The bet: openness as a competitive advantage
Nearly every accelerator covered in this cohort — Nvidia’s CUDA-tied architecture, AMD’s, Broadcom’s and Marvell’s custom designs — runs on proprietary, closed instruction-set architectures that lock a customer’s software investment to one vendor’s silicon roadmap. Tenstorrent is building its AI accelerators around RISC-V, an open, freely licensable instruction-set standard, betting that avoiding vendor lock-in is itself a meaningful competitive advantage for customers wary of over-depending on any single proprietary ecosystem [4].
The funding and backers
Tenstorrent has raised $800 million, with Fidelity and Bezos Expeditions among its backers [1]. That is a smaller raise than Cerebras’s pre-IPO valuation or Groq’s reported capital, but a serious commitment from institutional and high-profile investors nonetheless, reflecting real confidence in the open-architecture thesis even at an earlier stage of commercial maturity than some of its inference-focused peers.
The Rapidus connection
Tenstorrent is named among the lead early customers for Rapidus, Japan’s state-backed 2nm foundry effort covered elsewhere in this cohort, alongside IBM and Fujitsu [3]. That relationship is worth noting for what it says about both companies: Tenstorrent, still establishing itself as a chip designer, and Rapidus, still proving it can manufacture at the leading edge at all, are effectively betting on each other’s success simultaneously — an unusually mutual, symmetric risk relationship compared to the more established, one-directional customer relationships covered throughout the rest of this cohort.
Why RISC-V specifically, and why now
RISC-V has spent years as primarily an academic and embedded-systems architecture before becoming a serious contender in AI accelerator design. Its appeal for an AI-chip startup specifically is that it removes the need to either license a proprietary architecture from an incumbent or build an entirely bespoke instruction set from scratch — a startup can build on a shared, freely available foundation and differentiate at the level of the actual AI-specific compute engines layered on top of it, rather than reinventing basic instruction-set design work every other RISC-V-based company has already solved collectively [2].
The risk this strategy accepts
An open architecture is a double-edged strategic choice. It lowers the barrier for a customer to adopt Tenstorrent’s silicon without fear of long-term lock-in, which is a genuine sales advantage. It also lowers the barrier for a future competitor to build a similar open-architecture AI chip using the same RISC-V foundation Tenstorrent itself relies on — meaning Tenstorrent’s durable competitive advantage has to come from its specific AI-compute engine design and manufacturing execution, not from any exclusivity over the underlying instruction set itself, a different kind of competitive bet than the proprietary-architecture moats most of this cohort’s other accelerator companies are trying to build.
Why the open-source software ecosystem matters as much as the hardware
Tenstorrent’s bet is not purely about the instruction set architecture itself — it extends into the company’s approach to its software stack, which is likewise built with a stronger commitment to openness than most of the proprietary toolchains covered elsewhere in this cohort, Nvidia’s CUDA ecosystem chief among them. A closed software stack tightly coupled to specific hardware is one of the strongest lock-in mechanisms in the entire accelerator industry, arguably stronger than the hardware architecture itself in determining switching costs. Tenstorrent’s willingness to build more of its stack in the open is a bet that developers and enterprises increasingly value portability and auditability over the short-term convenience of an established, if closed, toolchain — a bet that will take years of adoption data to properly evaluate.
A useful comparison to the rest of the industry’s incumbents
Every major incumbent covered in this cohort’s earlier tracks — Nvidia, AMD, the Broadcom- and Marvell-designed hyperscaler chips — ultimately depends on some degree of proprietary lock-in to protect its market position, whether through a closed software ecosystem, an exclusive customer relationship, or accumulated non-public process expertise. Tenstorrent is one of the few companies in this entire cohort explicitly betting against that pattern rather than trying to replicate it at a smaller scale, which makes it a useful test case for whether an open strategy can compete with entrenched proprietary moats in a market this capital-intensive and this dominated by a small number of very large, well-resourced incumbents.
