The $1.7 Trillion Depreciation Trap: What Goldman's AI Capex Math Actually Requires

The number arrived wrapped in a single adjective: unprecedented. Goldman Sachs calculated that hyperscalers must generate unprecedented revenue growth to justify $1.7 trillion in AI capital expenditure. That is the complete payload the wires carried. No revenue figure. No time window. No depreciation schedule. No stated return on invested capital. No bull, base, and bear scenarios. Just a number so large it functions as its own argument. I have traded through five market cycles where a single round number was used to end an argument rather than start one. In late 2017, the round number was $1,000 ETH. In 2021, it was $100,000 BTC. In 2022, it was $1 UST. Each time, the number became a substitute for the mechanism, and the mechanism is always where money is either engineered or destroyed. Alpha isn't in the headline. It's in the second derivative — the depreciation schedule the headline buried. $1.7 trillion is not a demand signal. It is a liability issued against future cash flows. The market is treating it as the former. That distinction is the entire trade. Context The AI capital expenditure race is the largest coordinated industrial buildout since the transcontinental railroad, and it is being financed by four balance sheets. Microsoft, Amazon, Alphabet, and Meta are doing the bulk of the spending, with a secondary tier of Oracle, CoreWeave, and the unlisted labs pulling compute through structured commitments. Between 2024 and what most sell-side models project as 2030, the cumulative figure for AI-adjacent infrastructure lands somewhere in the $1.5 to $2 trillion corridor. Goldman's $1.7 trillion sits squarely inside that band. The framing matters more than the precision. This is a capex-to-revenue model. It asks one question: can the revenue generated by this infrastructure service the capital deployed to build it? That is not a technology question. It is a credit question wearing a technology costume, and credit questions do not care about narrative. Two prior frameworks established the terrain, and any honest reading of the Goldman item has to sit on top of them. In 2024, Sequoia's David Cahn published what became known as the six-hundred-billion-dollar question — his estimate that AI infrastructure needed on the order of $600 billion in annual revenue to cover the underlying spend, against actual AI revenue that was a fraction of that figure. In the same year, Goldman's own Jim Covello published a note asking, in substance, whether generative AI carried too much spend and too little benefit. That note questioned whether the return profile justified the capital intensity at all. The item I am working from is a third-generation artifact. Goldman's original research, compressed by financial media, re-compressed by a crypto outlet. Each layer of transcription strips constraints. A model that began with explicit scenarios — each carrying its own revenue threshold, depreciation assumption, and target return — arrives as a single number with one adjective attached. That is not reporting. That is a rumor with a citation, and I want to be explicit that everything downstream of this point is reconstruction, not quotation. The information decay is the first vulnerability to audit. When I stress-tested oracle-manipulation risk in DeFi Summer 2020, the failure mode was never the visible collateral — it was the stale feed behind it. Here, the stale feed is the transcription chain. Goldman's actual output almost certainly contained a revenue threshold number and the years attached to it. The wire dropped both. What remains is a headline calibrated for impact, not for decision. Treat it accordingly. So I am going to do the work the wire skipped. I will reconstruct the mechanism, name the variables, and mark exactly where the confidence degrades. Because a $1.7 trillion liability with no amortization schedule is not a thesis. It is a trapdoor, and someone is standing on it. The Core Model, Rebuilt from First Principles Let me build the model the way I would underwrite a yield strategy before committing a single dollar. In my four-hundred-transaction arbitrage run through the 2017 pre-sale chaos, I learned that the headline spread is never the trade. The trade is the cost of executing across the spread — gas, slippage, timing, counterparty. The same discipline applies here. The headline is $1.7 trillion. The trade is the cost of servicing it. Start with the standard capex-justification framework. Set cumulative AI capex at C, equal to $1.7 trillion. Apply a depreciation horizon, n. In hyperscaler accounting, server and accelerator lives have historically run three to six years. Take the midpoint of five years. Annual depreciation load equals C divided by n — roughly $280 to $340 billion per year. Now apply the revenue multiple. Capital-intensive infrastructure assets need gross revenue equal to a multiple of depreciation to cover power, land, labor, cooling, and the cost of capital itself. For high-margin compute leasing, the coverage ratio typically sits between two and four times depreciation, depending on the target return on invested capital. Run that range and the implied annual revenue threshold lands between $560 billion and $1.36 trillion per year. The working midpoint sits around $700 to $900 billion annually. Hold that number next to reality. Total global cloud infrastructure revenue — AWS, Azure, GCP, everything — was roughly $250 to $300 billion in 2024. Direct AI and generative-AI revenue inside that figure was, under most disclosure conventions, under fifteen percent. So the model is demanding that AI-specific revenue alone grow to a multiple of the entire current cloud market. That is what unprecedented means once you strip the adjective