Groq

Groq is a venture-backed semiconductor firm, not a fund; it has no investment committee. Jonathan Ross is the founder and CEO with a technical background as the original architect of Google's TPU. Product, chip architecture, and cloud engineering decisions run through his office and Chief Architect Dennis Abts. Operational scale is managed by COO Stuart Pann, who was recruited from Intel's supply-chain leadership.

No. Groq is a fabless chip designer that contracts fabrication to GlobalFoundries, not TSMC — a meaningful differentiator in the current geopolitically constrained wafer environment. The firm writes the LPU architecture, compiler, and software stack internally and deploys chips in its own managed cloud. It sells neither raw silicon nor chip-level IP to third parties.

Groq's Language Processing Unit is a deterministic, single-core architecture where the compiler decides scheduling ahead of time — no cache misses, no speculative execution, no variable latency. This is structurally different from Nvidia's CUDA GPU model, which uses massive parallelism and runtime scheduling. The result: higher and more predictable throughput on transformer inference, particularly on token-generation latency per user.

No. Groq does not build, train, or own foundation models. It provides the cloud infrastructure that runs open-source models like Llama and Mixtral. Its customers include the model builders themselves, enterprises deploying open-weight models, and developer platforms that need fast, cost-efficient inference. This places it in the picks-and-shovels layer, not the model layer.

All three attack GPU dominance with large-wafer or deterministic architectures, but their tradeoffs differ. Groq's LPU uses on-chip SRAM and a deterministic compiler optimized for transformer inference throughput, delivering best-in-class tokens-per-second on its publicly available GroqCloud. Cerebras makes the largest wafers on earth for training and inference. Sambanova's multi-core dataflow architecture targets high-batch throughput. Groq's developer-facing API makes it the most cloud-native of the three.

Meta's FAIR group, led by Yann LeCun, uses Groq's LPU cloud to serve Llama models publicly. LeCun has posted results showcasing 800+ tokens per second on smaller Llama variants. This is not a paid endorsement but a functional deployment by the world's largest open-source model team, giving Groq a public proof point that no media lab or demo-only benchmark can replicate.

Groq contracts with GlobalFoundries, the US-based semiconductor fabricator, rather than the overwhelmingly dominant TSMC. This choices hedges the Taiwan-strait concentration risk but constrains node size — GlobalFoundries does not compete at TSMC's 3nm and below. Hiring former Intel supply-chain COO Stuart Pann signals the company is treating wafer procurement as a C-level operational risk.