The Machines Are Trading Now: What a 47-Second Flash Crash Tells Investors About AI Risk
The Machines Are Trading Now: What a 47-Second Flash Crash Tells Investors About AI Risk
Every institutional client conversation I've had this quarter eventually circles back to the same underlying question, phrased in a dozen different ways: is the AI capital cycle creating value, or is it creating fragility? Having spent two decades watching capital cycles form and unwind, I've learned to distrust any answer that comes too easily. So this piece pulls together five threads I've been tracking separately — how AI agents are prompted for financial decisions, what verified performance data actually shows, how much hyperscalers are betting and what they're getting back, where the Federal Reserve's own economists disagree on AI's inflation effect, and a specific 47-second incident in June 2026 that regulators are still working through. None of these fit neatly into a single bullish or bearish story, and that's precisely the point.
- Five US hyperscalers now guide to a combined ~$725B in AI capex for 2026, up 77% year-over-year — yet free cash flow assumptions are being revised down across the sector.
- In June 2026, 23 uncoordinated autonomous trading agents produced a $500M loss in 47 seconds, reacting identically to the same data without communicating.
- The Fed's own leadership is split on whether AI is disinflationary: incoming Chair Warsh says yes; Governor Cook's own remarks point the other way, with CPI running at 3.8%.
The Case for Making AI Ask Before It Acts
Anyone using generative AI for investment research has hit this moment: you ask a model whether a position makes sense, and it answers immediately — confidently — without ever telling you what interest-rate path, holding period, or risk tolerance it silently assumed to get there. The model didn't ask. You likely didn't notice it should have.
The clearest treatment of this problem I've found isn't from finance at all. It's a 2023 paper out of Tsinghua University and Microsoft Research called "Asking Before Acting" (ABA), originally written for embodied agents — robot arms and virtual agents operating in unfamiliar environments. The finding was straightforward: agents that proactively ask for missing information before acting outperform agents that plow ahead on best-guess assumptions. I'd argue the same logic applies directly to investment workflows. Before an AI model produces a stock recommendation or a portfolio adjustment, it should be structured to surface — not silently assume — the stop-loss threshold, the holding period, and the rate scenario it's working from.
I want to be precise about what this is and isn't. ABA is not an established methodology inside asset management; it's a robotics and agent-research framework that I think transfers well to finance, but the transfer itself hasn't been rigorously tested in a market context. What it offers, even in this early form, is a discipline: force the model to expose its assumptions, and you regain a measure of control that a fluent, confident answer otherwise obscures.
What the Verified Numbers Actually Show
Set the theory aside for a moment — the more interesting evidence sits in South Korea's robo-advisor market, which happens to have unusually good disclosure. Algorithms on the Korea Exchange-affiliated Koscom testbed are required to publish audited, real-money performance, which gives us something rare: verifiable AI trading returns rather than backtested claims.
Quantec's domestic-equity algorithm returned 35% over the year to April 2025, while the KOSPI fell 9% over the same period. Fint's ISAAC Korea Equity solution, run by December & Company, posted a cumulative return of 45.41% through mid-2025, versus 20.81% for the KOSPI200 benchmark over the same window.
| Algorithm | Period | AI Return | Benchmark |
|---|---|---|---|
| Quantec Domestic Equity | 1yr, to Apr 2025 | +35% | KOSPI -9% |
| Fint ISAAC Korea Equity | Cumulative, to Jun 2025 | +45.41% | KOSPI200 +20.81% |
I'd caution institutional readers against reading too much alpha into this. These are specific algorithms over specific windows, not a market-wide average for robo-advisory performance. And the more precise explanation for the outperformance isn't necessarily superior forecasting — it's that the AI executed pre-set rebalancing rules without hesitation during a drawdown, while human investors, as a rule, delay cutting losses. That's a genuine edge, but it's a discipline edge, not necessarily an information edge, and the two carry very different implications for how durable the outperformance is likely to be.
