ARK Invest sees AI infrastructure spending surging by 2026.\n\nThat sentence crossed crypto media in a form that looked like news. Crypto Briefing and similar outlets carried it because capital flows into compute now move digital asset markets. But the coverage produced something closer to a religious tract than an investment thesis. No base figure. No endpoint. No assumed growth curve. No definition of which infrastructure category mattered. A statement with a date and a direction, unaccompanied by constraints, is not an input. It is a headline looking for a portfolio.\n\nI have seen this shape before. During the 2024 spot ETF wave I audited custodial wallets marketed to institutional clients with words like cold storage and multi-party computation. The safest-sounding products often had the weakest key-share distribution logic. The claim was the product. The implementation was the actual risk. Forecasts are the same species. They are claims that allocate other people's money. It is reasonable to ask what verifies them before the allocation happens.\n\nThis article is not a bet against the surge. The direction of the prediction is probably right. Cloud providers have announced aggressive capital plans for years. Electricity and accelerator supply remain constrained. A further increase in AI infrastructure spending by 2026 has the texture of a foregone conclusion. That is exactly why the forecast deserves scrutiny. Consensus claims travel furthest when no one checks their assumptions.\n\nThe Forecast Without a Witness\n\nStart with what ARK Invest is and is not. It is not a chip manufacturer. It is not a data center operator. It is a research and asset management firm known for directional bets on disruptive technologies. Nothing about that status invalidates its view. But it means the forecast is a byproduct of an investment narrative, not a byproduct of a build plan. The difference matters.\n\nThe timing also matters more than it seems. A forecast aimed at 2026 is close enough to audit. We are not waiting for a distant decade. By that date, actual power interconnection data, accelerator delivery schedules, and utilization reports will exist. If the claim is published without a detailed model now, it can still be checked later. But later is too late for portfolios that acted on the headline.\n\nThe rough arithmetic behind such calls is not secret. The largest cloud operators have moved beyond a quarter of a trillion dollars in combined annual capital expenditure. Data center demand has stretched electrical grid interconnection queues to multi-year backlogs. Accelerator vendors sell every unit they can produce. From that starting point, a surge is not a bold thesis. It is the default view of anyone reading ear nings transcripts.\n\nSo why does the forecast deserve a forensic look? Because the word surge hides the difference between spending as a decision and spending as a result. The first is voluntary. The second is enforced by demand.\n\nForecast Is a Transaction\n\nA market forecast is a commitment device. It asks readers to convert an assertion into a position. That makes forecasting a cousin of cryptography. In each case, an assertion about the world is transmitted and acted upon. The difference is that cryptography has developed proof systems. Forecasting has not.\n\nMy own education on this point came in the 2022 bear market. I spent six months building a minimal zkSNARK proof generator in Rust from scratch. I implemented the Groth16 proving system to understand the constraints of zero-knowledge proofs. The proof generation itself was not the hard part. The hard part was designing the constraint system. A circuit with missing constraints runs perfectly. It produces valid proofs of claims that prove nothing at all.\n\nThat lesson transfers directly. A forecast is a constraint system over the real economy. The constraints are the assumptions: capital availability, accelerator supply, power delivery timelines, model efficiency gains, and actual paid demand for inference. Remove the constraints and the forecast compiles. It even sounds persuasive. But it proves no predicate about the world. The 2026 surge claim, as distributed, is a circuit with no exposed gates.\n\nIn my audit experience, the same issue appears in institutional custody products. A multi-signature wallet with poor threshold logic does not fail because the cryptography is weak. It fails because the implementation does not match the claimed security model. Forecasting needs the same honesty. The vision of a spending surge should be released with the assumptions that make it true.\n\nThe absence of those assumptions is not an implementation detail. It is the product.\n\nWhere the Surge Comes From\n\nNow look at the structure of AI infrastructure demand. It is remarkably concentrated. A small set of large firms controls most frontier model training. The constellation of GPU clouds that emerged over the last cycle rents capacity to a customer base that remains narrow. This resembles the layer-2 narrative in crypto more closely than most observers want to admit.