Apr 4, 2026 · Technology · Hardware

AI Chips: The Next Big Industry?

Envisioning a future where hyper-specific chips beat out the GPU monolith.

We are witnessing the birth of what could be the next trillion-dollar industry.

It's no secret that the world is changing a lot, and very quickly. Historically, these changes kill off expired industries and create the next global giants. AI changes the game completely, however. It's not one benefactor or loser, rather a whole transformation of the economy. And one niche area that seeks to soak up the benefits is the AI Chips industry.

okay, but what do you mean by AI Chips?

From what I'm saying, it's easy to assume that I'm talking about things like GPUs or TPUs and the like. They've already had explosive growth and it doesn't seem to be slowing down. Their biggest win is the necessity in all areas of AI: You need RAM to train, run, and store cache for LLMs. However, there may be a new player on the scene.

As I said earlier, GPUs are great at doing everything when it comes to AI. But there have been quite a few bets that it could perhaps be too great. In other words, many see GPUs as a "jack of all trades, master of none" when it comes to LLMs. As a result, startups are setting out to find what the "master" actually looks like.

Enter AI Chips. They're basically new hardware designed specifically to run AI. TPUs as I mentioned earlier could be classified into this category, but the specializing runs deeper. You have the LPUs by Groq, Cerebras' infrastructure, and my personal favourite, the Taalas chips. These chips aim to fix the inefficiencies GPUs face due to their more generalist roles.

The results so far are promising, but there's still plenty of room to grow. These providers have become capable of providing LLM tokens at lower costs and higher speeds than competitors still relying on GPUs. This gives them an early advantage for those who prioritize minimal costs and fastest output. However, the fact that these chips are still early in their development means making sacrifices. Often, these providers are unable to run the most recent, bleeding-edge AI models due to the loss of flexibility without GPUs. Due to this, they are still "behind" when it comes to the LLM-providing competition.

The most extreme case is visible with Taalas, where their HC1 chip can generate up to 17,000 tokens per second (!) but at the limitation of using Llama 3 8B at an 8,192 context window, which in 2026 is a service no one really wants. Why? Because now you have many trillion-parameter, million-token context window options. Even if Taalas' token costs are practically nothing, people are not willing to use models that are this far behind. Therefore, for AI chips to become mainstream, they've got some catch-up to do. But once they do so, they are positioned far ahead of the GPU competition.

The Future for AI Chips

AI chip companies may be behind; still, they're innovating fast. But what does the landscape look like if (or when) they're able to compete with frontier models? Well, the answer totally comes down to how willing these companies are to play ball.

Right now, the industry seems to be closed off, with companies keeping their innovations to themselves and instead resorting to be LLM providers through APIs and the like. It's understandable why they're doing so right now: they don't want people taking their half-baked tech and turning it against them. However, this is still a very profitable direction for these companies. Right now, it's no secret that LLM usage is very heavily subsidized. These companies are taking massive losses on compute to capture the market. A strategy that works now, but won't work forever. Down the line, they will have to stop the subsidies and start making profits. This is where shock will hit the AI community hard as their habits end up costing them millions, potentially. Here, the AI chip companies will be able to provide their models at prices far, far lower than other datacenters. At that point, people would be willing to sacrifice the model quality for the orders of magnitude in savings. This way, these companies win the AI race with minimal LLM training work actually done by them.

Going in another direction, a whole realm of possibilities opens up when you consider those companies potentially selling these chips. The first customer group they may target are corporations. Companies that need to use AI but are confused on what the best option is. Here, the AI chips become the best solution by far. The first advantage is the cost. By buying the chips directly instead of using datacenters, the companies save big on compute and it becomes an investment with a ROI that would make accountants tear up. Another big one is privacy. Every time someone uses cloud AI providers, they are sending their data to external companies. For the large amount of companies that would keep their proprietary things their own, this will be a big factor. The last major win is the customizability. They don't have to rely on a generalist model thousands of miles away when they can get their own custom chip to work exactly the way they want it to.

Another event that could take place is those AI chips moving to consumer markets. This is where everything changes, and how AI chips could truly become the dominant industry in the future. With the way these chips are designed, it only takes so much to make them truly compatible with everything. Your car? AI. Your fridge? AI. Your phone? AI. Your knife? Believe it or not, AI. (This one I'm only slightly skeptical about) The great thing about this is the cost is almost nothing when it comes to the consumer. You're getting all those insane features at the cost of peanuts. And the privacy aspect means that there's nothing to worry about. We're looking at a future where models aren't one size fits all; instead, you'll have a fridge model, a TV model, a home cleaning assistant model, and a stove model. This can even start whole sub-industries with things like maintenance shops and "Upgrade your AI chip here!" stores in every neighborhood. If things go just right, AI chips may become as core as electricity in the new world order.

Of course, this is all very far away. Which presents now as the single best opportunity to be watching closely. There's still a lot of work to be done until we get here, but momentum is favoring these companies. Just recently, the TurboQuant research paper (that I wrote a blog about) gave these companies a huge relief, as they now have one less factor to worry about when it comes to competing with GPUs.

Anyway, that's it for this blog post. Who knows what the future holds for this industry. Maybe one day, your sink will be reading out this exact blog post for you as to how it all happened.