Tenstorrent vs SK Hynix

Compute vs. Memory: Different Layers of the Same Stack

The most important thing to understand about Tenstorrent and SK Hynix is that they are not direct competitors — they occupy different layers of the AI hardware stack. Tenstorrent designs the processors that execute AI workloads, while SK Hynix manufactures the memory that feeds those processors with data. In practice, an AI server could contain both Tenstorrent accelerators and SK Hynix memory chips.

That said, their strategic trajectories are converging. SK Hynix is pushing compute closer to memory with technologies like Processing-In-Memory (PIM) and its AiMX accelerator card, which embeds computation directly into GDDR6 memory. Tenstorrent, for its part, is deeply focused on memory bandwidth efficiency in its Tensix architecture. As the lines between compute and memory blur in the age of AI, these companies increasingly influence each other's addressable markets.

Open vs. Proprietary: Divergent Business Models

Tenstorrent has made openness its defining strategy. The company's RISC-V-based CPUs use a royalty-free instruction set architecture, its software stack is fully open-source, and its new Open Chiplet Atlas (OCA) ecosystem promotes vendor-neutral modularity. Jim Keller has stated his belief that RISC-V will dominate data centers within 5-10 years, and Tenstorrent is positioning itself to be the Arm Holdings of the RISC-V era — licensing high-performance IP rather than only selling finished chips.

SK Hynix operates the opposite model. Its competitive advantage comes from proprietary manufacturing processes, advanced 3D packaging technology for stacking HBM dies, and massive capital expenditure that creates barriers to entry. The company's 321-layer NAND and 16-layer HBM4 products represent manufacturing feats that only a handful of companies worldwide can replicate. This proprietary edge has allowed SK Hynix to command premium pricing and secure long-term supply agreements with NVIDIA through 2026.

AI Data Center Footprint

In the AI data center market, SK Hynix is already a dominant incumbent. Its HBM chips are inside virtually every NVIDIA H100 and H200 GPU deployed in hyperscale data centers. With NVIDIA reportedly driving 27% of SK Hynix's revenue in the first half of 2025, and the company sold out on all DRAM, NAND, and HBM supply through 2026, its data center position is firmly established.

Tenstorrent is the challenger. Its Galaxy Wormhole server, developed in partnership with Moreh, offers what the company calls a cost-efficient alternative to GPU-based systems for both inference and training. Unlike many AI accelerator startups that focus solely on inference, Tenstorrent supports both workloads — a critical differentiator. However, the company is still in the early stages of commercial deployment and faces the enormous challenge of building a software ecosystem to rival NVIDIA's CUDA.

Edge and On-Device AI

Tenstorrent has a clear edge in edge AI strategy. Its partnership with Razer produced a compact Thunderbolt 5 AI accelerator shown at CES 2026, and the TT-QuietBox 2 workstation (starting at $9,999) is the first RISC-V AI workstation capable of teraflop-class inference. These products target developers who want to run LLMs, image generation, and other AI workloads locally without cloud dependency.

SK Hynix's edge play is less direct but still significant. Its LPDDR6 memory, optimized for on-device AI, offers improved data processing speed and power efficiency for smartphones, laptops, and automotive applications. The company's SOCAMM2 module format targets AI-capable edge servers. SK Hynix enables edge AI by supplying the memory that on-device processors need, while Tenstorrent provides the processors themselves.

Financial Scale and Risk Profile

The financial gulf between these companies is vast. SK Hynix posted record annual profits in 2025, surpassing Samsung for the first time, with an HBM revenue run-rate approaching $8 billion annually. It is a publicly traded company with a market cap exceeding $460 billion as of March 2026. Tenstorrent, at a $3.2 billion private valuation with $1 billion raised, is roughly 1/140th the size.

This disparity translates directly to risk. SK Hynix's primary risk is cyclical: memory markets are notoriously boom-and-bust, and its heavy reliance on NVIDIA (and on HBM specifically) means a slowdown in AI infrastructure spending could hit hard. Tenstorrent's risk is existential: it must prove that its open-source approach can gain meaningful market share against NVIDIA's entrenched CUDA ecosystem and Arm's established data center presence, all while managing the capital-intensive process of bringing new chip architectures to market.

The NVIDIA Factor

Both companies' futures are deeply intertwined with NVIDIA, but in opposite ways. SK Hynix is NVIDIA's most critical supplier — it has secured an estimated 70% of NVIDIA's HBM4 orders for the Rubin platform in 2026. This relationship is enormously profitable but creates dependence. Tenstorrent, conversely, is positioning itself as an NVIDIA alternative for customers who want lower cost, more openness, or freedom from vendor lock-in. Jim Keller has explicitly targeted markets "not well served by NVIDIA," including cost-sensitive inference and edge deployment.

For enterprises evaluating their AI hardware strategy, these opposing relationships with NVIDIA are instructive: SK Hynix benefits from NVIDIA's dominance, while Tenstorrent benefits from dissatisfaction with it.