When Chipmakers Become Shareholders: Inside the New AI Silicon Alliances

Column Overview

Two announcements landed within days of each other this July, and together they say more about where the AI industry is headed than any earnings call. On July 22, AMD signed a deal to supply Anthropic with up to 2 gigawatts of its MI450 chips starting in early 2027, while committing as much as $5 billion in equity to the company it was selling to. Three days later, Nvidia and South Korea's SK Group unveiled a partnership worth more than $500 billion, spanning a 2-gigawatt AI data center built on Nvidia's Vera Rubin platform and a deepened alliance with SK hynix to co-develop next-generation HBM4 memory. Neither deal was really about a single transaction. Both were about something harder to unwind: chipmakers no longer just selling silicon to AI labs, they're becoming financial stakeholders in the outcome.

Two Deals, One Pattern

Set side by side, the AMD-Anthropic and Nvidia-SK agreements look like they're solving different problems. AMD needed a marquee customer willing to bet on its Instinct line at a scale that could make it a credible second source against Nvidia; Anthropic needed compute capacity it wasn't fully dependent on Nvidia or Google TPUs to get. Nvidia and SK, meanwhile, are tackling something further upstream — the physical scarcity of high-bandwidth memory, without which even the fastest GPU is a very expensive paperweight. SK hynix is the current leader in HBM production, and Nvidia's willingness to lock in supply at this scale tells you the bottleneck in this race isn't compute logic anymore. It's memory, power, and the ability to physically build data centers fast enough.

But the mechanism is the same in both cases: money moves in a loop. AMD invests in Anthropic, and Anthropic spends heavily on AMD chips. Nvidia's arrangement with SK Group similarly binds a data-center build-out to a memory supply commitment, with Nvidia's own platform sitting at the center of both. This isn't new — Nvidia's roughly $100 billion commitment to OpenAI's data-center expansion, and its earlier $10 billion stake in Anthropic, set the template well before this month's headlines. What's new is the sheer number of companies now doing it simultaneously, and the size of the numbers involved.

Why Chipmakers Are Choosing to Become Investors

There's a straightforward commercial logic here, and it's worth taking seriously before reaching for the word "bubble." A frontier AI lab burning tens of billions of dollars a year on training runs needs financing partners who understand its trajectory better than a typical bank would. A chipmaker sitting on that lab's cap table gets visibility into demand years out — exactly the kind of signal that justifies building new fabs, booking foundry capacity, or, in SK hynix's case, committing scarce HBM output years in advance. Equity stakes align incentives: if AMD owns a piece of Anthropic's future, it has every reason to make sure Anthropic's MI450 deployment actually works, down to the software stack.

There's also a competitive angle that's easy to miss if you only look at Nvidia. AMD's move to put real equity behind an AI lab isn't just about selling chips — it's about buying credibility. Frontier labs have overwhelmingly built their stacks around CUDA and Nvidia hardware; getting Anthropic to commit 2 gigawatts of MI450 capacity, with engineering collaboration attached, is AMD's clearest shot yet at proving its silicon is viable at the scale that actually matters. Money is the price of admission to that conversation.

The Circular Financing Question

None of this comes free of scrutiny, and it shouldn't. Industry estimates now put the total scale of these vendor-financed arrangements — chipmakers and cloud providers investing in AI labs that then spend the money on those same vendors' products — well above $800 billion when you add up Nvidia's stakes in OpenAI and Anthropic, AMD's warrants and equity deals, and the cloud commitments layered on top from Oracle and others. Jim Cramer's comparison to the late-1990s telecom equipment boom, when vendors financed their own customers' purchases only to watch the arrangement collapse once those customers ran out of cash, is the one every serious observer keeps coming back to.

The honest answer is that circular financing isn't inherently fraudulent — it's a known and legal way to accelerate capital-intensive buildouts when conventional financing is too slow for the pace of demand. But it does concentrate risk in a way that's hard to see from outside any single deal. If AMD's stock, Anthropic's model performance, and the eventual return on that $5 billion equity stake all depend on each other's success, a stumble in one place doesn't stay contained. Credit markets have already reacted: default swaps tied to chipmaker debt have shown some of their sharpest moves since active trading began, which is a reasonable proxy for how nervous fixed-income investors are getting about exactly this structure. And the gap between what the sector spends on infrastructure — reportedly on the order of $400 billion a year — and what it currently earns in revenue is large enough that "growing into it" is doing a lot of work as an assumption.

What This Means Going Forward

The practical takeaway for anyone watching this space isn't that a crash is imminent — it's that the AI chip market has quietly restructured itself into something closer to a small number of deeply interlinked conglomerates than a conventional buyer-supplier market. Nvidia, AMD, OpenAI, Anthropic, SK Group, Oracle, and Microsoft are no longer simply doing business with each other; increasingly, they own pieces of each other's success. That changes how you should read every future announcement in this space. A "record deal" headline is no longer just a revenue event — it's also a disclosure about who now shares exposure to whom.

For enterprises and smaller AI companies watching from outside this circle, the implication is more mundane but still real: chip supply, especially HBM, is being pre-allocated years in advance to a handful of the largest labs. If you're not one of the parties inside these equity-linked agreements, your position in the queue for cutting-edge memory and compute just got a little further back. That's arguably the least discussed consequence of this wave of deals — not whether the financing model is sound, but who gets left waiting while the biggest players lock up the world's HBM supply for themselves.

The AI chip race was always going to concentrate around a few essential resources — advanced packaging, HBM, and raw power capacity chief among them. What July's announcements make clear is that ownership, not just contracts, is becoming the tool of choice for securing a place at that table. Whether that proves to be smart capital allocation or a rerun of a familiar cycle will depend less on any single deal and more on whether AI revenue growth actually catches up to the infrastructure being built to support it.