Consider that a non-disclosure agreement is a liability that has not been marked to market yet. Most read AWS's decision to walk away from confidentiality agreements with government agencies as an ESG gesture — a responsible infrastructure company voluntarily opening the books. That reading is comfortable and almost certainly incomplete. When a hyperscaler holding roughly a third of global cloud infrastructure decides that secrecy is no longer worth the carrying cost, it is not making a moral statement. It is repricing a financial instrument. The instrument is the tax abatement, the subsidized power contract, the discounted land lease — all of which live inside the same envelope as the NDA. I have spent enough hours reading the fine print of incentive packages to know that confidentiality is rarely about privacy. It is about leverage.
To understand why this matters, you need the mechanics. A hyperscale data center is a capital-intensive, long-horizon asset. Power typically represents 30 to 50 percent of operating cost, and the difference between a market-rate utility tariff and a negotiated industrial rate can swing the internal rate of return of a multi-billion-dollar campus by several percentage points. That gap is not closed in public. It is closed in negotiation, and negotiations are governed by mutual confidentiality. The NDA is the wrapper around the deal — it protects the developer's competitive positioning and the jurisdiction's political cover.
Add AI. The current compute buildout is the largest concentrated infrastructure program since the interstate highway system. Demand for training clusters has pulled forward siting decisions, accelerated interconnection queues, and pushed data center electricity consumption onto grids that were never designed for it. Local governments — not federal regulators — are the ones absorbing the externalities: water draw for cooling, diesel backup emissions, substation load, noise. The push for transparency is coming from that layer, from county boards and state utility commissions, not from Washington. That is the crucial structural point. The regulatory pressure here is local, fragmented, and increasingly non-negotiable, and that is a harder constraint to manage than any single federal rule.
Strip the NDA to its function and it behaves like a settlement layer. Two parties exchange commitments they cannot verify externally, and the confidentiality clause is the consensus mechanism that keeps the ledger private. It works — until it does not. The moment a third party, a journalist or a county auditor or a community group, demands visibility into the state, the entire structure becomes a liability. This is the defining property of any system optimized for opacity: efficient in equilibrium, catastrophic in transition.
I have audited smart contracts with exactly this failure mode. In 2017, working through the Uniswap V1 core contracts during the ICO boom, I found an integer overflow in the price calculation path — a vulnerability invisible in normal operation and lethal at the boundary. The lesson generalized: systems fail at the edges, and the edges are precisely where disclosure is weakest. Government NDAs are an edge. They sit at the boundary between a private company and a public jurisdiction, and they were never designed to be stress-tested by a transparency campaign.
Now map the interdependencies, because the cost of dropping these agreements does not land in one place. It cascades.
The confidentiality layer is first. If the NDA is severed, the terms it protected — the abatement schedule, the power rate, the land basis — become discoverable. Discoverability removes the asymmetry that made the deal favorable to the developer.
The incentive layer follows. Local incentives are politically fragile. Once one jurisdiction's package is public, neighboring jurisdictions face pressure to match or beat it in the open. The race-to-the-bottom conducted in private becomes a race conducted in daylight, and daylight is expensive.
Then the unit economics layer. Recompute the campus IRR without the preferential tariff and the model degrades. Not catastrophically — AWS can absorb it — but enough that the marginal siting decision shifts.
The siting layer absorbs the rest. Capital reallocates toward jurisdictions with stable, predictable disclosure regimes. Volatility in the incentive environment is itself a cost, independent of the headline numbers.
And beneath all of it, the social license layer — the one the industry systematically underprices. Community opposition can delay a project by years. A transparency commitment is a cheap option purchased against that risk.
Here is where my domain intersects. Crypto's competitive thesis against hyperscale compute is not cost. It is verifiability. The pitch from decentralized compute networks is essentially: you should not have to trust the operator's benchmark, the operator's uptime claim, or the operator's pricing — you should be able to verify it. AWS walking away from NDAs is a partial, grudging concession to that same logic, applied to a different variable. The hyperscalers are being dragged, slowly, toward the verifiability standard that on-chain systems were built on from day one. That is not a victory for decentralization. It is evidence that the standard is correct and the incumbents are adopting it defensively.
I want to be precise about the data availability analogy, because it is overused. There is a tendency to frame every disclosure problem as a DA problem and reach for modular DA layers as the answer. Most rollups do not generate enough data to justify dedicated DA, and the same discipline applies here. AWS does not need a blockchain to disclose a tax abatement. It needs a policy and a filing. Reaching for cryptographic overkill where a spreadsheet would do is a category error, and this industry makes it constantly. Composability is a double-edged sword — and so is the reflex to solve every trust problem with a proof system.
Where the crypto analogy genuinely holds is the verification of AI compute itself. In 2026 I worked on a framework for verifying AI model outputs on-chain using ZK-SNARKs, cutting proof generation time by 40 percent. The institutional demand driving that work is identical to the demand driving data center transparency: capital allocators and regulators who no longer accept unverifiable claims. The data center NDA fight and the verifiable-compute fight are the same fight. One is waged in county planning meetings, the other in circuit constraints.
The consensus framing is that AWS did something good. The blind spot is that transparency at scale is a moat, not a gift. Here is the mechanism. AWS operates at a scale where absorbing a cost increase is survivable. A tier-two cloud provider or an emerging AI compute startup is not. If voluntary disclosure becomes the de facto standard — and the incentive to codify it into local legislation is strong once one large player complies — then every operator inherits the compliance burden. The burden is fixed cost. Fixed costs favor incumbents. This is not transparency as altruism; it is transparency as a barrier to entry, and the industry is applauding the construction of its own wall. The second blind spot is that not every NDA can be dropped. Some protect critical infrastructure security, and the line between being transparent and exposing attack surface is the real unresolved question. Architects build, auditors break — and the auditor's first question is always which of these agreements actually protects a secret that matters.
The signal to watch is not the announcement. It is whether Azure and Google Cloud follow, and whether a state legislature converts a voluntary gesture into a mandatory filing requirement. If both happen, the cost curve for everyone building compute — hyperscaler, neocloud, and decentralized network alike — reprices upward within eighteen months. The question is not whether AWS meant it. The question is what happens to the projects that cannot afford to mean it. Speculation audits the soul of value, and right now the market is pricing transparency as a virtue when it may be pricing it as a moat.