The $725 Billion Question: Is the CapEx Cycle Paying For Itself?
The most direct channel from AI into markets remains capital expenditure. Amazon, Microsoft, Alphabet, Meta, and Oracle have now guided to a combined $660–725 billion in 2026 AI infrastructure spending — up roughly 77% from about $410 billion in 2025. Amazon is the most aggressive at around $200 billion, Microsoft is tracking toward $190 billion at a $37.5 billion quarterly pace, Alphabet has guided to $175–185 billion, and Meta to $115–135 billion.
| Company | 2026 CapEx Guidance | 2025 Actual | Note |
|---|---|---|---|
| Amazon | ~$200B | $131.8B | Mostly AWS data centers |
| Microsoft | ~$190B | ~$88B | $37.5B quarterly run-rate |
| Alphabet | $175–185B | ~$91B | Cloud backlog +55% QoQ |
| Meta | $115–135B | ~$72B | $600B committed through 2028 |
Demand signals are genuinely strong — Alphabet's cloud revenue grew 35% year-over-year and its backlog jumped 55% quarter-over-quarter to $240 billion. But the analyst tone has shifted noticeably. Several banks now model negative free cash flow at Amazon, and there's a broader expectation that sector-wide FCF compresses further into 2027–2028 before recovering. The market's attention is migrating from "how much is being built" to "how fast the build converts into cash" — and that gap is where I'd expect the next round of equity repricing to concentrate.
Where Korea Sits in the Bottleneck
For global allocators, the structurally important point is that the real constraint on this entire capex cycle isn't GPU count — it's high-bandwidth memory (HBM), and that bottleneck runs directly through Korea. SK Hynix held 58% global HBM market share as of Q1 2026, well ahead of Samsung and Micron at roughly 21% each, and the company has effectively sold out its 2026 DRAM, NAND, and HBM capacity. Whichever way the CapEx-versus-ROI debate resolves, the physical chokepoint the money is flowing through doesn't move — and that's a structural variable global portfolios exposed to the AI trade should be pricing explicitly, not treating as a regional side note.
The Fed's Disinflation Debate Is Not Settled — and That Matters for Rates
A lot of AI-and-macro commentary treats "AI drives productivity, productivity is disinflationary" as a settled chain of logic. It isn't, and the disagreement is happening at the top of the Federal Reserve itself, which is precisely why global rate-sensitive portfolios should care.
Incoming Fed Chair Kevin Warsh calls AI "the most productivity-enhancing wave of our lifetimes" and argues it gives the Fed room to cut rates without reigniting inflation — a framing he's compared to the internet's early disinflationary effect. Governor Lisa Cook has pointed to the opposite mechanism: AI investment demand is itself pushing up prices for chips, advanced equipment, construction labor, electricity, and water. The data currently sits closer to Cook's read — CPI and PCE inflation were both running at 3.8% year-over-year as of mid-2026, well above the Fed's 2% target, and a University of Chicago Clark Center/Financial Times survey found 60% of surveyed economists expect AI's effect on inflation and the neutral rate over the next two years to be close to zero.
| The Optimist Case (Warsh) | The Skeptic Case (Cook, most economists) | |
|---|---|---|
| Core claim | Productivity gains create structural disinflation | AI investment demand adds near-term price pressure |
| Evidence cited | Parallel to early internet adoption | CPI/PCE at 3.8%; power, chip, labor cost inflation |
| Market implication | More room to cut rates | Elevated-rate regime persists longer |
My own read, after two decades of watching similar productivity narratives get invoked to justify rate paths, is that this is fundamentally a timing mismatch rather than a factual dispute. The capex bill — for power, land, chips, and labor — arrives now and shows up in prices immediately. The productivity payoff, if it materializes at the scale Warsh expects, typically takes years to show up in aggregate statistics. Right now we are holding the invoice while still waiting on the return.