and put a number under it. Every assumption in that chain is a lever. Change the depreciation horizon from five years to six, and the annual load drops by nearly a sixth. Change the revenue multiple from three to two, and the threshold falls by a third. Change the target return, and the whole output shifts. Goldman almost certainly modeled all of these as scenarios. The wire collapsed them into one number. That collapse is where the misinformation lives, and it is why I refuse to treat any single figure in the report as decision-grade. The Depreciation Mismatch Now the part that almost no one is pricing: the mismatch between the accounting life of an accelerator and its competitive life. A high-end GPU has a book life of three to six years. Its economic life — the period during which it is frontier-competitive — is eighteen months to two years. Every model generation resets the performance baseline. A cluster that was state-of-the-art at purchase is a discount asset two years later. This creates a specific, measurable vulnerability. The asset sits on the balance sheet at a value it can no longer command in the market. The depreciation charge is spread across a horizon longer than the period of relevance. Reported earnings look healthier than the underlying economics. This is not fraud. It is the ordinary friction between accrual accounting and technological obsolescence — and it is the single most important variable in whether the $1.7 trillion is fairly priced. I have seen this exact pattern before, at a smaller scale. During DeFi Summer 2020, I audited under-collateralized debt positions in Compound and flagged the CKP oracle-manipulation surface before the market recognized it. The position was collateralized on paper. The collateral was stale in practice. I shorted the exposure using ETH collateral and returned forty percent during the subsequent mini-crash. The lesson that compounded across my career: when the marked value of an asset diverges from its realizable value, the divergence is the trade. Accelerators are carrying the same kind of stale-collateral risk, only at a scale measured in hundreds of billions. If the effective economic life of a GPU is closer to two years than five, then the true annual depreciation load on $1.7 trillion is not $280 billion. It is closer to $850 billion. That changes everything. The revenue threshold I computed becomes insufficient. The required growth rate becomes not unprecedented but physically implausible within the stated window. The bull case has a response, and it deserves a fair hearing. Some workloads tolerate older silicon. Inference at the edge, batch processing, and non-latency-sensitive tasks can run on depreciated hardware economically. If a meaningful share of the $1.7 trillion is deployed into long-lived, latency-insensitive capacity, the effective life extends. That is a legitimate counter, and I will not pretend it away. But it does not rescue the frontier-training portion of the spend, which is precisely the portion most exposed to obsolescence — and it does not appear anywhere in the wire report. A thesis built only on the favorable subset of workloads is not a thesis. It is a selection bias with a spreadsheet. The Accounting Escape Hatch Here is the lever that institutional desks discuss in private and almost never put in print: depreciation policy is a management choice, not a physical constant. A hyperscaler can extend the estimated useful life of its servers from four years to six. The accounting consequence is immediate and mechanical. The annual depreciation charge falls by a third. Reported operating income rises. Earnings per share rises. And critically, the apparent cost of servicing the AI capex falls, which makes the Goldman-style revenue threshold look more achievable without a single incremental dollar of actual cash flow. This is not hypothetical. It is an established pattern in capital-intensive industries. Telecom operators did it. Airlines did it on fleet lives. Utilities do it on plant. The move is legal, disclosed in the footnotes, and almost always under-read by the market. Why does it matter for the $1.7 trillion question? Because it decouples the reported return from the real return. A company can justify its capex in its own financial statements by lengthening asset lives, while the physical asset loses competitive value faster than the schedule implies. The justification becomes an accounting artifact rather than an economic fact. And the escape hatch has a cost. Extending useful life defers depreciation into future periods. It front-loads reported profit and back-loads impairment risk. When the asset finally retires — or when a new model generation renders a cluster obsolete — the write-down arrives in a single, ugly quarter. This is leverage of the most dangerous kind: hidden, deferred, and denominated in accounting judgment rather than cash. Alpha isn't in the capex number. It's in the footnote that changes it. The signal to track is not the capex guidance. It is the depreciation policy disclosure. When a hyperscaler quietly extends server lives, read it as a confession that real returns are under pressure. That is the canary, and it sings in the least-read section of the filing. The Reflexive Paradox There is a deeper trap, and Goldman's framing cannot escape it. Suppose inference efficiency improves dramatically. Suppose a model that cost a dollar per million tokens last year costs ten cents this year. This is the stated goal of the entire stack — cheaper intelligence, broader access. The intuition is that cheaper inference means more usage, more revenue, a healthier capex picture. The math says the opposite can be true. If the unit cost of inference falls faster than the unit price, the total addressable revenue pool shrinks even as volume grows. This is the classic deflationary dynamic of any compute market. In the 1990s, the cost of a MIPS of computing collapsed, and hardware revenue consolidated into a handful of winners even as usage exploded. Bandwidth followed the same curve — capacity grew a thousandfold and the carriers that overbuilt it went bankrupt. Run it forward on AI. If the cost of a capable token falls by ninety percent over three years, and price falls by eighty percent, then gross margin per token expands while the aggregate revenue needed to service $1.7 trillion becomes harder to reach, not easier. The very efficiency the infrastructure enables erodes the revenue base required to pay for it. This is reflexivity — the asset's success undermines its own justification. It is the single most under-modeled risk in the entire AI capex complex, and it is the one that finance departments are culturally least equipped to see, because their models assume price stability and volume growth. The efficiency-improvement narrative and the capex-justification narrative are, at the aggregate level, in tension. The wire report presents the capex requirement in isolation, as if the technology roadmap were a neutral bystander. It is not. It is the variable that determines whether the threshold is reachable or asymptotic. When I modeled NFT floor prices in 2021, I watched the same inversion: as the technology enabled broader minting, supply expanded, floors compressed, and the holders who had treated the floor as a floor discovered it was a slope. Supply dynamics do not care about community narratives. Neither does token pricing math. The Revenue Attribution Problem Then there is the question of what counts as AI revenue at all. The threshold assumes a clean category — AI and generative-AI revenue. In practice, that category is porous. When Microsoft sells Copilot seats, how much is incremental AI revenue and how much is a repackaged Office subscription? When AWS sells Bedrock usage, how much is new demand and how much is existing compute relabeled? When Google bundles Gemini into Workspace, what is the marginal contribution? When Meta reports AI-driven ad gains, those gains never leave the advertising line at all. This matters because the justification math requires a multiple of depreciation in genuinely incremental revenue. If a meaningful share of reported AI revenue is AI-flavored repricing of legacy cloud, then the incremental figure is smaller than the headline, and the threshold is further away than it appears. The disclosure convention compounds the problem. Unlike a hardware segment, AI revenue is rarely broken out with a clean definition. Under my audit discipline — the same discipline that made me refuse unverified yields in 2020 — any line item without an auditable definition is a line item I discount. I apply a haircut to all AI revenue that cannot be cleanly separated from pre-existing cloud. On conservative assumptions, that haircut is not small. Understand what I am not saying. I am not saying the AI revenue is fake. I am saying it is un-auditable at the precision the $1.7 trillion question demands. And a thesis that depends on a number nobody can cleanly measure is a thesis with a hole in it. In my experience, holes like that are where capital quietly leaks out. The Financing Structure and the Counterparty Chain The threshold math assumes the capital is funded at some cost. The wire report omits the funding mix entirely, and the mix changes the answer. Cumulative capex of $1.7 trillion splits into two sources: operating cash flow and external financing. Hyperscalers generate enormous free cash flow — but not $1.7 trillion over the relevant window, especially while simultaneously returning capital to shareholders and funding non-AI businesses. So a portion is financed through debt issuance and, in some cases, through structured vehicles that move obligations off the primary balance sheet. Two consequences follow. First, the true hurdle rate is not the internal cost of capital but the weighted cost of the debt issued to fund it. If rates stay elevated, the revenue required to clear that hurdle rises. Second, off-balance-sheet structures — compute commitments, joint ventures, special-purpose vehicles, capacity prepayments — obscure where the risk ultimately sits. When an unlisted lab commits to compute it cannot yet afford, the obligation shades from the lab's problem into the cloud provider's problem. The $1.7 trillion may be booked on four balance sheets while economically distributed across a longer chain of counterparties. That distribution is a transmission mechanism. If end-demand fails, the failure does not stay with the end-buyer. It travels backward up the commitment chain. This is exactly the pattern we saw in 2022, when the Terra collapse transmitted through a web of leveraged CeFi counterparties who had all assumed the collateral was sound. I moved sixty percent of my book into Bitcoin and shorted LUNA derivatives via Deribit options as the contagion spread — not because I predicted the exact timing, but because I had stress-tested the counterparty chain and found it brittle. The AI financing chain has the same structural property: a cascade is only invisible until it fires, and then it is the only thing anyone can see. The Physical Ceiling No model of AI capex is complete without the constraint that money cannot buy: power. The bottleneck has migrated from silicon to electricity. In the United States, data-center interconnection queues in major markets run into multiple years. Transformer lead times have stretched from months to years. Natural-gas pipeline permitting is a multi-year process. The result is a widening gap between capital deployed and capacity energized. This gap directly extends the payback period. A dollar of capex