Today, $CIEN rose nearly 14%, forming a fairly good technical chart pattern. There wasn't any special news from CIEN itself. I think the main reason is that at today's investor day meeting, $MRVL raised its earnings forecast again and emphasized that one of the main contributions came from scale across. As we can see in the optical communications sector, several scale-across related stocks have seen good single-day gains, except for CIEN, where $NOK rose 7.5% today, $CSCO +4.5%

I am honestly disappointed with a lot of investors and popular accounts. They clearly didn't do their homework on $CBRS and were talking absolute rubbish. Let me clear this out. 1. Insider selling: The claim that the CEO sold 99.8% of his shares is flat wrong. Class A is what you and I trade. Class B is what founders hold, and it only converts to Class A when they sell. So the CEO shows just 17,659 Class A shares but still holds ~13.9M Class B. Did people actually read the second table? Since the IPO, insiders sold ~$400M, under 1% of the company. About 80% was pre-planned sales of old options bought at $1-8, and the rest came from VC board members and forced tax sales. Stake each still holds - CEO Feldman ~98% - CFO Komin ~97% - CTO Lie ~87% - COO Mallick ~50%, the only outlier The real short-term scare is the lockup, with ~39M shares (~16%) unlocking on Oct 14 and Oct 28. These are old shares that can now trade, not new ones. That's pressure on the stock, not dilution. 2. $NVDA competition: People are saying Nvidia just launched OpenAI's "Ultrafast" mode on Blackwell, so Cerebras is done. That's not the whole picture. Every AI chip trades speed against volume. Nvidia wins on volume. One GPU holds 288GB of memory, so it serves lots of users at once, the cheapest way to run big models for the masses. The catch with $NVDA is speed. Its memory sits next to the chip, not on it, so every token has to travel back and forth. To go fast anyway, Nvidia runs each GPU with only a few users, which wastes most of the chip. That's why Ultrafast costs 6x the normal price and still tops out around ~300 tokens/sec. Cerebras wins on speed per user. The whole model sits on the chip, and data moves thousands of times faster (21 PB/s vs ~8 TB/s on a GPU). Cerebras ran OpenAI's Ultrafast at 750 tokens/sec in August, on a smaller model, and the new CS-4 does 4,400+. The catch with $CBRS is memory and bandwidth. Each wafer holds 44GB, so big models get split across many wafers that talk to each other slowly, about a sixth of one Nvidia GPU. That's likely part of why the new tier went to Nvidia. But CS-4 links wafers 6x faster per rack, and more memory is coming. And speed is where a lot of the demand is heading, not just scale. AI agents work for one user or one task at a time, and every second they wait is lost work. Companies will happily pay for that. That's why $AMZN and $AMD pair Cerebras with their own chips instead of replacing it. Theirs reads the question, and Cerebras writes the answer. 3. OpenAI and concentration: Yes, OpenAI is the biggest customer, with a $20B+ deal through 2028 and about a third of last quarter's revenue. One launch on Nvidia doesn't cancel a multi-year contract though. And the list is growing with Amazon, AMD, $META, Mistral, Perplexity, Cognition, $CRWD and more. Wall Street already expects ~$3B of revenue and the first real profit in 2027, then ~$7.5B of revenue and ~$3B of free cash in 2028, with ~60% gross margins long term. 4. The $450M loss: $377M was stock comp, and $274M of that was a one-time IPO catch-up. It's non-cash, and without it the business lost only ~$7M. Stock comp keeps engineers from leaving for Nvidia, OpenAI or Google, and should run ~$100M+ per quarter until ~2028. The cost is dilution from new grants, roughly 1-2% a year, not cash. Where I land: Long term, Cerebras owns a specific job in the AI stack that gets more valuable as demand for fast AI grows. I'd love to see more chip sales, since ~61% of revenue is cloud today, plus the memory and bandwidth upgrades. The risks are real, but none of those break the story. - OpenAI is most of the backlog today, but it won't stay the only big customer. - Nvidia could get "fast enough" with the Groq technology it licensed, but the market is too big for just one winner. - Dilution is real, and the lockup plus Q3 earnings could bring short-term volatility. Do your homework before you panic. Not financial advice.

