Huawei makes 1.4nm chip??

追加: 2026-05-31

[00:00:00]Huawei just announced what's called Tao scaling law at the chip and infrastructure level and it's China's own way of competing with the rest of the world in the hardware space. Now while most of us that live in the application layer building cool apps and using agents might not feel the impact of what's happening in the lower hardware layer, the tension underneath is really interesting to learn about. So how exactly is this new scaling law proposed by Huawei impact the technological race between the US and China? Welcome to Caleb Bright's Code, where every second counts. Quick shout out to Centra. More on them later. It's an understatement to say that Huawei was set back technologically from the series of bans that the US placed on China since 2019. Huawei was not only put on the entity list, which was pretty challenging at the software level since Google ecosystem was no longer accessible, but more importantly, they lost access to the means to produce their hardware through TSMC and equipments like EUV lithography. For reference, this is where we are in the curve where the red line is the size of the technology in producing transistors measured in nanometers and the blue line shows the power density that's measured in power per squared nanometers. And this blue line becomes important later in the video. But historically speaking, semiconductors got a free lunch in scaling by making transistors smaller following the popular Moors law that we all know by observing all the way back to 1965. And this trajectory was formalized more clearly later in what's called the den scaling where you practically reduce the voltage by the same proportion as the transistor got smaller. And in 2019 and 2020, China was effectively kicked out from participating in this trajectory largely because the equipment to efficiently produce chips below 7 nanometers called EUV lithography had been banned which made it tricky for China to continue on this path without this very equipment.

[00:01:54]Now there's one nuance to add here which is that the old scaling law famously broke around 2005 as you can see here where the blue line actually starts to explode as we approach around 100 nanometers which is around the size of a virus as it went closer to the limits of physics which meant simply making transistors smaller had its own limits as it got smaller than the size of a virus and currents started to leak largely due to just how small the gate was getting. And the result of hitting this wall meant we couldn't increase frequency as transistors got smaller at the same time without the power getting out of hand. One example is Pentium 4 from Intel, which I remember was a pretty big deal back then. And Intel set their target on getting up to 10 GHz in frequency, which is incredibly fast, banking on the continuation of this chart. But they ended up canceling this plan after topping at around 3.6 GHz.

[00:02:48]And even 20 years later, the CPU that I used today for my workstation is still around the same base frequency. Now, Intel ended up switching their strategies and expanding out by adding more cores among many other improvements instead of just pumping up the frequency. Now, when we look at GPUs like Ruben GPU from Nvidia, they're built on three nanometers while consuming up to 2300 watts and operating at a lower clock speed than CPUs. And in order for China to compete with the US where models like Deep Seek V4 will eventually run inference in Huawei's Send Pods, China has to chart their own path to continue their race. And while at the moment DeepS models are still trained on Nvidia cards still, this will likely change as chips get more established and even for other labs in China as well, which is where we now get to Tao scaling law. How does Tao scaling law change things? How does China plan to work around the EUV ban to catch up to the rest of the industry? Or is China just cooked entirely? And this tower scaling law is nothing but a marketing scheme. But first, let's talk about CRA, who's sponsoring this video. There are so many AI assistants and apps out there. But CRA is one that is meant to work with you on a day-to-day basis. It comes with 12 AI helpers out of the box, ranging from Buddy, who works on business development, all the way to Gigi, who helps on personal development.

[00:04:09]And you probably don't need all of these, but a select few, which is fine.

[00:04:13]Personally, my favorite is Buddy, who I can easily send a prompt to check for breaking news and events around an AI and just send me a Slack message when it's done. And here I got a message notification from Buddy. Now, you can certainly extend this to how you want CRA to help you on a day-to-day basis.

[00:04:28]And if you want to create your own tasks, CRA has the ability to create new tasks easily with recommendations on what kind of things CRA can help you with. CRA also comes with a ton of integrations as you can see which means you can leverage your existing platform and give CRA access to granular control and permissions to start working with it. You can get a 72% discount on all plans for a limited time. Link in the description below. Thankfully, Huawei released a 16page paper outlining in detail the background and specs on their plan. And sadly, the paper reads more like a road map than an actual plan on how Huawei is planning to execute on their ambitious plan. For clarity, what Huawei revealed was getting to 1.4 nanometer equivalent while achieving zeraflop level infrastructure by 2030 in their super pod GPU cluster and reaching the density of 400 million transistors by squared nanometers and hitting 5 GHz at the same time. And a lot of their plan rests on this logic folding which goes beyond a simple geometric scaling in the 2D space but literally folding the logic upon itself like an origami or actually more like the scene in the movie Inception and Doctor Strange where the city is folding on itself. And using logic folding they can technically continue to compete in how the transistor density measured in the 2D space and that's why they're framing it as 1.4 4 nanometer equivalent emphasis on the word equivalent. And while all this sounds pretty ambitious, the growing question is how how exactly does Huawei really plan to execute on this rather ambitious plan. There's a reason why we prefer to use EUV over DUV when it comes to manufacturing semiconductors. For example, if I use this thick brush like a DUV to paint a drawing that needs to be precise, while technically I can paint it in smaller precision, it's not only trickier to actually manufacture this drawing, it leads to a higher error rate and the entire process is just really tricky.

[00:06:27]Meanwhile, a much smaller brush like the EUV allows me to be more precise and accurate, which makes manufacturing a lot smoother. And that's exactly the criticism when it comes to just how Huawei plans to manufacture their design and package them in a way that can mass prodduce their hardware to make millions and millions of phones and graphics cards that will be needed. in the entire 16-page paper that's outlined. While it sounds cool and honestly really exciting to read, given how it reads like a classic underdog trying to find an alternative path to remain competitive, we really haven't yet to see how they plan to go beyond theory and demonstrate the actual production. But if they do end up executing on this plan and and eventually they replicate the EUV technology in China internally, it could mean that given how the model architecture is going that China is positioned to take over this band of market right here. This section of the market is where models are served at around 50 to 100 tokens per second for an average level intelligence model. And while the US and inference providers here can also target this as well, the US is pushing the boundaries in more and more capable and intelligent model and China will be well positioned to use much more affordable ascent line GPUs from Huawei and plenty of work lives right in this range right here like checking emails, small bug fixes, correspondences, daily routines and reminders, organizations. All of these are ripe for China to take as the aggregate demand for the models grow in all fronts and all levels of intelligence. So the narrative is unfolding in a way where cutting off China could lead to an unorthodox way of continuing in the hardware race. And while this entire ordeal is more of a forced innovation rather than a true breakthrough, we could see from Huawei and China a new way of scaling their infrastructure as they focus away from geometric space and now to time instead.