Why did Tenstorrent just buy Blue Cheetah?

Is it 5D chess, or just a good price?

News broke last week that Tenstorrent, the AI startup run by the legendary CPU architect Jim Keller, was acquiring the analog design startup Blue Cheetah. I wasn’t surprised to see Blue Cheetah take a bit of a soft landing; while their product portfolio is genuinely impressive, the chiplet market simply isn’t mature enough to build a huge company there yet. I was much more surprised to see Tenstorrent as the buyer.

As recently as 2024, most of Tenstorrent’s sales haven’t come from selling their Wormhole and Blackhole PCIe cards, or selling chiplets, but instead have come from licensing their RISC-V-based AI chip designs to other companies, especially in the automotive and robotics industries. So if most of Tenstorrent’s revenue comes from licensing IP, and Blue Cheetah’s IP isn’t valuable enough due to the immature chiplet market to meaningfully add to Tenstorrent’s portfolio, why would they buy Blue Cheetah?

More broadly, I know a lot of people have questions about Tenstorrent’s business strategy. They have so many products: RISC-V IP, AI IP, two different PCIe cards, two different versions of a desktop workstation, and both datacenter and cloud product offerings as well. And this is in a world where most other AI chip startups are simply selling LLM inference by the token! What’s even going on here?

Part of this product confusion has to do with Tenstorrent’s complicated corporate history -- Jim Keller, the current CEO, was brought in as CTO years after Tenstorrent was founded, and re-molded the company based on his vision. But it’s unclear how Blue Cheetah’s IP portfolio will fit into Tenstorrent’s product portfolio, or if they simply made the acquisition opportunistically because Blue Cheetah was a company with great technology going through financial struggles.

Why was Blue Cheetah struggling?

Blue Cheetah Analog Design was founded in 2018 by Dr. Elad Alon and Dr. Eric Chang, two UC Berkeley researchers who developed the Berkley Analog Generator, or BAG. Normally, analog design is a highly manual process, with engineers picking the size and shape of every transistor in a design by hand. BAG allowed designers to quickly produce custom analog designs without all of the manual effort.

But instead of trying to sell BAG as a tool, Blue Cheetah decided to build a company that used BAG internally to build analog blocks, and sell the analog designs that were produced. This sort of business model meant that they could deliver products that semiconductor companies were already comfortable licensing and integrating into their designs, while maintaining much lower engineering costs.

The problem arose when they decided which kind of analog designs they wanted to target BAG at. Blue Cheetah decided to primarily focus on die-to-die interfaces for chiplets. I’m entirely speculating here, but I think they probably focused on chiplet interfaces due to the lack of competing IP in the chiplet space. If they were selling more conventional off-die interface IP, they’d be competing with giants like Broadcom and Cadence.

At the time, many people thought a chiplet revolution was coming soon, but it hasn’t entirely played out that way. Chiplet-based architectures are very common, but they usually consist of multiple chiplets designed by a single company, rather than a heterogeneous set of chiplets connected like LEGOs. High bandwidth memory is the one major exception, but HBM uses a completely different die-to-die interface than the Bunch-of-Wires and UCIe interfaces that chiplet IP vendors like Blue Cheetah were selling.

There are a lot of reasons why we don’t see the chiplet market growing significantly. There are concerns about which manufacturer is liable if one chiplet fails. Companies are worried about giving up some of their unique competitive advantages if they sell chiplets on the open market. The number of buyers for chiplets is still small. These are all structural challenges with the market that Blue Cheetah wasn’t equipped to solve, even though they had good technology.

I’m personally still optimistic about both automatic analog circuit generation and chiplets, and I’m especially optimistic about the idea of building silicon IP companies using internal tooling to supercharge your margin structure. Unfortunately, Blue Cheetah tied their fate to the success of the chiplet ecosystem before it was developed enough to support a company of their scale.

Now, what about Tenstorrent?