I just chatted with $AMD's SVP of AI Vamsi Boppana. Here are some incremental adds to the previous post. -Scope is AMD-wide. Not just physical AI. He called the team a nucleus to attract more talent. The insight is meant to come from model development at large, and the team was already heading into multimodal models. -No frontier competition, by design. He said competing with customers at the frontier is "not the intent at all." He framed the non-language focus as a deliberate choice to stay complementary to customers' LLM work. -Not an open-model push. He declined to frame it as an open-model strategy. AMD wants the capability for a world of many specialized models; one customer told him it needs 10,000. That's softer than Fei-Fei's public commitment to providing open models and platforms. -Horizontal, with no lead segment. He named content and media, robotics and embedded, and data center prediction and reasoning, calling it "a horizontal models capability for the company." He didn't name a first product or a robotics customer. -Built on AMD, deployed anywhere and on anything. Training will lean toward AMD, with more compute behind it. Deployment stays multi-platform because World Labs already has products running on other platforms. He gave no Instinct share. -The roadmap link is in-house experimentation. His explanation was that he has often wanted people who could "model this and actually train something" to test hardware ideas. That's the mechanism; he gave no MI500, MI600 or Kria specifics and I didn't expect him, too. -Talent and roadmap together drove the price. AMD saw World Labs' undisclosed roadmap and was "blown away." It already has a models team behind its Instella models and five NeurIPS papers, which this deal takes further. -Commercial status quo for now. For Marble, the API and Autodesk, nothing changes near term, and how they evolve is TBD.

Interesting $FSLY finding from testing Meta Muse. Through a series of hundreds of tests, it appears that the $FSLY role with Meta Muse is greater than analysts have noted Over the past week, I created several test websites and ran 738 controlled tests to send traffic from Muse agents to my websites I set up analytics on my website, and I was able to track things like IP address, user agent, operating system, etc From the IP addresses, across the 738 tests, Muse sent traffic from: -Cloudflare: 37.0% -Fastly: 29.3% -Meta: 19.4% -AWS: 14.4% Cloudflare is not a surprise – as there has been a lot of analyst commentary about Muse using Cloudflare. Fastly, on the other hand, was a surprising result I ran many different situations and tasks to try to understand why, or when, Muse sends traffic through Fastly – or others. And when I looked deeper, Muse used Fastly for 100% of the tests when browser navigation was required (e.g., “go to my favorite restaurant’s website and find me the menu”). And when Muse was requested to navigate on a website (e.g., make me a doctor appointment), Fastly was used in 100% of tests Then in other situations, for example when an API is involved or a curl request required, Cloudflare was used in 100% of tests Lastly, when requesting Muse to search the web, Muse split traffic through Fastly, Cloudflare, Meta directly, and AWS

Dario has written that we need to “pace the frontier,” and Sam has agreed. People may be surprised by my response: go ahead. You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier intelligence. You’ve also claimed the lead is widening because of recursive self-improvement. I don’t see what you see in the lab. If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible. But stop pretending you need anyone else’s permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. Stop pretending METR is independent when it is intertwined with Anthropic’s investors and staff. Stop pretending you need those same evaluators to police competitors who aren’t even at the frontier. Most of all, stop pretending the motivation to slow down is purely altruistic. You face massive product-liability exposure if your products enable a truly damaging cyberattack. The market already punishes models that behave in unpredictable or unauthorized ways. After the Hugging Face episode, it is simply good business for OpenAI and Anthropic to trade some raw power for reliability and predictability. Call it alignment if you want. It is also just giving customers what they want. Pacing the frontier would also create breathing room for a more intelligent conversation about regulation than Bernie Sanders’ “shut it all down.” China is very unlikely to join a global agreement, as you know, and that has to be taken into account as well. So go ahead and pace the frontier. You are the ones setting it. The easiest way not to build superintelligence is for you to agree not to build it. Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system. So just do it. If you do, you’ll buy goodwill for the next conversation. If you don’t, we’ll know this was just another bid for regulatory capture — or an election-season psyop.

