In April 2016, I threatened to climb over @andrewdfeldman's fence to give him his first term sheet for @cerebras. It was April Fool’s day, but I wasn’t fooling around. The story started in October 2007, when Andrew and his co-founder Gary Lauterbach had just started SeaMicro. Even then, Andrew was a force of nature. He was extremely intense and miswired in all the right ways. You could feel the sparks flying off him. We didn't invest in SeaMicro, but we stayed in touch. Andrew and the team built SeaMicro then sold it to AMD in 2012. When AMD acquired SeaMicro, I had a hunch Andrew wouldn't last long inside a big company. He has, as I've said many times, immense ambition and a heart full of disobedience. By early 2014, he was looking for an escape hatch. Over the next year and a half, Andrew and I met 6 or 7 times. Sometimes in our office. Sometimes at a coffee shop in Portola Valley. Sometimes at our local tennis and swim club. We kept coming back to one thing: deep learning workloads were growing exponentially, and traditional compute architectures couldn't keep up. GPUs had become the default for neural network training, mainly because researchers had accidentally discovered they were less terrible than CPUs. Andrew, Gary and Sean saw the GPU for what it was: a battlefield promotion of a chip optimized for graphics. Better than a CPU, but not what anyone would design starting from a blank sheet of paper. Their key insight was that memory bandwidth, not raw compute, was the real constraint on what neural networks could achieve. So Andrew, Sean Lie, Gary Lauterbach, Jean-Philippe Fricker and Michael James set out to do something nobody had pulled off in the 75-year history of semiconductors: Build a wafer-scale chip the size of a dinner plate. In April 2016, I asked Andrew if we could be his first term sheet. @ericvishria at Benchmark and I co-led the round along with Pierre Lamond from Eclipse. Then the hard work began. In the 75-year history of computing, no one had made wafer scale work. Which meant no one had ever had to solve the problems that came from trying. How do you power a chip that large? How do you cool one? How do you maintain electrical continuity across tens of thousands of connection points on a single piece of silicon? To get there, Cerebras had to invent in nearly every modern computing discipline at once: semiconductors, systems, data fabric, software, algorithms. Each was a startup in its own right. Their first wafer self-destructed on initial power-up and Andrew and the team were back in the lab the next morning, identifying what didn’t work and coming up with approaches to solving it. Yesterday, Cerebras went public. 19 years after our first meeting, 10 years after that April Fool's term sheet, they’ve built a generational AI company. From a coffee shop in Portola Valley to ringing the bell at the NASDAQ. What a journey. Proud to have been Andrew's first partner in Cerebras. Even prouder to call him my friend.

Crypto is in an Arctic winter right now. Two hours of daylight, freezing cold, and it feels like nobody wants to go there. Most of the attention and capital has moved to AI. But even in the dark, people are still running cool experiments. One example is memecoins tied to real companies. Take the movie theater chain AMC. Robinhood lets people trade tokens that track AMC’s stock. Last month AMC’s CEO publicly objected, and Vlad Tenev pushed back. Within days, traders had launched more than 20 AMC-themed meme coins. (One was even called “A Mad CEO”) A startup we met wants to take the idea further. Imagine earning a McDonald’s token just by walking into a McDonald’s, or an AMC token for going to the movies. It may seem a little silly. But where you physically go says something about you. If you’re showing up at Chick-fil-A every week, maybe you should get a Chick-fil-A meme coin. You’re quite literally voting with your feet. Crypto is here to stay. DeFi, stablecoins and real-world assets already have clear use cases. And the weird ideas - the ones that aren’t skeuomorphic copies of something that already exists - are still being explored. Midsummer is a long way off, but it will come.

Honored to be named to @Forbes' list of the country’s 250 Greatest Living Optimists. I didn’t even know there was a list of the world’s greatest optimists 😅 but when you back extraordinary founders for a living, it’s hard to see the world any other way.

Trust is built in how you handle the commitments you can’t keep. I learned this from @LynnJurich, co-founder of @Sunrun. She calls it “impeccable agreements.” I backed Sunrun in 2008 and watched Lynn take it from 4 people to a public company. Her standard for accountability was this: we can renegotiate any agreement but it can't just disappear into thin air. This was an “impeccable agreement.” Be specific about who will deliver what by when. If you’re going to miss it, come back and renegotiate before you miss the deadline. You promised a draft by Friday. On Wednesday, you know it won’t happen, so let’s have the conversation about it on Wednesday. It forces a lot of uncomfortable conversations. But it gives the other person time to adjust instead of discovering the problem when they need your work. Nobody has to chase anybody. Your team can work with the confidence that you’ll either deliver or raise the problem while there’s still time to act.

