Loading summary
Transcript1 lines
- [00:03]
A
I visited the Bay Area again after a year for another round of talks and visits. The AI boom is entering its third year. In some ways that is young to compare. The mobile boom at three years had yet to see the iPhone 4 and the App Store was just taking off. Year three for the AI boom feels different. It feels like things are becoming for sure. Certain assumptions and feelings are locking in at the same time. The OpenAI dream has fired up a new generation of builders and for that reason I think Silicon Valley is more exciting than it has been for many years. As always, this write up is all about vibes. No data. But the thing to realize is that we must embrace the vibes since that's how things seem to operate nowadays. I visited the Bay Area to attend Hot Chips, the same reason I came last year, and as usual I was not able to attend too many of the sessions because I flew in rather late and ended up talking to people around the show. Which I guess is better because you can just watch the sessions online. Anyway, last year the big theme was the plethora of AI accelerator teams. With Nvidia turning huge margins on its chips, startups and big companies alike were trotting out their own AI ASICs, Intel's Gaudi 3 Meta's, MTIA, Furiosa's RNGD, Microsoft Maya, etc. This year I got the sense that the window for new AI chips has closed. With the exception of Google's Ironwood tpu which looks to be very impressive, nobody is really close to what Nvidia has. The headlining event, in my opinion was on the morning of day two when four photonics companies, Celestial AI, AI Labs, Light Matter and Nvidia, went back to back to present their solutions for AI. I wasn't there, but apparently shade was thrown. As I somewhat mentioned in a prior video, the power demands of modern AI accelerator chips have grown to such an extent that it is finally dragging the photonics industry out of a decades long slump. After the telecom bubble burst, the photonics lineup caught the buzz and to me it is a sign of the AI industry's insatiable needs for scale permeating through the semiconductor supply chain. The grid energy limitations are so insane that it is funding entire new technologies. Silicon Photonics is known to be a very hard manufacturing problem and is notorious for having a relatively small market outside of a few niche applications, which leaves me somewhat concerned about the potential fallout in the case of an AI winter. But I go ahead of myself. Another presentation that I attended was Rapidus, the Japanese Foundry Experiment out in Hokkaido, the CEO gave a presentation remarking that it was the first time in 20 years that a Japanese company had done so at hotchips. Ever since Ken Kutaragi came on stage to do a presentation about the PlayStation, Rapidis is working on a 2nm process node using technology license from IBM. This IBM process is a research process that is clearly not ready for high volume and fortunately they know that. But they are going in a different direction than what I thought they would do. Instead of adapting the 2nm process to high volume manufacturing, they instead are co optimizing their fab end process for the world's fastest possible turnaround time. They presented in house research data that shows that they can cut standard turnaround to as little as 15 days if everything is just single wafers. They call it the all single wafer concept. Doug Laughlin of Semianalysis famously dubs it fine artisanal wafers. To be honest, I'm not sure what is going on here. My understanding from this reading is that their tactic for achieving these short turnarounds looks to literally just be AI. Fabless companies like Nvidia must spend hundreds of millions of dollars to develop AI chips for a particular process node. Wouldn't the Fabless customers want massive volume to pay for those costs? Why would they want a single batch done very fast? Maybe for prototyping concepts or ideas? Then where is the scale going to come from? I need to think on this more. Maybe I'm not hearing it right. Irrational Analysis, who I first introduced to you guys last year, as usual, has a brutal take that says it all. During the 2010s the last time I was in SF, the thing in vogue was software as a service or SaaS. SaaS describes a software delivery and billing mechanism. The software runs on the vendor company's own hardware and is sold as a cloud subscription. Some of the BEI's largest companies run this SaaS model, like Salesforce, Slack, Dropbox. The market richly valued these companies and their consistently recurring revenues. SaaS Economics were so good and compelling that you couldn't throw a stone in the bay without hitting a software startup founder chasing that familiar metric annual recurring revenue or ARR. And to be honest, I personally hated it. Fast forward a decade or so later and there are still SaaS startups being funded, many with an AI flavor to them. Heck, ChatGPT itself is a SaaS product. But the SaaS funding landscape post ChatGPT has somewhat changed. AI makes certain software teams productive at certain coding tasks, and that is a double edged sword has helped certain companies grow their revenues very rapidly. However, profits are a different matter. Something that I have heard several people bring up unprompted has been the Vibe coding thread that one day a broccoli haired Gen Z kid pops up with an 80% identical SaaS product that he vibe coded over the weekend. Okay, that's a bit tongue in