The premise is right about the pattern and wrong about the risk
Three equity-linked or prepaid supply arrangements landed in the week of 8 September 2026. The instinct was correct: this rhymes with 1999. But the rhyme is not where it looks, and the count is not three.
The five findings, in order
This is not a September cluster. The template deal is AMD and OpenAI, October 2025. Astera Labs granted Amazon a warrant in February 2026. Nvidia bought into Coherent, Lumentum, Marvell and Synopsys in March. Corning sold Nvidia warrants in May. Marvell granted Google a warrant in August. Qualcomm granted Amazon one in September. Twelve discrete arrangements touch the names examined, spanning eleven months.
Lucent and Nortel lent their customers the cash to buy the equipment, booked the revenue, and carried the receivable. When the carriers failed, the vendor lost the loan and had to restate the revenue. The 2026 structure extends no credit and creates no receivable. The supplier gives away equity instead. There is no restatement channel here.
Equity conveyed does not unwind if demand disappoints. It is permanent. And because several of these warrants vest against the supplier's own share price as well as against purchases, the dilution bites hardest exactly when the thesis works. The 1999 vendor was hurt when demand failed. The 2026 vendor is hurt when demand succeeds. That is a different risk, and it belongs in the analytical framework, not the solvency discussion.
Nvidia provides residual-value support on up to 25% of project costs across a $500bn securitisation platform. Broadcom's reported backstop on the $35bn Anthropic chip-lease facility runs to $29bn, or 83% of the facility. These are contingent vendor obligations on the value of the vendor's own product. That is precisely the 1999 structure, it is off balance sheet, and no one aggregates it.
Nvidia's $1.5bn to Amkor is a customer funding a supplier's capex at a cost of capital Amkor could not obtain unsecured. It is 11.8% of Amkor's market capitalisation and 20.1% of trailing revenue. That is a credit improvement, not a warning. The exposure it creates is concentration, not leverage: Nvidia becomes largest customer and largest creditor at the same moment.
Partially confirmed. The pattern is real, larger than reported, and worth a discrete metric. The risk it carries is dilution and upside capture, not credit. The credit risk the premise was reaching for exists, but it sits in the residual-value layer, in different names, at different sizes.
Stating the premise so it can be falsified
A premise that cannot fail is not a premise. This section fixes the claim, names the evidence that would confirm it, and names the evidence that would break it, before any data is examined.
The claim as originally put
Customer-financed supply, equity-linked, occurred three times in one week: Qualcomm and Amazon warrants, Marvell and Google warrants, Nvidia and Amkor prepayment. This echoes 1999 to 2000 telecom vendor financing. Aggregate equity-linked and prepaid supply financing across the complex should be tracked as a discrete analytical metric.
Decomposing it into three testable assertions
Claimed: equity-linked and prepaid supply arrangements are clustering, at a rate that is new.
Test: count every disclosed arrangement across the names examined and date them. Confirmed if the rate is rising. Broken if the September events are isolated.
Claimed: these arrangements are the same thing as 1999 to 2000 vendor financing.
Test: compare the mechanics side by side. Who advances what, who carries the risk, what unwinds on failure. Broken if the direction of the money differs.
Claimed: the aggregate is large enough to be an analytical risk in its own right.
Test: measure equity conveyed against contracted revenue and against shares outstanding. Broken if it is rounding.
In plain language
A chipmaker wants a very large customer. The customer knows it. So instead of simply paying a price, the customer says: give me a slice of your company, and the slice grows as I buy more. The chipmaker agrees, because the order is enormous and the shares are not cash. The question this note answers is what that slice actually costs, whether the practice is spreading, and whether it is the same mistake the telecom equipment makers made in 1999.
In technical terms
The supplier writes the customer a long-dated call on its own equity, struck at or near spot, with vesting contingent on the customer's own purchase volume and, in some cases, on the supplier's share price. The premium is not cash; it is the incremental probability of winning the design. Economically the supplier has sold volatility and upside on itself to fund customer acquisition, and has done so off the income statement. The exposure never appears as an expense, only as a share count.
