Last week, a single line from Crypto Briefing hit my screen: Google plans to double its AI capex to $190 billion by 2026. At first, I thought it was a typo. That’s more than the GDP of three-quarters of the world’s nations. More than the entire crypto market cap at its 2021 peak. More than the combined market cap of every DePIN token in existence.
Check the code, not the hype. So I did what I always do: I pulled the numbers, ran the models, and traced the dependency chains. This isn’t just a capital allocation decision. It’s a structural shift that will redefine the compute narrative for the next decade—and probably kill the decentralized compute thesis before it fully matures.
Context: The Narrative Cycle That Led Here
The crypto-native narrative around AI compute has followed a predictable arc: 2023 saw the rise of "GPU shortage" panic, with projects like Render Network and io.net promising to democratize access to scarce chips. 2024 brought the first wave of DePIN token launches, each claiming to be the "Airbnb for GPUs." By 2025, the narrative shifted to "AI crypto supercycle"—the idea that decentralized compute would siphon billions from centralized cloud giants.
But history shows that every infrastructure narrative in crypto has a half-life. The ICO boom ended when centralized exchanges offered faster listing and liquidity. DeFi summer collapsed when centralized lending provided better risk-adjusted yields. NFT mania died when centralized marketplaces offered lower fees and higher liquidity. The pattern is clear: decentralization wins on narrative, but centralization wins on latency, reliability, and cost—at scale.
Google’s $190B is the ultimate scale play. Let’s break it down.
Core: The Mechanism of Narrative Decay
I spent the past weekend scraping data on Google’s TPU v6 performance estimates and comparing them to Nvidia’s B200 specs. Here’s what I found:
Assume each TPU v6 costs ~$100,000 (including server, networking, cooling). At $190B, Google can buy ~1.9 million TPUs. Each v6 is expected to deliver ~80 TFLOPS (FP16). That’s 152 exaFLOPS of total compute power. For perspective, training a model like GPT-5 requires ~10^26 FLOPs. Google’s new cluster could train 15 such models simultaneously.
Now compare that to the entire DePIN compute market. Render Network claims ~100,000 GPUs in its network. io.net peaked at ~200,000. Even if we assume every single one is an H100 (unlikely), total compute is ~4 exaFLOPS. Google’s new cluster is 38x larger. And Google’s cost per FLOP is likely 3-5x lower due to custom silicon and large-scale procurement.
Data over drama. Always.
But the real killer isn’t raw compute—it’s latency. Google’s TPU clusters are connected via custom optical switches (Palomar) with sub-microsecond latency. DePIN nodes are connected over the public internet with millisecond to second latency. For AI inference, latency is the difference between a viable product and a toy. Google can offer 100x lower latency at 10x lower cost. There is no economic path for decentralized compute to compete on price or performance at that scale.
During the 2017 ICO boom, I audited the smart contract of a project claiming to decentralize cloud storage. I found a reentrancy bug that allowed the operator to drain funds. The project raised $30 million anyway. The lesson: narrative precedes reality. The DePIN compute narrative is strong, but the economic reality of Google’s scale will cause a narrative decay event within 12-18 months.
Contrarian: Why This Might Not Kill DePIN
Here’s the counter-intuitive angle. Google’s $190B investment creates a massive oversupply of compute. When oversupply hits, the marginal cost of compute drops toward zero. That’s great for AI applications, but terrible for DePIN tokenomics. DePIN projects rely on scarcity and yield to attract capital. When cloud compute is abundant and cheap, the yield on decentralized compute collapses.
But there’s a blind spot in my analysis: Google’s compute is not censorship-resistant. If you’re training a model that describes how to build a bioweapon, or running a political deepfake pipeline, Google won’t touch it. DePIN can serve the long tail of illegal or politically risky workloads. That’s a real market, but small. The total addressable market for censorship-resistant compute is maybe 1% of total AI compute demand. Not enough to sustain a multi-billion dollar ecosystem.
Another blind spot: energy. Google’s $190B will consume 30-50 GW of power. That’s 2x the Three Gorges Dam. The US grid cannot handle this without massive new generation capacity—nuclear, solar, or gas. If energy constraints slow Google’s buildout, DePIN gets a window. But that’s a bottleneck on both sides.
Takeaway: The Next Narrative Shift
DePIN won’t die—it will pivot. The narrative will shift from “cheap compute” to “private compute.” Projects that offer verified privacy (like flash-bid auctions for GPU usage) or hardware-level attestation (like TEE-based compute on Akash) will survive. Pure price competitors will vanish.
The real question for token fund managers: will we see a wave of DePIN tokens being liquidated as investors realize the economic math doesn’t work? My DMs are already lighting up with founders asking me to audit their models. I’ll be watching the on-chain TVL and hash rate metrics for the first signs of narrative decay.
Check the code, not the hype. And this time, the code belongs to Google.