cheek, but the core idea is that you start a software startup now. You enter a race to build out the product and get customers to use it. But if the concept works then it is likely to beget a swarm of AI accelerated teams building the same thing. Doug o' Laughlin nailed it in a recent Fabricated Knowledge newsletter when he said that AI represents peak software. AI makes software programmers so much more productive that startups without a head start in the form of proprietary data representation, regulation, brand or an advantage in enterprise access will have a hard time. And I think founders should start considering what that means for their own plans. Investors natural response to software's cheapening has been a growing willingness to fund efforts to solve very hard problems and I admire that. A highlight of this trip has been meeting companies trying to solve cool problems and a major focus point for me on this trip was the general concept of AI accelerated materials discovery. The right material can revolutionize both economy and society. Think about how steel helped enable the industrial revolution or how lithium cobalt oxide, a key cathode for today's modern lithium ion batteries, helped enable today's tech gadgets and EVs. Yet the process of discovering these new materials has hasn't much changed. Brutal trial and error or sheer dumb serendipity materials can sometimes sit for decades before we discover in them a useful property. The idea behind this latest raft of AI fueled science project startups looks to see whether or not trained models can accelerate or even automate this massive near infinite dimensional space of materials discovery. Some are quite high profile. For instance Periodic Labs co founded By Liam Fettes FedEx is well known in the AI community as one of the co creators of ChatGPT. They raised about $200 million from a 16Z at a billion dollar valuation and are in the midst of building a big automated wet lab. I've done a little reading, not too much about this landscape for a potential video, and there might be some potential in using traditional compute intensive methods like DFT to build up training data and then using that training data to train a model that can shortcut the process. And I'm somewhat encouraged by the use of agentic systems that can sort of utilize a variety of tools but at the same time, I wonder if it is really possible to use prior data to predict the behaviors of entirely new branches of items. Another trip highlight has been seeing some of the companies developing alternate forms of computer. The first was Psiquantum. I made a video about their unique approach to light based quantum computing a while back. They turned single photons into qubits and then smashed them together using silicon photonics components to run calculations. Founded by a bunch of well known British professors, Psiquantum has raised something like a billion dollars. And as they like to say they are not spending the money on Corvettes, they're burning their cash on building something from the ground up. At an industrial park in the city of Milpitas, the company is busy converting a former analog device's semiconductor fab into an assembly facility for their scaled up photonic quantum computer. I put on a hardhat and marveled as we walked through an active construction site as teams convert former clean rooms into something new. The day I arrived the company was receiving a massive steel enclosure, something akin to a freezer box, that would hold bunches of their silicon photonic quantum chips at superconducting temperatures. Then we went over to a different location where I saw a massive molecular beam epitaxy tool that they spent millions customizing to deposit thin layers of a special light refractive material called barium titanate. The intention is to eventually put three of these chonkers into the assembly facility and perhaps as soon as next year. The second company was SNOCAP Compute. They're commercializing a new superconductor based approach to low energy compute. In a prior video I spoke about IBM's Josephson technology project. That one failed to beat silicon CMOs, but the approach has evolved over the years. In the 1990s a different approach to superconductor computing emerged, one that appears to be far more sustainable and energy efficient compared to what IBM did. Moreover, SNOCAP spent years making their approach CMOS compatible, which makes it far more manufacturable at scale. I shall save the full write ups for these visits for a future day, but I am fascinated to see how these two parallel efforts towards alternate forms of compute will develop. The final big thing seems to be physical AI or robotics. I spoke with some robotics people in SF and it seems like there is progress. I visited some of the humanoid robotics companies last year and at the time I walked away thinking that the Chinese companies, unitree most of all, were going to roll them like a log in water. Fast forward a year later and I'm starting to get second thoughts about that assessment. 