無法驗證的人工智慧資料中心交易

Buy the Rumor; Sell the News

The AI data center deals that no one can verify

When there's no external validation of multi-billion-dollar deals, how can the market price them?

Over the past year, the AI industry has announced more than half a trillion dollars of infrastructure commitments.

In mature infrastructure markets, you don’t need to read every contract to evaluate these claims. You get a verification layer from markets and standardization: forward electricity prices, standardized capacity definitions, project finance diligence, credit spreads, and, critically, derivatives that force comparable terms. A megawatt means something precise because it has to.

In AI infrastructure, that verification layer barely exists. The result is a strange regime: numbers large enough to move trillion-dollar markets, paired with disclosures too non-standard to value.

This is not a moral complaint about secrecy. Big corporate contracts are often confidential. The problem is structural: the economic terms required for valuation are not publicly observable, and the units themselves aren’t standardized. That combination systematically invites the market to price optionality as if it were commitment.

What’s Actually Observable

After hours combing through SEC filings, earnings materials, and press releases, here’s what an outside investor can anchor on, i.e., what is publicly verifiable without privileged access:

Nvidia-OpenAI (~$100B). What exists publicly is a press release, plus some generic risk language. Nvidia’s CFO later indicated no definitive agreement exists. The market got a headline number and directionality, but little enforceable structure.

Oracle-OpenAI (~$300B). The $300B figure is associated with reporting in the Wall St Journal. rather than a named counterparty in Oracle’s SEC filings. What is visible: Oracle’s remaining performance obligations surged, and Oracle disclosed four multi-billion-dollar contracts with three different customers. But the filings don’t let you map the headline to a specific customer, economics, or specific schedule.

AMD-OpenAI (~$100B). This is the closest to a real contract story: a definitive agreement exists, signed, with warrants for up to 160 million shares. Vesting is tied to deployment milestones, and the structure indicates a high price-level condition (AMD stock at $600). What’s missing is the operative commercial contract: payment terms, milestone definitions, and remedies. All of these terms are opaque to the market.

Broadcom-OpenAI (~$10B). Broadcom referenced a ~$10B order on an earnings call from an unnamed customer; the stock moved before the customer was publicly identified. A joint announcement came later. But again: the term sheet, pricing mechanics, and contingencies are not something outside investors can standardize and model.

Taken together, this is not “zero paperwork exists.” It’s that the public surface area is too thin to price what these numbers mean economically.

The Unit Problem: “Gigawatt Deployed” Is Not Yet a Thing

The most revealing piece of this entire cycle is the recurring language around “gigawatts deployed” or “gigawatt-scale”. In mature infrastructure, a megawatt is a unit with contractual meaning: capacity rights, availability, interconnect terms, often an implied duty cycle or performance framework.

In AI infrastructure, “a gigawatt” can refer to at least five different things:

A press release: an aspirational target or planning envelope.

A grid interconnect: a right to draw power, not a built facility.

A constructed site: buildings and substations complete, but not loaded.

A commissioned facility: racks installed and powered, but not productive.

A sustained load curve: actual utilization over time.

Those are different worlds. The gap between them is not semantics; it’s 12-24 months of execution risk, and billions of dollars of working capital timing.

A gigawatt can be a headline, a substation, or a sustained load curve. Until the industry standardizes what it means, outside valuation is mostly story driven.

What Nobody Can Price From the Outside

Even if you accept that confidentiality is normal, the missing details here aren’t edge cases. They’re first-order drivers of value:

When cash actually changes hands. Are there deposits? Milestone payments? Take-or-pay commitments? Usage-based billing? Each implies radically different cash flow timing and balance sheet impacts, on both sides.

What’s binding versus optional. “Up to $100B” is an option, not an obligation. Warrants tied to milestones are incentives not necessarily purchase commitments. The market repeatedly treats optionality as capex.

What happens when conditions diverge. In the AMD structure, what if deployment milestones are achieved but the stock price condition isn’t? What if the stock hits the level, but deployment is delayed? Who bears that? Without definition, the instrument can’t be modeled.

Priority and enforceability. If the buyer faces constraints, such as capital, power, regulators, or internal reprioritization, who has senior claims? A definitive agreement is not the same as a press release. A press release is not the same as a take-or–pay contract with penalties.

These aren’t exotic hypotheticals. They’re basic questions that need answers, in order to decide whether “$100B” is a project or press strategy.

How This Works in Mature Infrastructure

Traditional infrastructure finance has an unfair advantage: it has markets that force comparability and leak reality. You don’t need to reach every power purchase agreement to price solar. Electricity forwards reveal wholesale prices. Standardized equipment costs and interest rates put bounds on what contracts must look like. Project finance lenders and rating agencies see the contracts and express the risk through credit spreads. REIT pricing creates continuous discovery for physical asset cash flows. Derivatives markets demand standard units, and standard units make verification cheaper.

In other words: mature infrastructure has multiple external feedback loops that prevent headlines from floating too far above economics for too long.

AI infrastructure, at least at this scale and pace, operates outside these loops. And when you remove the verification substrate, you get the weird spectacle we’ve been watching: massive market moves on announcements whose economics outsiders can’t triangulate.

Why Opacity is Rational

There are benign reasons these terms stay private:

Pricing and supply guarantees are completely sensitive.

Negotiations may still be fluid; disclosure can create liability.

Execution depends on scarce constraints (power gear, interconnect queues, permitting, chips), so conditionals are real.

Structures can rough through subsidiaries or special purpose entities that blur disclosure.

All of that can be true.

But note the implication: those same reasons also mean the headline numbers should be discounted like options, not priced like executed capex.

This is the core mistake the market keeps making. It sees a number and hears “commitment.” It should hear “a bundle of contingencies with asymmetric disclosure.”

A More Interesting, and More Damning, Interpretation Than “Promotion”

The cynical story is that these announcements are promotional instruments. The more structurally interesting story is that they’re coordination technology. Given that GPUs, transformers, and interconnect capacity are all scarce, and that lead times are long, a giant public number can signal to external parties, including suppliers, regulators, politicians, captial markets, and talent that these deals are real.

That doesn’t make the numbers fake. It makes them reflexive. The announcement is part of the mechanism that tries to make the world conform to the plan.

But reflexivity increases valuation risk. It increases it. Coordination claims should still be priced like options: high upside if constraints are unlocked, and real probability of delays, rescoping, or quiet expiry.

The Bet

My working theory is simple:

These headline figures are being priced as if they represent binding, time-certain capex commitments. In reality, the disclosed language and the missing definitions make many of them look more like optionality dressed up as commitment.

If I’m wrong—if these are real, binding commitments that will produce infrastructure on roughly the announced timelines—verification will show up fast. You’ll see shipments at scale, interconnects and permits matching the narrative, and financial statements reflecting the cash flow mechanics.

But the refusal to standardize terms, define units, and allow any market-based verification suggests the participants prefer a world where scrutiny is expensive. And the cleanest rational reason to prefer expensive scrutiny is that scrutiny would force the market to price these headlines as what they are: contingent claims, not executed projects.

Half a trillion dollars in headlines. Thin, non-standard disclosure. The market is being asked to trust, without the tools to verify.

If you enjoy this newsletter, consider sharing it with a colleague.

Share

I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

follow me on Twitter,

connect with me on LinkedIn, or

send an email to dave [at] davefriedman dot co. (Not .com!)

No posts

Ready for more?