On Tuesday, Tether CEO Paolo Ardoino dropped a bombshell that ricocheted through both AI and crypto circles. During a private investor call whose transcript was obtained by our team, Ardoino warned that the AI industry’s current capital expenditure model – massive GPU purchases financed by debt or equity while offering subsidized compute to users – is structurally unsound. "Assets depreciate in 3 to 5 years, but the profit cycle is mismatched. You’re selling compute below cost to buy growth, and the debt maturity doesn’t align with the asset’s useful life." For a man who runs the world’s largest stablecoin issuer, his words carry weight – and an uncomfortable mirror for crypto’s own history of burning capital to chase users.
Ardoino didn’t name specific AI companies, but the subtext is clear: OpenAI, Google DeepMind, Anthropic, and Microsoft are spending tens of billions on NVIDIA H100 clusters while their API revenue growth lags behind the depreciation curve. Open-source models like Llama 3 and Mistral continue to erode pricing power, compressing margins further. "It’s a classic capital structure mismatch," he said. "The same mistake we saw in crypto lending in 2022 – long-term assets funded by short-term liabilities, or in this case, revenue streams that don’t cover the cost of capital."
The timing is notable. Tether itself has been expanding into AI – investing in mining and compute infrastructure – and recently announced a partnership with a Northern European data center operator. Some analysts speculate that Ardoino’s warning serves a dual purpose: positioning Tether as the prudent voice while potentially lowering competitor valuations for future acquisitions. But even if self-interest is at play, the underlying logic deserves rigorous examination.
Hook Paolo Ardoino, CEO of Tether, told investors in a closed-door briefing that the AI industry’s "subsidized computing power" strategy mirrors the flawed capital structures that led to crypto’s 2022 credit crisis. He highlighted the rapid depreciation of GPU assets – 3 to 5 years – against revenue cycles that may never align. The comment triggered a 2.4% drop in NVIDIA shares during after-hours trading and renewed debate about whether AI giants are repeating the mistakes of DeFi’s yield-farming era.
Context Tether, with over $100 billion in stablecoin market cap, holds significant reserves in U.S. Treasuries, Bitcoin, and gold. Ardoino has increasingly positioned himself as a macro commentator, often comparing crypto and traditional finance risks. His latest target is the AI sector’s capital allocation. The AI industry’s capex has exploded: NVIDIA alone reported $22.4 billion in data center revenue last quarter, much of it from hyperscalers. Yet user pricing for inference APIs has declined by over 40% year-over-year as open-source models commoditize basic capabilities. This dynamic – falling revenue per compute unit combined with fixed hardware depreciation – creates what Ardoino calls a "profit cycle mismatch."
My own experience auditing DeFi protocols during the 2020 insanity gave me a front-row seat to similar dynamics. Compound Finance’s governance model, which I analyzed in "The Illusion of Infinite Yield," showed how token incentives could mask structural insolvency until liquidity dried up. The AI industry is not yet at that point, but the pattern is familiar: subsidize to gain market share, then struggle to raise prices when the subsidy ends.
Core Let’s break down the three components of Ardoino’s warning, backed by on-chain and financial data where available.
First, the asset depreciation problem. A single NVIDIA H100 GPU costs approximately $30,000 and has a typical useful life of 3-5 years before becoming obsolete for leading-edge AI workloads. Hyperscalers like Microsoft and Google have bought hundreds of thousands of these units. Using straight-line depreciation, that’s $6,000 to $10,000 per GPU per year in cost, not including power, cooling, and staffing. Meanwhile, the revenue generated per GPU varies wildly – some estimates suggest that an H100 used for inference can generate $10-15 per hour, but utilization rates are rarely above 60% in practice. At a conservative 40% utilization and $10/hour, annual revenue per GPU is ~$35,000 – which seems to cover depreciation. But that assumes the operator captures full price. In reality, subsidized pricing means selling compute below market. OpenAI’s GPT-4o API costs $2.50 per million input tokens, which translates to roughly $0.08 per hour of compute. At that rate, you need massive volume to approach break-even. Ardoino’s point is that the industry is effectively selling compute at a loss to build user base, banking on future price increases that may never materialize due to open-source competition.
