Consumer agents are the next frontier of AI. But right now, they're slower than a person with an iPhone. Consumers are incredibly impatient. Google learned this years ago. Where they showed delays of a few milliseconds changed how often people searched and how long they stayed. If I ask you to wait 8 seconds for a website to load, you’re gone. Now look at today's AI assistants. Our team asked the leading ones to book a dinner reservation. A person can do it by hand in 37 seconds. The assistants took up to 7 minutes and 40 seconds. So we built a prototype assistant on @Cerebras, tuned for this task. It booked the same table in 22 seconds, faster than a person could. It checked every restaurant in parallel, rather than sequentially. We taught it how the booking site worked ahead of time, so it didn't have to guess. And every call to the model came back in a fraction of the time because it ran on the Cerebras Wafer-Scale Engine. This was just one task. There is a lot more work to do before agents can handle every errand this well. But it shows what's coming. The leverage these assistants can give is enormous. The real breakthrough will come when agents are faster than we are. Then consumers will hand them everything. Fast inference is a big part of getting there. That's the future we're building for at @cerebras.

Last week, a short seller, @CulperResearch, published a report arguing that @cerebras wafers wear out after about two years, and based on that premise the report suggested that the economics of our business are fundamentally challenged.  That number is wrong, and given that the basis of the entire report is built upon that assumption, all of the assertions that follow are also wrong. We do agree that the question of how long our systems last is a fair question, so we want to address the topic plainly.   If you have a few minutes, I would encourage you to click on the link below, which will provide you with a better understanding of the pioneering work done by our engineering team over the past decade to ensure that the operating lifetime of our solutions is many, many years. Link Here:

Chipmakers work together more than people think. In the history of compute, there's a repeated pattern. As each workload matures, we divide it into stages. Then we build specialized hardware for each stage. In inference, it's called disaggregation. There are 2 distinct stages: Prefill - processing the prompt. Decode - generating the answer. The 2 stages need very different things. Prefill is highly parallelizable. Decode is strictly sequential. Each stage is best served by a different machine. Already, chipmakers are working together to use the best architecture for each stage. At @cerebras, we’re partnering with AWS and AMD to do exactly this. Today, Tony Kim and I spoke about this and much more at the Future of The Data Center and Compute Summit 2026. Thanks for having me, Tony.

Good to be back at @Stanford for the HAI Conference. This time, I was on a panel about how fast AI is moving. At @cerebras, we've been building chips for 10 years. And the industry has never moved faster. In the last 90 days, the lead at the frontier changed hands 4 times. More than 10 interesting open models came out. And we only expect it to accelerate from here. Good to see you @YejinChoinka, @dawnsongtweets and @rwmcmorrow.

Interesting look under the hood of @AlphaSenseInc's SuperAnalyst. Lots of thoughtful engineering in how they give agents the tools to do useful research while keeping access tightly controlled. And when speed matters, they route requests to open-weight models on Cerebras. Worth a read if you’re building agents:

The fastest inference in the world is coming to @OracleCloud. @cerebras and @Oracle partner to enable blisteringly fast inference on OCI. Cerebras brings fast AI. Oracle brings deep expertise building and running AI infrastructure at scale. Enterprises have trusted Oracle with their most important data and applications for decades. Partnering with Cerebras gives Oracle customers more choice to match each AI workload to the compute that suits it best. Coming soon, they’ll be able to build agents with true super fast AI in the cloud they already use. Thank you to @MThiagarajan_ and the @Oracle team. Excited to build this with you. Full announcement here:

Last week, I caught up with Cheddar TV’s Justine Miller during @TechCrunch Disrupt. We talked through what the future holds for Cerebras. This year, we doubled our speed. That’s something for the team to be proud of. Still, a few things on my mind for next year: - Building more data centers - Deploying vastly more equipment in them - Driving down the cost of our tokens - Doubling our speed again - next year, and every year after - Building technology that breaks the status quo. That’s the engineering we love. It’s all about scale, and speed.

Today’s AI is the slowest you'll ever use. In 5 years, it will look like dial-up. Last week, I joined @Vercel’s panel on fast AI with @CompleteSkeptic, founder of @typesafeai. We discussed where AI goes next. We both agreed - people keep underestimating the demand for faster AI. There is no ceiling. 10x, 100x, 1000x. In the 1950s, computing experts guessed a handful of computers could cover a whole country’s needs. For 70 years, we have delivered more compute. And builders have used all of it. Every time. Demand for AI will be no different. Every gain will be spent on bigger problems: -Websites that rebuild their UI for each visitor, the moment they arrive. -Apps you can build as fast as you can describe them. In 10 years, we will look back at today’s AI the way we look at black and white TV. The future will be fast. Thanks for having us, @rauchg, @cramforce and @vercel.