1x Technologies is aiming to get their robots into the household and that is their key focus. The machines have been designed seemingly from the ground up for that. Employing a unique cable and pulley system said to be safer for the indoors. Their go to market is like Tesla's. Their first machine is going to be very expensive so it will only appeal to the rich. During the day, while you're at work, the robot will be teleoperated by someone remote cleaning up the house. This would generate rich physical data for future training. At night the robot will be autonomously controlled where it does a few things like getting the door or picking up this simple thing. I think the key thing is shipping. If there is one lesson that these companies can take from Unitree is that they have to ship more often. They have to get stuff out there even if that stuff is might not be the best. Get the robots out there so that people can start responding to them. But there's progress being made and dates are actually being unveiled. 1x for instance says 2026. But I don't think the robotic future will be what you imagine it will be. The industry dynamics are nothing like a consumer software product like ChatGPT. If anything, the rollout seems to be more like cars. It will take years and be very localized to certain areas for certain people. Nothing will change for a very long time. I don't think the ChatGPT boom one to one leads to a robotics boom. The tech aren't the same yet at the same time I might argue that the rise of LLM fueled software has eroded the traditional moats of software products and has directed a torrent of investment money into harder things like robots. Again, I like this. For the final leg of my trip I return to San Francisco. I don't think there is anything like San Francisco, which most acknowledge as the epicenter of the AI world. Inside the AI world there is still talk about scaling. There remains talk of timelines and takeoffs. Talk about finding the gigawatts of power and squeezing more energy efficiency. It is becoming increasingly clear that coding is one of AI's killer apps. The ferocious rise of coding assistants like anysphere, cursor, replit, lovable and Claude code demonstrates it. There is little denying it. At this point, part of the AI economic story will be stealing a significant percentage of the wages of software engineers around the world. Which back of the envelope is estimated at anywhere between 800 billion and to some $2 trillion a year. But what else is out there? My Main question returning to SF this time was whether or not there is a second AI killer app. After coding and after discussions with various people across the board, the answer so far has been no. Yeah, sure, these models are getting better at math. The models can win international Math Olympiad competitions and all that, but who does that really help? Does the average office worker care about that? And then agentic AI products. The products seemed rather brittle, apt to go off the rails without close intervention. I remember when I first came across Deep Research. I was stunned by the seeming depth of the work. I thought that I was out of a job. But as I studied the output closely, I found it heavily padded, off topic and most importantly, not that interesting. And over time, my personal use of Deep Research has declined in favor of O3 or I guess GPT5. Thinking now it takes less time and the answer tends to be more useful. The key target for these agentic tools seemed to be the business process outsourcing industry. But the total revenues of India's IT outsourced services industry last year was about $70 billion. And not all of that is going to get automated away. It doesn't feel big enough. That brings me to the talk percolating outside of San Francisco. Outside of sf, there is increasing acceptance that there is indeed an AI bubble and discussion of the looming economic challenges that the AI ecosystem must weather later when that bubble does inevitably pop. Few people have tried to do quantitative analyses of the returns that the hyperscalers are making on their AI capital expenditures, but several who have say that it is not good. The reasoning behind why the popping of the AI bubble won't hurt so much was that these tech giants will fund these buildouts with pure cash flow. But now the information is reporting the presence of joint ventures, backstops and debt offerings with third party investors. I cannot help but think back to the deals that intel did with financial investors Brookfield and Apollo. If these are actually making money, why bring in the partners now? If AI solves one of the millennium problems like the Navier Stokes existence and smoothness problem as someone off handed predicted to me might happen over the next year, then, well, dynamics might change. And proof of such real breakthroughs will likely reignite the hype to a new level. AI materials discovery, Alt computing, agentic AI and robotics, and whatever else is going out there that I wasn't able to meet, it all seems rather crazy. Will any of it work? Or is it just a bunch of batty scientists just indulging their fantasies? Or worse yet, will these end up just another Theranos to this, more than a few people brought up the example of OpenAI itself. In 2019, the then unknown and quirky startup raised a billion dollars from Microsoft, and nobody had any idea why this company needed all that money, and few believed their approach would work. And then guess what happened? I think the OpenAI situation has really reverberated throughout the rest of Silicon Valley. Suddenly there seems to be a belief in commercializing hard technologies again. How will this end? I have no idea. But this is way better than funding another Uber for x or another SaaS. Let the money flow. All right, everyone, that's it for tonight. Thanks for watching. Subscribe to the channel, sign up for the Patreon, and I'll see you guys next time.