Switch's $50 Billion Reentrancy: The Unaudited Assumption at the Heart of AI Infrastructure

In 2017, I was nineteen, a student of cryptography in Jakarta, manually auditing the smart contracts of five mid-tier ICO projects. Reentrancy was the structural flaw that kept appearing. A function calls an external contract before updating its internal state, allowing the caller to re-enter, drain the balance, and repeat. The DAO hack had made this pattern famous a year earlier. It did not stop dozens of startups from deploying the same broken logic. The explanation, after the fact, was always identical: the narrative outpaced the engineering. “Code executes logic; humans execute fear.” The market priced the story. Later, it paid for the bug. Now it is 2025. I am a macro strategy analyst in Jakarta, and I am watching a new kind of contract being submitted to the SEC with a structural flaw that deserves the same label. Switch, a data center operator headquartered in Las Vegas, has filed confidentially for an IPO with a reported target valuation near 50 billion dollars. The company was taken private in 2022 by a DigitalBridge-led consortium for approximately 11 billion dollars including debt. In thirty-six months, the total consideration has increased, on paper, by a factor of roughly four and a half. Five bulge-bracket banks underwrite the deal: Bank of America, Citigroup, Goldman Sachs, JPMorgan, and Morgan Stanley. Ben Horowitz of Andreessen Horowitz has joined the board and is leading a new funding round. The IPO is targeted for November 2025. The confidential filing means there is no public S-1. No revenue. No EBITDA. No net debt. No customer concentration. No contracted megawatts. No power purchase agreements. No PUE metrics. No GPU counts. No technical description of the supply chain or the cooling architecture. Every material fact remains in the shadow of a single Bloomberg-sourced sentence. An unnamed person said the target is “nearly 50 billion dollars.” That sentence is the load-bearing wall of the entire valuation structure. This is not an IPO in the classical sense. It is an experiment in the capital market pricing of an entirely unverified assumption. It is also—and this is the point—a textbook obligation to remember that volatility is the tax on unverified assumptions. The tax is coming. The only question is which side of the ledger will hold the position when it is levied. Let me step back and describe what Switch actually is. Switch is an infrastructure company, not an AI company. It is a multi-tenant data center colocation provider and wholesale developer of customized data center projects. In plain language, it sells land, power, cooling, security, and physical floorspace to companies that need to house GPU clusters for AI training and inference. The customers are hyperscale cloud providers and AI research companies. There are no algorithms, no foundation models, no APIs in this story. The “AI” in “AI infrastructure” describes the tenants, not the product. The operating model is older than the AI hype cycle. Sign long-term leases, usually five to fifteen years. Pre-commit electrical capacity. Build the shell. Install power distribution equipment in volumes that would have shocked a 2019 engineer. Upgrade the cooling system to handle the thermal load of dense AI clusters. Then monetize through monthly recurring fees and the amortized cost of the construction. The model is understandable. It has high revenue visibility per contract, and it does not depend on the success of any individual model or algorithm. That is the draw for institutional investors. The Switch campus portfolio spans Nevada, Michigan, Georgia, and Texas. The geographic choices are strategic. Nevada offers solar and geothermal resources, low land costs, and a relatively empty interconnection queue. Michigan has access to freshwater for cooling and a strong regional grid. Georgia has become a data center corridor in its own right. Texas has a deregulated power market where large industrial consumers can sign wholesale agreements and manage their own risk. Each location carries a distinct set of regulatory and operational advantages. In the data center industry, geography is the deep moat. First prize is not scale or brand. First prize is the ability to obtain 100 megawatts of firm power within the next eighteen months at a price that leaves room for a spread. Switch's portfolio suggests that its internal team understands this rule and is building accordingly. But the market is now being asked to pay a price for that portfolio that has never been seen in the publicly traded data center sector. Let me lay out the arithmetic. Assume, based on industry benchmarks for comparable wholesale providers, that Switch generates 2025 revenue in the range of one to one and a half billion dollars. Assume an EBITDA margin of 45 to 55 percent, which is achievable if the portfolio is mostly leased and operated efficiently. That yields EBITDA of roughly 500 to 800 million dollars. Against the targeted enterprise value of around 50 billion including debt, that is an EV/EBITDA multiple of 60 to 100 times. For comparison, public data center REITs trade between 12 and 20 times forward EBITDA. Equinix, the largest and most diversified interconnection company in the world, has