What the data shows: twelve arrangements, five categories
Every disclosed equity-linked, prepaid or vendor-backstopped supply arrangement touching the names examined, sorted by category. The categories matter more than the count: they are not the same risk.
Category A: Equity-linked supply grant
The supplier grants warrants or shares to a customer, vesting on that customer's purchases. Equity is conveyed for no cash. This is the structure that appeared three times in the week of 8 September 2026.
| Issuer | Counterparty | Date | Instrument | Conveyed | Strike | Spot | Linked |
|---|---|---|---|---|---|---|---|
| MRVL (Marvell Technology) | Alphabet (Google) | Aug 2026 | Warrant, 58,970,907 sh | 58.97m sh | $206.58 | $234.41 | $120bn |
| QCOM (Qualcomm) | Amazon (AWS) | Sep 2026 | Warrant, up to 25,000,000 sh | 25.00m sh | $161.26 | $176.00 | $60bn |
| ALAB (Astera Labs) | Amazon | Feb 2026 | Warrant, 3,262,299 sh | 3.26m sh | $142.82 | $296.00 | $6bn |
| AMD (Advanced Micro Devices) | OpenAI | Oct 2025 | Penny warrant, up to 160,000,000 sh | 160.00m sh | $0.01 | $520.34 | $100bn |
MRVL. The cleanest read of the whole regime: 240,042 shares vest for every $500m of revenue. Marvell is also a Category B recipient of $2bn of Nvidia convertible preferred, making it the only name in the set entangled on both sides.
QCOM. First Western hyperscaler win for Qualcomm's data-centre franchise. The warrant struck below spot means Amazon was in the money on day one.
ALAB. Seven months earlier than the September cluster and almost entirely unremarked at the time. Evidence the regime did not begin in September.
AMD. The template deal and still the most extreme. Roughly 10% of the company for one customer relationship, struck at a cent. Because vesting is partly conditioned on AMD's own share price, the dilution lands hardest precisely when the thesis works.
Category B: Vendor equity into supplier
A customer or ecosystem sponsor buys equity in its own supplier for cash. Dilutive to existing holders but cash-positive to the issuer. Economically the opposite of Category A.
| Issuer | Counterparty | Date | Instrument | Conveyed | Strike | Spot |
|---|---|---|---|---|---|---|
| GLW (Corning) | Nvidia | May 2026 | Warrants, 15,000,000 traditional plus 3,000,000 pre-funded | 18.00m sh | $180.00 | $167.50 |
| COHR (Coherent) | Nvidia | Mar 2026 | Direct equity investment | $2.0bn | -- | $302.67 |
| LITE (Lumentum) | Nvidia | Mar 2026 | Series A convertible preferred, conversion $695.31 | 2.88m sh | $695.31 | $985.00 |
| MRVL (Marvell Technology) | Nvidia | Mar 2026 | Series A convertible preferred, bundled with NVLink Fusion | $2.0bn | -- | $234.41 |
| SNPS (Synopsys) | Nvidia | Mar 2026 | Equity stake | $2.0bn | -- | $394.02 |
GLW. Frequently miscounted as a giveaway. Nvidia paid $500m for these warrants. Cash in, dilution out. Sits inside a programme reported at up to $3.2bn covering three new US plants.
COHR. Share count is up 26.3% year on year, the largest increase in the optical names examined. The Nvidia subscription is visible in the denominator, not just the narrative.
LITE. Converts at $695.31 against a $985.00 spot. Nvidia is roughly $830m in the money on the conversion feature alone. Approximately 3.2% dilution on conversion.
MRVL. The second leg of Marvell's double entanglement: takes $2bn of Nvidia preferred in March, grants Google a $12.2bn warrant in August.
