Tenstorrent
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v3 ยท 3,709 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v3 ยท 3,709 words
Add missing citations, update stale details, or suggest a clearer explanation.
Tenstorrent is a North American artificial intelligence hardware and intellectual property company that designs processors for AI training and inference on the basis of the open-standard RISC-V instruction set architecture. Led since January 2023 by veteran chip designer Jim Keller, the company builds AI accelerators branded Grayskull, Wormhole, and Blackhole, plus high-performance RISC-V CPUs marketed as Ascalon, all paired with a fully open-source software stack. In December 2024 Tenstorrent closed a Series D round of more than $693 million at a pre-money valuation of $2 billion (roughly $2.6 billion post-money), led by Samsung Securities and AFW Partners with participation from Jeff Bezos, positioning it as a prominent challenger to NVIDIA's dominance in the AI accelerator market [1][2]. In June 2026 multiple outlets reported that Qualcomm was in talks to acquire the company for up to $10 billion [3].
Founded in 2016 in Toronto by former AMD and NVIDIA engineers, Tenstorrent's central thesis is that open standards beat proprietary lock-in: standard RISC-V cores, an open chiplet specification, and an open-source compiler and runtime, deliberately contrasted with NVIDIA's closed CUDA ecosystem. Keller has framed the bet bluntly, telling EE Times that the strategy is "whatever Nvidia does, we'll do the opposite" [4].
Tenstorrent was founded in 2016 in Toronto, Ontario by three engineers with backgrounds in graphics and CPU design: Ljubisa Bajic, Milos Trajkovic, and Ivan Hamer. The trio had previously worked at AMD and NVIDIA and shared a thesis that conventional GPUs were a poor architectural fit for the dynamically sparse, conditional computation patterns of modern neural networks. Their stated goal was to attack what they termed the memory wall and to build processors that could exploit fine-grained sparsity and conditional execution rather than treating every neural-network layer as a dense matrix multiplication.
The company raised approximately $5.9 million in seed funding led by Real Ventures and Eclipse Ventures and operated relatively quietly during its first several years, accumulating IP licensees and developing its first prototype chips. Tenstorrent later established its principal U.S. office in Santa Clara, California, while maintaining significant engineering operations in Toronto, Austin, Belgrade, Tokyo, Bangalore, Singapore, and Seoul [1].
The company's early business model centered on selling AI accelerator IP and discrete hardware to system builders, with most of the company's bookings in its first several years coming from IP licensing deals rather than chip sales. That dual model of selling both finished accelerators and licensable IP blocks remains a defining feature of the company today.
A pivotal moment for Tenstorrent came in December 2020, when Jim Keller joined the company as Chief Technology Officer, President, and board member. Keller's reputation in the semiconductor industry was already extraordinary. He had been a lead architect on the AMD K7 and K8 (Athlon and Opteron) CPUs, the original Apple A4 and A5 SoCs, the AMD Zen architecture that revived the company in the late 2010s, and Tesla's first generation of full self-driving silicon. He had also served as a senior vice president at Intel from 2018 to 2020. Upon joining, Keller publicly described Tenstorrent's Tensix architecture as "the most promising architecture out there" [5].
Founder Ljubisa Bajic stepped down as CEO and assumed the role of Chief Architect, and in January 2023 Keller was promoted from CTO to Chief Executive Officer [3]. Bajic departed Tenstorrent in early 2023 to launch a new AI hardware company called Taalas [23]. Under Keller's leadership, the company sharpened its positioning around open hardware, open software, and an explicit contrast with NVIDIA's proprietary CUDA ecosystem. In a widely quoted interview, Keller summarized the strategy as "whatever Nvidia does, we'll do the opposite" [4]. He has been equally direct about the long arc of the instruction set itself, telling EE Times in 2023, "My belief is in the next 5 to 10 years, RISC-V will take over all the data centers" [4].
Keller has also used his profile to recruit veteran silicon engineers from Apple, Tesla, Arm, AMD, and Intel into Tenstorrent's CPU and AI cores teams, in particular for the in-house Ascalon RISC-V CPU project. The company's executive team and board have since expanded to include former executives from Samsung, Hyundai, LG, and a number of Wall Street and Silicon Valley investors.
Tenstorrent's processors are built around a tile of small, programmable compute units called Tensix cores. The Tensix is conceptually closer to a parallel many-core processor than to a traditional GPU streaming multiprocessor. Each Tensix tile contains:
The Tensix natively operates on 32x32 tiles of data, a design choice that aligns naturally with the matrix multiplications and convolutions that dominate deep learning. Tiles are routed between Tensix cores over the on-chip NoC mesh, which in turn extends seamlessly over Ethernet links between chips, allowing a multi-chip system to be programmed as if it were a single large mesh [8]. Tenstorrent describes this property as the basis for its "scale-out" architectural pitch: rather than building one giant die and then bolting on external interconnect, every chip is a tile in a larger fabric from day one.
The vector and matrix units have evolved across generations. Grayskull processed 64-element vectors at 19-bit floating point precision, while Wormhole moved to 32-element vectors with full FP32 support, and Blackhole introduced a Tensix++ variant with broader data type support including FP8 and block floating point formats [7]. All three generations share the same programming model exposed through the company's open-source low-level SDK, TT-Metalium.