historically commanded around 20 to 25 times at peak. Digital Realty trades in the mid-teens. A 60 to 100 times multiple for a pure-play wholesale data center operator is a number from a completely different asset class. It is the multiple normally reserved for hyper-growth software companies whose revenue compounds at 30 percent annually without requiring massive physical capital. Even under an aggressive scenario—Switch generating 2 billion dollars of EBITDA in the current year—the multiple is 25 times. That is at the absolute upper extreme of the historical sector range. To sustain it, the company must double EBITDA within three years, entirely through new site deliveries and new contracts. That is possible. It is not verifiable with the information available. The gap between “possible” and “verifiable” is exactly where speculative excess lives and ultimately dies. I have seen this pattern before, in different clothes. In DeFi Summer 2020, I spent four weeks reverse-engineering the yield farming mechanics of Compound and Uniswap. I built simulations to test the liquidity curve under volatile conditions. I found material inefficiencies in early AMM pricing, and I published a five-part technical series on liquidity fragmentation and capital efficiency. What struck me at the time was not the technical detail; it was the divergence between what the market was pricing and what the code was doing. Protocols with high total value locked were being valued as if TVL were revenue. The market treated the TVL number as an accrual asset, ignoring that the tokens underlying it were subject to dramatic supply expansions and that the yield was coming from token inflation, not business profits. The market eventually learned this lesson. The lesson took a form called “volatility.” The same structural illusion is present in the Switch deal. The “contracted megawatts” in the company's pipeline are the token emissions of this asset class. They represent future revenue, but future revenue materializes only if three conditions hold simultaneously. First, the customer's AI program remains funded and strategically relevant throughout the lease. Second, the customer's own cost of capital does not rise enough to make renegotiation or exit attractive. Third, the facility is delivered on schedule and at close to the contracted price. Each condition is plausible by itself. All three holding together for a decade is a much stronger claim. That is the unverified assumption. This is the same structural process I analyzed in my 2024 ETF macro thesis. In the first ninety days of Bitcoin spot ETF inflows, I identified a 12 percent correlation between Nasdaq volatility and Bitcoin spot price stability. Institutional flows, once a trickle, were becoming the marginal price setter. The same mechanism is operating here. The IPO is not simply a funding event for Switch. It is a mechanism for global institutional capital to price the AI capex supercycle, using a physical infrastructure asset as the vehicle. The five banks are essentially packaging the spread between yield-hungry long-duration capital and the bond-like cash flows of a data center lease. They are selling the AI narrative as a new asset class. Customer concentration is the single most important vulnerability. In the custom data center industry, one or two anchor tenants typically drive occupancy for each facility. AI labs and hyperscale cloud providers sign pre-committed mega-contracts binding floorspace and power for five to ten years. In many mid-sized wholesale data center portfolios, the top three customers account for fifty to seventy percent of annual recurring revenue. Switch may be larger and more diversified. It is not likely to be immune to concentration. When the S-1 is published, the customer concentration table will be the correct first page to read. A single customer representing twenty percent of revenue is a warning sign. Thirty percent is a structural red flag. Thirty-five percent, combined with a 25 times multiple, is the blueprint for a violent repricing event. Then there is the leverage question. The 2022 acquisition was an LBO. The DigitalBridge consortium did not pay 11 billion in cash from a single wallet. The deal involved a substantial debt layer, syndicated across institutional lenders. Switch has continued its capital expenditure program since 2022, adding even more financial commitments. Net debt is almost certainly in the billions. The 50 billion figure is ambiguous: it could refer to enterprise value, equity value, or a headline number designed for media consumption. If enterprise value is 50 billion and net debt is 15 billion, the implied equity value is 35 billion. That distinction changes the investment analysis completely. The Bloomberg-based report did not specify. That is not an incidental omission. It is a material omission that will determine whether the company is being priced at a fair multiple or a fantasy multiple. The timing of the a16z round adds another dimension. A private round led by Ben Horowitz, announced after the confidential filing, is a price discovery event. If the a16z round clears at the 45 to 50 billion level, investors will interpret the IPO target as an extrapolation of private market reality. If the round prices lower, the 50 billion range becomes