transLog1c

Great point. I think a lot of people are missing what Gavin is saying. Just look at what Jensen keeps showing at GTC. Tokens per watt are improving dramatically, and better hardware and prefill/decode optimization are driving down the cost of producing tokens. Meanwhile, GPU rental prices are actually going up. Open-weight models are driving a huge increase in token usage, and all that demand still needs compute. For CSPs and neoclouds, this creates a pretty interesting opportunity. Higher GPU utilization, strong rental pricing, and for those running their own inference services, potentially much better margins as production costs fall faster than API prices. Even existing A100 and H200 fleets can become more economically valuable as inference software improves. What’s really interesting is the growing gap: the cost of producing each token keeps falling, token demand is exploding, and GPU rental prices are still holding up. Cheaper tokens don’t necessarily mean less money for AI infrastructure. They could actually make the whole market much bigger.

WTF🥲

One thing I find funny about the memory bear case right now. Everyone seems to assume that CSPs can just walk away from their LTAs, delay data center projects, and leave memory makers stuck with all the new capacity coming online next year. But if those contracts can really be canceled that easily, what's stopping memory makers from saying, "Fine, then we won't expand capacity either"? How does the bear case play out then? It's almost as if the whole narrative assumes memory suppliers have no say in any of this. They just have to sit there and take the hit.

I agree with Damnang on this. The whole discussion about CSPs canceling their memory LTAs doesn’t seem to start with supply and demand. It starts with the assumption that a lot of these data centers won’t get built, and then works backward to the conclusion that memory will be hit the hardest. But what has actually changed in the fundamentals? Not much that I can see. What’s changed is the stock prices and the narrative the market is trading. And somehow that’s completely changed how people think about these memory LTAs. Back in May or June, when everyone was talking about the LTAs memory makers had signed, the discussion was about tens of billions of dollars in customer deposits and how CSPs could lose a huge amount of money if they walked away from those contracts. Now people talk as if all these agreements can just disappear overnight. Maybe the better question to ask these CSPs is: do they regret signing those LTAs in the first place?

Why does this feel like a scene straight out of Oppenheimer?

Optical networking stocks are down more than 20% this week, and $CRDO is probably my favorite name for a rebound. What's interesting is that the fundamentals have actually been moving in the right direction lately. Part of the selling may be tied to concerns around SpaceX/xAI. Some investors worry that Democratic pushback against new data center projects could delay deployments and slow growth. I'm not sure how much of that will actually play out. If those concerns start to fade, I still think CRDO looks like one of the cheapest names in copper and optical interconnects at these levels. I've also been seeing a lot of industry data pointing to a massive shortage of 1.6T optical transceivers next year. BofA estimates demand could reach 90–100 million units in 2027, while DSP shortages may limit shipments to just 60–70 million. I think this could be a real opportunity for Credo to gain share with its own 1.6T ZeroFlap optical transceivers, especially if other suppliers struggle to keep up with demand. Separately, there have been rumors that SpaceX may adopt Credo's ALC technology. I hope management has more positive updates on ALC during the December earnings call. The market still seems skeptical that ALC can become a major revenue driver.

Great read! MicroLED and micro-VCSEL deserve more attention. Really solid work covering both technologies.

Following up on my earlier post about the divergence in AI sentiment across Asian markets. I think this shift in sentiment is particularly visible in Chinese optical stocks. Innolight is probably one of the best examples. Back in June, I remember hearing some pretty aggressive earnings expectations from Chinese investors. Some were looking for close to RMB 100B in net profit next year, largely based on capacity expansion and contributions from new businesses such as NPO. At its peak, Innolight reached a market cap of around RMB 1.4T. I even heard some investors putting forward more optimistic numbers. Today, with its market cap down to around RMB 530B, that implies roughly 5-6x next year's earnings if the RMB 90B-100B profit scenario is achieved. Just to be clear, I'm still very bullish on Innolight's long-term prospects. I continue to see significant growth opportunities across optical interconnects, NPO and other emerging technologies. But my own view of the business and how the market chooses to price it are two different things. I think it's worth understanding what has changed in investors' minds. I've been thinking about this, and I believe two major factors are behind the derating. The geopolitical discount has never really gone away. I've been following the Chinese optical module industry for around four years. For much of that time, investors were still debating how sustainable the industry's growth and profitability could be. Optical modules were often viewed as relatively low-tech manufacturing businesses, and there was considerable skepticism about how much value Chinese manufacturers could ultimately capture. There was also a geopolitical concern in the background: if Chinese optical module companies eventually became hugely profitable, would that attract more restrictions from the US? Innolight's roughly 10x rally has largely happened over the past year. It wasn't an easy journey. Even as earnings continued to surprise on the upside, it took time for the market to become comfortable underwriting those numbers. Then came the FCC-related reports and rumors in October. To me, this brought an old concern back into the spotlight. It was almost like the Sword of Damocles that had been hanging over the sector for years suddenly became much more tangible. I suspect this geopolitical overhang will remain an important factor in how Chinese optical companies are valued going forward. The other issue is how much of today's profitability the market believes these companies can actually sustain. For most of the past three years, Innolight and Eoptolink generally traded at around 10-15x forward earnings. Looking back at 2025, I noticed that whenever their valuations approached 15x next year's earnings, the stocks would often struggle to move higher and eventually pull back. That level had almost become a reference point for the sector. This finally changed in May-June this year. Both companies broke through that valuation range and started trading above 20x forward earnings, as investors became more comfortable looking beyond 2026 and pricing in 2027 and even 2028 earnings. But following the recent correction, sentiment has shifted considerably. In conversations I've had recently, some Chinese investors are now questioning whether these companies can even sustain a 10x multiple going forward. The $MRVL $LITE deal has added another layer to that discussion. Investors are increasingly concerned that greater vertical integration and control over critical components could eventually put a ceiling on module manufacturers' margins. Even with new opportunities from NPO, OCS and other technologies, there is growing skepticism about whether Innolight can maintain its current margins, let alone expand them further. I don't necessarily agree with all of these concerns, but I do think they are important to understand if we're trying to make sense of the current valuation. What stands out to me is just how dramatically expectations have changed in such a short period of time. Back in June, investors were willing to price in earnings growth well into 2028. Today, much of the discussion has shifted towards whether current margins are sustainable and whether the sector deserves even 10x earnings. I remain constructive on Innolight's long-term business prospects. But being bullish on a company's fundamentals doesn't mean ignoring the risks the market is pricing in. That's a pretty dramatic change in how the market is thinking about this industry.

