-
5 days ago | 
semianalysis.com | Dylan Patel |Kimbo Chen
In our AI Scaling Laws article from late last year, we discussed how multiple stacks of AI scaling laws have continued to drive the AI industry forward, enabling greater than Moore’s Law growth in model capabilities as well as a commensurately rapid reduction in unit token costs. These scaling laws are driven by training and inference optimizations and innovations, but advancements in compute capabilities transcending Moore’s Law have also played a critical role.
-
2 weeks ago | 
semianalysis.com | Dylan Patel |Daniel Nishball |Wega Chu |Ivan Chiam
For the past six months, AMD has been in a Wartime stance. They have been working hard and working smart towards their goal of being competitive with Nvidia. At its Advancing AI 2025 event, AMD launched the MI350X/MI355X GPUs which could be competitive to Nvidia’s HGX B200 solutions for inference of small to medium LLMs on a performance per TCO basis.
-
2 weeks ago | 
semianalysis.com | Tanj Bennett |Dylan Patel
Standard Ethernet intially lost significant market share to Nvidia’s InfiniBand in the early days of the GenAI boom. Since then, Ethernet has started clawing back market share, largely driven by cost, the various deficiencies of InfiniBand, as well as the ability to add more features and customization on top of Ethernet.
-
2 weeks ago | 
semianalysis.com | Dylan Patel |AJ Kourabi
The test time scaling paradigm is thriving. Reasoning models continue to rapidly improve, and are becoming more effective and affordable. Evaluations measuring real world software engineering tasks, like SWE-Bench, are seeing higher scores at cheaper costs. Below is a chart showing how models are both getting cheaper and better. Reinforcement learning (RL) is the reason for this progress.
-
1 month ago | 
semianalysis.com | Dylan Patel |Daniel Nishball |Ivan Chiam
It has been long claimed that AMD’s AI servers can achieve better inference performance per total cost of ownership (TCO) than Nvidia. We have spent the past six months investigating and validating this claim through a comprehensive analysis and benchmarking of inference solutions offered by both Nvidia and AMD. We expected to arrive at a simple answer, but instead the results were far more nuanced and surprising to us.
-
1 month ago | 
semianalysis.com | Dylan Patel |Jeremie Eliahou Ontiveros |AJ Kourabi |Ivan Chiam
The US has signed two landmark agreements with the United Arab Emirates and Kingdom of Saudi Arabia (KSA) that that will noticeably shift the balance of power. The deals have economic, geopolitical, and national security implications and taken together we believe these bilateral deals reinforce America’s claim to AI leadership. They also raise real concerns – but none that the US can’t address.
-
2 months ago | 
semianalysis.com | Dylan Patel |Jeremie Eliahou Ontiveros |Maya Barkin |Gigawatt Dreams
Over the last few months, there have been a number of headlines raising concerns about Microsoft’s reduction in datacenter leasing activities including a few datacenter leasing cancellations. We called out Microsoft’s exit of multiple datacenter leasing contracts to our datacenter model clients on December 17th (two months before the Wall Street headlines). This story isn’t as simple as it seems and there is a lot more nuance to Microsoft’s actions.
-
2 months ago | 
chinatalk.media | Jordan Schneider |Dylan Patel |Doug O'Laughlin
Dylan Patel and Doug O’Laughlin of SemiAnalysis explore Huawei new rack and its broader strategy of leveraging structural advantages to resist, circumvent, and nullify export controls — as well as how the U.S. government should respond. We also get into chip tariff policy. Semiconductors are exempt from Trump’s reciprocal tariffs for now, but the administration has signaled that chip-specific tariffs are coming soon.
-
2 months ago | 
semianalysis.com | Dylan Patel |Daniel Nishball |Myron Xie |Wega Chu
Huawei is making waves with its new AI accelerator and rack scale architecture. Meet China’s newest and most powerful Chinese domestic solution, the CloudMatrix 384 built using the Ascend 910C. This solution competes directly with the GB200 NVL72, and in some metrics is more advanced than Nvidia’s rack scale solution. The engineering advantage is at the system level not just at the chip level, with innovation at the accelerator, networking, optics, and software layers.
-
2 months ago | 
semianalysis.com | Dylan Patel |Jeff Koch |Lithos Graphein
“Lithos Graphein” is back as a guest contributor to this report. Follow him on twitter @lithos_grapheinIf TSMC is “the most important company in the world,” according to President Trump, then EUV lithography tools are surely “the most important machines in the world.” Does that make SPIE Advanced Lithography & Patterning the most important conference in the world? Not exactly, but it’s at least a good opportunity to see the latest progress in EUV and advanced logic.
