AI Compute Infrastructure

3/748 → 8/649; +0.83 percentage points of formation share.

The eight companies here split cleanly along the compute stack, and that split is the real story: some, like SF Tensor and Cumulus Labs, are building the software layer that lets researchers and engineering teams skip infrastructure chores entirely, while others, like Piris Labs and Sygaldry Technologies, are attacking the physical bottleneck underneath, whether that's photonic networking to stop idle chips or quantum-accelerated servers to speed training. Wafer and Luminal occupy the middle, treating GPU kernel optimization and compiler-level speedups as the lever, with Luminal claiming a 10x speed multiplier and Piris Labs citing up to 80% chip idle time as the problem it exists to solve. That idle-compute statistic is telling: even as the formation grows at 3.07x, a huge share of existing hardware sits unused, meaning the bottleneck has shifted from acquiring GPUs to orchestrating and networking the ones already deployed. This formation is growing almost in lockstep with the broader "Cloud Computing" category (3.10x) but far outpacing generic "Infrastructure" (1.48x), suggesting the demand is specifically for AI-tuned compute plumbing rather than infrastructure in general. Partcl and Aemon point to a further layer of specialization, applying GPU-accelerated physics models to chip design itself and using autonomous agents to iterate on engineering problems, which signals the category is moving from renting raw compute toward automating the labor of making compute efficient. None of the adjacent Hugging Face or arXiv clusters are genuinely about this infrastructure layer, they're closer to model and agent research, which reinforces that this formation is a distinct, hardware-adjacent lane rather than a software trend riding the same LLM wave. The timing tracks with the post-2023 GPU scarcity triggered by the ChatGPT-driven compute rush, which made every inefficiency in training and inference expensive enough to justify entirely new companies built just to close the gap.

No comparable outside activity is connected yet.

These share a word with “AI Compute Infrastructure” but are tracked as separate categories. The real overlap below is the actual share of AI Compute Infrastructure’s companies also found in that category — not a guess.