Second, the profit cycle mismatch. AI companies typically raise large rounds of equity or debt to fund capex. When assets depreciate faster than revenue growth, the company must raise more capital to maintain the same level of service. This is akin to a Ponzi-like dynamic, though not fraudulent. In crypto, I saw this firsthand during the 2022 Terra collapse: the anchor protocol promised 20% yields, funded by new depositors. When withdrawals exceeded deposits, the peg broke. Here, the "yield" is cheap compute, and the "deposits" are venture capital. If investor sentiment sours, funding dries up, and the subsidy stops – causing user churn and further revenue decline.
Third, open-source erosion. Ardoino specifically called out open-source AI as a threat. "Open source keeps eroding revenue, because the marginal cost of serving a user with an open model is near zero, and they don’t need to recoup billions in R&D." This is a variant of the same problem crypto faced in 2021: L1 blockchains like Solana and Avalanche offered subsidized transactions to attract developers, but Ethereum’s larger ecosystem and L2 rollups eventually provided cheaper, more secure alternatives without the subsidy hangover. The AI market is similar: companies like Meta release Llama 3 for free, and startups like Mistral provide competitive models at low cost. The AI giant’s only moat is scale – they can afford to burn cash longer than their competitors. But that’s not a moat; it’s a tolerance for pain.
To quantify the risk, I built a simplified model using public data from Microsoft’s 10-K and OpenAI’s reported revenue. Microsoft disclosed $44.6 billion in property, plant, and equipment (PPE) additions last fiscal year, much of it for AI. Assuming a 5-year useful life, annual depreciation is ~$8.9 billion. OpenAI’s annualized revenue is about $3.4 billion. Even if all revenue came from Microsoft’s Azure AI services, that’s less than half the depreciation. Of course, Microsoft uses the GPUs for internal products too, but the subsidy to OpenAI – $13 billion total investment – means the ROI is back-loaded. The ledger doesn’t lie: current revenue doesn’t cover hardware costs.
Contrarian Despite the grim picture, there’s a counter-argument that Ardoino’s warning is either premature or self-serving. First, the AI giants have diversified revenue streams beyond API calls. Microsoft sells Copilot subscriptions ($30/user/month) with high margins; Google integrates AI into its cloud and search advertising. The GPUs are not stranded assets – they also power internal automation and product improvements. Second, the depreciation cycle may be overestimated. If the next generation of chips (Blackwell) enables 5x performance improvement, older H100s can be redeployed for non-critical inference at lower cost, extending useful life. Third, Tether itself is a controversial messenger. The company has faced regulatory scrutiny over reserve transparency, and its recent push into AI mining gives it an incentive to talk down competitors. Some analysts argue that Ardoino’s comment is a strategic attempt to scare capital away from AI giants and toward Tether’s own infrastructure partnerships.
Moreover, the comparison to crypto’s 2022 credit crisis is flawed. Crypto lenders like Celsius and BlockFi had opaque balance sheets and no real revenue. AI companies have actual paying customers – even if subsidized. The risk is more akin to Amazon’s early years, where the company burned cash for over a decade before turning profitable. Bezos famously said, "Your margin is my opportunity." AI giants may be betting that when the subsidy ends, the scale will allow them to raise prices without losing users, similar to how AWS eventually raised prices after dominating the cloud market.
However, that analogy breaks down when you consider open-source competition. AWS’s competitors (Azure, GCP) also had to build similar infrastructure; there was no free, open-source AWS equivalent. In AI, the open-source models are good enough to run on commodity hardware. If a startup can self-host Llama 3 for a fraction of the API cost, the AI giant’s pricing power is capped permanently. This is the "commoditization of intelligence" that many analysts predicted.
Takeaway Ardoino’s warning should not be dismissed as FUD, nor should it be treated as a death knell. The market will ultimately punish those who underprice risk. For crypto investors, the lesson is clear: when a sector relies on continuous capital inflows to sustain artificially low prices, the first whiff of a funding winter can trigger a cascade. I will be watching NVIDIA’s next earnings report and the API pricing movements closely. If open-source models continue to close the performance gap, the AI giants will face a choice – raise prices and lose users, or keep burning and hope for a technological miracle. Either way, the ledger will tell the story.
Ledgers don’t lie. The code is the contract. Data is the only truth.