Proven Silicon
Edem Kumodzi 🇬🇭🇹🇬@edemkumodzi
−5.6%
since Sep 30
The idea
I believe correctness becomes the bottleneck as AI writes more code and hyperscalers design their own chips, so the verification layer sits alongside the custom-silicon boom as a toll booth on the whole thing. The basket spans proof engines, physical verification tools, custom-chip designers, the hyperscalers building silicon and proof research in house, and the foundry and IP backbone underneath it all.
What it owns
- These companies sell the formal verification tools every chip design must pass before manufacturing, so they get paid on every new AI chip regardless of whose chip wins. · Cadence Design SystemsIt is the other half of the design-software duopoly, and every custom accelerator from the hyperscalers passes through its verification engines., SynopsysIt sells the formal proof tools chip designers run before committing hundreds of millions to a tape-out, making it a toll booth on the bet that correctness becomes mandatory for custom AI chips., Keysight TechnologiesIt builds the test and measurement layer that checks designs actually behave as specified, the practical end of the correctness story.22%
- The companies the hyperscalers hire to design and connect their in-house AI chips, the direct beneficiaries of the custom-chip wave. · BroadcomIt designs and delivers Google's TPUs, the biggest custom AI chip program running, making it the most direct way to own the custom-silicon expansion., Marvell TechnologyIt is Amazon's partner on Trainium and other custom chips, the second big hyperscaler design win in the basket., Astera LabsIt sells the connectivity chips that let custom accelerators talk to memory and each other inside AI clusters, growing with every new custom chip deployment., Credo TechnologyIt makes the high-speed cabling and links that wire AI clusters together, which custom chips from hyperscalers need just as much as external alternatives.21%
- Google, Amazon, Microsoft and Meta are designing their own AI chips and also run the deepest formal-methods research in software. · AlphabetIt owns the TPU, the most mature custom AI chip anywhere, and runs serious formal methods research on its own software., AmazonIt builds Trainium and runs one of the deepest automated reasoning groups in industry, using formal proof to verify its own systems., MicrosoftIt built Z3, the proof engine much of the formal verification world runs on, and is scaling its own Maia AI chip with deep internal correctness research., Meta PlatformsIt is scaling MTIA chips as its bet on running AI on its own silicon instead of paying for external processors forever.20%
- Once a chip is made it has to be inspected and measured to prove it matches the design, and these firms make the machines that do it. · KLA CorporationIt dominates the market for machines that inspect chips physically once they are made, confirming that a proven design was actually built correctly., Onto InnovationIt measures and inspects chips through advanced packaging, the step where AI accelerators get stitched together and where a hidden flaw ruins an expensive part., Nova LtdIt sells the measurement tools fabs use to confirm each layer of a chip came out as designed, providing physical proof that matches the correctness requirement., CamtekIt inspects AI chips and their packaging for defects, so its orders track directly with how many custom accelerators get built and deployed.14%
- Nearly every custom AI chip is built on TSMC and licenses Arm's designs, so they collect on the whole wave. · Taiwan SemiconductorIt manufactures every custom accelerator in the take, from TPUs to Trainium to Maia to MTIA, so it collects from all of them without picking a winner., Arm HoldingsIt licenses the processor designs that sit inside most hyperscaler custom chips, earning a royalty on the same boom the rest of the basket rides.12%
- Cash11%