When the AI trade wobbled this week, the one asset that is supposed to zig while equities zag did neither. Gold slid below $4,100 — not because the metal lost any shine, but because investors hemorrhaging money on technology stocks sold their bullion to cover losses everywhere else. In the same stretch, South Korea's benchmark fell more than 10% in a single session, dragged down by chipmakers, while the Nasdaq shed 2.2% on renewed fears that the AI-driven equity rally has run too far. By the next morning, Asian markets were already rebounding, with traders waving it off as "technical factors."
Markets gyrate; that is their job. But the mechanics of this particular selloff carry a lesson that has little to do with your portfolio and a great deal to do with your risk register. When one thesis grows large enough — here, the durability of the AI boom — it stops being a single risk among many and becomes the variable that prices everything else. Diversification quietly stops working. And the same dynamic is now reproducing itself inside organizations, several layers down, in exactly the places most governance frameworks were never designed to inspect.
When your hedge and your bet move together
The textbook definition of concentration risk is having too much exposure to a single factor. The textbook fix is diversification: spread the bets so no one failure can sink you. What this week demonstrated, in public and at speed, is how AI has quietly become the common factor across assets that used to be unrelated. Korean semiconductors, U.S. megacaps, and a traditional safe-haven metal all moved together because they were all, directly or indirectly, expressions of the same wager.
That correlation is the tell. When the thing you own to protect yourself falls in lockstep with the thing you own to grow, you never actually held two positions — you held one position twice. Governance professionals should sit with that image, because it is not confined to trading desks.
Concentration didn't stay in the market. It moved into the stack.
Look at how enterprises actually build with AI today, and the same single-factor exposure appears — just dressed as architecture instead of allocation.
Consider the steady consolidation of the AI developer toolchain. This week Cursor quietly acquired Continue, one of the more prominent open-source alternatives to GitHub Copilot. Open-source options exist in part to be a hedge — a way to avoid being captive to a single commercial vendor. When the hedges themselves get absorbed into larger players, the menu of genuinely independent choices shrinks, and more of the industry ends up standing on the same few foundations.
That matters more than it would for ordinary software, because AI fails differently. As Gray Swan's Matt Fredrikson and OpenAI board member Zico Kolter argued in a recent conversation on AI red-teaming, securing these systems is not "cybersecurity with AI" bolted on — the attack surface is native to the model itself: adversarial inputs, jailbreaks, and failure modes with no equivalent in traditional software. The governance implication is uncomfortable. If a single model underpins your coding assistant, your customer service, and your internal copilots, then a single adversarial weakness is not contained to one app. It is a shared failure mode across everything built on top of it. One model, many doors.
Your third parties are making the same bet you are
Now widen the lens to your vendors — and to the half of AI exposure that lives outside your walls entirely.
The third installment of Michael Volkov's series on AI and compliance makes the point bluntly: if you concluded that AI governance is mostly about what your employees do with AI inside your organization, you have identified roughly half your exposure. The rest lives in your third parties — the suppliers, processors, and service providers quietly embedding AI into the work they do for you. Most compliance programs cannot see it, because vendor questionnaires were written for a world where you asked who held your data, not which model wrote your contractor's code.
Here is where concentration becomes genuinely invisible. You can run a mature third-party risk program, diversify across a dozen suppliers, and congratulate yourself on resilience — while every one of those suppliers quietly standardizes on the same two foundation models and the same handful of cloud providers underneath. On paper you are diversified. In reality you have twelve vendors and one point of failure. The correlation that took down gold and chipmakers in the same session is the correlation hiding in your supplier base, and almost no third-party risk register is structured to surface it.
A few questions worth adding to your next vendor review:
- Which foundation models and inference providers does this vendor depend on — and do those overlap with our own stack or our other vendors'?
- If that model or provider were unavailable, degraded, or compromised for 48 hours, what of ours stops working?
- Does the vendor's own AI usage fall inside the contractual and regulatory commitments they've made to us?
Prevention was never the plan. Resilience has to be.
This is why a quiet shift in the security world is the right instinct for governance more broadly. As one widely shared argument put it this week, cybersecurity is "no longer about protection — it's about survival": the mature posture assumes the breach will happen and optimizes for surviving it, not for the fantasy of preventing every one. That mindset transfers directly to concentration risk. You will not prevent the consolidation of the AI market, and you cannot force your vendors to diversify their model choices. What you can govern is your own ability to absorb a failure when the common factor cracks.
A word on false comfort, too. Elon Musk's SpaceX spent this week adding billions in debt while simultaneously cutting its annual interest cost — a feat Bloomberg fairly called financial alchemy. It is a useful reminder that risk can be restructured to look smaller on the surface while remaining exactly as large underneath. The same illusion shows up in AI governance: a tidy vendor list, a signed model card, a passed audit. None of it tells you whether your apparent diversity collapses to a single dependency one layer down.
What governance teams should do before the next wobble
Seventy years into the field — IEEE marked the anniversary this month by noting that AI's rate of adoption has been "unprecedented" relative to earlier general-purpose technologies — the governance problem is no longer whether to adopt, but how to avoid quietly concentrating everything on the same handful of providers without noticing.
Three moves separate the organizations that will weather the next correlated shock from those that will discover their exposure during it:
- Map the common factor. Build a single inventory that crosses the boundaries your org chart usually keeps apart — which models, chips, and clouds sit beneath your products, your tools, and your critical vendors. Concentration only looks small when you measure each domain in isolation.
- Stress-test the dependency, not the device. Ask what breaks if one model family or one provider goes dark, the way you already ask what breaks if a data center fails.
- Govern for survivable failure. Define, in advance, the manual fallbacks and second-source options that let critical processes limp along when the shared factor cracks.
The selloff will be forgotten by next quarter. The structure it exposed will not. When the asset you bought for safety and the asset you bought for growth fall together, the lesson is identical whether you sit on a trading desk or in a compliance function: you do not get to claim diversification you have not actually verified. The teams that internalize that now — and rebuild their inventories around the common factor rather than the org chart — are the ones still standing when the AI thesis gets tested for real.
Sources
- Cybersecurity is no longer about protection. It’s about survival. — CSO Online
- Cursor quietly acquires Continue, an open-source alternative to GitHub Copilot — The New Stack
- The Third-Party AI Blind Spot — The Risk Your Compliance Program Is Not Addressing (Part 3 of 3) — Volkov Law — Corruption, Crime & Compliance
- Red-Teaming after Mythos — Zico Kolter & Matt Fredrikson, Gray Swan — Latent Space (swyx & Alessio)
- Commemorating 70 Years of Artificial Intelligence — IEEE Spectrum
- Tech Selloff Reverberates Through Global Markets — Bloomberg Markets
- Korean Stocks Rebound After 10% Rout With Focus Back on Earnings — Bloomberg Markets
- Markets Recoil in Global Sell-Off Driven by Tech Stocks — NYT Business
- Musk’s SpaceX Adds Billions in Debt While Cutting Interest Costs — Bloomberg Markets
- Gold Drops Below $4,100 as Tech-Led Selloff Spurs Liquidation — Bloomberg Markets