Cheap Solar Panels and AI Race

Cheap Solar Panels

All market prices are fundamentally governed by supply and demand forces. The market for solar panels today is far oversupplied, and that’s creating serious financial instability for suppliers. Access to cheap energy is the bedrock of economic growth, so while this is problematic for the suppliers, it’s generally beneficial for everyone else.

To add context, 80% of the world’s solar capacity is produced in China. There is a longstanding fight within the US on solar panel import tariffs, as US producers cannot compete with Chinese manufacturing. The US uses tariffs to help maintain domestically competitive manufacturing and protect jobs. This seems obviously shortsighted when we have an intense focus on reducing inflation. We all benefit tremendously from cheap energy, and that includes access to cheap solar panels. I’m not sure we need domestic supply chain redundancy with solar the way we do with critical medicines or semiconductor fabs.

The energy cost of a standard monocrystalline solar panel is quoted in $/watt. In 2010, it was $1.50/watt. Heading into Covid, it was getting closer to $0.20 with a brief jump to just below $0.30 in 2022. Today, that cost has fallen to a record below $0.10. Again, we all benefit tremendously from access to cheap energy, especially when dealing with inflation.

AI Race

Someday, a book will be published recounting the daily drama of the early stages of the AI arms race. Meta’s latest open-source model, Llama 3.1, is competitive with ChatGPT, which adds to the commodification of AI tools. I’m still not entirely clear on the merits of open and closed-source models, other than “open” has been generally useful in the past. Some say open source is not genuinely open for various technical reasons like you can’t see the dataset it was trained on, and you can’t modify the weights the way you would be able to if you were modifying code.

You can test drive this model on Groq.com and Meta.AI. The image generation on the Meta.AI site updates in real-time as you’re typing. This is the first time I’ve seen a model able to generate images this fast, but maybe it exists elsewhere, and I just haven’t seen it. Real-time image generation is more of a peak at the future than a useful tool today.

A lot of AI tools are like this, which is why Wall Street is starting to get concerned and/or upset that there isn’t a more immediately obvious return on capex. Investors want to see more evidence, but that may be impossible to isolate given how flexible these tools are and how they can help in subtle, minor ways throughout the day.

Are companies overspending in an AI arms race? Probably. I don’t think that inherently means there is an AI bubble overall. Stock market bubbles are traditionally associated with a parabolic price for simply changing the company name. Long Island Iced Tea corp renamed as Long Island Blockchain during the 2017 crypto bull market boom and bust. The famous example is any company that added “.com” to its name in the early 2000s. Even the parabolic NVDA stock price is supported by a generally reasonable PE multiple based solely on their historical results.