Nobody is going to win the AI chip race by making a faster chip. That sounds absurd given the last twelve months, but stay with me. Cerebras’ CS-4 claims 30x faster inference than GPUs. Google’s Ironwood pushed past Nvidia’s Blackwell on performance. OLIX pulled in $312 million in Series B funding. Every one of those headlines is about compute. And every one of those products depends on a layer of resin-coated laminate that almost nobody in this industry can name.
I review tools for a living. The pattern I keep running into is that the flashy spec on the box is rarely the thing that determines whether you can actually get the tool. The AI accelerator market has the same problem, and it lives in the substrate.
The part of the chip nobody puts on a slide
Underneath the silicon that does the math is a package substrate. In data center accelerators, that substrate is typically ABF — a build-up film laminate that carries thousands of connections between the die and the board. It is unglamorous. It is also, in 2026, the reason a chip either ships or sits in a slide deck.
Tom’s Hardware ran a piece this year on the state of ABF substrates in data center silicon, which is roughly the least viral topic imaginable next to a 30x inference claim. But that is exactly why it matters. Nobody writes trend pieces about the bottleneck until the bottleneck bites.
Here is the structural reason it deserves attention. Compute performance is something a company can improve with better architecture, better software, better memory hierarchy. It scales with engineering talent. Substrate supply scales with factories, and factories scale with years. Those two curves do not move at the same speed, and when they diverge, the slower one sets the pace.
What this means for the funding headlines
OLIX raising $312 million is a real signal that money is still flowing into chip startups. Good. But when I read a Series B number in this category, my first question is not “how fast is the chip.” It is “who is packaging it, and did they get in line early.”
A startup with a brilliant design and no substrate allocation is a company selling futures. A slower design with locked-in supply ships product. In toolkit terms, this is the difference between a library with a beautiful API and one you can actually install.
Some questions worth asking when a new accelerator gets announced:
- What is the actual availability timeline, not the announcement date?
- Is the performance claim measured on shipping hardware or a reference design?
- Who else is competing for the same packaging capacity?
- Does the vendor have a second source for anything in the package?
None of those show up in a benchmark chart. All of them determine whether you can buy the thing.
Why the big players look different through this lens
Ironwood beating Blackwell is a genuinely notable result. Google building silicon that outperforms Nvidia’s flagship changes the competitive math. But look at who is positioned to convert that into volume. Google has scale, capital, and a long history of manufacturing relationships. Cerebras took a different route entirely — its whole approach is built around unusually large silicon, which is its own packaging story.
That’s the pattern. The companies making noise at the top of this market are the ones that solved the boring supply problem, or built around it. The chip design is the visible half.
How I’d read the next six months
Expect more performance claims. The 30x number from Cerebras and the Ironwood result will pull competitors into a specs war, because specs are cheap to announce. What I’ll be watching instead is delivery. Which accelerators actually reach customers in quantity, and which ones quietly slip a quarter, then two.
If you’re evaluating AI infrastructure right now, my honest advice is to weight availability over peak throughput. A chip you can deploy at 60% of a competitor’s benchmark is worth more than one you’re on a waitlist for. That is not a thrilling recommendation. It is the one that keeps projects on schedule.
The AI chip story in 2026 gets told as a contest between architectures. It is closer to a contest over who secured capacity in a supply chain most people never think about. The substrate is not the interesting part of the chip. It is the part that decides whether the interesting part matters.
🕒 Published: