Tenstorrent and the RISC-V Open Hardware Revolution

1. Financial Overview and Market Momentum

Tenstorrent has solidified its position as one of the most disruptive companies in the global semiconductor industry by closing its Series E funding round, raising $1.5 billion. The investment pushed the company's market valuation to $6.2 billion, a significant jump from the $2.6 billion valuation recorded in December 2024.

This acceleration in valuation reflects the rapid conversion of its technology pipeline into commercial traction, driven by global demand for Sovereign AI initiatives and the market's urgency to find viable alternatives to the proprietary GPU ecosystem, bypassing supply bottlenecks and prohibitive infrastructure costs.

The composition of the funding round demonstrates diverse backing spanning venture capital, institutional investors, and industrial ecosystem partners:

The convergent presence of automotive OEMs (Hyundai), consumer electronics giants (Samsung, LG), financial and crypto infrastructure players (Tether, Morgan Stanley, ARK), and long-term institutional funds in the same cap table validates the thesis that demand for custom silicon free from vendor lock-in is cross-sectoral, extending from data centers to the edge.

2. Jim Keller's Vision and RISC-V Philosophy vs. Traditional Models

Tenstorrent's strategic direction is under the leadership of its CEO, Jim Keller, an influential microprocessor architect with stints on projects that redefined the industry at Apple, AMD (where he led the Zen architecture), Tesla (FSD system), and Intel. Keller argues that artificial intelligence infrastructure requires a transition toward open, transparent, and highly efficient architectures in terms of cost and power consumption.

The foundation of this strategy is the adoption of the RISC-V ISA (Instruction Set Architecture), which establishes a direct contrast to dominant proprietary models:

• RISC-V: An open ISA completely free of royalties. It enables instruction set customization, reduces decoder complexity in the CPU, and eliminates contractual restrictions or vendor lock-in risks directly at the silicon level.

• x86 (Intel / AMD): Proprietary and closed architecture, characterized by high legacy complexity in the instruction pipeline and centralized control by a few corporations, preventing free modification of silicon by third parties.

• ARM: Model based on restrictive commercial licensing, subject to ongoing royalty payments and contractual limitations that prevent deep microarchitecture modifications by licensees.

3. Blackhole Processor Architecture (6nm)

The Blackhole is Tenstorrent's flagship processor, designed specifically to process intensive AI model training and inference workloads. Manufactured on a 6nm process, the component features an architecture optimized for high throughput and local control:

• AI Processing: 140 Tensix++ cores dedicated to vector and matrix processing, operating in 32x32 tile arrays.

• Integrated Control Capability: 16 high-performance RISC-V CPU cores (SiFive architecture) integrated on the die itself for orchestration, operating system management, and local flow control.

• Computational Performance: Delivers up to 745 TFLOPS of computing capacity in FP8 precision.

• Memory Subsystem: Bandwidth of 512 GB/s through GDDR6 memory, offering superior thermal and financial balance compared to HBM memory.

Asynchronous Data Orchestration and Local SRAM Memory

Each of the 140 Tensix++ cores integrates 5 RISC-V microprocessors (called "baby RISC-Vs") into its structure. The function of this structure is to perform asynchronous data orchestration, decoupling data movement in silicon from computational execution of matrix units. This arrangement prevents pipeline stalls during matrix multiplications. The microprocessors manage data transfer directly to the 1.5 MB local SRAM memory per core, ensuring very low latency access and maximizing energy efficiency by minimizing recurring accesses to external DRAM memory.

4. Connectivity and Integrated Networks (NoC)

The Blackhole processor has an interconnect architecture that integrates 10 native 400 Gbps Ethernet ports directly on the die through a Network-on-Chip (NoC).

This approach brings a relevant architectural advantage by eliminating dependence on external PCIe switches and proprietary network adapters. By embedding Ethernet connectivity natively in silicon, Tenstorrent bypasses proprietary scale-out interconnects like Nvidia NVLink or InfiniBand. Thus, standardized Ethernet switches with commoditized market prices can be used to connect and scale silicon clusters directly, reducing capital costs at large scale.

5. Infrastructure Scale: Galaxy Servers

To meet large-scale processing demand, Tenstorrent developed Galaxy, a rack-level solution designed for high computational density:

• Composition: Aggregates 32 Blackhole processors natively integrated through the chip's connectivity network.

• Processing Power: Delivers approximately 23 Petaflops of performance in FP8.

• Thermal Flexibility: Offers support for liquid or air cooling configurations, enabling deployment in both legacy datacenters and modern high-density facilities.

• Open Stack for Sovereign AI: Optimized for running and training open language models (such as Llama 3 and DeepSeek) and video generators. The pairing of open-weights models with an open ISA (RISC-V) and open-source software creates a fully sovereign technology ecosystem unencumbered by commercial locks, with infrastructure costs lower than traditional GPU-based platforms.

6. Overcoming the "CUDA Moat" Through Open Software

Nvidia's proprietary ecosystem, anchored in the CUDA platform, represents one of the main barriers to entry in the AI hardware market. To overcome this limitation, Tenstorrent developed a 100% open-source software stack, designed to offer transparency and direct control to developers, divided into two complementary layers:

TT-Metalium

Aimed at low-level programming and optimization. TT-Metalium grants developers visibility and control over data flow in the Network-on-Chip (NoC), enabling direct management of buffers in local SRAM memory and writing custom kernels to extract maximum performance from silicon without the intermediation of closed libraries.

TT-Buda

Acts as the platform's high-level compiler, bridging AI models and hardware. It connects directly and transparently to major market frameworks such as PyTorch and ONNX, translating computational graphs into hardware execution without proprietary "black boxes."

7. Hybrid Business Model and Market Expansion

Tenstorrent's commercial strategy combines direct product sales with technology licensing, operating on two complementary fronts:

• Hardware Product Sales: Commercialization of accelerators and complete systems, spanning PCIe cards (Wormhole and Blackhole lines), development workstations (TT-QuietBox), and high-density racks (Galaxy).

• Intellectual Property (IP) Licensing: Licensing of RISC-V CPU core IP blocks (TT-Ascalon architecture) and AI cores (Tensix). This front allows partners like Samsung and LG to integrate this IP into their own chip designs for applications in appliances, industrial automation, and mobility.

IP licensing complements capital-intensive hardware sales, creating recurring revenue with high gross margins and accelerating the diffusion of Tensix/RISC-V architecture across multiple industrial sectors.

8. Global Adoption and Sovereign AI Initiatives

Acceptance of Tenstorrent's architecture is reflected in global critical infrastructure projects:

• Sovereign AI Infrastructure in Japan: Japan's national ai& initiative selected Tenstorrent technology for building its computational center, planning deployment of more than 120 Galaxy systems.

• Adoption in Corporate Data Centers: Global infrastructure and financial services companies such as Virtu Financial and Equinix have begun integrating Tenstorrent systems into their data centers to optimize AI operational costs.

9. Analysis Conclusion

Tenstorrent establishes an alternative model for the evolution of artificial intelligence infrastructure. By integrating the open RISC-V ISA, native Ethernet connectivity directly in silicon (NoC), and a fully open-source software stack, the company builds a technically scalable and financially efficient architecture. Supported by a hybrid business model of hardware and IP licensing, Tenstorrent positions itself as a disruptive force against proprietary GPU solutions in the global semiconductor market.