At @cerebras, family matters. Last week, @communidiyi hosted a toy swap in our cafeteria. For a lot of our team, life outside work revolves around their kids. We don’t ask people to leave that at the door. We invite families to our company events. Spouses, kids. I often bring my parents too. Diyi is our community lead, and a parent herself. She asked a few parents if they’d want a toy swap. They enthusiastically said yes. All week, parents dropped toys off at her desk. Lots of puzzles. Even a circuit board. On the day, parents met other parents. They traded notes on raising kids. And their kids’ old toys found new homes. In Diyi’s culture, hand-me-downs bring the next kid good luck. A lot of Cerebras kids just got a little luckier. That is the culture we want at Cerebras. Where people bring their whole lives to work, family included.

At @cerebras, we like building the platforms on which other people’s ideas take flight. Yesterday at @TechCrunch Disrupt, I was asked how my grandkids will use AI. I told the room I don’t know. In the late ’90s, I was building some of the first hardware-based switches and routers. A whole community of us were doing it, and between us we made IP networking almost free. None of us knew the result would be WhatsApp or Viber. And that people around the world, even the poorest among us, could talk to and see loved ones far away.  Once moving data was cheap, other people built great things. I am not a builder of consumer products. I have a rule: if my mother buys it or uses it, I don’t want to make it or sell it. I love building infrastructure that makes other people’s ideas take flight. Most of my career has been about making networking faster or compute faster. At Cerebras, we make inference 15 to 30 times faster. And we make it fast so other people’s ideas and products can shine. Thanks to TechCrunch for having me.

AI is moving at the speed of software. Data centers are moving at the speed of real estate. That is why AI is so supply constrained. Demand for compute has grown exponentially. Almost nobody expected it to rise the way it has. But every part of the supply chain that delivers compute is a multi-year physical bet. Start with chips. A leading-edge fab costs $40 to $50 billion and takes about 5 years to build. Only a few companies in the world know how to build one, and even they can't build a dozen at once. So by the time a new fab opens, demand has already outgrown it. Then we need data centers. Even with a good builder, going from raw land and permits to a 50-megawatt data center takes about 18 months. And each part of the supply chain has its own supply chain. The building itself goes up fairly quickly. Turning it into a high-powered hotel for compute takes electricians, cooling, chillers, generators and electrical transmission switches, and many of those have long lead-times right now. Under all of it sits a grid built on technology from the 1940s and '50s. Builders are even buying up old paper mills and factories in the Rust Belt, because those sites already have the power and the permits. 5 years from now, the industry will still be racing to build enough compute to meet demand.

Silicon Valley often talks about long-term partners. There are few better examples than TSMC. In 2017, we flew to Taipei to ask TSMC to take a bet on us. For a product that didn't exist. Cerebras was 30 people. Pre-product. Pre-revenue. We were asking the greatest manufacturer in the world to collaborate on something that hadn’t been done before. And to work with a company smaller than the catering staff in their dining hall. TSMC’s fabs are our generation's pyramids - $40 to $50 billion, each. Their entire manufacturing process was built around producing chips the size of a postage stamp. We were proposing a chip 55x larger than anything ever made. It had been tried before. And it had failed every time in the 70-year history of computing. TSMC had a thousand reasons to say no. They were already a behemoth. The best in the world at making a completely different kind of chip. Why take the risk? Instead, they greenlit the project in the meeting. And they've been an extraordinary partner ever since. That decision sums up why TSMC continues to win: An openness to new ideas. And a willingness to collaborate, regardless of the size of the company.

“Is that AI thing going to kill us all in 10 years?" A carpenter working on my house asked. Then he shrugged and said “I hear you work in computers” The very question is our fault. As a community, we communicate poorly, and particularly poorly outside of silicon valley. We have failed to share the benefits and promise of our technology. We have failed to explain clearly what it can do for people and why they should be excited. To describe its power, we have focused on job loss and risk, rather than its benefits. Some have used scare tactics to illustrate the power of AI. And we have not treated the world outside Silicon Valley with the respect it deserves. We have not dignified their concerns with sufficient respect. The result is that many people are afraid of AI, including the people it will help most. This is a failure on the part of the AI community. And the result is that many in our nation are afraid. They are afraid of the very technology that will help them. We must combat the hysteria with careful, measured and honest communication. And with clear demonstrations of the good AI is already doing and will do in the future while being clear-eyed and articulating the risks. We are working on a powerful technology. The answer to powerful technology has never been choreographed tweets, performative statements from government officials, fear mongering, and panic. It has been patient work by serious people, dedicated to making it safe, and proving that the benefits outweigh the risks.

Proud to collaborate with @gimletlabs on their Inference Cloud. Planned speeds of up to 3,000 tokens per second. Gimlet's software breaks inference into phases and runs each phase on the silicon best suited to it. @cerebras handles the work where speed matters, and GPUs handle the high-throughput work. Users get faster AI and operators get better data center economics. Our teams have been working with joint customers since last year, and the integrated system is already serving tokens in private deployments. Gimlet is also a launch partner for CS-4. The first Cerebras-powered Gimlet Cloud data center is expected to come online later this year. Thank you to @zainasgar and the Gimlet team.