How not to respond to AI: - Start a special projects team - Staff it with your best people - Tell everyone else to keep chopping wood I’d love to say it works. In reality, I don’t think it ever does. You can’t afford to only have a handful of people working on the thing that might disrupt your entire business. It also creates a schism between the smart, special people who get to work on the secret moonshot project and the rest of the company, who have to carry on as usual. You’ll need every one of them when the new direction becomes the only direction. The best thing you can do to meet a moment like today is take your entire company and tell them this: We all need to move in this new direction. The waves are bigger and moving faster than I’ve seen in my career. You can’t turn the turret with half the crew in a different room.

Stanford Professor Karl Deisseroth just won the Nobel Prize in Physiology/Medicine. His path getting there was not conventional. Karl, Peter Hegemann, and Georg Nagel won the award for their work in optogenetics, a way to turn specific brain cells on and off with light. Peter and Georg discovered a light-sensitive protein in green algae. Karl figured out how to put it into neurons, which gave scientists a way to study how those cells influence memory, emotion, and behavior. In 2021, I interviewed Karl for my Missionary Misfits series in Forbes. He called his path "quite circuitous," which undersells it by a mile. Karl trained as a psychiatrist. Since he was a kid, he'd wanted to know how feelings were created in the brain, but medical school didn’t give him the answers he was looking for. So he followed his curiosity. He'd seen cellular research transform fields like cancer and immunology, and he wanted to bring that approach to the brain. He thought the answer could lie in bioengineering. As his psychiatry training wrapped up, he applied to become a bioengineering professor at Stanford. He did not have an engineering degree, but he still went for it anyway. He introduced himself to the head of the department, pitched the idea, and after a few rigorous interviews, got the job. Karl is the epitome of a missionary misfit. His passion for understanding where our emotions come from transcended any single discipline and he crossed into a field he'd never trained in to find the answer. There may be no bigger question than what makes us human. Karl has spent his career chasing it, and I'm thrilled to see the Nobel committee recognize that. Congratulations, Karl!

Today, startups can't be afraid to reinvent themselves, even at massive scale: - Replit: 20M users. Reinvention. - Linear: $100M+ ARR. Reinvention. - Genspark: shut down AI search with millions of users... then hit $36M ARR in 45 days All of these companies are valued in the billions. And they’ve been willing to take big leaps to find their next leg of growth. Being willing to pivot early has always mattered. A lot of great founders found their way through the fog that way. What's new is that the reinvention doesn't stop once something works. It used to feel like once you were through $50 or $60 million in revenue and still growing, you were done reinventing the product. That’s no longer the case. The technology is moving so quickly that that founders must turn their turrets faster than ever before. Be willing to found, refound, then refound your companies. That is how you’ll win.

Every startup talks about their Cash Runway. They should be more concerned about their Relevance Runway. There are a lot of companies right now sitting on plenty of cash. They raised a big round a couple of years ago and on paper they have years left. But the race they are actually in is against becoming irrelevant and it is more pressing than any financial pressure. Take data integration - workflow tooling and the long tail of specialized pipeline and orchestration products. That whole layer of tooling got heavily funded over a 2-3 year window, and a lot of those products were best-in-class when they shipped. Now the frontier models and the coding agents are encroaching on "good enough." You can have an amazing product, and your customer can still get to a good-enough version of it with a few Codex prompts. Categories that used to be categories are consolidating fast. So what do you do if you are running one of these companies? Three things come to mind. 1. Decide whether you're staying independent. If you can't see a path back to relevance, find a home while there is still a buyer. It is one of the best M&A markets I’ve seen in years. Look hard at whether you have enough momentum with customers and enough surface area of product to hoover up the adjacent companies in your space and be the biggest banana in the bunch. More surface area means more data flywheel and more proprietary understanding of how customers use you. 2. Gird yourself for battle. Assume you are going to have to reinvent the product, probably more than once. It didn't used to work this way. Once you were past $50 or $60 million in revenue and growing, you might think you could stop worrying that everything might change underneath you. That would be a huge mistake. A friend of mine runs a company with >$100 million ARR and I wouldn’t be surprised if they have another pivot ahead of them. 3. Rebuild your vision. Cash buys time. It does not buy motivation. If you are going to reinvent the product, your team needs to see where it's going before they'll follow you there. The people who can help you pull off a reinvention are also the ones with options elsewhere. @andrewdfeldman put it well at our summit last year: give them great teammates, give them great problems to work on. If your team can't see a future worth building, you won't get the chance to build it. All that said, nobody can tell you how long your relevance runway is. Figure it out before your best engineer does.