a stretch. The private capital transaction is effectively a calibration tool for the public narrative. I will be watching the announcement details closely. The valuation communicated may tell me more than the entire bank syndicate presentation. There is another layer to this. The IPO is not an isolated event. Since the start of 2025, multiple data center operators and related equipment suppliers have tapped the US public markets. The pattern forms a sustained wave: from service providers to cooling system manufacturers, from transformer suppliers to regional developers. Each listing strengthens a broader story in which the public equity market is the new financing engine for AI physical infrastructure. If Switch prices near 50 billion, it will become one of the largest listed data center companies on the planet, directly comparable to Equinix and Digital Realty. The consequence is a sector-wide rerating. Private data center developers will begin modeling IPO exits. Private equity sponsors will accelerate their holding period. New capital will chase new construction. The asset class will become financialized. Why should crypto observers pay attention? Because Switch's IPO is the most explicit evidence yet that traditional financial infrastructure is absorbing the same narrative mechanics that drive digital assets. The term DePIN—decentralized physical infrastructure networks—has been a crypto-native narrative for years. Here, the centralized equivalent is being executed by five banks, with an S-1 standing in for a white paper and a board seat standing in for a governance token. The RWA tokenization wave has focused on Treasury bills and bonds. Switch demonstrates that the equity market can tokenize AI infrastructure faster, with less transparency, and at a multiple that would embarrass an Avalanche subnet. This is the infrastructure-first-skepticism lesson. Now let me address the contrarian angle directly. The classic “pick-and-shovel” argument says that infrastructure is safer than the technology it serves. In the gold rush, whoever sold shovels made a reliable return. This argument is seductive and partially wrong. The shovel seller in an AI gold rush faces at least three structural dangers that the “pick-and-shovel” narrative ignores. The first is the customer's ability to reduce demand. An AI lab can pause a training run. It can switch to a more efficient architecture. It can negotiate a discount in exchange for early renewal. It can go bankrupt. The data center operator is the opposite: it has a load-bearing contract, pre-committed electrical capacity, and debt that must be serviced regardless of the lab's financial health. The lab's flexibility is the operator's risk. The relationship is reversed from the popular perception. The tenant controls the lion's share of the optionality. The second danger is regulatory. Data center energy consumption has become a public policy issue. Grid operators in Virginia, Texas, and Georgia have warned that data centers are straining interconnection queues. Environmental groups in Nevada and Michigan are questioning the water consumption of cooling systems. The US federal government is accelerating AI capacity through subsidies and tax incentives, creating a conflict with local utility limits and grid reliability concerns. A company like Switch sits precisely at the intersection of that contradiction. Pro-AI federal subsidies benefit it; local grid constraints limit it; ESG scrutiny burdens it; power purchase agreements tie it down. This is a complex regulatory lattice that no board seat can cut through. The third is the underwriting conflict. The presence of five top-tier banks assures financeability, not fairness. Banks earn fees from completed IPOs, not from accurate valuations. Their diligence process produces a document designed to enable the transaction, not to answer every question a skeptical macro analyst would ask. I have read enough prospectuses to know that the truth is in the caveats, not the celebrations. The S-1 will contain the disclosures that lawyers require and the silence that cleverness permits. Reading it will require the same forensic attention I applied to ICO contracts in my earlier career. There is also the question of the “AI label premium,” which is the most dangerous dynamic in this deal. The market has decided that “AI data center” assets deserve a higher multiple than “traditional data center” assets of identical physical quality. The logic is that the tenant profile differs. Yet a gigawatt of power behind a wire does not care where a token comes from. The physical asset is the same; only the narrative differs. When the market starts paying 2x for the same concrete, copper, and cooling towers simply because one facility hosts an OpenAI training cluster and the other hosts an enterprise SAP deployment, the pricing system has left the domain of fundamental analysis and entered the domain of cultural belief. Cultural belief is not stable. It changes at the speed of Twitter. Now let me discuss energy, because energy is the true moat in this business. The long-term competitive advantage of a data center operator is not its brand or its client roster. It is the ability to secure large quantities of reliable, low-cost electricity, with acceptable green credentials, in