The Chip Was Never the Moat. The Software That Boots It Is.

A San Francisco startup called Infinity raised $15 million in seed funding at a $100 million post-money valuation on the claim that its agent, Ignition, can automatically generate optimized inference software for any new AI chip.

That is a small round in a month full of enormous ones. It is also the only deal I saw in July that attacks the actual lock-in.

Here is the thing the capex headlines keep obscuring. The reason a frontier lab cannot casually switch silicon is not that competing chips lack transistors. It is that a mature inference stack, the kernels, the compiler passes, the serving layer, the thousand small optimizations, takes years of specialist engineering per architecture. The chip is the visible asset. The software that makes it usable is the moat.

Watch how the rest of the month lines up against that.

Nvidia is putting roughly $5 billion of equity into Safe Superintelligence alongside Vera Rubin access worth about a 10x lift in the lab's compute. Oxmiq is selling a licensable GPU architecture to chipmakers that want their own accelerator without the full design cycle. China's ChangXin Memory raised at least $8.6 billion in its Shanghai listing and rose more than 450 percent on debut. Samsung, in the middle of a quarter where semiconductor profit jumped more than 250-fold, told the market the AI chip shortage gets worse through 2027 and 2028.

Read those together and you get an odd picture. Capital is pouring into alternative silicon and alternative memory. The incumbent is deepening its hold on the frontier labs through equity and early hardware access. And the supplier with the best view of physical capacity says the shortage runs for two more years.

Alternative silicon only relieves that shortage if someone can actually run production workloads on it. Which is a software problem.

The model layer is already moving in the same direction. Moonshot released Kimi K3's 2.8 trillion parameter open weights for public download. A coalition of about 25 companies including Nvidia and Microsoft signed a letter asking Washington not to restrict open-weight models, and Nvidia then convened an Open Secure AI Alliance with more than 30 members to secure the open stack. If the weights are portable and the serving stack becomes portable, what is left of the lock-in is physical supply, and physical supply is the one thing Samsung just said is constrained.

I would not bet the thesis on one seed round. Automated kernel generation has been promised before, and "works on any chip" tends to mean "works on any chip, at 60 percent of hand-tuned throughput", which is not good enough when you are paying for the other 40 percent in power. Infinity is at a $100 million valuation, not a $100 billion one, and that is the market's honest confidence interval.

But the direction is worth pricing. The strategic question for the next 18 months is not who has the fastest accelerator. It is how expensive it stays to move off one.

**What to do with this**

Watch for one specific event: a named lab running a production inference workload on non-Nvidia silicon using a stack it did not hand-build. Not a benchmark, not a demo, a production workload. That is the tripwire, and it will show up in a hiring page or an engineering blog long before it shows up in a market note.

And when you read the next enormous compute deal, ask what the porting cost is. Our deal tracker shows the money moving in two directions this month, toward the incumbent's roadmap and toward everyone trying to route around it. The second group only wins if the software gap closes, so treat every automated-compiler round as a small, cheap option on that outcome.