47 Seconds, $500 Million: A New Category of Market Risk
The 2010 Flash Crash was caused by one large, badly timed sell order from a single algorithm that triggered a roughly 9% intraday plunge in the Dow. What happened in June 2026 was structurally different. Twenty-three autonomous AI trading agents, with no communication or coordination between them, reacted almost identically to the same market data and produced a $500 million loss in 47 seconds. The SEC's characterization was pointed: this wasn't a story about one faulty algorithm, but about a market ecosystem in which thousands of independent, correlated agents can amplify each other's actions on a millisecond timescale.
Regulators are moving, unevenly. At the ECB's Sintra forum in late June 2026, Bank of England Deputy Governor Sarah Breeden confirmed the Bank is running simulations of AI traders executing correlated trades simultaneously and is exploring market-wide "kill switches." The Financial Stability Board has published a 12-point draft framework for responsible AI adoption in finance, open for comment, with a final version due to G20 finance ministers in October 2026. The US, notably, has no equivalent framework in place or announced — and a Wolters Kluwer survey of 230 US banking professionals found 72% could not confirm their institution has the ability to shut down a malfunctioning AI model on demand.
What Twenty Years in This Business Tells Me to Keep
Pull these five threads together and one throughline emerges: AI is generating real, measurable edges across the investment process, but none of those edges survive being taken at face value. The robo-advisor numbers are genuine but sample- and period-specific. The capex build is enormous but its monetization timeline is only starting to be tested by the market. The disinflation narrative isn't settled even inside the institution that would act on it. And the correlated behavior of autonomous trading agents is a risk category the market infrastructure built after 2010 was never designed to handle.
If I had to distill this into two working rules for an institutional desk, they'd be these. First, treat an AI model's confident output as the least interesting part of the answer — the assumptions underneath it are what actually matter, and a model that won't surface them isn't giving you an edge, it's giving you a black box. Second, never conflate "AI makes markets more efficient" with "AI makes markets more stable." Those are different claims entirely, and the 47 seconds in June 2026 are the clearest evidence yet of the gap between them.
[LINK: SK Hynix and the HBM Supercycle — A Deep Dive]
[LINK: Big Tech's Power Crunch and the AI CapEx Cycle]
[LINK: Reassessing Korean Semiconductor Valuations]
This article is published for informational and analytical purposes only. Nothing contained herein constitutes financial, investment, legal, or tax advice. All investment decisions involve risk, including the possible loss of principal. Past performance is not indicative of future results. Readers should conduct their own research and consult a qualified financial advisor before making any investment decisions. The views expressed are solely those of the author based on publicly available information.
Sources & Further Reading
- Futurum Group, "AI Capex 2026: The $690B Infrastructure Sprint" (Feb 2026)
- CNBC, "Tech AI spending approaches $700 billion in 2026" (Feb 6, 2026)
- Axios, "Fed officials warn AI's economic costs may arrive faster than benefits" (Jun 1, 2026)
- Federal Reserve, Speech by Governor Lisa Cook (May 27, 2026)
- The Motley Fool / Charles Schwab, coverage of the Warsh AI-disinflation debate (Jun 2026)
- University of Chicago Clark Center / Financial Times economist survey (Feb 2026)
- Yahoo Finance / 24/7 Wall St., "An AI-Powered Flash Crash Is Coming" (Jul 2026)
- TechTimes, Bank of England AI kill-switch proposal (Jul 2, 2026)
- informedclearly.com, "AI Flash Crash 2026: 47 Seconds, $500M Lost, SEC Responds" (Jun 13, 2026)
- Chen et al., "Asking Before Acting" (arXiv:2305.15695, Tsinghua University & Microsoft Research)
- Hankyung, "35% vs -9%: A Well-Trained AI Beat KOSPI" (Apr 6, 2025)
- rabby.kr, 2026 Korean robo-advisor fee and performance comparison
- The Motley Fool, "Prediction: SK Hynix Will Become the Next Nvidia" (Jul 15, 2026)
Comments
Post a Comment