SNPS. Share count up 21.1% year on year. Brings the design-automation layer inside the Nvidia capital perimeter.
Category C: Customer prepayment and capacity prefunding
A customer advances cash to fund supplier capex, repaid in product. No equity conveyed. Credit-positive to the supplier; creates counterparty dependence.
| Issuer | Counterparty | Date | Instrument | Value |
|---|---|---|---|---|
| AMKR (Amkor Technology) | Nvidia | Sep 2026 | Cash prepayment against future packaging and test capacity | $1.5bn |
AMKR. The benign category, and the one the premise misfiles. $1.5bn is 11.8% of Amkor's market capitalisation and 20.1% of trailing revenue, supplied by a customer at a cost of capital Amkor could not raise unsecured. This is a credit improvement, not a warning. The real exposure is concentration: Nvidia becomes largest customer and largest creditor at once.
Category D: Vendor residual-value or lease backstop
The vendor guarantees the residual value of, or leases on, its own product to enable third-party financing. Contingent, off balance sheet. This is the true structural analogue of 1999 to 2000 vendor financing.
| Issuer | Counterparty | Date | Instrument | Value | vs. facility |
|---|---|---|---|---|---|
| NVDA (Nvidia) | Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR | Aug 2026 | Residual-value support, up to 25% of project cost | $125.0bn | vs. $500bn platform |
| AVGO (Broadcom) | Anthropic, via Apollo/Blackstone SPV | Jun 2026 | Vendor backstop on chip-lease SPV | $29.0bn | 83% of facility |
NVDA. The item that matters. A contingent obligation on hardware residual value, sized at up to $125bn against a $5.4tn market capitalisation. Small relative to equity, structurally identical to what sank the 1999 vendors, and disclosed nowhere in a form that aggregates.
AVGO. An 83% vendor guarantee on a single-counterparty facility. As a share of market capitalisation it is 1.7% and easy to wave away. As a share of the facility it is the highest vendor-guarantee ratio in the complex.
Testing the 1999 analogy, mechanism by mechanism
The premise rests on a structural claim. Structural claims are settled by comparing mechanics, not instinct. Six questions, asked of both eras.
| Question | 1999 to 2000 vendor financing | 2026 Category A warrant | 2026 Category D backstop |
|---|---|---|---|
| What does the vendor advance? | Cash, as a loan or guaranteed credit line to the buyer. | Newly issuable shares in itself. No cash leaves. | A promise to make good a shortfall in its own hardware's residual value. |
| Is revenue booked against it? | Yes. Shipments funded by the vendor's own loan were recognised as revenue. | Yes, but the revenue is genuinely paid for in cash by the customer. | Yes. The lease that the backstop enables funds real purchases. |
| What sits on the balance sheet? | A receivable that had to be reserved against, then written off. | Nothing. The obligation is a share count, disclosed in an 8-K. | Nothing, until the guarantee is called. Contingent and off balance sheet. |
| Who bears the demand risk? | The vendor. Carrier failure destroyed both the loan and the revenue. | The customer. It has already paid cash for what it took. | The vendor, on the residual value, exactly as in 1999. |
| What happens if the cycle turns? | Write-offs, restatements, covenant breaches, insolvency. | Unvested tranches simply never vest. The structure self-limits. | Guarantees are called against depreciated hardware in a market with no bid. |
| What happens if the cycle runs hot? | Nothing bad. The loans are repaid and the vendor keeps the margin. | Maximum dilution. The customer's slice is largest when the equity is worth most. | Nothing bad. Residuals hold and the guarantee lapses unused. |
What the comparison settles
Rows three, four and five break the analogy for Category A. There is no receivable, the customer carries the demand risk, and unvested tranches expire harmlessly. A Marvell warrant that never vests costs Marvell nothing. A Lucent loan that was never repaid cost Lucent everything.
