Cerebras vs Tenstorrent

Architectural Divergence: Monolithic vs. Modular

Cerebras and Tenstorrent represent opposite ends of the semiconductor design spectrum. Cerebras' wafer-scale approach is maximalist — the WSE-3 uses an entire 300mm silicon wafer as a single processor, packing 4 trillion transistors and 44 GB of on-chip SRAM into a chip the size of a dinner plate. This eliminates the inter-chip communication overhead that plagues distributed GPU clusters, enabling a single Cerebras CS-3 system to replace hundreds of GPUs for certain workloads.

Tenstorrent takes the opposite approach: modularity and openness. Built on the RISC-V open instruction set, Tenstorrent's Tensix cores use a mesh-based architecture that scales horizontally. The company's Open Chiplet Atlas (OCA) ecosystem, announced in late 2025, aims to create an industry-wide open standard for chiplet interconnects — an ISA-neutral and IP-neutral framework that lets chip designers mix and match components without vendor lock-in. Where Cerebras bets on the power of one enormous chip, Tenstorrent bets on the flexibility of many composable ones.

Inference Performance and the Token Economy

In the emerging AI inference economy, speed and cost-per-token are the metrics that matter. Cerebras has demonstrated extraordinary inference throughput: over 2,100 output tokens per second on benchmarks, with models like OpenAI's gpt-oss-120B reaching 3,000 tokens per second and Qwen3 235B running more than 10x faster than leading GPU clouds. The March 2026 partnership with AWS introduces inference disaggregation, where AWS Trainium handles the prefill stage and Cerebras WSE handles decode — a novel hybrid architecture that could reshape cloud inference economics.

Tenstorrent's inference story is different. Rather than chasing raw token throughput, Tenstorrent's Tensix architecture uses conditional execution — the ability to skip unnecessary computation during inference. This is particularly powerful for sparse models and workloads with variable-length sequences, where a significant fraction of compute can be avoided entirely. Combined with Moreh's MoAI framework for data center deployment, Tenstorrent targets cost-per-inference rather than peak speed, making it attractive for enterprises optimizing total cost of ownership.

Ecosystem and Partnership Strategy

Cerebras has secured partnerships that validate its technology at the highest tier. The multi-year deal with OpenAI — worth over $10 billion to supply 750 megawatts of compute capacity through 2028 — is arguably the most significant non-NVIDIA AI hardware deal ever signed. Combined with AWS Marketplace availability and adoption by national labs and pharmaceutical companies, Cerebras has built an ecosystem centered on hyperscale cloud and enterprise AI.

Tenstorrent's ecosystem strategy is broader and more grassroots. The Razer partnership for a Thunderbolt 5 compact AI accelerator, unveiled at CES 2026, brings AI development hardware to individual developers. The Open Chiplet Atlas initiative invites the entire semiconductor industry to participate in an open standard. Partnerships with Samsung, LG Electronics, and AutoCore position Tenstorrent across consumer electronics, automotive, and enterprise markets. Tenstorrent's DevCloud also provides remote access to its hardware for developers — a lower barrier to entry than Cerebras' enterprise-focused sales model.

Open Standards vs. Proprietary Performance

The philosophical divide between these companies mirrors a recurring tension in technology: open ecosystems versus proprietary optimization. Tenstorrent's RISC-V foundation means its chip designs are built on an open instruction set architecture, free from the licensing fees and restrictions of ARM or x86. Jim Keller has explicitly positioned Tenstorrent as the "open hardware revolution," with the Ascalon RISC-V CPU core designed to compete with ARM in data center workloads — a play that extends well beyond AI accelerators.

Cerebras' approach is inherently proprietary — no one else makes wafer-scale chips, and the entire software stack is tightly integrated with the hardware. This vertical integration enables extraordinary performance but creates dependency. For organizations that prioritize sovereignty and flexibility over raw speed, Tenstorrent's open architecture may be the more strategic long-term investment, particularly as RISC-V adoption accelerates across the industry.

Market Positioning and Financial Trajectory

The valuation gap tells a story about market perception. Cerebras' $23 billion valuation and imminent IPO reflect investor confidence in its hyperscale inference play — particularly after the OpenAI and AWS partnerships. The company has positioned itself as the premier alternative to NVIDIA for organizations that need the absolute fastest inference at scale.

Tenstorrent's $3.2 billion valuation reflects a broader but earlier-stage market opportunity. The company's edge AI strategy, automotive partnerships, and open chiplet ecosystem represent a more diversified bet across multiple segments of the semiconductor market. However, the 7.5% layoff reported in early 2026 suggests Tenstorrent is sharpening its focus — potentially prioritizing the developer and edge markets where its compact accelerator can gain traction quickly over data center workloads where Cerebras and NVIDIA dominate.