I see it the same way If you believe AI capex is peaking and want to sell AI stocks, stay away from the sector, or even short AI, that’s completely fair. I can understand that argument But what I don’t buy is the idea that money rotating out of AI should automatically pour into consumer stocks or other sectors with mediocre, or even deteriorating, fundamentals To me, that’s more of a short-term narrative and a hedge fund pair trade than a genuine reason for these companies to deserve higher valuations Let’s not forget that EPS growth across many of these non-AI names hasn’t been particularly impressive. Yet their stocks are being pushed higher largely through multiple expansion, rather than actual earnings growth Selling AI because you think the growth outlook is deteriorating is one thing. Paying increasingly higher multiples for companies with weak earnings growth is another I don’t see why the first should automatically justify the second

Yes, this is exactly what I was talking about yesterday. Sentiment in China is extremely bearish right now, especially after the TRS channels were reportedly shut down, making it much harder for domestic capital to gain exposure to US stocks through total return swaps. Meanwhile, A-shares are also getting hammered. It’s a pretty painful situation.

One concern that has persisted among Chinese investors is whether Innolight’s margins can remain at current levels as the optical supply chain evolves. Following the $MRVL $LITE acquisition announcement, Innolight shares fell nearly 8% today. I think this deal could plant the seeds of a longer-term concern around margin sustainability. To be clear, nothing has actually changed yet. Just a couple of days ago, Innolight’s management said its DSP supply remained stable. From what I understand, MRVL hasn’t raised DSP prices for Innolight this year, and the relationship between the two companies remains very stable. The market’s concern is really about what happens after the merger. Will MRVL continue supplying Innolight on the same terms? Could the combined company prioritize its own optical module business when allocating critical components? And could this eventually put pressure on Innolight’s margins? Innolight still has significant advantages in manufacturing scale, yields, customer qualifications and system integration. But I think the market is beginning to question whether those advantages will be enough to sustain further margin expansion in a more vertically integrated supply chain.

Yeah, I’m also struggling to understand why this merger needs to happen. It almost feels like the two companies don’t particularly like each other, yet somehow they’re trying to merge. Pretty strange. I also have a feeling Jensen Huang wouldn’t be too happy to see this deal go through. Interestingly, I think Google could end up being one of the biggest beneficiaries of this merger.

RT @JonahLupton: Every week there’s a new FUD narrative for the memory companies and the neoclouds. Over the weekend a couple accounts cl…

I think $MRVL is about to begin its long-term challenge to $AVGO. Looking at Marvell’s recent acquisitions, especially Celestial AI, XConn, and now the proposed Lumentum deal, the strategy is becoming increasingly clear. MRVL is no longer just building custom ASICs and optical DSPs. It’s moving toward a much more complete AI infrastructure platform, spanning custom silicon, switching, SerDes, optical DSPs, InP lasers, CPO, OCS, and scale-up connectivity. In many ways, Marvell is starting to build something increasingly similar to Broadcom. Of course, AVGO is still far ahead, particularly in Ethernet switching, custom ASICs, and its established hyperscaler ecosystem. That gap won’t be easy to close. But I think this marks the beginning of MRVL’s attempt to challenge Broadcom at the system level, rather than competing in just a few individual product categories. The next few years should be very interesting.

yes！

A few things I've been hearing from some Hong Kong-based funds: They believe 2027 will mark the peak in AI capex growth. Spending may continue to rise in 2028–2029, but a slowdown in the growth rate is inevitable. Interestingly, I've been hearing this same argument from them for the past 2–3 years. Quite a few remain skeptical that CSPs will ever earn an adequate return on their massive AI investments. They also doubt that the current profitability of memory companies is sustainable. Their valuation argument follows from this: once capex growth starts slowing, these stocks can no longer justify 20–30x earnings. One of them put it this way: "Even if a stock looks cheap at 10–15x today, that valuation doesn't mean much if capex growth starts decelerating next year." They also believe the market needs a much stronger "new narrative" to trigger a meaningful reversal in sentiment. Personally, I find this reasoning a little amusing. It's almost the complete opposite of what I've been hearing from funds in Taiwan and the US. Same industry, same data, yet such different conclusions.