Tenstorrent has had a complicated history. Originally founded in 2016 by Ljubisa Bajic, they were one of the “first wave” of AI chip startups, along with Groq, Cerebras, and SambaNova, to recognize that specialized compute for AI would be a large market. Ljubisa and the original Tenstorrent team architected Grayskull, their first chip designed for AI. Grayskull, as well as its successors Wormhole and Blackhole, leveraged SiFive X280 RISC-V AI CPUs as control processors.

However, when Jim Keller joined Tenstorrent as CTO, he started pushing the company towards developing its own RISC-V cores in-house, rather than relying on RISC-V cores licensed from SiFive, an external vendor. This marked a clear departure from Tenstorrent’s original goals of building AI chips. And that departure became even more marked when Tenstorrent started licensing their RISC-V core externally.

This was a confusing move. Usually, companies don’t want to both license chip IP cores and also sell chips. When a company makes both chips and IP, the chips they make could end up competing with the chips their licensees make, which can scare companies away from becoming potential licensees. But because Tenstorrent was just so passionate about RISC-V, they did it anyway. They also licensed their AI cores externally as well.

While this move was confusing, it did seem to pay off for the company. When they closed a $600M+ fundraise in 2024, most of their large deals were IP licensing agreements, rather than sales of physical chips. And apparently, a big reason Tenstorrent claims they were able to close these licensing deals was because of their open source software stack. This, in turn, makes Tenstorrent’s foray into desktop workstations make a little bit more sense, at least.

If Tenstorrent’s IP licensees want an open-source compiler, there need to be open-source developers to work on that compiler. And what do those developers need? Workstations. Do most of them want workstations that cost twelve thousand dollars? Probably not, but apparently Tenstorrent thinks that this is the best way to get their hardware into the hands of open-source developers.

Tenstorrent seems to be running a weird strategy, but at least it makes some sense: make PCIe cards, desktop workstations, and an evaluation cloud to get developers to support your open source ecosystem, and then use that ecosystem to sell RISC-V and AI IP to customers in the automotive industry in Japan. It’s unclear if this strategy will pay off, but this pivot from dedicated chips to RISC-V IP may have been one of the reasons why Ljubisa Bajic and some of the original Tenstorrent founding team left to go found Taalas, an absolutely wild AI chip startup that could be the subject of an entire article on its own.

But even if Tenstorrent’s strategy pays out, it’s unclear how Blue Cheetah factors into that strategy.

Does Blue Cheetah add anything to Tenstorrent?

As far as I can tell, Tenstorrent’s actual PCIe chips aren’t all that impressive, and mostly serve as a platform for developers to contribute to the open-source compiler ecosystem that drives Tenstorrent’s IP sales. So if Tenstorrent primarily acquired Blue Cheetah to put Blue Cheetah IP inside of Tenstorrent’s chips, it doesn't make all that much sense, as chip sales aren’t a major driver of Tenstorrent’s revenue.

But it also doesn’t make sense for Tenstorrent to acquire Blue Cheetah just to continue licensing Blue Cheetah’s IP. As I just discussed, Blue Cheetah’s IP, despite being high-quality, isn’t driving enough revenue to support the company due to the immature chiplet market. So why did Tenstorrent buy Blue Cheetah? Was it just such a good deal due to Blue Cheetah’s financial struggles that they couldn’t say no?

Well, maybe. But also, Blue Cheetah’s technology could factor into Tenstorrent’s plans to sell chiplets rather than just selling IP. They’re working on selling off-the-shelf chiplets for both their RISC-V and AI cores that can be integrated into larger systems for automotive and robotics applications. These chiplets need die-to-die interfaces in multiple different process nodes including Samsung’s SF4X and Rapidus’s new 2nm node. Blue Cheetah’s analog design automation technology may be a key way for Tenstorrent to deliver these chiplet-based designs.

Could Tenstorrent just license Blue Cheetah’s IP to build their chiplets? Sure. But when Blue Cheetah is going through financial struggles, it may be easier to just buy the whole company than some of their products. It’s a weird reality of startups, but I think it’s the most likely motivation for this acquisition.