Remix.run Logo
▲ aurareturn 5 hours ago

My prediction:

In the AI/AGI/ASI era, the 4 most valuable things will be:

- Model

- Compute (chips/energy)

- Manufacturers

- Materials

You can already see it beginning at the chip design stage. If AI makes it so that it takes 2 months to design a chip instead of 2 years, then there will way more chips designed to accelerate applications that previously weren't viable economically.

Therefore, it's manufacturers like TSMC that will become more valuable. There will be more chip companies, more designs, more demand.

And in order to make things, you need materials.

In order to design the chip, you need the model and AI hardware.

▲hkalbasi 4 hours ago | parent | next [-]

Why you wouldn't ask the ASI to design a TSMC too? Then the silicon (sand) will become more valuable. Even raw materials may become less important if ASI manages to achieve space elevator or something like that. But my prediction is that the growth of current AI generation is hitting end of the S-curve, and an AI era where AI can design chips in 2 months and fabs in 2 years is not as soon as it seems to be.

▲aurareturn 4 hours ago | parent | next [-]

It's possible ASI can design and build a TSMC as well. But instead of that happening within 5 years, maybe that's 10 years down the line?

▲freakynit an hour ago | parent | next [-]

Instead of focusing on silicon based chips, maybe we need to point these AI's at chips that can be grown like life grows out of a single seed cell.

Put that "seed" in a soup of chemicals, and within few days a chip builds up.

▲freecodeio 2 hours ago | parent | prev [-]

I mean given the state of this conversation you might as well as just spin a wheel.

▲aurareturn an hour ago | parent [-]

Yes. But I'm sure that the days of many chip designs and fast are coming. I'm not sure if AI will be able to replace TSMC.

▲barrenko 4 hours ago | parent | prev [-]

One giant all-encompassing loop.

▲zarify 3 hours ago | parent [-]

Agent Centipede? Seems on brand.

▲audunw 2 hours ago | parent | prev | next [-]

I don’t think models will be valuable. It’ll be a shared commodity. The models all seem to converge in performance, and eventually fully open source models with open source training data will become “good enough”. We may end up with a “Linux” of the LLM model world, that just ends up dominating everything else.

It’s similar to how proprietary source code is rapidly losing its value. The most valuable code now is open source, and whoever is best at building processes around open source becomes the winners.

One caveat is that new AI architectures could change the equation, at least temporarily.

We’re converging on intellectual property losing its value in general. What ends up being valuable can be boiled down to culture, human processes and physical materials.

▲fedpost 4 hours ago | parent | prev | next [-]

Much like with agentic PCB design, the HDL isn't the hard part.

▲dlisboa 4 hours ago | parent | prev | next [-]

I’d say TSMC becomes less valuable because there will be multiple fabs trying to meet demand. There’s no way for a 10x increase in fab capability to be handled by just the very few companies that exist.

Also most likely these bespoke chips wouldn’t need the latest nanometer processes to work more efficiently, lowering the value of highly advanced EUV machines.

▲aurareturn 4 hours ago | parent | next [-]

  I’d say TSMC becomes less valuable because there will be multiple fabs trying to meet demand. There’s no way for a 10x increase in fab capability to be handled by just the very few companies that exist.
Fabs follow Rock's law. The cost of a fab doubles every 4 years. https://en.wikipedia.org/wiki/Moore%27s_second_law

Right now, it costs $20b - $40b to create one fab. And that's without the R&D time and cost for the actual node process such as N2.

So could AI recreate TSMC? Maybe in a more distant future. But it's a lot harder to recreate TSMC than to design a chip and send the file to TSMC.

I think it's more likely that AI recreates some chip manufacturing breakthrough than to compete against TSMC without one.

▲edg5000 3 hours ago | parent [-]

Maybe Rock's law will change now that AI makes technological development less laborious, which is where the money was going. All non-technical overhead also gets reduced; what a 10-person company can do today assuming they can run their agents 24/7 makes the situation from just 10 years ago look medieval. So we'll see more foundries and more production capacity pop up, nibbling at TSMCs feet.

▲petra an hour ago | parent | next [-]

It's just a guess, but given that it takes 20 years to take a chip manufacturing idea to market, ideas are not the bottleneck, but working in the physical world is.

▲aurareturn 3 hours ago | parent | prev [-]

If there is a breakthrough in chip manufacturing, IE, a completely new process, then maybe.

We don't really know that.

We do know that AI is already there or close to being able to design a chip, verify it, and send it to TSMC.

▲luckystarr 4 hours ago | parent | prev | next [-]

These companies operate under a strict risk modeling framework for their investments, trying to model 10 or so years ahead, because in order to stay profitable they have to shell out 10s of billions at a time. So competition may arise, but very slowly. Heavily simplified.

▲lazystone 4 hours ago | parent | prev [-]

> there will be multiple fabs trying to meet demand

From what I know, that's easier to say, than to implement. Creating a chip is so complex, that it works for TSMC because of the scale. On lower scale, production won't be profitable.

▲xattt 4 hours ago | parent | prev | next [-]

When design time is compressed to this extent, do subsequent steps that used to be insignificant become the new bottleneck?

▲petra 4 hours ago | parent | prev | next [-]

I would add "simulation/emulation platform" as one of those things.

▲DeathArrow 4 hours ago | parent | prev [-]

>Model

I think China is working towards making the models commodities.

The real breakers will be access to cheap raw materials and to cheap energy.