doubles roughly every seven months, and open-source alternatives trail by fewer than twelve months. What is genuinely scarce, and becoming scarcer, is the physical infrastructure required to run that intelligence at scale: electricity, cooling capacity, copper interconnects, and the uranium fuel rods that may ultimately power it all.

This report examines what we call the Atoms Over Bits Trade — the thesis that investment returns in the AI era will increasingly accrue to companies that control physical bottlenecks rather than digital ones. We analyze the power bottleneck constraining data center buildout, the nuclear renaissance driven by hyperscaler demand, the critical minerals facing structural supply deficits, and provide a valuation framework for distinguishing genuine infrastructure moats from momentum-driven speculation.

The core argument is straightforward: software margins compress under competition; atoms appreciate under scarcity. Investors who understand this inversion will be better positioned for the next phase of the AI cycle.

Part I: Why Software Margins Compress

The Commoditization Treadmill

The AI software layer is experiencing a dynamic that should be familiar to anyone who lived through the cloud computing buildout of 2010-2020: capability commoditizes faster than the market expects. Consider the trajectory:

• In March 2023, GPT-4 was the only frontier model capable of passing the bar exam. By March 2026, at least seven models from five different organizations match or exceed that benchmark.
• In January 2024, a state-of-the-art AI coding assistant cost approximately $0.03 per 1,000 tokens of output. By Q1 2026, equivalent capability costs less than $0.003 — a 90% decline in two years.
• Open-source models (Llama 3, Mistral Large, DeepSeek-V3) now perform within 5-10% of proprietary frontier models on most standard benchmarks, eliminating the pricing power that closed-source providers initially enjoyed.

This commoditization has direct implications for margins. Microsoft disclosed in its Q1 FY2027 earnings that Azure AI gross margins declined approximately 400 basis points year-over-year, driven by competitive pricing pressure and rising infrastructure costs. Google faced similar dynamics — Cloud AI revenue grew 45% but operating margins for the segment compressed from 8% to 5% as the company invested heavily in TPU capacity and offered aggressive pricing to win enterprise contracts.

The pattern echoes what happened to cloud infrastructure providers between 2015 and 2020: revenue grew rapidly while margins compressed as competitors matched capabilities and customers gained bargaining power. Amazon Web Services margins peaked at approximately 31% in 2018 and have oscillated between 24-30% since, despite revenue more than tripling.

The Implication for Investors

Software companies in the AI stack face a structural challenge: their product improves continuously, but so does everyone else's. The result is a treadmill where companies must spend more on compute infrastructure simply to maintain competitive parity. This dynamic transfers value from the software layer to the infrastructure layer — from bits to atoms.

The companies that recognize this have already begun repositioning. Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart the Three Mile Island Unit 1 nuclear reactor. Amazon acquired a nuclear-powered data center campus from Talen Energy for $650 million. Google signed agreements with Kairos Power for small modular reactor (SMR) deployment. These are not PR gestures — they are strategic acknowledgments that power is the binding constraint on AI scaling, and securing it is worth billions in upfront capital.

Part II: The Power Bottleneck

Data Center Electricity Demand

The numbers are staggering and still underappreciated by most investors. The International Energy Agency (IEA) projects that global data center electricity consumption will reach 945 TWh by 2030, more than doubling from approximately 460 TWh in 2024. To put this in perspective, 945 TWh exceeds the total electricity consumption of Japan — the world's fifth-largest economy.

In the United States specifically, data centers are projected to consume 12-15% of total electricity generation by 2030, up from approximately 4% in 2024. This is not gradual growth — it is a step-function increase driven almost entirely by AI training and inference workloads.

The constraint is not just generation capacity but grid interconnection. New data center projects in Northern Virginia — the world's largest data center market — now face grid interconnection wait times of 4-7 years. Dominion Energy, the primary utility serving the region, has stated publicly that it cannot meet projected demand growth with its current generation and transmission infrastructure.

Similar bottlenecks are emerging in:

• Central Ohio (driven by Amazon and Google buildouts): AEP Ohio has requested regulatory approval for $2.7 billion in grid upgrades.
• Phoenix, Arizona: Salt River Project has paused new large-load interconnection requests pending infrastructure upgrades.
• West Texas: The ERCOT grid is processing over 150 GW of interconnection requests, more than the grid's total current capacity.

