It was 2 a.m. in Paris when I read the piece for the third time, and by then I had stopped looking for the summit.
The article carried a headline about AI safety raising alarms at a conference in Oxford. It had a summary. And the summary was the body — word for word, character for character, a perfect mirror. No byline. No date. No venue. No speaker. No number. No quote. One verifiable fact buried inside five paragraphs: an AI summit happened at Oxford. Everything else was assertion wearing the costume of reporting.
I hold a degree in cybersecurity and I have spent twenty-one years watching this industry dress its own noise up as signal. So I did what I always do. I tried to find the event. Oxford Martin School? The Internet Institute? Saïd Business School? The Union? A third-party commercial conference renting the word "Oxford" the way a startup rents a Mayfair address? I could not confirm a single one of them.
And then there was the sentence that kept me awake: safety concerns "may affect market valuations." No evidence. No mechanism. No subject — valuations of what, exactly?
That gap is the entire story. Because the sentence is directionally correct and structurally empty at the same moment, and the distance between those two things is precisely where your money is sitting right now.
Let me give you the ground truth the ghost summit skipped.
The EU AI Act entered into force in August 2024. Prohibited practices landed in February 2025. General-purpose model obligations arrive in August 2025. Penalties climb to 7% of global turnover or €35 million, whichever bites harder. Alongside it, NIST's AI Risk Management Framework and ISO/IEC 42001 have already turned "governance" into something a third party can audit, a contract can reference, and a regulator can inspect on a Tuesday afternoon.
That is genuinely new. For a decade, "AI safety" was a conference word — a promise, a panel, a glossy whitepaper, a handshake. Now it has a technical form. It has paperwork. It has a line item with a decimal point.
And the liability is no longer theoretical. In 2024, Air Canada was held responsible for misleading answers generated by its own chatbot — the company, not the model vendor, paid. Also in 2024, an engineering consultancy in Hong Kong lost roughly $25 million to a deepfake video conference. Read those two cases side by side and you find the only empirically proven link in the entire "safety to valuation" chain: when AI fails, the financial consequence lands on the enterprise balance sheet, not on the supplier's.
Meanwhile the global picture is divergence, not convergence. Europe tightens. The US federal layer pivoted in 2025 toward deregulation and an "AI Action Plan," with EO 14110 revoked, while states like Colorado and California push their own frontier rules. China keeps its dual track of model filing and content safety. For a multinational, compliance cost is the sum of the strictest jurisdiction, not a single standard. That explains boardroom anxiety far better than any summit communiqué ever could.
And here is the reverse test the ghost summit never ran. Enterprise adoption of generative AI kept climbing through 2024 and 2025 — McKinsey's surveys show most organizations already using it in at least one function. At the same time, Gartner put post-proof-of-concept abandonment near 30%. Both numbers are true at once. Which means safety is not a brake. It is a filter — it kills the use cases that cannot pass a compliance review, not the overall spend.
Now the part that actually matters in a bear market, where the only honest question is which protocols are bleeding and whether your assets are safe.
The word "safety" in that article is a container holding four completely different things: frontier catastrophic risk; content and brand compliance; data privacy and IP infringement; and plain, unglamorous IT security — prompt injection, model supply chain compromise, agent overreach. Each of those four has a different buyer, a different budget line, a different regulator, and a different time horizon. Compressing them into one word is how you manufacture a headline. It is not how you build a valuation.
The real migration nobody wrote about is that governance is moving out of the compliance department and into procurement and legal. Enterprises now demand model cards, evaluation reports, penetration tests, and data provenance statements from every vendor they touch. Governance capability is becoming a B2B sales credential — an admission ticket — rather than a cost center. That single shift is the most underreported industrial fact of the year, and the ghost summit never grazed it.
Which brings us to the four channels through which AI safety actually reaches a valuation. None of them operate through sentiment.
Channel A — compliance cost hits operating margin. High-risk system obligations under the EU AI Act bite in 2026 and 2027. Compliance spending converts from a project expense into a recurring expense. That is a permanent margin haircut for exposed companies, and it does not reverse when the fear fades.
Channel B — liability exposure is an unpriced tail. A single deepfake fraud or a model-induced misstatement can cost eight figures, as the 2024 cases already prove. Yet almost no enterprise carries an AI liability reserve, and almost none has insurance. This is the tail that nobody has bothered to price, and tails are exactly what wipe out leveraged books in a bear market.
Channel C — procurement cycles delay revenue recognition. Longer sales cycles, with large enterprises stretching from months to more than half a year, are a cash-flow weapon aimed at application-layer companies. Infrastructure barely feels it. Safety risk is distributed wildly unevenly inside the AI stack.
Channel D — the capex narrative is fragile. Compute valuations rest on the assumption of accelerating enterprise demand. If adoption shows a visible inflection, the first thing repriced is the compute chain, not the application layer — precisely because compute has been priced as perpetual growth and has no room to disappoint.
