Warrants vs Equity vs Prepay: AI Chip Deals in 2026
The Short Answer
2026’s AI infrastructure agreements keep getting described as “deals worth $X billion” when the structures underneath are completely different instruments with completely different meanings.
| Warrant | Equity Stake | Prepayment | |
|---|---|---|---|
| Cash paid up front | ❌ None | ✅ Full amount | ✅ Full amount |
| Ownership transferred | Only on exercise | Immediately | Never |
| Typical direction | Customer → supplier equity | Supplier → customer | Customer → supplier |
| Trigger | Milestones (usually purchases) | Closing | Contract signing |
| What it buys | Upside participation | Influence + upside | Capacity + priority |
| Circularity risk | Low | High if supplier funds customer | Moderate |
| Headline inflation risk | High — quoted at full exercise | Moderate | High — quoted at total contract value |
| 2026 example | Google–Marvell, Aug 18, 2026 | Supplier investments in AI labs | Multi-year compute contracts |
Verified August 24, 2026.
Warrants: Ownership You Have to Earn
How it works: the issuer grants the right to buy a fixed number of shares at a fixed strike price. The right vests against milestones. No cash moves until exercise.
The 2026 reference case: Marvell issued Google a warrant dated August 18, 2026 for up to 58,970,907 shares at $206.58 — roughly $12.2 billion if fully exercised — vesting in tranches reported as tied to each $500 million of Google chip purchases. Marvell stock rose about 10% on the news.
Why it is the cleanest of the three: the customer spends its own money on products it wants, and the equity is a reward for spending that would have happened anyway. There is no recycled capital, no manufactured demand, and no cash outlay until the option is in the money.
Where it misleads: headlines quote the fully-exercised figure. “$12.2 billion” is a ceiling contingent on a very large amount of purchasing, not a payment. The honest measure is how many tranches actually vest — which is why disclosure at earnings matters more than the announcement.
Equity Stakes: Ownership Bought Now
How it works: cash for shares at closing. Immediate ownership, immediate balance-sheet impact, often board or information rights.
Why it is used: it is the strongest possible commitment signal and it works in both directions — a customer securing supply relationships, or a supplier locking in a strategically important buyer.
Where the circularity problem lives: when a supplier invests in a customer that then buys the supplier’s product, part of the reported revenue is the supplier’s own capital returning as sales. The demand is not fake, but it is not independent either, and from outside the two are hard to separate.
That is the core of the 2026 debate around Nvidia. Bloomberg reported in late July 2026 that Nvidia was working on AI infrastructure deals potentially worth more than $750 billion, including an SK Group partnership valued above $500 billion and discussions to provide a guarantee of as much as $250 billion to help OpenAI lease compute. Nvidia’s counterargument — made publicly by its leadership — is that these are ordinary strategic investments expected to return capital on their own merits. Both things can be true; the difficulty is that only one is visible in a revenue line.
Prepayments: Cash for Certainty
How it works: the buyer pays in advance for future delivery. No ownership changes hands. The buyer gets capacity, priority and price certainty; the seller gets working capital and a hard demand signal.
Why it dominates compute contracts: in a supply-constrained market, guaranteed allocation is worth more than a discount. Prepaying is how you get to the front of the queue.
Where it misleads: the headline is usually total contract value across many years. A “$10 billion compute deal” may be $10 billion of committed spend over five years, contingent on delivery, with termination provisions. It is also demand pulled forward — capacity booked today is capacity not available to book next year, which can make a future period look softer than the underlying business is.
How to Read Any 2026 AI Deal Headline
Four questions, in order:
- Which direction does the money flow? Customer paying supplier is ordinary commerce. Supplier funding customer is where circularity concerns belong.
- Is the headline number paid, committed, or contingent? Warrants quote ceilings. Prepayments quote multi-year totals. Only equity stakes usually quote cash actually moving.
- What is the milestone? A structure that vests on purchases is telling you the purchases are the real event. A structure with no milestone is telling you the announcement is the event.
- Who bears the downside if demand disappoints? In a warrant, nobody — it simply never vests. In an equity stake, the investor. In a prepayment, the buyer, who has already paid.
Why This Matters Beyond Finance
For anyone building on this infrastructure, the structures are a demand forecast in disguise.
Warrants tied to purchase tranches are the most trustworthy signal available, because they only pay out if a sophisticated buyer keeps spending. When Google structures its Marvell relationship so that upside depends on continuing to buy chips, it is making a multi-year statement about accelerator demand that it cannot quietly walk back.
Supplier-funded demand is the signal to discount. Not because it is dishonest, but because it measures the supplier’s conviction rather than the market’s, and those diverge exactly when it matters most.
Sources
- Marvell’s stock pops 10% on AI chip deal that lets Google buy up to $12.2 billion in shares — CNBC, August 19, 2026
- Marvell, Google to co-design custom AI chips for TPUs — RCR Tech
- Nvidia partners with Wall Street giants on $500 billion AI financing — Reuters, August 10, 2026
- Nvidia reignites “circular” AI concerns as it weighs OpenAI financing guarantee — Axios, July 27, 2026