Three Wall Street desks looked at the same set of disclosures yesterday and produced the first three-way sell-side spread of AMD's AI-capex cycle. JPMorgan's Harlan Sur stayed Neutral and warned the Meta warrant could "represent $30 billion to $100 billion in equity value" while reducing gross margins by "200 to 400 basis points." [1] Goldman Sachs's Toshiya Hari raised his price target to $240 from $210 but kept a Neutral rating. Bank of America's Vivek Arya held Buy at $280 and modeled $20 of EPS by 2030 on the combined Meta-and-OpenAI deal stack. The standard read of Tuesday's print and Meta announcement was that Q2 guidance and the 6-gigawatt deal were the news; today's standard is that the warrant structure is the news the desks cannot agree on.

The structure is a $0.01-strike warrant for up to 160 million AMD shares — about 10% of the company — issued to Meta on Feb. 24. [2] It vests in tranches as Meta's cumulative purchases scale to 6 gigawatts and as AMD's stock price hits performance thresholds running up to $600. [3] The full award is good through 2031. AMD inked a structurally identical warrant with OpenAI last year for the same 160 million shares at the same penny strike. [4] If both counterparties exercise, existing holders are diluted by 20%.

That is the math the three desks are pricing differently. JPM's Sur is treating the warrants as economic compensation that has to be subtracted from gross-profit dollars before earnings flow through. His 200-to-400-basis-point margin haircut, applied across an AMD revenue base running toward $40 billion in 2026, is a $1.6-billion-to-$3.2-billion annual gross-profit deduction the consensus model does not yet carry. [1] Goldman's Hari raised the target on the volume story but stayed Neutral on the dilution. BofA's Arya is treating the warrants as customer-alignment incentives that pay for themselves through revenue scale: at 6 gigawatts of Meta-purchased silicon and 6 gigawatts of OpenAI-purchased silicon, the cumulative revenue overwhelms the dilution.

The consequence frame X is running with is the circular-deal frame. Each gigawatt of compute is, in CEO Lisa Su's language, "worth double-digit billions." [5] Meta has a Nvidia deal worth a reported $50 billion next to the AMD deal. [3] OpenAI has the Cerebras 750-megawatt contract worth $20 billion-plus, the Broadcom XPU collaboration, and the AMD warrant. The buyers fund the chipmakers; the chipmakers issue equity to the buyers; the equity vests on stock-price performance. The structure is legal. It is also visible enough that it now appears as a risk factor in two prospectuses on the same May tape — the Cerebras S-1/A and, by reference, the AMD 10-Q.

The $20 EPS-by-2030 number is the Buy thesis's mathematical anchor. BofA gets there by combining a low-end estimate of 6-gigawatt-Meta plus 6-gigawatt-OpenAI silicon revenue with current AMD Instinct gross margins, holding the warrant tranches at the maximum dilution case, and scaling out share count. [3] If any of the three inputs — gigawatts shipped, gross margin, or warrant exercise — softens, the EPS number softens with it. The Neutrals are not buying the product of all three at the upper bound.

What the spread tells the next quarter is that AMD is no longer trading on its own results. It is trading on the warrant structure's interaction with hyperscaler capex and a stock-price target list that runs to $600. The Q2 guide of $11.0-$11.6 billion in revenue is the fundamental number; the warrant disclosure in the 10-Q, due within the next several weeks, is the document the three desks will reconcile against. [3]

Two warrants on two prospectuses on the same week is the new normal of AI capex disclosure. The JPM-vs-BofA spread is the first time the sell-side has admitted, in print, that nobody has marked the structure to a single price.

-- THEO KAPLAN, San Francisco