How AI Model Costs Are Crushing Hardware Spending

Lucas: There's a really interesting split happening right now in the AI market. If you look at the stock moves over the past week, hardware names are all over the map - NVIDIA's basically flat, AMD is up nearly twelve percent, Super Micro is down almost ten. But the software and platform companies - Snowflake, ServiceNow, Adobe - they're all up solidly. Luna: So the narrative that 'AI equals NVIDIA' is starting to fray?

Lucas: I think it's more nuanced than that. The hardware story isn't broken, but the market is starting to price in a shift. For the last two years, the big spend was on training - building bigger and bigger models. That required massive clusters of GPUs.

But now, with models like Claude Sonnet 5 launching today at a much lower price point, the economics of inference are changing fast. Luna: Right, Sonnet 5 is supposed to be Anthropic's cheaper agent-friendly model. I saw they're positioning it as a way to run AI agents without breaking the bank. Lucas: Exactly.

And that's the key. When inference costs drop, it unlocks new use cases - but it also means that the hardware required to run those models might not need to be as expensive or as specialized. You can run a lot of inference on mid-range chips or even CPUs for some tasks. So the huge capital expenditure cycle that benefited NVIDIA and AMD for training might not repeat at the same scale for inference.

Luna: That would explain why Snowflake is up over twelve percent in the last five days. They're an inference platform, essentially - companies use Snowflake to run models on their own data. Lower inference costs mean more queries, more usage, more revenue. Lucas: Snowflake is a great example.

They're up twelve point six percent in the last week. ServiceNow is up nearly six. Adobe up four. Meanwhile, Broadcom - which is heavily tied to custom AI chips - is down.

Oracle, which is building out AI data centers, is down seven percent. The market is starting to reward the application layer over the infrastructure layer. Luna: But AMD's rally suggests there's still appetite for hardware - just maybe not NVIDIA specifically. Lucas: That's a good point.

AMD's eleven point eight percent gain this week likely reflects some rotation out of NVIDIA into a cheaper alternative. AMD's mi series accelerators are gaining traction, and their price-performance ratio is attractive for inference workloads. Plus, there's a startup called Etched that hit a five billion dollar valuation today with a billion in sales for its AI chip - so the hardware story isn't dead, it's just fragmenting. Luna: And Super Micro dropping almost ten percent - that's the server maker that rode the AI wave.

Maybe the market is worried about oversupply? Lucas: Could be. If inference doesn't require as many high-end servers, the demand for Super Micro's liquid-cooled racks might not grow as fast as expected. There's also the broader enterprise adoption story we talked about a few episodes ago - a lot of companies are hitting a wall with AI deployment.

The cost isn't just the model, it's the integration, the data pipeline, the talent. So even if inference gets cheaper, the total cost of ownership might still be too high for many. Luna: That's why you see Salesforce and Adobe doing well - they're embedding AI into existing software that companies already pay for. No new infrastructure needed.

Lucas: Right. The winners in this next phase are likely to be the platforms that make AI easy to consume, not the ones that sell the shovels. Snowflake, ServiceNow, Adobe - they all have large installed bases and can layer on AI features with incremental pricing. That's a much more predictable revenue stream than selling chips that might get obsoleted in eighteen months.

Luna: How does Claude Sonnet 5 fit into this? I mean, Anthropic is a model provider, not a platform per se. Lucas: Right, but they're enabling the platform play. By making agents cheaper to run, they're essentially expanding the total addressable market for inference.

And that benefits the platforms that host those agents. So Anthropic's pricing strategy is actually a tailwind for Snowflake and others. It's a virtuous cycle: cheaper models drive more usage, which drives more platform revenue, which funds more model improvements. Luna: But it also means the model providers themselves might struggle to monetize directly.

If Anthropic keeps cutting prices, their margins get squeezed. Lucas: That's the tension. OpenAI and Anthropic are in an arms race to offer the cheapest API, but they need to keep spending on R&D and compute. That's why you see them raising huge rounds and also launching consumer products like ChatGPT - to diversify.

The real money might be in the distribution, not the model itself. Luna: So for an investor, the takeaway is: look at the companies that sit between the model and the end user. Lucas: I think that's a smart framework. And it's playing out in the data this week.

The tech ETF IGV is up over five percent - that's broad software. So the rotation is real. But I don't think it's a straight line. Hardware will have moments - like AMD this week - and some chip companies will find niches.

But the big, sustained returns over the next few years might come from the application layer. Luna: And if inference costs keep dropping forty percent a year, as we've seen, that only accelerates the trend. Lucas: Exactly. So if you're building a product or a portfolio, you want to be asking: does this company benefit from cheaper AI, or does it get commoditized by it?

That's the question. Luna: Speaking of building - if today's conversation gave you something useful to think about, we'll just say that keeping this show ad-free and independent is something we value. If that resonates, you can support us at buy me a coffee dot com slash fexingo. No pressure, just a way to keep the conversation going.

Lucas: And we appreciate everyone who listens and engages. The feedback we get helps shape episodes like this one. Luna: So back to that question - who benefits? Snowflake's rally this week suggests the market is starting to answer it.

But we'll be watching how the next few earnings calls play out. Lucas: Yeah, the Q2 reports in July and August will be telling. If software companies show accelerating ai related revenue, and hardware companies start guiding lower, that's the confirmation. Until then, it's a hypothesis - but a pretty compelling one.