拆解 Google 与 Marvell 传闻中的 1200 亿美元芯片合作

COMPUTE

Another day, another big deal that pairs a tech titan with a chip shepherd or a chip supplier – sometimes both at the same time – that covers tens of billions or more for GenAI infrastructure.

This particular deal, which was announced last week and which I have been mulling over while attending Hot Chips 2026 virtually, is between hyperscaler and cloud provider Google and chip shepherd and chip maker Marvell, which very much wants have its many faceted chip business, which spans various kinds of controllers, switch ASICs, and photonics, and its relatively new custom chip shepherding business, which helps customers design custom CPUs or XPUs, gets them through the foundry, and into packages.

This particular deal was announced in an 8-K filing by Marvell, and it does not say a lot. Google did not say anything publicly about the deal. It is important to look at the share warrant agreement the two have signed as of this week, which you can see here. It does tell us a few things.

Here’s the deal: Google is expanding its chip partnership with Marvell and in exchange, Google has warrants that let it convert met revenue goals into tranches of Marvell stock. The headlines all read that Google is investing $12.5 billion into Marvell stock and Marvell is getting $120 billion in chippery sales and work in exchange from now until 2033, but the deal is a bit more slippery than it has been reported.

To be specific, the 8-K filed with the US Securities and Exchange Commission says the two companies “entered into a commercial agreement relating to the Company’s development of custom semiconductor products to Google (the “Custom Products”).” And added that: “The expanded partnership spans a comprehensive range of custom silicon programs that attach to the TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute.”

All of the trade press and business press jumped on board with a rumor that the deal included Google collaborating on a specialized AI chip code-named “Merope” and which no one outside of Google and maybe Marvell know very little about. This Merope chip is being characterized as a language processing unit, or LPU, akin to the specialized chips that are good at the decode portion of GenAI inference – this is the processing related to providing a response to a query rather than the prefill portion of the workload, which is the processing of the prompt and its context to figure out what a response should be.

The Information reported back in April that Google was in talks with Marvell to develop two chips: the LPU as the current rumors are talking about an refer to as Merope as well as some kind of memory processor, which could be a memory area network based on CXL technologies to allow its TPU systems to share DRAM as a caching tier to reduce the amount of DRAM needed in the TPU host systems. (I would have been doing this long before the memory price spikes, but now it is absolutely necessary with DRAM memory prices already up 7.1X in the past year.

Neither Google nor Marvell have commented on these rumors. And the $18.6 billion rise in Marvell’s stock price the day the deal was announced more than covers the $12.5 billion value of the warrants, so in a way, the deal has already paid for itself and not a single component has shipped under the deal. By the way, those warrants for 58.97 million Marvell shares would give Google 6.3 percent of the company once the deal is completed. 1.36 million of the shares vest across the first four quarters and the rest vest in 240 buckets, with each bucket tied to $500 million in revenues for Marvell for stuff sold to Google.

This deal is another kind of roundtrip in this GenAI frenzy.

Google pays for $500 million a quarter for Marvell stock over six years to have a partnership with Marvell, and the very deal it covers helps drive the stock up year after year, which can help cover the costs of actually getting TPUs and LPUs and other custom stuff as well as off the shelf components covering the gamut inside of its servers and networks. This gear is used in Google Cloud, and some of it will no doubt appear in TPU systems being sold to the likes of OpenAI and Anthropic. The increasing share of the rise in market cap that Google will gain – presuming, of course, there is one – is a kind of rebate on paying Marvell to do chip shepherding. Google pays real money for the chippery and gets a kind of funny money that strengthens its balance sheet back, and that helps Google be better able to raise private debt to further invest in hardware for the GenAI buildout, which helps Marvell.

Here is the funny bit about this deal. If you read the fine print in the addendum, Google doesn’t have to buy the shares. The exercise of converting the vested shares happens on Google’s own timetable, and it can wait until August 18, 2033 to do it if it feels like it. The next day, if Google doesn’t exercise its rights, for instance if the shares fall below the floor that was set at $206.58 per share, the shares vanish. Poof! I doubt very much that Google will do that, wanting both influence on the Marvell board and also wanting to help cover the costs of its AI buildout in the next five years with Marvell funny money.

Here is the other funny bit: the Merope chip is not mentioned by name, but a product code-named “Kestrel” is. This could be the codename for the memory server, and it looks like Google wants it to be qualified by no later than November 10, 2027.

Just to be confusing – or maybe not – Google has another Project Kestrel, which is the codename for a series of datacenters being built in its Kansas City, Missouri region, which runs to around 430 acres and which will hold up to five hyperscale datacenters worth around $100 billion and weighing in at around 2.5 million square feet of aggregate capacity. (Earlier rumors said six datacenters with 1.8 million square feet of capacity.)

So the Kestrel product launch mentioned in the addendum could be a market basket of storage, memory, and network controllers, memory expanders, and a custom AI inference chip to be used alongside its TPUs (which are good for AI training and prefill but maybe not so good at decode at low latency, much as GPUs and Trainiums are also not) or a memory area network based on CXL.

I think Kestrel is the custom chip that goes into the Kestrel datacenters. In essence, the Kestrel datacenters are a massive clustered AI inference accelerator. It is reasonable for the Kestrel accelerators to be qualified for deployment in late 2027, and they roll into the completed datacenters in 2028 as Marvell shepherds them through the Taiwan Semiconductor Manufacturing Co foundries and packaging. But that is just human intelligence inference, called a hunch.

SPONSORED FEATURE: HPE and NVIDIA help organizations apply accelerated computing, software, enterprise infrastructure, networking, control plane, and services to the workloads and business outcomes that matter most.

October 1, 2026

September 30, 2026

AMD Buys World Labs For $8.2 Billion To Get AI Model Cred