\n\nDozens of layer-2 networks launched claiming to solve scalability. The result was not a larger user base but the same user base divided into smaller pools. That is not scaling. It is slicing scarce liquidity into fragments. Something similar is happening in AI compute. New infrastructure providers announce themselves as if supply creates demand. What they have actually created is fragmentation of a compute market with a short list of credible buyers.\n\nThe spending surge forecast implicitly assumes this concentration is temporary. It assumes the buyer set will widen. That may happen. But the forecast does not explain when or why. It does not identify the applications that will make inference demand broad enough to absorb new supply. Without that demand layer, infrastructure spending is not an investment thesis. It is a subsidy extended by capital markets to hardware supply chains.\n\nThis is where my institutional auditing work sharpens the analysis. In 2025, I worked with a legal-tech team to integrate zero-knowledge compliance proofs into a DeFi lending protocol. The circuit verified borrower creditworthiness without exposing personal data. The interesting part was not the cryptography. It was the legal requirements translated into constraints. We had to rework the circuit until the compliance predicate matched the legal claim. The words and the math had to agree.\n\nForecasts have the same obligation. The word surge is legal language in that context. It has enforcement consequences because investors act on it. The math behind it should be published with the same care. It usually is not.\n\nSpending Is Not Demand\n\nThe deepest issue with the 2026 surge claim is contained in a single metric: capital expenditure.\n\nCapital expenditure measures how much money is being spent. It does not measure how much value is being generated. In crypto, the community learned this lesson with total value locked. A protocol could lock billions of dollars and produce no revenue. The money was parked, not productive. The metric was treated as demand when it was often just supply seeking yield.\n\nAI infrastructure spending risks the same confusion. A data center under construction is an expense. It becomes an asset only when the compute inside it finds buyers. If utilization stays low, the expense is not infrastructure for the future. It is a warehouse for rapidly depreciating hardware. The forecast of a spending surge says nothing about which scenario will dominate. It counts the construction, not the utilization.\n\nThe LUNA collapse taught me to look for this specific category error. In 2021 I spent three weeks tracing the Anchor Protocol contract logic after the crash. What looked like a demand for stablecoin yield was actually an emission schedule paying users for holding the token. The system used a metric that felt like market demand. In reality, it was protocol supply moving in a circle until the mechanism broke. I found the exact point of failure in the withdraw logic and redemption oracle design. The lesson was not about stablecoin mechanics. It was about distinguishing internal flows from external demand.\n\nA similar discipline applies to AI. Accelerator purchases can be funded by equity that expects future demand. The purchase itself does not create the future demand. It creates a debt to that future. If applications arrive, the debt is repaid. If they do not, the spending surge becomes a spending glut.\n\nThe forecast would be far more credible if it tracked the demand-side constraints with the same enthusiasm it gives to supply-side construction.\n\nThe Verification Stack\n\nThere is no shortage of data that would make such forecasting verifiable.\n\nPower interconnection queues are public in many regions. The number of new data centers waiting for grid approval is a leading indicator. Transformer lead times are measurable. Accelerator delivery schedules appear in supply chain reports. Most importantly, the cost per unit of inference is becoming a public price in real time. Model providers charge per token. Compute markets list prices per GPU hour. None of this is proprietary.\n\nA serious infrastructure forecast should be organized around those constraints. Instead, most forecasts are organized around a belief about human adoption of AI tools. That creates an unverifiable claim. The claim becomes a matter of sentiment rather than arithmetic.\n\nThe irony is that AI itself is now subject to verification research. I wrote a whitepaper on verifiable inference in 2026, proposing a standard for proving that an AI output was generated without tampering. The goal was to verify off-chain model outputs by proving that input data and weights were authentic. The problem was not that such proofs are impossible. The problem is that most market participants do not even think to ask for them.