Behind-the-Meter Solutions

The grid bottleneck has spawned a parallel infrastructure trend: behind-the-meter generation, where data center operators build dedicated power plants that connect directly to the facility without relying on the public grid. This approach bypasses interconnection queues, avoids transmission losses, and provides the 99.999% uptime that AI workloads demand.

Behind-the-meter configurations include:

Natural Gas: The fastest path to dedicated power. Companies like Bloom Energy (fuel cells) and various gas turbine manufacturers are supplying behind-the-meter natural gas generation for data centers. Amazon has deployed natural gas-powered data centers in Mississippi and Oregon. The advantage is speed — a natural gas plant can be permitted and built in 18-24 months. The disadvantage is carbon emissions, which conflict with hyperscaler sustainability commitments.

Nuclear: The long-term play. Nuclear provides carbon-free baseload power at scale — exactly what AI data centers need. The economics have shifted dramatically: a 1 GW nuclear plant can power approximately 500,000 high-performance GPU servers continuously. At current AI compute pricing, the revenue per MWh from powering AI workloads exceeds the revenue from selling the same electricity into wholesale power markets by a factor of 5-10x.

Battery + Solar/Wind: Viable for supplemental power but insufficient for baseload. AI training workloads require continuous power for weeks or months — intermittent renewables with battery storage cannot economically provide this at the scale required. Battery storage costs would need to decline by an additional 60-70% to make 24/7 renewable-powered AI data centers competitive with nuclear or natural gas.

The Nuclear Renaissance

The convergence of AI power demand and decarbonization commitments has triggered what industry observers are calling a nuclear renaissance. The evidence is substantial:

Reactor Restarts: Constellation Energy is restarting Three Mile Island Unit 1 (835 MW) under a 20-year PPA with Microsoft, with an expected online date in 2028. Holtec International is pursuing the restart of the Palisades nuclear plant in Michigan (800 MW), backed by a $1.5 billion DOE loan guarantee. These restarts represent the fastest path to new nuclear capacity — existing infrastructure reduces construction timelines from 10+ years to 3-4 years.

Small Modular Reactors (SMRs): NuScale Power received NRC design certification for its 77 MW SMR module in 2023 — the first and still only SMR design with full NRC approval. Other SMR developers including Kairos Power, X-energy, and TerraPower (backed by Bill Gates) are progressing through the regulatory pipeline. SMRs are particularly attractive for behind-the-meter data center applications because they can be sized to match facility demand (typically 100-500 MW per campus) and potentially factory-manufactured, reducing construction costs and timelines.

International Activity: The UAE's Barakah nuclear plant (5.6 GW) reached full commercial operation in 2024, demonstrating that new nuclear construction can be completed on time and on budget outside of Western regulatory frameworks. South Korea, Japan, and the UK have all announced expanded nuclear programs with explicit references to AI-driven electricity demand.

The investment implications are significant. Companies with existing nuclear fleets — Constellation Energy, Cameco, Centrus Energy — have seen their valuations re-rate dramatically. Constellation Energy's stock price tripled between January 2024 and March 2026, driven almost entirely by the market's recognition that its nuclear fleet is a scarce, irreplaceable asset in the AI era.

Part III: Critical Minerals — Structural Supply Deficits Meeting AI Demand

Uranium

The uranium market is experiencing a structural supply deficit that predates AI but is being dramatically amplified by it. Global uranium production in 2025 was approximately 145 million pounds U3O8, against demand of approximately 180 million pounds — a deficit of 35 million pounds covered by drawdown of secondary supplies (government stockpiles, underfeeding, recycled material).

AI-driven nuclear demand will add an estimated 25-50 million pounds of annual uranium demand by 2035, according to UxC (the uranium industry's primary market research firm). This incremental demand arrives into a market where:

• Mine supply has been in structural deficit since 2016. The uranium price collapse following Fukushima (2011) led to widespread mine closures and a decade of underinvestment in new production capacity.
• Secondary supplies are depleting. The U.S. and Russian governments have drawn down Cold War-era enriched uranium stockpiles over the past two decades. The Megatons to Megawatts program, which converted Russian warheads to reactor fuel, ended in 2013. These one-time supply sources cannot be replenished.
• New mine development takes 10-15 years. From discovery to first production, a new uranium mine in a Western jurisdiction requires environmental permitting, licensing, construction, and ramp-up timelines that span a decade or more.
• Kazatomprom, the world's largest uranium producer (approximately 25% of global supply), announced production guidance for 2026 that came in 10% below previous targets due to sulfuric acid shortages and construction delays.

