Thanks BlueRaphus, your insights have always been very helpful.
I agree with you that Tesla will spend a lot of money on AI infrastructure no matter whether they could eventually reach AGI or not. The stake is just way too high to miss the opportunity and let someone else own it exclusively (which musk hates)
Microsoft mentioned in their earning call yesterday that their capex last quarter reached $14b, next quarter at 16b, and guided “significant higher capex in 2025”. That’s a lot of money to be spent on GPU, ASIC, and networking equipments, maybe reaching 65b annual runrate . Google and META guided slightly lower but similar capex level too. It almost seemed that in order to compete in the “AI league”, one would at least spend $10b per quarter just to qualify to play. Most importantly, we haven’t heard from Amazon and Apple (the one with the deepest pocket) about their capex plan yet. The TAM might be even bigger than we initially thought.
This is perhaps a once in a lifetime tailwind for companies like Alphawave to thrive. So from this lens I can understand a little bit why the management sounded "I don't get out of bed for less than $10,000 a day" :) but I agree with what you said, the competition is always fierce. And business opportunities don't translate to financial success so easily. Execution is key and we might need to wait a bit longer to see the results.
As for the 3nm IP, I just checked and I think you were right - cadence has them too and so does Synopsys. I got the impression that AWE is the only supplier to Alchip because my industry friend told me at the time the only other comparable UCIe IPs was from GUC, which is a competitor of Alchip so they likely do not collaborate.
He also told me that AWE’s D2D PHY and controller is the best in class, in terms of speed, stability, energy efficiency etc. Maybe even better than the designs from Broadcom and Marvell(Inphi), who design custom chips for Google and Microsoft respectively.
We could also workout this "$500m of potential lifetime silicon revenue based on 2023 design wins"
A typical AI ASIC costs $1500 to make by TSMC.
HBM is a must and that costs about $2000.
Packaging (Cowos) costs somewhere like $500-1000.
Add the costs up you get $4000-4500 per chip. Alphawave adds 50% gross margin on top, the end chip would end up selling at $8-10k. Yes that's much cheaper than Nvidia's H100 at $30k which is why these hyperscalers are building their own
$500m / $8-10k = 50,000 - 60,000 chips.
Latest large language models such as GPT3.5/4 were typically trained on computing cluster that contain about 10,000-50,000 GPU/ASICs. So the above calculation looks about right.
An educated guess is that this customer is AWS.
Hxulcolrdoh, I am equally looking forward to the optical product roll-out, they are expensive , e.g. 1.6G optic module costs $1-2k and coherent ones cost even more, and they typically ramps in high volume within cloud service provider (1000s to 10,000s of them). I know from industry insiders that this "$300m non-binding contract" is with Amazon. But I don't know how many they will ship to them this years, likely in small volume so that Amazon can test the products
Now I am more interested to see how their custom silicon line of business grows. For a couple of consecutive quarters now, the management's emphasis had been their 3-5nm IP design wins and custom silicon wins. Q1 trading update mentioned "1 IP win for a leading North American Automative company". As far as we know there's only one such company building custom chip in 5nm domain, and it's Tesla, and the chip is dojo2 chip for FSD. If we reached a point where Alphawave's management is able to share this name to the general public, I think that's gonna be pretty good news, because it confirms AWE's technical leadership.
I have some info from the supply chain that AWS's Inferentia(3nm), Trainium(3nm) are all done in collaboration with Alchip, a Taiwanese bank-end chip design company. Alchip doesn't have IPs, so it's almost certain that Alchip acquired those IPs from AWE for Amazon. Why? Because there's no IP vendor other than AWE that has connectivity IPs in 3nm. This TSMC process node is unimaginably expensive to develop anything, which explains AWE's high R&D costs this year.
As for Tesla, I know their Dojo 1(7nm) is with Alchip, and their Dojo 2(5nm) is with TSMC Direct. Oh by the way, Amazon's latest Graviton 3 chips, Google's tensor(3nm) are also developed with TSMC Direct. TSMC Direct is an alliance of multiple chip design companies that include Alphawave. You can find more info here:
https://www.tsmc.com/english/dedicatedFoundry/design-center-alliance
One interesting thing I note from the list of companies on that website, only Alphawave has 2.5D Cowos or 3D chiplet design capabilities (see the tick on TSMC 3DFabric column). That technology is an absolute requirement if you want to build an AI chip, using which you bind the computing unit with High-bandwith-memory unit together. Therefore, I highly suspect Tesla's Dojo2 (which is an AI chip), Google's Tensor, AWS's Graviton could at some point be developed with Alphawave's OpenFive unit, be it logic design, physical implementation, or the shipping of actual end-product. I remember Tony the CEO said they are working with "all hyperscalers" in the US, so I wouldn't be surprised Alphawave is involved in these high profile ASIC projects.
Thank you BlueRaphus, unfortunately I saw your message late and missed the webcast. Have you attended and found out anything meaningful in the meeting?
I am personally puzzled that how come on one hand the company is "shifting away from low margin China business" but on the other hand guided lower than expected EBITDA margin at 20%? The only explanation would be that the higher margin licensing business revenue is dropping even faster, which is possible because they mentioned about recognizing licensing & NRE revenue less aggressively. If that's the decision by the new CFO from Rambus, that's actually a good thing, because the mess around AWE's finances is finally over, hopefully.
I feel that the old CFO really did an awful job to be fired that way.