$340m of Warrants Against $8bn of Generators: Amazon Has Started Paying Its Suppliers in Their Own Shares

On 3 September 2026 Qualcomm issued Amazon a warrant over 25 million of its own shares. Thirteen days later Generac issued Amazon a warrant over 1,693,745 of its own shares. Same counterparty entity on both — Amazon.com NV Investment Holdings LLC — same vesting logic, same absence of any cash changing hands at signing. What looks like two unrelated supply deals is one financing structure used twice, and the structure is worth understanding on its own terms, because the numbers in the headlines are not the numbers in the filings.

What the two filings actually say

Both disclosures are short, and both were filed under Item 3.02 — unregistered sales of equity securities — which is the first clue about what they are. This is not a procurement announcement with a warrant attached. It is a securities issuance with a procurement contract attached.

Two things stand out. The first is that the dilution is nearly identical — 2.4% against 2.9% — even though one warrant is twelve times the size of the other in dollars. The second is that the reaction was not proportional to the headline. Qualcomm's 3.2% added about $5.6bn of market value; Generac's 18.3% added about $1.9bn. The bigger deal moved the smaller percentage and the larger dollar amount.

That is not a market inefficiency. It is the ratio doing its work.

$8bn is a vesting ceiling, not an order book

The sentence in Generac's 8-K rewards a careful read. Warrant shares vest "in multiple tranches contingent upon aggregate gross payments, net of certain offsets, received by the Company and its global affiliates from or on behalf of Amazon and its affiliates for backup power generators for Amazon data centers, up to a total of $8 billion."

The $8 billion is the point at which the warrant is fully vested. It is a measuring stick for the ladder, not a commitment to climb it. The only delivery figure the filing actually states is the next sentence: initial deliveries "are expected to total $2.4 billion in 2027 and 2028."

Why the buyer can ask for equity at all

Here is the arithmetic that explains the structure better than any commentary.

Amazon spent $54.208bn on property and equipment in the single quarter ended 30 June 2026, and $173.0bn in the twelve months to that date, per its quarterly report filed with the SEC. Against that, an $8bn generator ceiling spread over seven years is about 15% of one quarter's capex, and the $340m warrant is about 0.6% of it.

Generac reported $4.209bn of net sales in the year to 31 December 2025. Against that, $8bn is 1.9 times annual revenue, and the contracted $2.4bn over two years is about 28.5% of a full year's sales in each of them.

A supplier asked to build for a single customer at that scale is taking on real concentration risk: capital spending and long-lead component commitments that only pay back if one buyer keeps buying. The warrant is what makes that trade acceptable to both sides. The supplier gets a counterparty with a direct financial interest in its share price; the buyer gets a claim on the equity value its own order helps create, and pays for it with paper rather than cash.

The strike the announcement moved through

Generac's exercise price is $200.9266. Its closing price the day the warrant was signed was $175.11. So the strike was set 14.7% above the spot price — out of the money on day one.

That looks strange until you look at the quarter. Generac closed at $279.15 on 18 June 2026 and $175.11 on 16 September — a 37.3% de-rating over three months, through the same stretch in which the whole AI-capex complex was repriced. The average of the prior thirty closing prices was $198.68, which sits within about 1% of the strike. The filings do not disclose how either strike was set, and it would be guessing to claim a formula. What the tape shows is that a strike near the trailing average came out well above spot precisely because the stock had fallen so far, so fast.

Then the announcement closed the gap by itself. Generac opened at $229.50 on 17 September, traded as high as $231.89 and as low as $202.26, and closed at $207.23 — $6.30 above the strike, on 8.0 million shares against recent sessions of well under 2 million. The warrant was struck out of the money and was in the money by the end of the first full trading day, on the strength of its own disclosure.

Qualcomm's strike of $161.26 sat about 4.3% below its 3 September close of $168.57, which is the more conventional shape. Both stocks also faded their opening gaps — Qualcomm gave back 54% of its, Generac 41%.

