Blockquote
First, just to be clear, I own Astera and have small positions in Credo and Arista.
Second, however, many things can be true at the same time.
I don’t think that’s correct. The $8.2B for a quarter prediction is for all of their AI revenue, which includes their custom AI accelerator (XPU) business.
OTOH, Marvell recently acquired Celestica AI, which, according to this report, was the #3 Ethernet Data Switch vendor last year (Artisa was #1, Cisco #2, and Nvidia itself was #5). Of course, ethernet switches are used in more than just AI data centers, and AI networking is more than just switches.
The networking space is complex, always changing, and bifurcated between Scale-Up and Scale-Out, with a further bifurcation between Nvidia and ASICs. I don’t know all of the products from all of the companies in this space, because, well, keeping track of all of them as they change and leap frog each other is close to a full time job.
Nvidia itself has products for both (NVLink for Scale-Up and InfiniBand for Scale-Out). Astera and Credo also have both as well. Arista is Scale-Out only.
Astera had a good business with PCIe retimers back in the Nvidia Hopper days. For Blackwell, Nvidia redesigned and optimized their boards so fewer of them are needed, and has replaced Astera retimers with their own ConnectX-8 NICs.
Furthermore, Nvidia has been progressing from up the value chain, not just selling boards with GPUs on them, and not just selling servers, but now whole Exa-scale racks. Their NVL72 (great link) is not just a reference design, it’s an actual Exa-scale rack in a (big) box, heavily leveraging NVLink to connect those 72 GPUs (and 36 ARM CPUs) together, so no opportunity for third-party Scale-Up solutions inside. A NVL72 costs betweeen $3million and $3.5million according to multiple sources, and are available from Hewlett-Packard, Quanta, Dell, SuperMicro, Asus, and others.
This post describes Meta’s custom version of the NVL72, called “Catalina,” which ups the CPU count from 36 to 72, among other changes, such as removing legacy PCIe boards. I recommend scrolling down to the “Investment implications” section in that long post, which itself has several conclusions, but here are the ones related to networking:
• For Scale-Up networking, it’s all Nvidia’s NVLink.
• However, for Scale-Out, it’s third-party Ethernet instead of InfiniBand.
That keeps Broadcom’s Tomahawk line and Cisco Silicon One deeply embedded in Meta’s AI networks and tilts the broader industry toward scheduled Ethernet rather than IB for scale‑out. Arista and other open‑networking system vendors ride this with 400G optics today and 800G ports configured as 2×400G for breakouts tomorrow. The practical read‑through is that Spectrum‑X wins at select customers, but the hyperscale center of gravity remains merchant Ethernet, where Broadcom, Cisco, Marvell, and system OEMs capture sustained silicon and system share.
I didn’t see either Astera nor Credo mentioned, just Arista, Broadcom, Cisco, and Marvel. Maybe that’s an oversight on the post author, or maybe he knows something about what choices Meta is making. Meta matters because of its multi-billion dollar spend, but of course there are other high-spenders as well.
Another aspect is the non-Nvidia AI build-outs. Amazon through Anthropic has used almost exclusively its own Trainium ASICs, and, of course, Google has TPUs, and even Broadcom has XPUs.
Astera knows its retimer business is on the down slope, as PCI-e gets replaced (they hope) by UA-Link. Amazon has reportedly talked about switching to UA-Link with their 2027 and later build-outs.
while NVIDIA’s reference designs have reduced retimer content, Astera has seen increased content in hyperscaler customizations of platforms such as Grace Blackwell and expects similar customization trends for Vera Rubin. He added that PCIe-based scale-up deployments represent a larger opportunity for retimers.
There’s a lot more in that article, including ALAB acknowledgement of slower CXL sales, and competition from Marvel with the Celestica AI acquistion of CPO (Co-Packaged Optics) technology compelling ALAB to start to develop its own “optical engine to add optical I/O to the Scorpio family, supported by the company’s acquisition of Xscape for packaging capabilities.”
For Taurus Ethernet AECs, [Astera’s] Mohan said second-half 2025 growth was driven by one lead customer, supported by both increased deployments and share gains. He expects additional growth as the industry transitions from 400G to 800G, with Astera spending the first half of 2026 in qualification and expecting volume ramps in the second half.
I think the investing thesis for ALAB remains:
Discussing go-to-market strategy versus competitors, Mohan said Astera supplies the smart cable module and firmware, while partners build the full cable—an approach he said provides hyperscalers supply chain diversity, avoids “margin stacking,” and reduces requalification burdens across applications.
How important a differentiator that is in terms of sales, is, unfortunately, beyond my knowledge or expertise.
This is not a simple space to understand, and I feel trying to predict who will or won’t be successful based on how good and appropriate their products are is a really tough game. This might be yet another case where the Saul “I don’t understand, nor need to understant, the tech” because he/we focus on the business numbers is what matters.
So, in that regard, how are Astera, Credo, and Arista doing, business-financial numbers wise? Who’s growing, who’s stalling, etc.?