Row six is where Category A bites, and it is the inversion worth holding onto. Every other risk in this analysis is a risk that the thesis fails. This one is a risk that the thesis works and the shareholder does not collect the whole of it. It is a claim on the residual, which is the equity holder's only claim. That makes it a question of per-share economics, and per-share questions need a number.
The right-hand column is a different matter. Every row matches 1999. Contingent, off balance sheet, borne by the vendor, called precisely when the collateral is impaired and unsaleable. The premise was correct that 1999 is repeating. It was looking at the wrong instrument.
Building the metric from the mechanics
A metric earns its place by changing an analysis. Three components, each answering one question the per-share case actually depends on.
The question it answers: how much of the revenue is being paid for in equity?
Gross value of conveyed shares at spot, over the contracted revenue the vesting runs against. Marvell vests 240,042 shares per $500m of Google revenue. At $234.41 that is $56.3m of stock for every $500m booked, an 11.5% equity yield-up. Every dollar of that revenue arrives having already surrendered eleven and a half cents of the company. Score is the lesser of 40 and 3.2 times the yield-up percentage.
The question it answers: how much of the company has been conveyed?
Conveyed shares over shares outstanding. Marvell 6.7%, AMD 9.8%, Qualcomm 2.3%, Astera Labs 1.9%. Reported diluted EPS excludes unvested tranches, so the published figure understates the claim in every case. Score is the lesser of 35 and 3.5 times the dilution percentage.
The question it answers: how good is the counterparty the equity is riding on?
Counterparty share of the financed revenue, multiplied by how that counterparty funds its own capex. One times self-funded from operating cash flow, 1.5 times part debt-funded, two times sponsor or SPV funded. A warrant vesting against a buyer who is themselves equity-financed is worth less than the same warrant against a cash-generative buyer. Score is the lesser of 25 and share times multiple times 12.5.
Grade bands
| Grade | Composite | Reading |
|---|---|---|
| A | Below 20 | No disclosed equity-linked supply, or equity conveyed for full cash consideration. Per-share analysis unaffected. |
| B | 20 to 34 | Structure present but immaterial to per-share economics. Re-examine on each 8-K. |
| C | 35 to 49 | Dilution or linkage material enough to move the per-share case. Use fully-conveyed share count in every per-share metric. |
| D | 50 to 69 | The equity conveyed is a first-order claim on the thesis. New grants are a re-underwrite event. |
| F | 70 and above | The customer owns a controlling share of the upside. The residual must be assessed with an explicit view on why it is still worth owning. |
Why this is not just a tally
A tally counts deals. This counts the claim each deal places on the residual, per unit of the revenue it buys, adjusted for the quality of the buyer. Two names can announce the same headline number and score forty points apart. Corning and Marvell both appear in the register with Nvidia and Google respectively; Corning scores 14 and Marvell 79, because Nvidia paid $500m for its warrants and Google paid nothing for its.
Where the complex actually stands
Eleven of fifty-eight names examined carry a disclosed arrangement. The distribution matters more than the count.