The spot uranium price has risen from $48/lb in January 2024 to approximately $85/lb in May 2026. Term contract prices — more relevant for long-term investment decisions — have risen from $60/lb to approximately $80/lb over the same period. Industry analysts project that sustained prices above $100/lb will be necessary to incentivize sufficient new mine development to meet demand growth through the 2030s.

Key investment vehicles include Cameco (the largest Western uranium producer), Kazatomprom (via London-listed GDRs), NexGen Energy (developing the Arrow deposit in Saskatchewan, one of the highest-grade undeveloped uranium deposits globally), and the Sprott Physical Uranium Trust (which holds physical uranium, providing direct commodity exposure).

Copper

Copper is the metal that makes electrification possible, and AI infrastructure is extraordinarily copper-intensive. A single hyperscale data center requires approximately 30,000-60,000 metric tons of copper for electrical wiring, busbars, cooling systems, and grid connections. For context, this is equivalent to the annual copper production of a mid-sized mine.

The structural supply picture for copper mirrors uranium but at a much larger economic scale:

• Global copper demand is projected to reach 35 million metric tons by 2035, up from approximately 26 million in 2025. AI data center buildout accounts for an estimated 1.5-2.5 million metric tons of this incremental demand.
• Mine supply is constrained by declining ore grades at existing operations (the average copper ore grade has fallen from 1.5% in 1990 to approximately 0.6% today), water scarcity in key producing regions (Chile, Peru), and lengthy permitting timelines for new projects.
• The supply gap is projected to reach 5-8 million metric tons annually by 2035 under current development trajectories, according to S&P Global's copper outlook.

Copper prices have responded accordingly, rising from approximately $3.80/lb in early 2024 to $5.20/lb in May 2026. Analysts at Goldman Sachs and Bank of America project copper prices above $6.00/lb by 2028, driven by the convergence of AI demand, electric vehicle adoption, and grid modernization.

For investors, the copper thesis is best expressed through major producers with long-life, low-cost assets: Freeport-McMoRan (the largest publicly traded copper producer), Southern Copper (lowest-cost major producer), and BHP (diversified miner with significant copper exposure and pending acquisition of Anglo American's copper assets).

Silver

Silver occupies a unique position in the AI infrastructure supply chain. It is the most electrically conductive metal, making it irreplaceable in high-performance electrical contacts, connectors, and — critically — photovoltaic cells. The AI buildout drives silver demand both directly (through data center electrical components) and indirectly (through the massive solar installations being built to power data centers).

Industrial silver demand reached a record 680 million ounces in 2025, driven primarily by solar panel manufacturing (which consumed approximately 200 million ounces, up from 100 million in 2020). Total silver supply was approximately 1.01 billion ounces against total demand (industrial + investment + jewelry) of approximately 1.18 billion ounces — a structural deficit of 170 million ounces, the fourth consecutive year of deficit.

Silver mine supply is constrained by the fact that approximately 70% of silver is produced as a byproduct of copper, lead, and zinc mining. This means silver supply does not respond directly to silver price signals — it depends on the economics of other metals. New primary silver mines are rare and typically small-scale.

The silver price has risen from approximately $24/oz in January 2024 to approximately $38/oz in May 2026. Given the structural deficit and growing industrial demand, several analysts project silver prices above $50/oz by 2028. Investment vehicles include Pan American Silver, First Majestic Silver, and the iShares Silver Trust (SLV) for physical exposure.

Part IV: Contracted Revenue vs. Momentum Trades

Not all "atoms" investments are created equal. The critical distinction for investors is between companies with contracted revenue — long-term agreements that provide visibility into future cash flows — and companies riding momentum — benefiting from sector enthusiasm without the underlying business fundamentals to justify their valuations.

Characteristics of Moat Businesses

Companies with genuine infrastructure moats in the AI power complex share several characteristics:

Long-duration contracted revenue: Constellation Energy's 20-year PPA with Microsoft is the gold standard. This contract provides inflation-adjusted revenue visibility through the mid-2040s, backed by a creditworthy counterparty with existential need for the power. Similarly, uranium producers with long-term supply contracts at prices above their cost of production have locked in margins for 5-10 years.