Now the uncomfortable part. In mainstream valuation models, "safety" carries almost zero weight. Sell-side models obsess over compute supply, token prices, and ARR growth. Almost none assign a line item or a discount to AI liability and compliance. Which means the variable is priced all-or-nothing: nothing happens, zero impact; a landmark event happens, one-shot repricing. That is not a fundamental input. It is a tail hedge that nobody has bought yet.
The catalysts are specific, and not one of them is a summit. First: a major liability judgment against a foundation model supplier rather than an enterprise user. Second: a large enterprise disclosing material AI-caused losses at 8-K scale. Third: a regulator issuing the first truly large fine. Fourth: a major AI liability insurance denial. Watch those four, not the conference calendar.
And the tradeable mismatch is this: if governance is industrializing, the beneficiaries are evaluation, audit, security software, compliance consulting, and insurance — low-multiple, high-certainty assets — while the market's attention stays glued to compute and applications. That is a cognition gap, not a consensus gap, and those are the only edges worth chasing.
Let me bring my own scars to this table. In 2017 I pitched token models to three exchanges in a single year and learned that speed beats perfection at market entry. In DeFi summer I watched community hype lead price by weeks and wrote the beginner guide that crossed 50,000 views in seven days. In 2022 I watched Terra/Luna erase conviction faster than any model could update, and I learned that emotional resilience is itself a risk parameter. What all three taught me is the same lesson: the story moves first, the accounting moves second, and the people who confuse the two get liquidated in the space between. AI governance is in the story phase right now. The accounting phase — reserves, insurance, indemnity clauses, signing authority — is coming.
And this is where crypto actually connects, though not in the way the crypto media wants you to believe. The technical demands of AI governance — audit logs, provenance, verifiable inference, agent permission control — are precisely the primitives blockchain has been selling for a decade. Verifiability is the bridge. But be careful. I have watched RWA promises for three years and learned that traditional institutions do not need your public chain to solve a compliance problem they can solve with a database and an auditor. The same trap awaits anyone who believes AI governance will be settled on-chain. Institutions will procure verifiability from whoever sells it cheapest and most defensibly — and that is usually not a token.
The honest signal, the one the ghost summit never mentioned, is insurance. Lloyd's syndicates began writing AI performance and liability products from 2024. When underwriters start setting deductibles, exclusions, and mandatory controls for AI liability, that pricing becomes the truest measure of risk in the entire market — far truer than any declaration. Insurance is the place where risk stops being a narrative and becomes a number. It is also where the real cost of the indemnity gap shows up: most model suppliers attach generous exclusions to their compensation promises — "not for high-risk use" — while enterprises want precisely those high-risk uses. That gap is the litigation source of the next cycle.
Here is the angle nobody reported, and it is uncomfortable for the people spreading the "safety slows AI" story.
The biggest beneficiaries of that narrative are incumbent cloud vendors and large systems integrators. Rising compliance complexity pushes enterprises away from self-built, multi-model assembly and toward one-stop, pre-certified platforms. Ask who funds the alarmism. The answer is almost never the auditor.
Second, the squeeze is structural, and it hits small players hardest. Governance cost is close to a fixed cost. As a share of revenue it is several times heavier for an SME than for a giant. The real consequence is not slower AI adoption — it is further concentration of AI capability inside the firms that can afford to be compliant. That is the exact opposite of the democratization story the industry tells itself over drinks.
Third, the structural cause of "safety-washing": a polished framework document costs almost nothing, while a genuine red-team disclosure costs a great deal. So the whole industry rationally prefers the former. Frontier labs all publish tiered risk frameworks — Responsible Scaling Policies, Preparedness systems, Frontier Safety Frameworks — and almost none submit their thresholds, methods, or post-trigger behavior to mandatory third-party verification. Safety commitments are cheap signals, not binding constraints. That is the most fragile link in the entire competitive landscape, and nobody is pricing it. If regulators move toward mandatory third-party evaluation, the capability gap between tier-one and tier-two labs gets partly flattened by compliance standardization — an unintended consequence that reshuffles the whole board.
And the deepest irony. A crypto outlet published an AI safety piece with the body identical to the summary, no byline, no sources, no blockchain content at all. That article is itself a sample of the risk it describes — AI-generated low-quality content at scale. The narrative about safety was produced by the unsafe thing. I do not say that to mock it. I say it because it is the cleanest evidence in the whole story that the signal layer is broken, and that verifiability is not an abstract virtue — it is the difference between a fact and a mirror.
So stop asking whether AI safety will slow adoption. The growth data already answered that, and the answer is no.
Ask the question that actually redistributes money: how will governance cost reallocate profit and admission across the AI stack — and will crypto's verifiability primitives get procured, or quietly routed around?
Volatility isn't what kills you in a bear market. An unpriced tail is. Watch insurance pricing, the first supplier-liability judgment, and enterprise sales-cycle data. Those three will tell you the truth long before any summit ever does. And when the repricing comes — because it will — I may wince at the drawdown, but I won't regret the dance.