| Issuer | Pillar | Counterparty | Conveyed | Dilution | Value $bn | % cap | Score | Grade |
|---|---|---|---|---|---|---|---|---|
| AMD (Advanced Micro Devices) | Compute and Neo-Cloud | OpenAI | 160.0m | 9.82% | $83.3bn | 9.8% | 99 | F |
| MRVL (Marvell Technology) | Optical / Connectivity | Alphabet (Google); Nvidia | 59.0m | 6.72% | $15.8bn | 7.7% | 79 | F |
| ALAB (Astera Labs) | Optical / Connectivity | Amazon | 3.3m | 1.88% | $1.0bn | 1.9% | 65 | D |
| AVGO (Broadcom) | Compute and Neo-Cloud | Anthropic, via Apollo/Blackstone SPV | -- | -- | $29.0bn | 1.7% | 65 | D |
| QCOM (Qualcomm) | Compute and Neo-Cloud | Amazon (AWS) | 25.0m | 2.34% | $4.4bn | 2.3% | 50 | D |
| NVDA (Nvidia) | Compute and Neo-Cloud | Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR | -- | -- | $125.0bn | 2.3% | 37 | C |
| LITE (Lumentum) | Optical / Connectivity | Nvidia | 2.9m | 3.21% | $2.8bn | 3.2% | 17 | A |
| GLW (Corning) | Optical / Connectivity | Nvidia | 18.0m | 2.09% | $3.0bn | 2.1% | 14 | A |
| AMKR (Amkor Technology) | Packaging / Equipment | Nvidia | -- | -- | $1.5bn | 11.8% | 12 | A |
| COHR (Coherent) | Optical / Connectivity | Nvidia | -- | -- | $2.0bn | 3.4% | 6 | A |
| SNPS (Synopsys) | Agentic AI / Physical-AI Design | Nvidia | -- | -- | $2.0bn | 2.7% | 6 | A |
Pillar penetration
| Pillar | Names examined | Flagged | % | Worst grade | Headline value | Notable names |
|---|---|---|---|---|---|---|
| Optical / Connectivity | 12 | 5 | 42% | F | $19.1bn | MRVL, ALAB, LITE, GLW, COHR |
| Compute and Neo-Cloud | 8 | 2 | 25% | D | $154.0bn | AVGO, NVDA |
| Packaging / Equipment | 4 | 1 | 25% | A | $1.5bn | AMKR |
| Agentic AI | 10 | 1 | 10% | A | $2.0bn | SNPS |
| Power / Grid | 9 | 0 | 0% | -- | -- | Clean |
| Autonomous Systems and Defense | 8 | 0 | 0% | -- | -- | Clean |
| Data / Materials | 5 | 0 | 0% | -- | -- | Clean |
| Infrastructure | 2 | 0 | 0% | -- | -- | Clean |
Three things the distribution says
Five of twelve names flagged, one at grade F. This is not a coincidence of the names examined: optical is the layer where a hyperscaler's custom silicon programme most needs a named partner, and where the partner is small enough that equity is a meaningful inducement.
Power and Grid, Autonomous Systems and Defense, Data and Materials, and Infrastructure carry no disclosed equity-linked or prepaid arrangement at all. Their customers pay cash on ordinary commercial terms. That is a positive finding, and it argues for these pillars as the structural counterweight in the analysis, not merely a diversifier.
It takes $2bn of Nvidia convertible preferred in March and grants Google a $12.2bn warrant in August. It is simultaneously inside one sponsor's capital perimeter and paying another customer in its own equity. It carries the highest composite score in the set.
Analytical implications
A finding that does not sharpen an analysis is noise. Seven observations that follow from the data.
Two names can disclose the same headline and sit forty points apart on the composite. The metric is the claim on the residual per unit of revenue, not the size of the order.
Unvested warrant tranches sit outside reported diluted EPS. Per-share analysis on any grade D or F name should be re-run on the fully-conveyed count before it means anything. On Marvell that is a 6.7% haircut to every per-share figure currently published.
The stock typically rises on the contract announcement. The dilution arrives in the 8-K one to two days later, and the market prices the first before it reads the second. The filing, not the announcement, is the document that matters.
A customer prepayment funding supplier capex lowers the supplier's weighted average cost of capital. It is a credit positive. The exposure it creates is counterparty concentration, which is a separate and named risk, not leverage.
Owning several grade D and F names means owning the same structural risk through multiple wrappers, not a set of independent single-name views. The factor should be sized and named as such.
The level confirms the cycle is mature, which is already known. What is not known is whether the rate of new grants is still accelerating. Two consecutive quarters of rising complex-wide Equity Yield-Up is the signal that warrants a structural reassessment. That is why this register is a standing monitor, not a one-off study.
Alphabet holds the Marvell warrant. Amazon holds the Qualcomm and Astera warrants. Both are covered in this analysis. The customer position partially offsets the supplier position from a factor perspective. That offset should be made explicit before forming a view on either leg.
