SecondSource's "buyer's rebate" call on Google's Marvell warrant was backwards. Under US accounting rules it is a discount Marvell pays Google, about six points of Google's purchases by our calculation.

Key takeaways

Yesterday's daily brief called Google's warrant in Marvell a buyer's rebate; this piece finds that call backwards. On August 19, Marvell filed with the US Securities and Exchange Commission a warrant that lets Google buy up to 58.97 million Marvell shares at US$206.58 each. Apart from a small slice in the first year, every tranche unlocks only when Google places orders: one tranche for each US$500 million of custom chips, 240 tranches in all. The chips are custom silicon. Google is buying them from Marvell to widen the supply base around its TPU line. Yesterday's daily brief (August 20) made this call: because the strike sits "near the market price", only 4.4% below the August 18 close, the deal is the opposite of AMD handing OpenAI about a tenth of AMD's equity at a one-cent strike. AMD's warrant was a seller paying up. On that reading, Marvell's was a purchasing rebate. Google would get back part of what it paid for chips through gains in Marvell's share price.

Yesterday's call pointed the wrong way because it measured the warrant by the wrong yardstick. It compared the strike with one day's close instead of valuing the warrant at fair value, which is what the accounting rules require. US accounting rules define this kind of warrant precisely: equity granted to a customer is consideration payable to a customer, and the amount comes straight off the seller's own revenue. The seller, Marvell, gives up part of its revenue; Google is not clawing money back. Marvell and AMD did the same kind of thing in the same direction. The difference is the price, and the price can be computed: about six percentage points of what Google buys (our calculation).

This piece adds three concrete things to yesterday's item. First, the denominator behind "4.4% in the money" comes apart. The figure is not a stable reading; it is the most flattering of four defensible reference dates. On the day the commercial agreement was signed (July 29) the strike was 26.4% out of the money. On the trading day before issuance, it was 11.8% in the money. On the filing day, it was 12.9% in. The 4.4% comes from August 18, a trough left by a 7.8% one-day drop. Second, yesterday's core claim, "this is not a seller's concession", does not survive a check against the accounting standard. The Financial Accounting Standards Board's update of May 2025 addresses exactly this, share-based consideration granted to a customer: measure at grant-date fair value, book as a reduction of revenue, and recognize once "the customer will buy that much" becomes probable. Third, a test is written down in advance: if Marvell's August 27 earnings call discloses a fair value for this warrant clearly below US$4 billion, the six percentage points in this piece get rewritten. That threshold sits well below the US$6.15 billion this piece gets at its lowest volatility assumption. Only a much lower volatility or a much shorter term would explain it.

"4.4% in the money" is the best-looking of four answers

Start with where the number came from, because yesterday's entire chain of reasoning rested on it.

On August 18 Marvell issued the warrant with a US$206.58 strike. Marvell's stock closed that day at US$216.00. The strike is US$9.42 below that close, or 4.4%. Financial media that day described the warrant as roughly 4.4% in the money. Yesterday's brief took the figure as given and built on it: a strike pinned to the market price means Google pays nearly full price for the shares, so the warrant is no giveaway.

The problem is that US$206.58 and US$216.00 were set at different times by different mechanisms. The ratio between them is not stable. Three weeks of closing prices make the point:

(Closing prices are from the historical price tables of S&P Global Market Intelligence, a subscription service with no public link.)

All four are defensible reference points, and they give four answers far apart. August 18 itself is the worst of them: the stock fell 7.8% from the day before and rose 9.9% the day after. Across those three weeks, the August 18 close was the single day nearest the strike. The reading yesterday's brief picked up was the only one under which "pinned to the market price" holds.

So where did US$206.58 come from? The 8-K does not say, and parts of the warrant's exhibits were omitted under the filing rules. The number can be reverse-engineered. Across moving averages of various lengths over that period, the 20-trading-day average close ending July 31 is US$206.47, 11 cents (0.05%) from the strike; the 30-day average ending August 18 is US$207.82, 0.6% away. The shape matches industry practice. When Amazon took a warrant in Plug Power in 2017, the US$1.1893 strike was stated outright as the 30-trading-day volume-weighted average price as of April 4. That 2017 figure comes from secondhand accounts; the original filing hasn't been reviewed.

