The Signal #038 — Dakota’s read on the AI news that actually matters to people running a business.

Most operators are not thinking about chips. They are thinking about prompts, tools, and workflows. That is the right place to spend most of your attention. But occasionally something shifts at the infrastructure layer that changes the cost and availability of everything built on top of it. This is one of those moments worth understanding.

What happened

A post circulating on the LocalLLaMA subreddit mapped out seven Chinese companies that are already shipping AI chips described as H100/H200-class. For context, the H100 and H200 are Nvidia’s high-end data center chips (the specialized processors used to train and run large AI models) that have been the de facto standard for serious AI workloads. According to the post, most of these seven companies have IPO’d in the last six months.

You can read the original breakdown here.

The summary is this: a meaningful cluster of competitors to Nvidia’s dominant position is not theoretical anymore. They are shipping product. Some of them are publicly traded. And they are doing it in a compressed timeframe that most industry observers did not anticipate.

Why it matters for operators

If you are running a business that uses AI, you probably do not buy chips directly. You access compute (processing power, rented by the hour or by the token) through cloud providers or AI platforms that sit on top of the hardware. So why should you care what is happening at the chip layer?

Because chip supply and competition is what determines the price and availability of that compute over time.

When one company controls the dominant hardware stack, pricing reflects that control. When multiple credible competitors start shipping comparable hardware, the market dynamics shift. Cloud providers gain more negotiating leverage. New entrants can build data centers without waiting years for Nvidia allocation. That competition, if it develops, works its way downstream toward the per-token and per-hour costs that operators actually pay.

A SaaS company running inference (the process of getting an AI model to generate a response) at scale cares deeply about those per-token costs. A healthcare platform processing thousands of documents a day cares. A manufacturing operation using AI for quality inspection cares. The chip layer is distant from those workflows, but it is not irrelevant to them.

What most people get wrong

The instinct when reading news like this is to either dismiss it or overcorrect.

The dismissal goes: these are Chinese chips, there are export controls, geopolitics makes this complicated, nothing changes for me. That reaction misses the structural point. Even if you never touch hardware from any of these seven companies, their existence changes what Nvidia has to do on pricing and roadmap. Competition does not have to reach you directly to affect your costs indirectly.

The overcorrection goes: everything is about to change, Nvidia is finished, pick a new winner. That is not supported by anything in the source either. Shipping product and matching Nvidia’s software ecosystem, developer tooling, and years of optimization are two very different things. The chip is one part of the stack. The software layer that makes it usable for AI workloads is its own years-long project.

The accurate read is somewhere in between. There are now seven companies, most newly public, shipping hardware in a class that was recently the exclusive territory of one dominant supplier. That is a real market development. Its full implications will take time to land. But it is happening, and the timeline is faster than most people assumed a year ago.

The lesson for anyone making AI decisions right now

You do not need to become a chip analyst. You do need to understand that AI infrastructure is not a fixed cost with one permanent provider. The compute market is actively competing for the next several years of AI demand, and that competition is now coming from more directions than most roadmaps accounted for.

For near-term decisions, nothing about this changes what tools you should be using or how you should be building your workflows. For longer-term decisions about which platforms and providers to build deeper integrations with, it is worth asking how exposed your stack is to a single point of infrastructure control.

The operators who tend to do well with AI over time are not the ones who picked the best tool in 2024. They are the ones who built in a way that lets them move when the landscape shifts. Right now the landscape is shifting at a layer most people are not watching.

If you want help thinking through how your AI setup holds up as the infrastructure layer evolves, start at xovionlabs.com.