What would change this analysis
Three things. First, a Category A warrant struck materially above spot and vesting only on revenue well beyond consensus: that is a cheap option sold at a reasonable price and the grade should fall. Second, evidence that a grant genuinely wins share that would otherwise have gone elsewhere, in which case the equity yield-up is a customer acquisition cost with a measurable return, not a giveaway. Third, a Category D backstop that is disclosed, sized and reserved against in the accounts, which would move it out of the 1999 analogue entirely. Each refresh of this monitor should look for all three.
Glossary
- Warrant
- A contract giving the holder the right to buy newly issued shares at a fixed price for a fixed period. Unlike buying stock in the market, exercising a warrant creates new shares, so every existing holder ends up owning a smaller slice of the same company.
- Penny warrant
- A warrant with a nominal exercise price, typically one cent. In substance a gift of shares rather than a right to buy them. AMD's warrant to OpenAI is the example.
- Pre-funded warrant
- A warrant whose purchase price is paid almost entirely up front, leaving a token exercise price. Used to convey economic ownership without crossing a reporting or control threshold. Three million of Corning's eighteen million are this type.
- Vesting tranche
- A block of warrant shares that becomes exercisable only when a defined milestone is hit, here almost always a level of purchases by the counterparty. Marvell's warrant has 240 of them.
- Moneyness
- How far the current share price sits above the exercise price. In the money means the holder profits by exercising today. Amazon's Qualcomm warrant was in the money on the day it was signed.
- Convertible preferred
- A senior class of shares paying a fixed return, exchangeable into ordinary shares at a set price. Dilutive when converted. Nvidia holds this form in Lumentum and Marvell.
- Customer prepayment
- Cash advanced by a buyer to fund a supplier's capacity, settled later in product rather than in money. Nvidia to Amkor, $1.5bn.
- Residual-value support
- A promise by an equipment maker to make good any shortfall if its hardware is worth less than assumed at the end of a lease. Contingent and off balance sheet, until it is called.
- Securitisation and SPV
- A separate legal entity that holds assets and issues debt against their cash flows, keeping that debt off the sponsor's own balance sheet.
- Vendor financing
- The 1999 to 2000 practice of an equipment maker lending its customer the money to buy its own equipment. Revenue was booked, a receivable was carried, and when the carriers failed both disappeared.
- Equity Yield-Up (EYU)
- CFS-1. The share of contracted revenue conveyed as equity.
- Dilution Overhang (DO)
- CFS-2. Conveyed shares over shares outstanding.
- Financing Linkage Concentration (FLC)
- CFS-3. Counterparty concentration adjusted for how that counterparty funds its own capital expenditure.
- CFS composite
- The sum of CFS-1, CFS-2 and CFS-3, on a 0 to 100 scale where higher is worse, graded A to F.
Sources
Deal terms are taken from SEC 8-K filings and company press releases where available, and from primary financial press reporting where a filing was not accessible. Share counts and market capitalisations are 9 to 10 September 2026 reported figures. Prices are IBKR intraday snapshots taken 10 September 2026.
- AMD investor relations, 6 Oct 2025; Tom's Hardware
- Amkor investor relations press release, Sep 2026
- Axios, 10 Jun 2026; Benzinga; Capacity
- Axios, 11 Aug 2026; MonitorDaily; the-decoder
- CNBC, 2 Mar 2026
- CNBC, 2 Mar 2026; Photoncap deal summary
- Photoncap deal summary; Nvidia disclosures
- Qualcomm/CNBC, 8 Sep 2026; TheStreet, 9 Sep 2026
- Reuters/Yahoo Finance, Mar 2026
- SEC 8-K, Astera Labs, 5 Feb 2026
- SEC 8-K, Corning Incorporated, May 2026; CNBC, 6 May 2026
- SEC 8-K, Marvell Technology, filed 19 Aug 2026