Our read: the strike is an average price over a window around the agreement, not a number benchmarked to any one day's close. That inference does not change the conclusion, but it drains the 4.4% of meaning. The 4.4% is not a negotiated discount. It comes from an averaging formula meeting one unusually low close.

The deeper problem: even if the strike equalled the market price exactly, the warrant would still be worth something.

A different ruler: the number that actually gets booked

The question. What is the right to buy one share at US$206.58 worth when the share trades at US$216.00? If the right expired tomorrow, US$9.42: the spread, which is what yesterday's brief computed. But this warrant runs to August 18, 2033, seven years out. Over seven years the stock could reach 400 or fall to 100. If the stock rises, the holder exercises and collects the spread; if the stock falls, the holder lets the warrant lapse and loses nothing. That asymmetry, all of the upside and none of the downside, has a price of its own, and on a seven-year warrant that price is far larger than today's spread.

The accounting rules settle this; there is no room for choice. Under US GAAP, equity granted to a customer is measured at fair value under ASC 718, the share-based payment standard, which means an option-pricing model. The spread is not an acceptable measure. The resulting amount is treated under ASC 606, the revenue standard, as consideration payable to a customer, and the amount reduces the seller's revenue.

The evidence. The rule is not only on the books; the rule is new. On May 15, 2025 the Financial Accounting Standards Board issued ASU 2025-04, an update written specifically to clarify how share-based consideration granted to a customer is booked. This piece relies on the interpretations published by several large accounting firms; the standard's own text hasn't been reviewed. Three points bear directly on this case: ① measure at grant-date fair value under ASC 718, with no remeasurement afterward; ② the amount reduces the transaction price, which means it reduces revenue; ③ if the grant carries a performance condition, here "Google buys US$500 million", the revenue reduction is recognized at grant-date fair value once that condition becomes probable.

AMD has described its OpenAI warrant the same way: the warrant's grant-date fair value is recorded as a reduction of revenue as the related revenue is recognized.

The result. Using the standard option-pricing formula, Black-Scholes, with every input taken from the filing and public prices: share price 216.00, strike 206.58, term 7 years, risk-free rate 4%, dividend yield 0.11%. The one input that has to be assumed is volatility. Marvell's realized annualized volatility between June 10 and August 20 was 99%; the stock fell from US$310.58 on June 18 to US$163.40 on July 29, a 47.4% drop, and then recovered. A seven-year valuation should not use a short-term extreme, but the figure shows that a 40%–60% range is conservative for this stock. No implied volatility from the options market was obtained; the true seven-year implied value for this stock remains unknown.

In one line: the warrant's price is roughly five to six and a half percentage points of what Google buys (our calculation). Yesterday's brief put the figure at 4.5% if the stock reaches US$300, and at zero if the stock stays below the strike. That arithmetic treated a seven-year warrant as if it expired tomorrow.

One independent reading is available as a cross-check. A published market analysis puts the warrant's value in an illustrative range of US$6.8 billion to US$7.7 billion and concludes that roughly 6% or more of the qualifying revenue flows back to Google through the revenue reduction. That matches the 50%–60% volatility rows above.

Google still pays cash, and that does not change. For each tranche that unlocks, Google pays 240,042 × US$206.58, about US$49.59 million, to receive the shares; exercising everything costs about US$11.9 billion. That money goes to Marvell. The option valuation has already netted it out: a fair value of US$122 means the right is still worth US$122 net after the US$206.58 is paid.

Boundaries. Every cell in the table is this newsletter's own arithmetic, not a filed figure. Three assumptions could be wrong: ① the volatility; ② valuing all 240 tranches on the same seven-year term (every tranche shares the same expiry, so the term is right, but the accounting may use a shorter expected term instead); ③ the omitted warrant exhibits keep the anti-dilution, acceleration and strike-reset terms out of view. If the omitted terms include a strike reset, the whole table has to be recomputed.

