Google Just Took a $12 Billion Position in Its Chip Supplier. Everyone Yelled Bubble. Founders Should Yell Something Else.

On Wednesday, Marvell announced it had granted Google a warrant to buy up to 58.97 million of its shares, worth as much as $12.2 billion if fully exercised, as part of an expanded custom silicon deal running through 2033. Marvell stock jumped double digits. Broadcom, the incumbent king of custom AI chips, fell about five percent on the news. And right on cue, the commentary machine produced its favorite word of 2026: circular. Big Tech is investing in its own suppliers, the money is just going around in a loop, this is how bubbles look right before they pop. I have read a dozen versions of that take in the past 24 hours and I think almost all of them are looking at the wrong thing.

The circular framing is lazy

Yes, there is real circularity in parts of the AI economy, and yes, some of it should worry you. When a model lab takes investment from a chip vendor and immediately spends it back on that vendor's chips, revenue quality gets murky and everyone should squint. But that is not what this deal is. Google is not a startup recycling its Series D into GPU invoices. It is one of the most profitable companies in history using a warrant structure to lock in supply, align a supplier's incentives, and get paid in equity upside for the demand certainty it brings. Vendors have handed big customers warrants for decades. Airlines do it. Retailers do it. The structure is boring. What is interesting is why Google wants it and what it does to the market underneath every AI founder.

What Google is actually buying

Google is not buying chips with this warrant. It is buying a second source. Broadcom has been the dominant partner for custom AI accelerators, including Google's own TPUs, and dominance in a critical input is exactly the kind of dependency Google spends money to destroy. The five percent drop in Broadcom on the announcement is the whole story in one number: Google just told its most important silicon partner that it has options, and the market repriced Broadcom's pricing power in an afternoon. This is the biggest company in search running the same playbook I keep telling founders to run against model vendors. Never let a critical supplier believe they are irreplaceable. Google will happily spend twelve billion dollars of warrant paper to make that sentence true at the silicon layer.

The same commoditization, one layer down

Here is the pattern worth internalizing. For two years this blog has argued that models commoditize, that the frontier premium collapses, and that moats live in workflow, data, distribution, and trust. That argument is now consensus enough that Stripe paid seven billion dollars for a model routing layer. But the identical dynamic is playing out one layer down the stack. Nvidia's margin is a tax on everyone above it, and every serious buyer of compute is now funding the competition: Google with Marvell and its TPUs, Amazon with Trainium, Meta with its own accelerators, OpenAI reportedly working with multiple silicon partners. Custom silicon deals like this one are how the Nvidia premium erodes, not through some dramatic dethroning but through big customers methodically buying themselves alternatives. Compute is being commoditized the same way models were: by customers who refuse to accept a single supplier's margin as a law of nature.

And commoditized compute is not a spectator sport for founders. Every point of margin squeezed out of the silicon layer eventually shows up as cheaper inference. Cheaper inference is what made open weight models economically dominant for routine workloads. The Marvell deal is not a bubble signal. It is a deflation signal for the single largest cost line in AI, and deflation in your cost of goods is about the best macro news a software founder can get.

Where the worry is legitimate

I will grant the bears one thing. The warrant only vests as Google actually buys chips, which means Marvell's headline number is contingent revenue dressed up as a stock pop, and public markets are currently pricing contingent AI revenue as if it were certain. If you are an investor, that gap should bother you. If you are a founder, it mostly should not, because you are not buying Marvell shares, you are buying the compute that all this capacity war produces. Overbuilt capacity and margin wars among your suppliers are bad for their shareholders and wonderful for you. The people who got hurt when the telecoms overbuilt fiber were telecom investors. The people who got rich were the companies that built on top of absurdly cheap bandwidth for the next twenty years.

The founder takeaway

So skip the bubble discourse and steal the tactic. Google, with all its leverage, still pays real money to avoid single supplier dependency at every layer it considers critical. Most founders I talk to have less leverage than Google and more dependency, and they treat that as fine because switching is annoying. It is not fine. If a two trillion dollar company thinks second sourcing is worth a twelve billion dollar warrant, your startup can afford the two weeks it takes to make your inference layer portable.

Here is the test I would run this quarter: name your single most concentrated dependency, the one supplier whose pricing change or capability hold would hurt most, and tell me concretely what your second source is. If the answer is a shrug, you now know your next infrastructure sprint. Google just showed you what it is willing to pay to never have to shrug.