Irreplaceable assets: A permitted, operating nuclear plant cannot be replicated on any reasonable timeline. A high-grade uranium deposit in a stable jurisdiction cannot be discovered on demand. A copper mine with 30+ year reserves and first-quartile costs represents decades of geological luck and billions in sunk capital. These assets have replacement costs that far exceed their current market valuations.

Regulatory barriers to entry: Nuclear plant licensing, mining permits, grid interconnection rights — these regulatory approvals take years to obtain and represent durable competitive advantages. A competitor cannot simply "disrupt" a nuclear plant with a better app.

Physical constraints on supply response: Unlike software, where a successful product can be copied and scaled in months, physical infrastructure has irreducible construction timelines. A new uranium mine takes 10-15 years from discovery to production. A nuclear plant takes 7-12 years. A copper mine takes 8-15 years. These timelines create extended periods of supply scarcity that support elevated prices.

Characteristics of Momentum Trades

Conversely, several categories of "AI infrastructure" investment lack these moat characteristics and are better understood as momentum trades:

SPACs and early-stage nuclear developers without NRC approval: Several companies have attracted significant market capitalization on the promise of novel reactor designs that have not yet received regulatory approval. The history of nuclear technology development suggests that the path from concept to commercial operation is measured in decades, not years. Investors should demand a clear regulatory pathway and funded construction timeline before assigning significant value to these ventures.

Exploration-stage mining companies: The mining sector is rife with companies that own prospective mineral deposits but lack the permits, infrastructure, or capital to bring them into production. While some of these will ultimately become producing mines, the attrition rate is high — historically, fewer than 1 in 1,000 mineral exploration projects becomes a producing mine.

Data center REITs without power security: Some data center operators have seen their valuations re-rate on AI demand expectations without having secured the long-term power supply necessary to fill their development pipelines. A data center without power is a warehouse. Investors should scrutinize the power procurement strategy of any data center investment.

Part V: Valuation Framework

Valuing physical infrastructure assets in the AI era requires adjusting traditional frameworks for the structural demand shift. We propose the following approach:

For Power Producers (Nuclear, Gas)

Replacement cost analysis: What would it cost to build an equivalent asset today? For nuclear plants, replacement costs are estimated at $8,000-$15,000 per kW of capacity, compared to current market valuations that often imply $3,000-$6,000 per kW. This replacement cost gap provides a margin of safety — even if AI demand disappoints, the assets are worth more than the market currently implies because they cannot be replicated at current valuations.

Contracted revenue multiple: Value long-term PPAs as annuity streams, discounted at rates appropriate for the counterparty credit quality. A 20-year PPA with Microsoft should be discounted at investment-grade rates (4-6%), while merchant power revenue should be discounted at higher rates (8-12%) reflecting price volatility.

Optionality on AI demand: Beyond contracted revenue, nuclear and large-scale power assets have optionality on incremental AI demand that has not yet been contracted. This optionality should be valued using scenario analysis rather than DCF, reflecting the uncertainty in demand timing and magnitude.

For Mining Companies

Net asset value (NAV) using incentive pricing: Traditional mining valuations use current commodity prices or consensus forecasts. For uranium and copper, where structural deficits are well-documented, we recommend using incentive prices — the commodity price necessary to incentivize sufficient new supply to meet projected demand. For uranium, this is approximately $80-100/lb U3O8. For copper, this is approximately $5.00-6.00/lb. Valuations based on incentive pricing provide a more realistic assessment of long-term asset value.

Reserve life and cost position: In a rising-price environment, the most valuable mines are those with the longest reserve lives and lowest cost positions. These assets generate free cash flow across the widest range of commodity price scenarios and have the longest duration of exposure to the structural demand shift.

Jurisdiction risk adjustment: Not all mineral deposits are equal from an investment perspective. Assets in stable, mining-friendly jurisdictions (Canada, Australia, parts of the U.S.) command valuation premiums over assets in jurisdictions with higher political or regulatory risk (parts of Africa, Central Asia, Latin America). This premium is justified by the lower probability of adverse regulatory changes, tax increases, or resource nationalism.

Part VI: Portfolio Construction Considerations

The Barbell Approach

We advocate a barbell approach to AI infrastructure investing: high-conviction positions in contracted, moat-protected assets combined with smaller, speculative positions in earlier-stage opportunities with asymmetric return profiles.