What is co-design? Why does it matter for AI? What are examples? What are the tradeoffs? 1. What is Co-Design? In traditional computing design, one company designs a general-purpose chip and other companies write software to run on it. Nvidia ships the GPU. Everyone else adapts their models to it. Co-design means that the chip and the software are designed together for a specific job at the same time. You start with the actual programs you need to run, measure where they spend their time, and then change whichever side fixes it: the chip, the model, the memory layout, the compiler. Nothing is fixed in advance except the objective. The payoff is much better performance on the jobs you optimize for. 2. Why does it matter for AI? Because AI workloads are becoming huge. Google now processes over 3.2 quadrillion tokens a month, 7x more than a year ago. When a workload like inference gets that big, even small efficiency gains from co-designing the hardware and software are worth billions. 3. What are examples of co-design? Google's first TPU is one of the earliest examples I know. Google did not build a chip for every AI workload. They picked the ones that were most important for them (RankBrain for Search, the LSTM behind Google Translate, Inception for image recognition, and AlphaGo) and designed the first TPU around what those programs needed. MLPs and LSTMs were 90% of Google's inference demand. CNNs, the thing everyone benchmarked on, were 5%. In 2015, TPU v1 goes to the data center and they found the chip was sitting idle most of the time, waiting for data to arrive. For RankBrain and Translate, the bottleneck was memory. In TPU v2 they swapped the memory for a much faster kind (HBM), which had roughly 18x the bandwidth. That fix exposed the next one. Models like Translate had outgrown a single chip, so the slowest link became the wiring between chips. In v3 they built a dedicated network that wired 1,024 chips together into one machine. By v4 even that wiring was the constraint, because different models want the chips connected in different shapes. So they replaced fixed copper links with optical switches - tiny mirrors that route data at the speed of light, so the network can be rewired in software to match whatever model is running on it. Another great example is @cerebras. They co-designed their chip for AI training workloads, which were bottlenecked on memory, the same problem Google hit with the TPU. 4. What are the tradeoffs? - The tighter you fit the chip to the job, the worse it does on everything else. - It only works if you have a job big enough to design around. - The job can change under you - chips can take years to build. TPU v1 went into the data center in 2015, 2 years before the transformer paper. All that said, the biggest AI jobs are now big enough to design around. I expect a lot more of this to come.

Phase 1 of the AI infrastructure build out is over. Initially there was this frenzy for compute at all costs. Availability mattered more than anything else. As we've moved from chat to research to long-running agents, token consumption has gone through the roof. Workloads are getting both larger and more varied. So businesses are starting to ask a new question: how do I make every token cost-effective? Anyone running a data center is looking for ways to squeak out more performance per watt and per dollar. Now, we are beginning Phase 2: Heterogeneous hardware. We will see an ensemble of hardware over the coming years. Inference is no longer a single workload and different workloads have different objective functions, so they need different infrastructure. Prefill wants compute and decode wants memory bandwidth. @cerebras has announced disaggregated inference with the likes of AWS and AMD. AMD’s Helios takes prefill, the wafer-scale engine takes decode. That pairing gets 5x the tokens per second per watt. Nvidia would likely see similar uplift pairing its GPUs with a wafer-scale engine on decode. And OpenAI just launched Jalapeño with Broadcom too. Demand continues to inflect and the metric driving all of this is useful intelligence per dollar, per megawatt, per second. No single chip wins on all three for every workload. That's why the ensemble is coming.