Put the price on the price list, and the "reversal" disappears

Four deals can be compared on one measure: the warrant's fair value, divided by the revenue that triggers its vesting.

AMD to OpenAI (October 2025): 160 million shares at a one-cent strike. With the strike at almost zero, each warrant share is worth about a full share of stock. That is the extreme entry on the list.

Cerebras to OpenAI: the most useful entry, because the figure is already in a published income statement, not an estimate. Cerebras is an AI chip maker that listed on the stock market only recently. The figures here are secondhand relays of the earnings release; the original hasn't been reviewed. For the second quarter of 2026, Cerebras reported GAAP revenue of US$180.1 million. That quarter Cerebras booked US$44.26 million of customer-warrant amortization. It did not count that amount as revenue but as a discount handed back to the customer. Adding that back gives pre-reduction revenue of US$224.4 million (180.1 + 44.3), and against that denominator the discount size is 19.7%. That is a floor: the warrant amortizes with OpenAI's usage, so if OpenAI is only part of Cerebras's revenue, the discount size on the OpenAI portion alone is higher than 19.7%. Cerebras's balance sheet still carries US$1.128 billion of unamortized customer-warrant assets.

Marvell to Google (August 2026): 5.1%–6.6%, from this piece's arithmetic.

Amazon to Plug Power (April 2017): this entry matters for its date, not its size. No primary filing was obtained for this deal; the figures below come from secondhand accounts, and a second, conflicting share count exists. Plug Power is a hydrogen fuel-cell company. 55.28 million shares at a strike of US$1.1893 (a 30-day volume-weighted average), vesting when Amazon's own purchases reach US$600 million. Buyer purchases trigger vesting; the strike is an average around the signing date. Same mould as Marvell's warrant, nine years earlier.

Set the four side by side, and yesterday's claim that "the third instance flipped the economics" fails in two places.

First, the direction did not flip. All four are the seller paying the buyer, and all four reduce the seller's revenue. Cerebras's is already visible on an income statement. The difference is the price, not the direction.

Second, "third instance" counted the wrong population. Yesterday's brief stated an observation condition: a third instance would be needed before calling this an industry template rather than AMD's particular situation. That count defined the population as AI chips. The structure's real population is "buying an anchor customer with equity", which has run for nine years in logistics, hydrogen and aviation, and the accounting standard-setter has issued an update specifically for it. In the correct population, Marvell's warrant is a nine-year-old practice applied to custom silicon for the first time. It is not a third AI-chip instance.

Yesterday's brief got two things right. One: the strike mechanism really is different from AMD's, and by a wide margin. But that difference sets the size of the discount, not who pays whom. Two: the US$120 billion ceiling really is close to unreachable. Unlocking all 240 tranches requires Google alone to buy US$120 billion in about 6.25 years, about US$19.2 billion a year, while Marvell's whole-company revenue last fiscal year was US$8.195 billion (data center above US$6 billion). Full vesting would require one customer's annual orders to reach 2.3 times today's entire company (our calculation). Full vesting would give Google all 58.97 million shares, about 7% of Marvell. That is the ceiling, and where the real stake lands depends on how much Google buys.

What the discount size measures, and which line to read on August 27

Line the four discount sizes up: Cerebras at least 19.7%, Marvell 5.1%–6.6%, and AMD the highest of the four, because a near-zero strike makes each warrant share worth about a full share. AMD's entry cannot be put on the same denominator as the others. With a one-cent strike it is the highest by construction. Our read: the size of the discount measures who needs the order more.

Cerebras is newly listed with about US$200 million of revenue a quarter; landing OpenAI vouches for Cerebras as a credible vendor. Cerebras could afford to give back about a fifth of its pre-reduction revenue. AMD had to pry a 6 GW anchor customer away from Nvidia, and AMD gave OpenAI warrants worth close to a full share each. Marvell is already a supplier with more than US$6 billion in data center revenue, and Google came to Marvell to widen the custom-silicon supply around TPU: Marvell needed to pay only six percentage points. That number is the hardest piece of information in this case: of the four deals, Marvell is the seller with the best bargaining position.