The core positions (60-70% of allocation) should be in:

• Established nuclear operators with long-term PPAs and AI-related contracts
• Tier-1 uranium producers with contracted order books and expanding production
• Major copper producers with first-quartile costs and 20+ year reserve lives

The speculative positions (30-40% of allocation) might include:

• SMR developers with credible regulatory pathways and funded projects
• Junior miners with high-quality deposits in favorable jurisdictions approaching production decisions
• Infrastructure technology companies (grid management, power conversion, cooling systems) with contracted revenue from hyperscalers

Risk Factors

Investors should monitor several risks to the atoms-over-bits thesis:

AI efficiency improvements: If model inference becomes dramatically more energy-efficient — for example, through neuromorphic computing or algorithmic breakthroughs — the power demand projections underpinning this thesis could prove overstated. Current trends suggest efficiency improvements of 2-3x per year, but demand growth has consistently outpaced efficiency gains. For additional context on how a forced seller cascade in AI stocks could temporarily depress infrastructure valuations, see our related analysis.

Demand timing risk: Data center buildout plans can be delayed or cancelled. If AI revenue growth disappoints expectations, hyperscalers may slow their infrastructure expansion, reducing near-term demand for power and minerals. The counter-argument is that infrastructure assets retain value independent of AI demand — nuclear plants sell power, copper is used in electrification broadly — but near-term price volatility is a real risk.

Regulatory risk: Nuclear energy faces ongoing public opposition in some jurisdictions. Changes in mining regulations, royalty rates, or environmental requirements could affect project economics. The current regulatory environment is broadly supportive of nuclear and critical minerals development, but this could change.

Correlation risk: In a broad market selloff, infrastructure assets are not immune to price declines. As our analysis of how correlations go to one in crisis periods demonstrates, even fundamentally sound assets can experience significant drawdowns when liquidity contracts. Physical infrastructure stocks declined 25-40% during the 2022 market correction despite having no change in their underlying business fundamentals.

Scenario uncertainty: The range of outcomes for AI infrastructure demand is wide. Our scenario matrix maps the key variables — capability growth, deployment speed, regulatory response — and their implications for infrastructure valuations. Investors should stress-test their positions against the full range of scenarios rather than anchoring to a single forecast.

Conclusion

The atoms-over-bits trade is not a prediction about the failure of AI software — quite the opposite. It is a prediction that AI software will succeed so thoroughly that the binding constraint on the industry shifts from intelligence to infrastructure. In a world where intelligence is abundant and improving rapidly, the scarce resources are the physical ones: the electricity to power it, the nuclear fuel to generate that electricity cleanly, and the copper and silver to wire it all together.

This is not a novel pattern in technology investing. The California Gold Rush made more millionaires among the merchants selling picks, shovels, and blue jeans than among the miners themselves. The internet boom of the late 1990s destroyed value in most application-layer companies while creating durable franchises in infrastructure (fiber optics, data centers, semiconductor equipment). The cloud computing era saw infrastructure providers (Amazon AWS, Microsoft Azure) capture more value than most of the application companies built on top of them.

The AI era appears to be following the same pattern, with one important difference: the physical infrastructure layer faces genuine scarcity constraints that cannot be resolved through software innovation. You cannot write code to create a uranium deposit, manufacture copper ore, or accelerate the construction of a nuclear plant. These constraints create a durable, structural advantage for companies that control physical assets — an advantage that widens as AI demand grows.

For investors, the practical implication is clear: allocate a meaningful portion of AI-related investment exposure to the physical layer. Look for contracted revenue, irreplaceable assets, regulatory moats, and structural supply deficits. Avoid momentum trades without these characteristics. And remember that in every technology revolution, the most durable returns have accrued to those who owned the infrastructure — the atoms, not the bits.

---

Disclaimer: This analysis is provided for educational and informational purposes only and does not constitute investment advice, a recommendation, or a solicitation to buy or sell any security. The authors and PitchGrade do not provide personalized investment advice. All investment decisions involve risk, including the potential loss of principal. Past performance is not indicative of future results. Commodity prices, equity valuations, and market conditions can change rapidly and unpredictably. Readers should conduct their own due diligence and consult with a qualified financial advisor before making any investment decisions. PitchGrade may hold positions in securities mentioned in this report.