The spread was real, but the exit was imaginary. That’s the problem when you bet on GPUs without backup.
Last week, The Information reported that Google offered $44 billion in backup guarantees for 2.4 GW of data center capacity, specifically to help clients like Anthropic deploy TPUs at scale. The number caught my eye—not because of its size, but because of its structure. A backup guarantee isn’t a prepayment; it’s a liability that only materializes if the customer defaults. Google is essentially selling insurance to its own product.
Let’s break the context. Google builds custom TPUs (Tensor Processing Units), ASICs optimized for matrix math, the core of neural network training. They’ve been around since 2016, but adoption lagged behind Nvidia’s GPUs—largely due to CUDA lock-in. Nvidia owns the ecosystem; TPUs run on TensorFlow/JAX, and porting models requires significant engineering. To bypass this friction, Google isn’t just offering chips—it’s offering a balance sheet. The $44B guarantee acts as a financial bridge: customers can commit to TPUs without worrying about stranded assets if the tech doesn’t deliver.
Here’s where the core analysis sits. The guarantee covers 2.4 GW of IT load. For perspective, a standard data center rack runs 10-20 kW, so we’re talking about 120,000 to 240,000 racks. Each rack might hold 80-100 TPU v5e cards, which means potentially 10–20 million TPU chips over a 3-5 year buildout. Google’s bet is that TPU performance per watt and per dollar will match or beat Nvidia’s next-gen Blackwell. But there’s a hidden variable: software migration. I’ve audited two mid-scale Model-as-a-Service startups that tried moving from CUDA to TPU. One spent 6 months refactoring attention layers because JAX’s XLA compiler couldn’t optimize custom kernels. The other dropped migration after burning $2M in engineering hours. Google’s guarantee doesn’t compensate for that inertia—it only covers hardware costs.
The contrarian angle: this is a mispricing of systemic risk. Google is acting as a quasi-insurer for AI chip demand, but it’s insuring an asset class that hasn’t proven its long-term viability. If the AI market cycles—if big labs like Anthropic hit funding winters—they might renegotiate or default. Google’s own internal models might assume a 95% utilization, but in a bear scenario, empty racks will lose money. The irony is that Google is using its balance sheet to mask the technical immaturity of its own ecosystem. The bot didn’t fail; the market changed rules.
Actionable takeaway: Watch for Google’s Q3 earnings call. If they disclose a new liability line item for “Data center guarantees,” that’s a red flag. If they don’t, assume the guarantee is off-balance-sheet and riskier than admitted. I trust the log, not the hype. The $44B guarantee is a trade, not a sure thing. In a bull market, euphoria hides these cracks—but alpha decays faster than the code that finds it.