Which Way the Money Flows
The AI backstop ran backwards, and the direction is the signal
Steve Williams · babnews.org
There's a mechanism running underneath the entire AI buildout that most people notice only as a headline number, and the number is the least informative part of it. It's the financing that increasingly ties chip suppliers and their customers together, warrants, guarantees, backstops, equity stakes woven into supply agreements. This week Qualcomm and Amazon added another node to that web, and it's worth studying closely, because it runs in the opposite direction from every major example that came before it, and the direction is the entire signal.
The deal
Qualcomm and Amazon Web Services announced a multi-generation partnership to co-develop custom AI chips, focused on inference, along with high-speed optical interconnect reaching 1.6 terabits per second for AWS data centers. As part of it, Qualcomm issued Amazon a warrant to buy 25 million Qualcomm shares at $161.26 each, roughly $4 billion worth, expiring in 2036. The warrant vests in tranches tied to Amazon actually buying Qualcomm's products, with the vesting conditions linked to up to $60 billion in purchases over the life of the deal. A small first tranche, 3.75 million shares, vested immediately on Amazon's initial commitments. Qualcomm's stock jumped, up as much as 10 percent early before settling.
Read quickly, that's "Qualcomm lands a $60 billion AI customer, gives Amazon a stake." Read carefully, it's stranger and more interesting than that, because of who is financing whom.
The arrow points the wrong way, on purpose
For the past year, the defining financial structure of the AI boom has been the vendor backstopping the customer. Nvidia committing enormous sums to support OpenAI's buildout. Broadcom providing guarantees to help the frontier labs finance the chips they can't yet pay for. In every one of those, the chip supplier lends its balance sheet to the customer, and the reason is always the same: the supplier is racing to lock in demand, and is willing to take on financial risk to make sure the orders keep coming. When the vendor finances the customer, the vendor is the anxious party.
This deal inverts that. Here the customer is financing the supplier. Amazon is receiving the equity upside, and Qualcomm is granting it, in exchange for Amazon's commitment to buy. And the inversion carries the opposite meaning. When the customer finances the vendor, the customer holds the leverage, and it's spending that leverage to bring a supplier into being. Amazon is not rescuing Qualcomm. Amazon is paying Qualcomm, in the currency of its own future stock appreciation, to become a credible merchant alternative to Nvidia and Broadcom in AI silicon.
Why would the most powerful buyer in the market do that? Because the single greatest strategic vulnerability a hyperscaler has is dependence on one supplier for its most critical, most expensive input. Amazon already builds its own AI chips, its Trainium and Inferentia line is now running at more than $25 billion annualized, but in-house silicon plus one dominant merchant vendor is still a thin supply base. A second credible merchant supplier is worth real money to Amazon purely as insurance, and the warrant is how it buys that insurance: by giving Qualcomm a direct financial incentive to succeed, and aligning Qualcomm's stock to Amazon's own purchasing. It's a customer manufacturing competition in its own supply chain, and paying for the privilege.
Why inference, and why that matters
The deal is specifically about inference chips, not training, and that detail is not incidental. It's the whole reason a Qualcomm can enter at all.
Training is where Nvidia's moat is deepest, the CUDA software ecosystem, the years of tooling, the switching costs that keep buyers locked in. Inference, running the trained models to generate outputs, is the contestable layer. The moat is thinner, the workloads are more standardized, and it's where the field is fragmenting as custom ASICs from Broadcom, Amazon's own chips, and now Qualcomm's proliferate. So when a hyperscaler wants to seed a new supplier, inference is the only place the seed can actually grow. Amazon is manufacturing optionality precisely where optionality is possible, and steering clear of the part of the market where a challenger has no chance. That's a deliberate, sophisticated read of where the incumbent is beatable.
The flattered number
The number everyone is repeating is $60 billion, and it's the least reliable figure in the story. It is not an order. It is not booked revenue. It is the ceiling of purchases against which Amazon's warrant fully vests, the amount Amazon would have to buy for all $4 billion of stock to come into Amazon's hands. Most of it vests only if Amazon actually places the orders, which happens only if Qualcomm actually delivers competitive silicon on schedule. And Qualcomm's history in adjacent attempts to move beyond smartphones gives investors a real reason to treat on-time, competitive delivery as an open question rather than a given.
So the honest way to read $60 billion is as the size of the incentive, not the size of the sale. It's how much business is on the table to motivate Qualcomm, structured so that Qualcomm captures the equity upside only by earning the revenue. That's a well-designed alignment mechanism. It is not a contract, and reporting that treats it as one is quoting the carrot as if it were the meal.
What Qualcomm is really chasing
It's worth being precise about the target, too. The popular framing is "Qualcomm takes on Nvidia," but the company Qualcomm actually has to catch is Broadcom, whose AI semiconductor revenue is already running in the tens of billions and growing triple digits by supplying custom accelerators to hyperscalers like Google and Meta. That custom-silicon-for-hyperscalers business is the one Qualcomm is trying to enter, and Amazon just became its first credible reference customer in it, a genuine validation for a company under pressure from weak handset demand and the eventual loss of its Apple modem business. So the strategic logic is sound on both sides: Amazon gets a second merchant supplier it can shape, and Qualcomm gets a marquee anchor customer in the business it most needs to break into.
The honest reading
None of this makes the deal a trick or a weakness. It's a smart arrangement for both parties, and the market's positive reaction is rational. Amazon buys supply-chain insurance and influence over a nascent competitor to its dominant vendors. Qualcomm gets a credibility-defining customer and a financial structure that costs it nothing unless the revenue materializes, though it does dilute existing Qualcomm shareholders, which is the real price Qualcomm is paying to secure Amazon's commitment. Both sides win if Qualcomm executes.
The point isn't that the deal is good or bad. It's that the direction of the financing is a signal most coverage is ignoring. The AI economy has become densely cross-financed, suppliers and customers stitched together by equity and guarantees, and in that web the arrow of the money tells you who holds the power in each relationship. Supplier finances customer: the supplier is scared, chasing demand. Customer finances supplier: the customer is dominant, buying options and shaping its own supply base. Same instrument, opposite meaning, and you can read the balance of power in any of these deals by asking which way the value flows.
This one flows from the customer to the supplier, which tells you Amazon is negotiating from strength, using its checkbook to make sure it is never again cornered by the handful of companies that sell the chips it can't do without.
You do not remove a bottleneck. You relocate it. Amazon just spent $4 billion in equity upside relocating its own supplier-concentration risk, by paying to build the alternative supplier it did not want to live without. The bottleneck wasn't chips. It was dependence, and the fix was to fund its way out of it.