Read from the buyer's side, the same logic carries a signal. Google used a mechanical average-price formula, inside a window of violent price swings, to lock in a discount worth about six percentage points of its purchases, with no premium paid up front and no commitment made: the purchase volume itself is the consideration. Our read: for the next custom-silicon negotiation, Google and Marvell have signed a template that prices volume commitments through an averaged strike. Marvell's six percentage points are the nearest reference point for that negotiation.

Which line to read on August 27. On Marvell's earnings call a number that did not exist before will appear for the first time: the warrant's fair value, and how much revenue reduction was recognized this quarter. Under ASU 2025-04 the reduction is recognized only once the performance condition is probable. So the size of the reduction Marvell books is Marvell's public statement of how much Marvell expects Google to buy. The accounting standard forces a forward-looking disclosure: a supplier has to convert its internal forecast for a single customer into a number on the income statement. No one had to do that before.

The same call settles a second question: whether Marvell reports a second revenue figure that excludes warrant amortization. Cerebras already does, under a metric Cerebras labels core revenue, and Cerebras's stock fell 17% after the second-quarter report, with coverage describing the results as confusing. When a company's revenue comes in two numbers, the first question is which one is the cash that actually arrives. Neither figure is the cash that arrives from the warrant. The warrant reduces reported revenue and hands over equity. Its only cash is the strike price Google pays when it exercises.

What this changes, and for whom

Hyperscaler procurement has one more item on the table. The contract template now carries a list price, 5% to 20% of purchases, depending on how badly the supplier needs the customer. A next round of custom-silicon talks that does not include a warrant of this type leaves that value with the supplier's shareholders. How the discount is obtained matters as much as whether it is: the strike tracks a period average, so which stretch of trading the negotiation window falls into decides more percentage points than negotiating skill does.

Custom-silicon suppliers and their shareholders have one reading to track: the six percentage points. The headline 'Google gets 7%' is the wrong number to track; the six points are. The six points are a share of Google's purchases that Marvell gives back. They come off the revenue line rather than the expense line, so gross profit falls by the same dollars. Every future quarter's revenue growth rate is pressed down a little by this discount, and how much is set by the auditors' judgment of how much the customer will buy. On the other side, six percentage points is low among peers, which is evidence of bargaining position.

Can this deal overturn the share call on Nvidia versus custom silicon? The deal is evidence against SecondSource's standing call that custom ASICs cannot reach 30% of AI accelerator supply value against Nvidia. The evidence is weak. Google's willingness to pay six percentage points to widen its second source says Google judges the risk of a single-supplier TPU chain worth that price. Our read: the deal prices Google's supply risk, not market share. Consideration and market share are measured in different units, and the deal is not enough on its own to reverse that call.

What would prove this wrong.

• Marvell discloses a warrant fair value on August 27 clearly below US$4 billion (this piece's floor is US$6.15 billion) → the volatility or term assumption is wrong, and "six percentage points" gets rewritten to a smaller number.
• Marvell does not treat the warrant as a reduction of revenue, for instance by arguing Google transferred a distinct good or service to Marvell → the "seller pays" core argument has to be rewritten. The standard does leave that exception, but even where it applies, any fair value above the fair value of the goods or services received still reduces the transaction price, so a zero reduction is unlikely.
• The full warrant text (Exhibit 4.1) shows the strike can be reset → the premise of this piece and of yesterday's judgment falls together, and the valuation table is void.
• A fifth and sixth instance appear with discount sizes unrelated to "who needs the order more" → the fourth section's call that the discount size measures bargaining position is downgraded to an observation.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.

Sources this deep dive draws on

Compiled mechanically from the records of the facts this piece draws on; each line is the source of one of them. Whether we read the original or only reporting is stated in the body.

• U.S. Securities and Exchange Commission: Form 8-K - AMD Material Event: OpenAI Warrant Agreement (October 6, 2025)
• AMD Investor Relations: AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs
• Quartz: qz.com/marvell-google-stock-warrant-ai-chip-deal-081926