The risk factor today, the constraint tomorrow
The AI infrastructure cycle is being friced as a question of scale. More chips, more data centers, more power. That framing assumes that inputs expand linearly with capital — that if the demand is there, the supply will follow. It is the assumption that has governed technology cycles for decades.
That assumption is already breaking.
At the level of investment now being deployed — Amazon approaching USD 200 billion in annual capex, aggregate hyperscaler spending converging on USD 700 billion in 2026 — the problem is no longer how much a company is willing to spend. The problem becomes what can actually be secured. Capital is necessary but no longer sufficient. The binding constraints are becoming physical, local, and increasingly resistant to being solved through markets alone.
Energy was the first place where this shift became legible. It moved from being treated as an operational cost to being treated as a constraint. And once it became a constraint, it stopped being bought on demand and started being secured in advance. Amazon now classifies its long-term energy purchase agreements as derivative financial instruments — not utility contracts, but financial instruments — due to their twenty-year duration and price volatility. Meta has committed USD 24.97 billion in renewable energy agreements, the majority due beyond five years (Note 12, Annual Report 2024). Microsoft has contracted 34 gigawatts across 24 countries. What used to be a variable input is now a precondition for growth, embedded in the balance sheet.
This transition followed a sequence that is now repeating itself in two other resources. First, a physical input appears in Risk Factors — disclosed because lawyers determined it was material, not because management was acting on it. Then it moves into MD&A and capital allocation decisions. Eventually it reaches the financial notes as a quantified commitment or a registered asset. By the time it becomes an asset class, the market has already adjusted.
Energy has completed that sequence. Water and critical minerals have not
The opportunity is in reading where they are in the arc — and understanding what that position implies for the assets that will be required when contracting mechanisms finally emerge.
This analysis is based on direct examination of SEC 10-K annual reports (FY2023–2025) and Q1 2026 earnings transcripts for all seven companies. All data points were verified against original document text. The signal heatmap accompanying this note maps 22 terms across 7 companies and 4 document sections.
The thermodynamic constraint
Water has not been financialized. It still appears in filings as a risk — cost volatility, regulatory exposure, climate-driven scarcity. It is embedded in operating costs under “utilities,” rarely isolated as a strategic variable. But this framing is becoming inconsistent with how AI systems are actually being built.
The architecture of compute is changing in a way that makes water structurally relevant in a manner it has not been before. As model density increases, the limiting factor is no longer only how much power can be delivered, but how effectively heat can be removed. Air cooling, which defined earlier generations of data centers, is approaching its physical limits. NVIDIA’s Blackwell architecture marks the inflection point: the GB200 NVL72 system — the current state of the art for large language model training — is specified as rack-scale liquid-cooled. Not as an optimization. As a requirement.
This shift is visible in language before it becomes visible in contracts. Tesla’s FY2025 filing introduced a term that did not exist in any prior Mag 7 document:
“...innovation demands exponentially greater compute, memory, energy and thermal resources, which may prove insufficient in scale or affordability to meet our requirements. — ”
The significance is not the term itself. It is what the term implies: the scaling of AI is no longer purely electrical. It is thermodynamic. And thermodynamic systems introduce constraints that cannot be abstracted away through financial instruments. Energy can be transported across continents and contracted decades in advance. Heat must be dissipated where it is generated, continuously, within physical limits. This makes cooling capacity a local constraint in a way that energy never fully was.
NVIDIA goes further than any other company in the dataset. Its FY2024 Risk Factors frame water not as a cost variable but as a second-order demand constraint: regulations limiting water availability could impair its customers’ ability to expand data center capacity, which in turn limits NVIDIA’s own revenue opportunity. The logic chain — water regulation, customer capacity, NVIDIA demand — is the most analytically precise water-related disclosure in the entire set of filings examined.
The most structurally significant finding is not what appears in the filings. It is what does not. Tesla holds water rights for its Gigafactory Nevada as a registered asset in its financial notes — the only company in the dataset to own water as property rather than purchasing it as a utility service. But no company has done for water what Amazon, Meta, and Microsoft have done for energy.
