The Hard Truth Index: Reading the 2025 Tariff Drawdown Against History

On 21 April 2025, the S&P 500 had dropped 18% from its January peak. That made it the worst start to a year in any comparable set of recent drawdowns: worse than 2022's 16%, 2001's 12%, and the 2000 peak-to-trough 10%. By the second week of June 2026, with the index comfortably off the lows and the VIX having travelled from a peak of 52 back to 23, the question on every trading desk was no longer "is this a bear market" but "what kind of drawdown was it, and what does the answer tell us about what comes next."

A framework circulated by the CO2/Coinbase public-markets desk — and discussed at length on TBPN on 19 June 2026 — has begun to settle that question. It draws a sharp line between corrections and crises using two simple tests: the depth of the drawdown, and the depth of the earnings cut that accompanied it. Crises, on the CO2 read, come with peak-to-trough S&P drawdowns of 49% (dot-com), 57% (global financial crisis) and 42% (2022 inflation shock), paired with earnings-per-share declines of around 35% and recovery times measured in years. Corrections, by contrast, stay inside a 25% drawdown band, eat only single-digit percentages out of forward EPS, and tend to clear in months. The 2025 tariff episode, on the data, sits firmly in the second column.

That framing matters because the surrounding sentiment told a different story — loudly, and in real time. Polymarket's implied probability of a US recession in 2025 climbed from 22% in February to 66% the week of the trade-war escalation, before drifting back to roughly 30% by mid-June 2026. Survey-based bearishness across investors, businesses and consumers hit a 25-year peak. And the gap between that sentiment and the underlying hard data — payrolls, retail sales, credit-card volumes — widened to one of the largest readings on record.

What the hard data actually said

It is worth being precise about what the hard-data tape showed, because the prevailing narrative got ahead of it. Visa, Capital One, Bank of America and American Express all reported consumer spend holding up through the drawdown. Restaurant traffic in Los Angeles, anecdotally, did not behave like the consumer-confidence survey suggested it should. A new vocabulary — the "vibe session," a phrase coined by market commentator Kylo Scanlan — has emerged precisely to describe the gap between the mood music and the receipts.

The mechanism here is not mysterious. Tariffs are, in the language of the CO2 note, "government-manufactured" shocks: they can be imposed in a week and unwound in a quarter. The dot-com bust required a generation of capital to burn through. The 2008 financial crisis required the securitisation complex to actually break. The 2025 episode required, at most, an executive-order pen. The path-dependence of the shock is part of the shock. And on the CO2 read, an 18% drawdown on a 3% earnings hit does not meet the historical threshold for a crisis — even when the sentiment around it screams otherwise.

A stress test for the AI capex story

If the macro frame is a correction, the AI capex story is its own stress test. Nvidia's Q1 fiscal results, reported in late May 2026, set the bar: $44 billion in quarterly revenue, up 69% year-on-year, with another $2.5 billion of H20 chips sitting in inventory that the company could not ship to China and a projected $8 billion of forgone China revenue in the current quarter. The data point investors are still chewing on, however, is not the headline. It is the workload mix.

On the call, CFO Colette Kress and CEO Jensen Huang both described a "sharp jump" and a "step function" in inference demand. Ben Thompson, writing at Stratechery and cited on the TBPN discussion, pointed out that the actual unit mix — roughly 40% inference, 60% training — has not moved since the prior quarter. That gap is now the single most-watched data point in the AI infrastructure trade. Either the orders are coming but have not hit revenue yet, or the narrative is running ahead of the deployment curve. Microsoft's own quarterly disclosure — 100 trillion tokens processed, 5x year-on-year, 50 trillion in a single month, with 70% of the Fortune 500 now on Copilot, per Satya Nadella on the 30 April earnings call — argues for the first reading. The capacity-constraint warnings Microsoft issued for the back half of the year point the same way.

The broader capex tape reinforces it. Combined 2025 capital-expenditure guidance from Amazon, Microsoft, Google, Meta, Oracle and Tesla now stands at roughly $365 billion, up from around $213 billion the prior cycle — a 70% year-on-year increase, of which Tesla accounts for about $10 billion and Oracle around $16 billion. The remaining $300 billion-plus is concentrated in the four hyperscalers. Anthropic's run-rate revenue, by way of comparison, climbed from $1 billion to $3 billion in five months — a pace that would have been industry-shocking at Snowflake scale and is now treated as the baseline. Even allowing for hype, the underlying consumption curve is unlike anything the consumer-internet cycle produced at the equivalent months-post-launch. OpenAI's user base, per Sam Altman, doubled from roughly 400 million to 800 million in a matter of weeks following the deep-research and Operator launches. The CO2 report's framing — "tokens are more important than tariffs" — is not a slogan; it is a description of where demand is actually showing up.

The export question, and the moat that keeps moving

What the AI capex cycle is also doing, structurally, is dragging US-China technology competition out of the tariff lane and into the export lane. Jensen Huang used the Nvidia call to make the case directly: "The question is not whether China will have AI. It already does. The question is whether one of the world's largest AI markets will run on American platforms." That is a pro-export turn for a company that, four years ago, held 95% of the China accelerator market and today holds 50%.

The hardware moat, meanwhile, is being extended in two directions at once. CUDA remains the programming-layer lock-in. The new NVLink Fusion platform, unveiled alongside the earnings, lets customers stitch Nvidia GPUs together with non-Nvidia CPUs and accelerators — Trainium, Inferentia, TPUs, MTIA — through Nvidia's own networking fabric. The pitch on the call was aimed at hyperscalers. The real target, on the read of analyst Aaron Gin, is the fabric layer itself, and the longer-term goal is to keep Huawei out of it.

Gin — and Tan Khoshkish of Beacon Global Strategies, in a separate segment — made the strategic case bluntly: the United States must behave like Boeing did for the commercial-aviation stack and become the default AI exporter, or it will cede the global stack to Huawei by default. The dual-use argument against exports, in Gin's framing, is "an argument from ignorance" — China has functioning missile technology without US AI silicon. The current export regime, in this view, is already partially leaky: TSMC is alleged to have shipped roughly 2.9 million dies to Huawei — enough for around one million Ascend 910B/C chips — in violation of US controls. Huawei's yields on those dies, per Beacon Global's reading, sit around 20% of Nvidia's. The gap is real, but it is not eternal.

The bond-market ceiling

Even with a correction-grade equity drawdown and a runaway AI capex cycle, the macro frame has a fiscal ceiling. The Congressional Budget Office and the Committee for a Responsible Federal Budget have estimated that the "One Big Beautiful Bill" adds roughly $3 trillion to US debt over the next decade, and $5 trillion if its temporary provisions are made permanent. Federal interest expense, per recent Wall Street Journal reporting, is now running above the defence budget. Berkshire Hathaway's cash pile has climbed to $350 billion. The 10-year Treasury yield sits around 4.4%. Jamie Dimon, the JP Morgan chief executive, warned in April 2025 that "you're going to see a crack in the bond market." Ray Dalio, in a Bloomberg interview tied to his new book How Countries Go Broke, put a tighter clock on it: "I give America three years, give or take a year, to avert an economic heart attack."

Whether the bond market cooperates with the AI capex cycle is, on this evidence, the bigger question than whether 2025 was a correction or a crisis. The historical record says it was the former. The forward-looking tape says the next test will not be a drawdown at all — it will be a Treasury auction.

Kicker: A correction cleared. A capex super-cycle is still clearing its own milestones. The next stress test will not be in equities — it will be in the long bond, and the clock on that one is shorter than the equity market's mood suggests.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://www.youtube.com/watch?v=9ZGoLX_QjZM