The warrant is a discount, and it lands in revenue

This is the part that changes how the numbers should be read, and it is pure accounting mechanism.

Under US GAAP, equity issued to a customer is not an operating expense. Share-based consideration payable to a customer is measured at grant-date fair value under ASC 718 and recognised as a reduction of the transaction price under ASC 606 — the treatment the FASB clarified in ASU 2019-08 and refined again in ASU 2025-04. Functionally, the warrant is a volume discount denominated in shares, and it reduces revenue as the related sales are recognised.

So there are three different numbers in circulation and they are not interchangeable:

None of that makes the deal smaller than it is. It makes the reported top line a different quantity from the press-release headline — a distinction worth holding when the 2027 numbers start printing.

What the number does to the actual profit mix

Generac's own second-quarter 10-Q contains the detail that the headline omits: the segment this growth lands in earns roughly half the margin of the one it is displacing.

Data-centre generators sit in Commercial & Industrial — the segment that covers stationary generators with outputs up to 3,250kW. That segment grew 29.2% in the quarter while Residential shrank 2.2%, and it did so at a 14.6% adjusted EBITDA margin against Residential's 34.7%.

Run the contracted figure through that: $1.2bn a year of incremental C&I revenue at the segment's Q2 margin is roughly $176m of incremental adjusted EBITDA a year. That is arithmetic at today's mix, not a forecast — segment margins move with product mix, pricing and the capacity spending needed to deliver, and C&I adjusted EBITDA did grow 52.8% year on year, faster than its sales. But the direction of the mix shift is not ambiguous, and neither is the fact that revenue growth of this shape converts to profit at a different rate than the business it is replacing.

Where this touches what you trade

Be honest about the transmission here, because it is indirect.

A company with a $12bn market capitalisation is not what moves a US large-cap index. The read-across is not the single name; it is the fact that the AI buildout's binding constraint keeps migrating downstream. It was chips. Then it was interconnection and land — the story in Microsoft's $329.1bn of leases that haven't started. Now it is prosaic electrical equipment: switchgear, transformers, turbines and, in this case, diesel and gas backup sets with multi-year lead times.

The physical demand underneath is documented. The IEA's Energy and AI analysis puts global data-centre electricity consumption at about 415 TWh in 2024 and projects roughly 945 TWh by 2030, with the United States accounting for the largest share of both the level and the growth. Backup generation is not the same thing as grid supply, but it scales with the same buildout, and it is procured years ahead of the racks it protects.

For a currency, this is a growth-and-investment channel rather than a rates one: sustained domestic capital spending on this scale supports the dollar's growth factor, one of the five the meter scores, without saying anything about the next print. And the equity-market lesson of the last month is more specific than "AI is big" — as the AI slowdown selloff that priced a spending cut nobody made showed, the complex now reprices on disclosure structure as much as on demand.

What would change the picture

Three things, in rough order of how soon they resolve.

Whether the ladder gets climbed. The $2.4bn is contracted; the remaining $5.6bn of the ceiling is contingent. Watch cumulative warrant vesting in future filings — it is a disclosed, quantitative read on how much Amazon has actually paid, and it will tell the story before any guidance does.

Whether the structure spreads. Two warrants in thirteen days, from the same Amazon subsidiary, is a pattern rather than a coincidence. If equity-linked supply agreements become standard for scarce AI-infrastructure capacity, then dilution becomes a routine cost of winning hyperscaler business — and the headline "deal size" becomes a systematically less useful number across the whole supply chain.

Whether the margin holds. A 14.6% segment margin carrying nearly 30% growth is the number to watch, not the revenue line. Capacity for a $1.2bn-a-year customer has to be built before it is billed, and the spending comes first.

What none of this supports is a view on where any of these shares go next. The useful takeaway is narrower and more durable: when a filing says "up to," find the sentence that says what was actually agreed, and check which end of the structure the headline number came from. For more on how this site reads mechanisms rather than forecasts, see what pip theory is for.

Educational macro context only — not investment advice.