There are no Water Purchase Agreements. Not one, across all seven companies
That absence is the signal. It marks exactly where water is in the constraint formation arc: stress is being disclosed, but contracting mechanisms have not yet emerged. The company that structures the first industrial-scale WPA will mark the same transition that Amazon’s first classified energy derivative marked for the energy market.
The processing constraint
Critical minerals are following the same arc, with one important difference. The binding variable is not availability. It is processing capacity.
Tesla is the most advanced company in the dataset. It is the only one that names specific minerals — lithium, nickel, cobalt, copper — in its Risk Factors, and the only one that has responded with a physical asset. Its lithium refinery in Texas, operational since January 2026, is the clearest example in the dataset of a company moving from disclosure to ownership. The refinery is not an answer to a scarcity problem. It is an answer to a refining capacity and supply chain control problem that the filing itself identifies.
Apple's FY2025 filing contains the second most significant mineral disclosure in the dataset. The term "rare earths" appears four times — entirely absent from the FY2024 filing. It appears in Risk Factors and MD&A simultaneously.
Under SEC materiality standards, a company is not required to disclose every risk it faces — only those that a reasonable investor would consider significant. The introduction of a new, specific term across two sections of the same annual report is not a drafting decision. It is the conclusion of a materiality analysis. The rare earths exposure did not emerge in 2025. The legal determination that it met the threshold for SEC disclosure did.
The contrast with NVIDIA is instructive. NVIDIA names gold, tantalum, tungsten, and tin in its Annual Report — but in the context of its Responsible Minerals Policy, an ethical sourcing commitment required by SEC conflict minerals rules, not a financial risk assessment. The same regulatory framework, applied differently: one company disclosing minerals because it is required to for compliance, another disclosing them because it has determined they are material to its financial results. That distinction is the methodology.
What is not yet priced
The market has correctly priced AI capex growth, chip demand, and energy contracts. It has partially priced nuclear energy supply, grid infrastructure, and copper demand — assets close enough to the energy story to have attracted attention, but not yet fully reflected. What it has not priced is the layer beneath: water availability, thermal infrastructure, and mineral processing capacity.These are the inputs that appear in risk factors but have not yet reached financial notes. They are the constraint before the contract.
The filing evidence makes three questions unavoidable for any investor in real assets. Which geographies combine reliable water access, clean energy, and regulatory stability at the scale required by gigawatt-class AI infrastructure? Where is the refining and processing capacity for the minerals that the technology supply chain now considers material enough for individualized SEC disclosure? Who will be the counterparty to the first Water Purchase Agreement — and what does owning that position look like before the contract exists?
They are the questions that the filing evidence makes visible. The constraint formation arc is legible. The contracting mechanisms have not yet emerged. That gap — between what is disclosed and what is contracted — is where the work begins.
Under SEC disclosure rules, companies are required to report any information a reasonable investor would consider significant in making an investment decision. That standard — materiality — is not a narrative choice. It is a legal threshold. When a new term appears in a Risk Factor for the first time, it reflects a determination, made by the company's legal and financial teams, that the exposure it describes has crossed that threshold. The filing is not editorializing. It is complying.
This is what makes the filings, for the reader trained to look, something more than financial documents. They are a forward-looking map of physical constraint — produced before the market has priced what is being described, because the legal obligation to disclose precedes the market's ability to act. Disclosure precedes contract. Contract precedes valuation.
A note on method
This note does not analyze what technology companies say about their strategy. It analyzes what their filings are legally required to say about their risks.
The distinction matters. Earnings calls and investor presentations are communications — managed, optimistic, forward-looking by design. SEC filings are disclosures — governed by materiality standards that require companies to report what a reasonable investor would consider significant, regardless of whether management wants to highlight it. When a new term appears in a risk factor, it is not a narrative choice. It is the output of a legal determination.
The signal we track is the migration of physical resource language across four sections of each filing: Risk Factors, Business Description, MD&A, and Financial Notes. Movement from the first toward the last — from declared concern toward quantified commitment — describes the arc from awareness to financialization. We read that arc as a leading indicator of real asset demand.
This note is for informational purposes only and does not constitute investment advice.