Shennon Biotechnologies spent 5 years figuring out how to make better cancer drugs. Today, they made their most important hire yet: Cyril Konto joins as CEO. Once you’ve built a platform that can screen millions of single-cell interactions and find the rare few that inform the next-gen cancer drug, how do you get that drug into a patient? ShennonBio has spent the last five years solving the first part of that problem. Today, Cyril Konto officially joins as CEO to help take on the next: bringing better cancer drugs to life. Li Sun, PhD is one of the most brilliant people I know, and a consummate optimist. When she left Foundation Capital in 2021 to start ShennonBio, she was returning to a problem she’d been thinking about for more than a decade: how to turn the ultra-high-throughput functional single-cell screening technology she developed during her PhD at Harvard into better cancer medicines. Doing so has meant bringing together 3 disciplines that don’t typically live at the same address: Immunotherapy, to harness T cells against cancer. Microfluidics, to profile millions of single-cell interactions in a single experiment. And AI, to help identify targets, design therapeutics, and predict potential safety risks. Most companies specialize in just one. ShennonBio lives at the intersection of all three. Discovering a promising drug candidate takes one set of muscles. Getting it to patients requires another. An oncologist by training, Cyril has spent his career developing cancer therapies at Bristol Myers Squibb, Pfizer, Allogene, and most recently Ichnos, where he served as CEO. He brings deep expertise in clinical development, the FDA pathway, and working with big pharma. Li described their fit to me as the two of them “running toward each other.” She’ll keep pushing the science and technology forward. Cyril will help turn what she’s building into treatments for patients. Lung, ovarian, and liver cancer are first. From there, they plan to tackle other cancers and, eventually, diseases beyond oncology. Welcome, Cyril, and congratulations, Li! Excited for what you two will make possible as a team.

Everyone talks about the risk of buying from the frontier AI labs. Almost nobody talks about the risk of selling to them. I recently met a startup that landed one of the model labs as a customer. They saw it as a massive win, I wasn’t so sure. The company applies AI to a deeply technical engineering problem. Their original plan was selling to the incumbents in their industry. It didn't work. Then they realized they could sell their data to a frontier lab instead. Now they're effectively outsourced R&D for one of the biggest labs in the world, and they're thrilled about it. I get why. The labs might spend $100 million with a company like this over the course of a few years. That’s an enormous number. But understand what the lab is actually buying... it's the data and expertise they need to train your capability into their model. They are turning your intellectual property into theirs. Once their model passes the benchmark, the box is checked. "We now support fluid dynamics." Then they move on. They were never going to keep pushing as if they were purpose-built for that use case. The startup's theory is that the lab will always have questions, and they'll always be the right people to answer them. That is not proving true in my experience. Once they get what they want, they're out. Gone, on to the next thing. Once you get your 800 on the SAT, do you keep taking practice tests? No, you're done. If a frontier lab wants to write your startup a $100 million check, take it seriously. But remember that a one-off $100 million contract and a business are not the same thing.

Founders are constantly told: "don't think about selling." The ambition behind it is right but the advice is naive. It's killed more good outcomes than it's created. I've had this conversation twice in the past 3 weeks. Most people start their company thinking they're building the next Cerebras. That ambition is important. But many outcomes - even good ones - come from M&A. Sometimes a recovery of capital, sometimes 1.5x, hopefully 3x. Any relationship worth its weight in salt in an M&A conversation takes 18 to 24 months to build. Most founders start way too late. They pound the for-sale sign into the lawn with 3 months of runway left. At that point, the outcome is unlikely to be favorable. Instead, treat it the way you'd ask your head of business development to build pipeline: 1. Write down ~10 executives at companies that could ultimately acquire you. Not the head of Corp Dev at Datadog - operating execs who'd understand what you're building. 2, Get to know them. A conference, an intro, a demo. 3. Keep a dialogue going - a check-in every quarter. That way, if M&A turns out to be the right path, the relationship started 2 or 3 years ago. If you're not going to ring the bell at NASDAQ, you can still manufacture an outcome. And if it's organized and intentional, you're far more likely to get a good one for your team. So start now - build your list.