Here are some of the important takeaways from Goldman Sachs Communacopia + Technology Conference 2026 and Citi’s 2026 Global TMT Conference this week: > Demand for AI compute, infrastructure, and hardware is described as vastly outpacing supply, with multiple companies indicating no signs of a slowdown through 2027 and into 2028. The expectation remains strong that CSPs continue to ramp up CapEx next year. > All of the power semiconductor companies highlighted that inventory levels remain below historical average with no signs of inventory buildup. > Physical AI was mentioned by several companies (Nvidia, Sandisk, Teradyne, Lam Research) as an upcoming movement further strengthening semiconductor demand signal. > SPE providers building up capacity headed towards 2030, with cleanroom being the top constraint. CEO of Teradyne sees strong demand through 2031-2032, claiming it will be the longest semi cycle ever. > Memory makers (Sandisk, Western Digital) have discussed LTAs extending into 2031. CEO of Sandisk saying the LTAs are there to smooth demand and make it more predictable, but some customers still want a lot more supply than the LTAs can provide. > Supply remains intensely tight across advanced packaging (such as 2.5D and high-density fan-out), HBM, LPDDR/DRAM, wafers, connectors, voltage regulators, and specialized optical components (like calcium fluoride optics with lead times up to 18–24 months). > Amkor mentions advanced packaging capacity as extremely tight and in very high demand, same signal thrown out by SPE providers like Lam Research and Applied Materials. > Data center land, power availability, shell readiness, and cooling represent the primary operational bottlenecks for the industry rather than immediate semiconductor supply. Cleanroom construction lead times span 2 to 4 years, creating a structural constraint that is projected to worsen into 2027. > Michael Dell says working with neoclouds is the best because they offer the best demand signal due to their ability to get land, power, and shell. > Jensen noted that power requirements are substantial, citing that a single AI GPU system setup can scale up to 250,000 kilowatts and cost per gigawatt sits around $40 billion. > Microsoft reported that customer demand signals remain "healthy" and that they have established numerous LTAs across the supply chain to rapidly bring supply online. > Nebius reported that AI compute demand is vastly outpacing supply, pushing orders into 2028 and causing persistent price increases that CEO Arkady Volozh expects to last at least another 24 months. > Cisco noted that AI agent traffic has surged 14x since February, with agents now consuming 60% of inference capacity and requiring 450% more bandwidth than human users. > President of Global Business Units at Lumentum says laser supply will not catch up to demand for another couple years. Overall, 2027 looks to be a strong year for semiconductors with no sign of a spending slowdown, and an initial belief that the upcycle continues in 2028 as well. It shouldn't take too long for companies to start claiming demand visibility till 2030. $AMD $NVDA $DELL $LITE $NBIS $MSFT $SNDK

Citrini Research founder James van Geelen has sold the firm to Dylan Patel’s SemiAnalysis Van Geelen will remain CEO of Citrini for now & is also planning to launch a new fund This brings together 2 major independent research platforms focused on markets, AI & semiconductors

Credo $CRDO CEO at the Goldman Sachs Communicopia on Active LED Cables "we are going from 7m with AECs to 30m with an equal level of reliability and efficiency" also adding that "the cable volumetrically is 70% less bulky" than traditional AECs. Looking forward to the ALC demo at OCP.

Dan. I understand your thesis - I am not sure security will see the pressure you describe. There are pressures to change with AI, but not an existential threat. The LLM companies are all partnering because they don't have a perimeter strategy which is where prevention needs to happen. Hard to do low latency, small form factor edge security with frontier LLMs. They are useful for some security use cases, and will require the industry to refactor various products. In addition, over time all software products need to be more agentic and "opinionated" with AI. I don't disagree that productivity software, creative studios, systems of record that have embedded human business processes will feel pressure or will need to be rewritten.