\n\nThat is precisely where forecast infrastructure stands. The distribution channel expects narratives. The audience expects conclusions. Nobody asks for the constraint system behind the claim. In zero-knowledge cryptography, we say that trust is computed, not given. In financial media, trust is given without computation. The 2026 surge forecast is a case study in that failure.\n\nThe Overbuild Blind Spot\n\nThe contrarian angle here is not that infrastructure spending will disappoint by going down. The real risk is that it will go up exactly as predicted and still fail as an investment.\n\nMarket participants price the risk of underbuilding. They worry about not having enough compute when a killer application arrives. They rarely price the risk of overbuilding. That is the structural blind spot of capacity expansion cycles. Infrastructure decisions are made with a lag. Power contracts are signed years before utilization data is available. Chip orders are placed before model efficiency improvements reduce the demand for those chips. In every capital-intensive industry, the eventual crash comes after a period when the spending was perfectly visible and the demand was not.\n\nCode is law, but bugs are reality. The forecast is the code. Physical capacity is the reality. A spending surge that precedes actual inference demand creates an economy of stranded assets. The hardware will not disappear. It will just be repurposed at lower prices, making the original capital allocation unprofitable even though the headline was accurate.\n\nPrivacy adds another layer to the blind spot. Cloud providers and large AI labs keep utilization data private. That information would allow the market to verify whether new capacity is actually being consumed. Instead, it is shielded as commercially sensitive. For individuals, privacy is a feature, not a bug. But for markets, private utilization data prevents real-time validation of public spending claims. The result is that capital moves on the authority of forecasters rather than on the evidence of demand.\n\nNo one in the current cycle wants to say that the spending surge might be real and still be a trap. That is exactly why it needs saying.\n\nWhat to Watch Instead\n\nThe useful response to an unverifiable forecast is not cynicism. It is a verification protocol.\n\nI would rather watch the price-performance curve of inference than the total dollar amount of infrastructure spending. A healthy AI market shows the cost per unit of model output falling faster than capital expenditure rises. That spread is the true signal. It means efficiency is outrunning the capital investment. It means demand is real enough to reward marginal capacity. If capital expenditure rises while inference pricing stays flat, the market is building supply ahead of demand.\n\nPower interconnection data is the second signal. Announcements of new data centers are abundant. Completed grid connections are the hard evidence of actual buildout. The gap between announcement and energization is where overbuild risk lives. Treat announcements as intents, not facts. Treat grid connections as confirmations.\n\nAccelerator utilization is the third signal, even if it is private. If large fleets are idle, supply is ahead of demand. The narrative of a surge will not include those idles. The audit trail will.\n\nThe Takeaway\n\nA forecast without a witness is not a thesis. It is an invitation to trust a process the forecaster has not disclosed. The 2026 AI infrastructure surge may fully materialize. That would not make the forecast valuable. What would make it valuable is a constraint system that lets readers see the conditions under which it fails.\n\nMath does not negotiate. Either the assumptions are published and testable, or the forecast belongs in the same category as a smart contract with no circuit exposed to audit. The code compiles. The claim runs. But no one can verify what it actually proves.\n\nThe best position is not to bet against the surge. It is to refuse to act on the surge until the constraints are visible. Watch the price-performance curve. Watch power interconnection queues. Watch for the release of assumptions behind confident headline calls.\n\nIf the forecasters will not publish their gates, the rational response is to treat their output as a decorative statement, not a portfolio input. '\n\nAnd by 2026 we will find out who was running a real model, and who was simply running a narrative on hardware.\n\nIn an industry that claims to be data-driven, the demand for auditable forecasting should be the default. The claim is not that the surge is impossible. The claim is that it need not be believed on insufficient evidence. Make the constraints public. Make the witness available. Then let the capital flow to the forecast that verifies itself.