Last weekend, we lost one of the great engineering teachers of the past 50 years. Larry Leifer was my master's advisor at Stanford and one of my favorite professors. Larry had two really special gifts: First, he saw in people what they couldn’t yet see in themselves. Early in my time at Stanford, Larry saw my unbounded curiosity for how complex systems work, shaped by a childhood spent repairing my trusty snowblower and taking apart every broken appliance in the house. He believed I could pass that curiosity onto other students and encouraged me to set up a class with Margot Brereton, one of his PhD students. It became "ME191: Exploring Engineering Intuition". We got half a dozen lawnmower engines donated and built the course around taking them apart. Nobody had ever looked at my tinkering and seen a teacher. Larry did. Second, Larry had this unbelievable ability to beam the future into the present. In 1993, the idea that design was something you could research - that engineers should learn by taking things apart rather than sitting through another problem set - was fringe. Thirty years later, that fringe idea is orthodoxy. It has a name - design thinking - and its own building and program at Stanford. Every good product team on earth now builds, breaks, and rebuilds its way to understanding. Larry was already living there in 1993, inviting us into the future. Larry was an n-of-one. I'll miss standing over an open lawnmower engine with him, being taught without ever feeling like I was being taught. RIP Larry John Leifer, PhD 1940 – 2026

.@cerebras CTO Sean Lie - one of the most brilliant minds in AI - is now on X. Give him a follow!

Today, @cerebras launched CS-4. The fastest AI in the world just got much faster. Some highlights: - Up to 2x faster inference than CS-3 and up to 30x faster inference than GPUs. - Can deliver more than 1,000 tokens per second on models exceeding 10 trillion parameters. - Up to 10x more throughput per Watt than CS-3. - A new modular rack design (Nexus) that cuts deployment time from days to hours. An incredible achievement by @andrewdfeldman, Sean Lie and team. Still so much more to come!

My advice to early-stage founders trying to hire right now: Generously give out the “founding” title. Founding engineer. Founding GTM lead. Founding designer. Take it seriously and it works. It tells those early hires they're building, not "joining" a company. And it's a high prestige title, even in a 10-person company. It’s also something that one of the big labs, or Google, or Nvidia could never offer. Do not try to compete on comp, it is a losing battle. Top engineers and researchers are commanding $20, $30, $200 million. And use this as a filter. Look for the person who lights up at a founding title. They are the person who will run towards the fight. If the potential hire needs a title to come with clarity and a big team, they were never going to thrive at your stage anyway. Embrace your founderness and say it out loud.

Most heads of HR are barely qualified to carry the watermelon to the company picnic. Any truly great head of HR could have done any other job. Head of marketing, head of sales, head of business development. But they chose HR because amplifying human potential is their calling. Unfortunately, a lot of heads of HR position themselves as the company teddy bear. Or the human suggestion box. Bring me all your grievances, dissensions and issues. That's the wrong way to excel at the job. So what do great heads of HR do? First, they understand what is special about this company. Why did its founders start it? How did it get this far? They build a deep appreciation for that. Then, they figure out the infrastructure to scale what's working - and deal with what isn't. One of my favorite examples is Chad Herring at @Sunrun. He was super strategic and a little edgy. Chad was not a company teddy bear. The task at Sunrun was not straightforward. The mission attracted a lot of tree hugging, save-the-world kind of people. But the market was brutally competitive and @LynnJurich and Ed Fenster are two of the most competitive people I've ever worked with. So Chad scaled a performance oriented culture on top of the mission. He treated people - finding them, developing them, retaining them - as the biggest point of leverage we had. He helped instill a culture of accountability and excellence, without losing the missionaries. And that helped us win the market and build a multibillion dollar company. That's what a great head of HR does. They take what is special about your company and build the org that wins with it.

👀

Tokenmaxxing. Looksmaxxing. Benchmaxxing. We've basically entered the business equivalent of Jackass. The whole logic of that franchise was "whatever the last stunt was, top it." This is how tech talks right now. Product claims are dialed to 11. Almost unbelievable claims - about what your product can do, about why it's so much better than anything else out there. Every launch louder than the last. We have benchmaxxing, tokenmaxxing, comp-maxxing. Everything is maxxing. How do you stay ahead? You just have to be even more extreme. The temptation is to leapfrog everyone and make even bigger bolder claims. But in a market like this exaggeration stops working. When every company claims 10x, "10x" means nothing. Customers can't tell the real breakthroughs from the noise, so they do the rational thing - discount everyone. And trust starts to really, really matter. In a market where nobody believes anyone, being believed is the rarest thing you can be. The company whose claims keep checking out earns something no launch video can buy - customers who take its word, design partners who provide helpful early feedback, employees/teammates who stick around not just when things feel good. So make fewer claims… and deliver on every one.