The latest roadshow summary of Cerebras, official statement for reference. 1. OpenAI trading/production validation/revenue visibility OpenAI has reached hardware partnerships with multiple enterprises, and currently only Nvidia and Cerebras have put their models into production operation. We have signed a binding 'take or pay' contract with OpenAI, involving 750 megawatts of electricity, with a total contract amount exceeding 20 billion US dollars and a prepayment of 1 billion US dollars. The agreement will be implemented in three phases: deploying 250 megawatts each in 2026, 2027, and 2028. In 2026, 250 MW will be delivered through the Cerebras cloud platform and leased for three years. OpenAI has the right to renew the lease for the fourth and fifth years; From 2027 to 2028, OpenAI can choose to buy out the devices or continue to receive them through cloud platforms. In addition, OpenAI has an additional 1.25 gigawatt capacity option, which is one of the largest transactions in Silicon Valley history. The financial accounting and securities model both adopt the most conservative assumption: all three phases of capacity will be delivered through cloud platforms, with the longest revenue sharing cycle and the highest capital expenditures. 2. Why does OpenAI need Cerebras/Delay as a product issue? OpenAI is working with Cerebras to solve this problem, and this deal is worth over $20 billion in the coming years. Sam Ultraman summarized the value of speed: When AI responds quickly, users use it more frequently, stay longer, and apply it to more important scenarios. 3. The explosion of inference/intelligent agent computing: Currently in the era of inference, inference is a technology that generates better answers through computing power, and it is also the core reason for the huge amount of AI inference computation. This is extremely advantageous for Cerebras, which has a superior computing speed. The world's top models, including Gemini, Claude, OpenAI GPT, and various models from China and the United States, all use inference techniques. Single inference only requires basic computing power to generate answers, while critical reasoning consumes about 100 times more computing power. By disassembling problems, solving them step by step, verifying and integrating, drafting and reviewing, and outputting solutions, the quality of answers is higher, but the time and computing power consumption are significantly increased. Moving from inference to the stage of "intelligent agent+inference", a single query will initiate multiple inference streams, and the computational load will increase exponentially. 4. Core Speed Proposition/Cross Model Breadth. Cerebras has built the world's fastest inference infrastructure, which is about 15 times faster than the fastest GPU. This advantage is not an exception, covering various models in the United States and China, large and small models, trillion level and billion level parameter models, and achieving a 15 fold speed increase in all scenarios. The industry has widely recognized us as the world's fastest solution, thanks to our breakthrough in wafer level processor technology, which was once considered impossible to achieve in 75 years. Our processor size is like a dinner plate, while traditional processors are only the size of stamps. The chip area is 58 times larger than the largest GPU, and the memory bandwidth exceeds GPU 2500 times, making it the core foundation of high-speed inference. 5. Architecture/Why is wafer level technology crucial for inference? The core logic of inference is: to generate the next word, all model weights need to be transferred from memory to the computing unit and executed sequentially word by word. The 70 billion parameter model is a small model with a weight data volume equivalent to approximately 100 high-definition movies, and generating each word requires transmitting data of this scale. A 1000 word answer in the demonstration is equivalent to transmitting data from 100000 high-definition movies. The transfer speed from HBM memory of GPU to computing unit is about 8TB/s, while the transfer speed of Cerebras wafer level processor reaches 21000TB/s, which is more than 2600 times faster than GPU. We adopt a large chip design and are equipped with high-speed SRAM memory. By expanding the chip area, we have solved the traditional shortcoming of low SRAM storage density. 6. Expanding to Large Scale/Trillion Parameter Models GPU is facing industry challenges in model scaling. Based on the practice of OpenAI and other clients, we share a solution for carrying super large models: for the 1.1 trillion parameter model, our software will split the model and allocate it to different processors to run inference in a pipeline manner, which can support the world's largest open source and closed source models. We do not sell chips separately, but focus on the sales system and cloud platform usage duration. 7. Cost/power consumption during high-speed operation compared to GPU. The industry often misunderstands that GPUs have a unified pricing. GB200 NVL72 can output about 5 million tokens/second at a low speed of 35 tokens/second, supporting tens of thousands of users to use simultaneously. GPU efficiency is extremely high during low-speed operation; But at a speed of 270 tokens per second, it can only support one user, which is extremely expensive. Performance comparison needs to clarify the running speed. GPU efficiency is superior when running at low speeds. After 100-150 tokens/second, GPU costs soar and energy efficiency drops sharply. Our cost and power consumption are only a very small part of GPU. If our performance is included in the Nvidia performance chart, we would need to significantly adjust the horizontal axis scale. Our performance has exceeded the range of the chart, and no matter how many GPUs we have, we cannot achieve our speed. We have no competitors in this high-speed range. 8. Gross profit margin/pricing/cost leverage. We have achieved full production yield, with third-generation product yield better than B200. There is still room for compression in supply chain costs, and pricing strategies can be further optimized. Last year, the revenue was about 500 million US dollars, the gross profit margin was about 50%, and the supply chain investment was 250 million US dollars. The operational efficiency needs to be improved. After the expansion of supply chain investment scale, costs will continue to improve; In terms of pricing, we have raised prices by 30% in the past four months. Anthropic has launched a service with a speed increase of 2 times and a fee increase of 6 times, while our speed is 15 times faster than it. There is still a lot of room for exploration in the pricing ceiling. Bob added: There is room for profit margin improvement in both pricing and cost aspects, and the medium-term potential may exceed expectations. 9. The problem lies in the decline of gross profit margin/production capacity limitation in 2026, rather than product economy. The main reason for the decline in gross profit margin in 2026 is due to the low profit margin of self owned data centers under the demand for AI inference in the second half of 2025, resulting in a 4 percentage point decrease in gross profit margin in 2025 and further decline in 2026. With the expansion of self owned data center production capacity and the gradual exit from leasing mode, the gross profit margin will quickly rebound and return to the long-term goal of over 60%. The gross profit margin corresponding to the current backlog of orders has reached this long-term level. Once the production capacity of our own data center is sufficient and there is no need to lease external production capacity, the gross profit margin will quickly increase to the long-term target value. This article is sponsored by @ bitget_zh, titled 'Bitget Buying US Stocks: Instant Entry, Smooth Trading'