that cannot be powered does not generate revenue on schedule. The $1.7 trillion, if it is committed through 2030, assumes a power buildout that is not currently on a timeline to materialize. The physical constraint is the reason the revenue gap cannot be closed by simply spending faster. Capital is elastic. Electrons are not. I would rank this — not chip supply, not model capability — as the binding constraint on whether the capex can ever be serviced. The wire report does not mention it. That omission alone should lower your confidence in any conclusion drawn from the headline, because it means the analysis was conducted on a spreadsheet with no physics layer. The Supply-Chain Concentration Follow the $1.7 trillion down the stack and it lands, disproportionately, on a single vendor. Nvidia captures the largest share of accelerator spend, which means the hyperscalers' capex is, in large part, Nvidia's revenue. That concentration is both the source of Nvidia's extraordinary valuation and its principal fragility. The hedge is custom silicon. Google's TPU, Amazon's Trainium and Inferentia, Microsoft's Maia — each is an attempt to convert external dependency into internal capability and to compress the cost per unit of compute. The strategic logic is sound. The execution is incomplete. Custom silicon still handles a minority of frontier workloads, and migration is slow because the software ecosystems built around Nvidia's stack are sticky and self-reinforcing. The consequence for the $1.7 trillion question is a bargaining-power problem. When one supplier sets the price of the dominant input, the return on the entire buildout is partly determined by a third party's margin policy. The hyperscalers can justify their capex only if their own gross margins survive the vendor's. That is a structural vulnerability, and it sits outside any model that treats capex as an undifferentiated lump sum. The holder-concentration logic I applied to BAYC floors in 2021 applies here in a different costume: when a market depends on a concentrated base of holders or suppliers, its stability is a function of their behavior, not of its own resilience. Contrarian: The Crowd Buys Demand, the Desk Maps the Liability Here is where the crowd and the desk diverge. Retail reads $1.7 trillion as demand. The reflexive move is to buy the beneficiaries — Nvidia first, then the data-center REITs, the power names, the cooling vendors, the liquid-cooling specialists. The number confirms the thesis. More capex means more revenue for the supply chain. Momentum begets momentum, and the narrative compounds. Smart money reads the same number as a liability. $1.7 trillion of capital deployed against a revenue threshold that, on conservative depreciation assumptions, requires AI revenue to exceed the entire current cloud market. That is not a confirmation. That is a hurdle, and hurdles create dispersion. They separate the companies that can convert compute into high-margin, sticky product revenue from the companies that can only rent raw capacity into a price war. We do not chase pumps; we engineer the squeeze. In this market, the pump is the AI revenue narrative. The squeeze is on the depreciating asset — the accelerator whose book value outlasts its competitive value, the cluster that must generate returns before the next model generation renders it second-tier. The crowd is buying the pump. The desk is mapping the squeeze. There is a second-order signal the wire report leaks without intending to. The item was published by a crypto outlet. Why would an outlet covering digital assets spend column inches on hyperscaler depreciation math? Because AI and crypto are financed by the same pool of risk-appetite capital. When that capital gets nervous about AI returns, it does not stay in crypto and wait. It leaves both simultaneously. The crypto desk is watching the AI capex story because it is a leading indicator for its own liquidity. That editorial choice is itself a macro signal, and almost no reader will catch it. And note the messenger. Goldman is a lead underwriter and trading counterparty across the very AI ecosystem whose capex it is auditing. A skeptical report from a conflicted source is not automatically wrong — if anything, the conflict makes a bearish finding more credible, because there is no promotional incentive behind it. But the conflict is real, and neither the wire report nor its crypto echo disclosed it. Under my own audit standard, that omission is a flagged item, and flagged items get priced at a discount until they are resolved. Takeaway: Track the Variables, Not the Headline Do not trade the headline. Trade the variables. Track three things. The depreciation policy disclosures in hyperscaler filings — a lengthening of server lives is a confession, not a convenience. Nvidia's backlog concentration and its data-center customer mix — that is the upstream sensitivity gauge. And the unit economics of inference — the curve of cost per token against price per token, because that spread determines whether the addressable revenue pool expands or erodes. Set thresholds before the data arrives. If AI revenue growth decelerates while depreciation schedules extend, the justification narrative is being managed, not met. If inference pricing compresses faster than cost, the threshold recedes no matter how much is spent. If power interconnection timelines slip, the payback period stretches regardless of how strong the demand signal looks. The $1.7 trillion will be spent. The infrastructure will rise. The only open question — the one the adjective hid — is who services the liability when the revenue arrives late, or arrives at a lower price than the model assumed. That is not a question about AI. It is a question about who holds the bag. And the bag, as always, is leverage.