A chief investment officer at one of Asia's largest banks looked into a camera and said Nvidia trades at 17 times forward earnings, so AI stocks are nowhere near a bubble. It was a precise number, and precise numbers are supposed to survive checking. I checked it. Between October 2024 and October 2025, Nvidia never traded at 17x any standard twelve-month forward P/E. The anomaly isn't a rounding error — it's the truth screaming. That number is the load-bearing beam of a trillion-dollar narrative, and it does not hold.
I have pulled on figures like this before. In 2017 I manually traced 14,000 ETH out of the EOS pre-sale contracts and found a 23% gap between reported sales and on-chain liquidity. The lesson carried forward: when a figure cannot be rebuilt from primary sources, it isn't data — it is marketing. That same marketing is now being minted on-chain, and the wallets are telling a story the press release won't.
The claim came from DBS, Singapore's largest bank, and it reached readers through a blockchain aggregator with no author, no source link, and no date. That provenance is the first signal. The piece carries critical metadata gaps — no year, no original interview, no data provider — and those gaps are not incidental. They are the fingerprint of a claim engineered to be repeated rather than verified.
For crypto readers this matters more than it looks. The AI narrative and the crypto narrative have fused. AI-compute tokens, GPU DePIN networks, and agent coins now trade on Nvidia's halo without inheriting Nvidia's cash flow. When a bank CIO declares "AI is not a bubble," the reflexive bid lands on-chain within hours — in tokens whose only link to AI is a whitepaper and a rented cluster. So the valuation question is not academic. It is the upstream of a trade that flows directly into on-chain markets, and the same financing structure driving Nvidia's cycle is being replicated in token form. My 2024 ETF dashboard tracked institutional inflows against on-chain exchange reserves and flagged three corrections by watching that divergence. The mechanism here is identical: narrative at the top, flows underneath, and a gap between them that only wallets can settle.
Here is the reconstruction the source never provides. Nvidia's twelve-month forward P/E was roughly 30x in October 2024 at a ~$122 share price and ~$3.0 trillion market cap. It compressed to 19-22x in April 2025 at the tariff-driven low near $95. By October 2025 it sat at 28-31x on a ~$185 price and ~$4.5 trillion cap. For 17x to hold at $184, implied EPS must be about $10.8 — a figure exceeding every sell-side consensus for FY2026 and FY2027, appearing only in aggressive FY2028 bull cases. The 17x is not a current valuation. It is a conditional valuation: "if the growth all shows up." Setting that beside Cisco's 100x peak multiple is not a comparison. It is a category error.
A second contradiction sits inside the same sentence. If 70% profit growth and a 17x multiple were simultaneously true, the PEG ratio would be 0.24 — a price no functioning market sustains. At least one input is wrong, and the consensus points at the multiple. The quantitative spine of the "no bubble" thesis rests on a single unverified integer.
But the number is the small story. The big story is the financing structure — and that is where I stopped reading valuation and started reading wallets.
Nvidia has committed a $100 billion-scale investment to OpenAI. OpenAI has signed enormous compute contracts with Oracle. AMD struck an equity-linked agreement with OpenAI covering up to six gigawatts of deployment. Follow the arrows: a chipmaker funds a lab; the lab buys compute; the compute buyer commits to more chips. This is the Lucent-Nortel supplier-financing structure of 2000, rebuilt larger and dressed as strategic partnership. In 2000 the customers were debt-financed CLECs whose demand collapsed; Cisco fell roughly 80% from its March 2000 peak and took about sixteen years to reclaim it. In 2025 the customers are cash-rich hyperscalers — that is the bull case — but the incremental financing layered on top is increasingly debt and circular, and that layer is where fragility lives. The source never mentions it.
I have seen this shape before. Tracing the EOS pre-sale in 2017, I found wallets that were simultaneously "investors" and "market makers," recycling ETH to manufacture demand. Clustering exposed the loop because the same addresses sat on both sides of the trade. The AI-compute trade now has an on-chain twin: foundation treasuries fund ecosystem grants, grantees buy back the token, and the token's price validates the treasury. When the same cluster sits on both sides of a transaction, you are not looking at a market — you are looking at a mirror.
The competitive picture the source omits compounds the problem. Nvidia holds 80-90% of AI accelerators, but 2025 delivered three erosion vectors: Google's TPU v6/v7 at scale, AWS Trainium, Meta's MTIA, Microsoft's Maia — and critically, OpenAI partnering with Broadcom on a 10-gigawatt ASIC program, the first real challenge to the assumption that training must run on CUDA. AMD's MI355X and its equity-linked OpenAI deal give it a first top-lab commitment. China is effectively gone: near-zero data-center revenue, H20 impairments, a 15% revenue-share arrangement, with Huawei Ascend absorbing the share. The CUDA moat is shifting from "irreplaceable" to "expensive to leave" — and when your customer spends $100 billion a year, expensive is survivable.
Then the physical ceiling. The 2025 bottleneck moved from GPUs to power — grid interconnection queues of three to seven years, transformer lead times of two to four. Hyperscaler capex runs at $380-400 billion for 2025 and consensus above $500 billion for 2026, against global AI application revenue in the tens of billions. That order-of-magnitude gap is the strongest bubble evidence on the table, and the source never touches it. Depreciation makes it worse: GPUs have a real economic life of two to three years in training workloads while some balance sheets stretch them over five to six. That gap is not an accounting footnote. It is deferred loss.
Now the part I have to say carefully, because it is where analysts overreach. Correlation is not causation, and a closed loop is not automatically a fraud. Vendor financing predates 2000 and will outlast 2025. Some circular structures are rational — seeding demand you intend to supply is legitimate when end demand is real. The on-chain mirror I described is not proof of manipulation; it is proof of opacity. What I can assert more narrowly and more usefully: a low P/E on a cyclical hardware company at peak earnings is not a market error — it is the market pricing mean reversion. The "cheap" multiple is the market saying it expects FY2028-2029 capex to roll over. Using that multiple to argue "no bubble" is circular reasoning. And the barbell advice — gold near records, hedge funds, investment-grade bonds — quietly concedes what the headline denies. You do not hedge a thesis you believe; you hedge one you doubt. Add the structural blind spot: Nvidia is 7-8% of the S&P 500, and the top ten names sit near 40%, a historical extreme. "Nvidia is cheap" and "owning the index is safe" are different questions. And DBS earns from clients staying invested, adding fixed income, and buying alternatives. The recommendation is also a sales framework. Community safety is the ultimate metric of value.
Watch two signals next week. First, any AI-compute token whose top holders overlap with its "strategic partners" — cluster the wallets before you trust the narrative. Second, vendor-financing disclosures: the moment a chipmaker's commitment to a customer exceeds that customer's free cash flow, the loop has stopped being strategy and started being structure. Connecting the dots that others ignore or fear is not pessimism. It is the only way to tell a cycle from a collapse.