the shortest possible time. This is why geography is central. Nevada's solar and geothermal resources give Switch access to renewable energy at a time when public companies face increasing pressure to report and reduce carbon footprints. Michigan's water resources provide cooling capacity without the water stress that plagues data centers in the American Southwest. Georgia benefits from the regional transmission system and a regulatory climate that has welcomed hyperscale developers. Texas offers wholesale power markets where large consumers can hedge long-term supply. If Switch executes well on this energy procurement strategy, it creates a genuine moat that the market multiples will eventually respect. But the proof will be in the power purchase agreements, not in the IPO roadshow deck. There is one more energy-related risk that almost no one in the public market is discussing: the transformer bottleneck. The lead time for large power transformers has expanded from a few months to two years or more. Switchgear and switchboard delivery times are similarly stretched. Every new data center build is an exercise in supply chain management as much as construction management. If the transformer arrives late, the revenue recognition is delayed, the capitalized interest grows, and the whole financial model starts cracking. This material constraint is entirely absent from the public commentary around the IPO. It should not be. Let me now provide the assessment framework. When the S-1 becomes public, likely in late October or early November, I will follow a specific set of signals. First, the balance sheet: net debt, maturity schedule, and interest coverage. A long-dated fixed-rate debt stack is supportable. A near-term maturity wall under a potentially rising rate environment is not. Second, the revenue backlog: contracted megawatts under construction and their scheduled delivery dates. The pipeline is the true measure of future revenue; it is the equivalent of total value locked in DeFi, but with contractual clauses and penalties. Third, customer concentration: the share of annualized revenue from the top one, top three, and top five customers. A single customer above twenty percent demands a valuation discount. Above thirty percent, it is an existential concentration risk. Fourth, the power procurement structure: fixed-price agreements, green tariffs, and utility partnerships. Low-cost, renewable, firm power is the only durable moat. Market-indexed power prices are a perpetual margin risk. Fifth, the valuation language used by the banks. If the IPO is framed as “a multiple of 2026 EBITDA,” the implication is that 2025 numbers cannot support the target. That phrase was used extensively in 2021, right before a broad correction. It is a linguistic warning sign. In 2022, during the run-up to the Terra collapse, I had a framework that identified structural vulnerabilities in algorithmic stablecoins. The mechanism promised to maintain a peg by social proof and arbitrage, not by a robust collateral base. I positioned a hedge before the collapse by shorting the ecosystem and increasing stablecoin reserves. The hedge was not based on the view that the teams were malicious; it was based on the view that the assumption of equilibrium under stress was unverified. The same pattern applies here. Investors in Switch are being asked to believe that AI demand will grow at a rate that keeps a 60x or 25x multiple anchored. That is not an act of analysis. It is an act of faith. Now let me return to the beginning. Reentrancy. In the 2017 ICO audits, the flaw was a missing state update. The contract intended to transfer value but did it in a sequence that allowed the receiver to call back. The sequence was not an accident; it was an assumption about execution order. The same assumption is being made in the Switch IPO. The market assumes the sequence is: AI demand grows, tenant income grows, the operator's contracts reprice accordingly, and all this happens in a world where interest rates stay low enough and energy remains cheap enough and the grid gets built fast enough. Each step appears reasonable. The sequence as a whole is unproven. Switch may deliver. The company could have the strongest financials in the data center sector. It may have locked in the best power agreements and the most diversified customer base. The IPO could become a landmark vindication of AI infrastructure financing. But the data is not yet public, and the valuation is being socialized before the evidence. In that gap, the buyer is being asked to fund the narrative with no access to the source code. I have learned from years in both crypto and traditional markets that unverified assumptions eventually generate volatility. The tax collector does not announce the date of collection. The wise position is to define one's conditions before the event, not after. My condition is simple. I will wait for the S-1. I will read the customer concentration table, the debt schedule, the power agreements, and the backlog. I will quantify the gap between the 50 billion target and the assets that support it. If the arithmetic is sound, fine. If it is not, the market will eventually discover its own conclusion. “Volatility is the tax on unverified assumptions.” The tax is certain. The timing is not.