Grayskull was Tenstorrent's first commercial accelerator and the first member of its He-Man-inspired naming series. It introduced the Tensix core to the market and served as the workhorse for early developer cards.
Announced in 2021 and shipping in volume from 2024, Wormhole increased the Tensix count, added GDDR6 memory, and introduced 16 lanes of 100 Gbps Ethernet on the die itself for chip-to-chip scaling. Tenstorrent's RISC-V-based Wormhole accelerators were made available for pre-order in 2024, with pre-built developer workstations starting at $12,000 [9].
Blackhole was unveiled at Hot Chips 2024 and officially launched at Tenstorrent Dev Day on April 3, 2025 [6]. It is the first Tenstorrent accelerator designed to function as a standalone AI computer rather than purely as a PCIe co-processor: each Blackhole die includes a cluster of 16 SiFive X280 RISC-V CPU cores that can boot a Linux kernel and orchestrate workloads without a host CPU [7].
| Specification | Grayskull | Wormhole | Blackhole |
|---|---|---|---|
| Year introduced | 2020 (sampling) | 2021 announce, 2024 GA | 2024 announce, 2025 GA |
| Tensix cores | 128 | 80 (Tensix+) | 140 (Tensix++) |
| Process node | GlobalFoundries 12 nm | GlobalFoundries 12 nm | TSMC 6 nm |
| Die size | 620 mm2 | 670 mm2 | not disclosed |
| Peak performance | ~315 INT8 TOPS | 328 FP8 TOPS | 745 FP8 TFLOPS, 332 BF16 TFLOPS, 664 BLOCKFP8 TFLOPS |
| Memory | LPDDR4 (developer cards) | 12 GB GDDR6, 336 GB/s | 32 GB GDDR6, 512 GB/s |
| Host interface | PCIe Gen4 x16 | PCIe Gen4 x16 | PCIe Gen5 x16 |
| Ethernet | none | 16x100 Gbps | 10x400 Gbps (4x QSFP-DD 800G) |
| Onboard CPUs | Baby RISC-V only | Baby RISC-V only | 16 SiFive X280 RISC-V + Baby RISC-V |
| Card TDP | 65 W (chip), 75 W (PCIe card) | not disclosed | not disclosed |
At Dev Day in April 2025, Tenstorrent published a developer-oriented Blackhole product family with publicly listed prices, which is unusual for the AI accelerator market [6]:
| Product | Configuration | Cooling | List price (USD) |
|---|---|---|---|
| Blackhole p100 | 1x Blackhole, no Ethernet | Active | $999 |
| Blackhole p150 | 1x Blackhole with Ethernet | Passive / active / liquid | $1,399 |
| TT-QuietBox | Liquid-cooled workstation with 4 Blackhole | Liquid | $11,999 |
| Galaxy Blackhole | 6U server with 32 Blackhole accelerators | Liquid | $110,000 |
The Galaxy Blackhole system became generally available in April 2026 and is positioned by Tenstorrent as a direct alternative to NVIDIA's rack-scale Blackwell systems [10]. The company has claimed inference throughput of around 350 tokens per second on the DeepSeek R1 reasoning model in a Galaxy Blackhole configuration, with a total cost of ownership well below comparable NVIDIA GB300 deployments [21].
In parallel with its AI accelerator program, Tenstorrent has invested heavily in becoming a high-performance RISC-V CPU vendor. The flagship of this effort is Ascalon, which the company describes as the highest-performance RISC-V core platform in the industry [13].
The top-of-line variant, sometimes referred to as Ascalon-X, is an 8-wide decode, out-of-order, superscalar RISC-V core that implements the RVA23 profile and includes dual 256-bit vector units conformant with the RVV 1.0 vector extension. Tenstorrent has publicly cited a performance level of roughly 21 SPECint2006 per GHz, placing the core in the same range as Arm's Neoverse V3 and AMD's Zen 5 cores, a level of performance that was previously considered out of reach for RISC-V implementations [12].
Ascalon was officially productized as licensable IP in 2025, alongside a roadmap that includes a successor core called Babylon [12]. Tenstorrent licenses Ascalon and its Tensix AI cores as IP blocks that customers can integrate into their own SoCs, in addition to selling finished Tenstorrent-branded silicon. In 2023 the company licensed both its Tensix AI accelerator core IP and its Ascalon CPU core IP to LG Electronics.
In October 2025 Tenstorrent introduced the Open Chiplet Atlas (OCA), an open architecture specification intended to allow chiplets from different vendors to be combined into a single package without proprietary licensing fees at any layer. OCA defines a standard for physical interfaces, protocols, and packaging conventions and is positioned as a counterpoint to the more vendor-controlled chiplet ecosystems built around UCIe and proprietary interconnects [18].
The initial OCA coalition includes Tenstorrent itself, LG Electronics, Axelera AI, AheadComputing, AIDC, BOS Semiconductors, and a number of other partners. Tenstorrent's first OCA-compliant chiplets, codenamed Aegis, are scheduled to be manufactured on Samsung Foundry's 4 nm and 3 nm nodes.
A central pillar of Tenstorrent's strategy is its commitment to an end-to-end open-source software stack. The three named layers are TT-Metalium, TT-Forge, and the legacy TT-Buda [14].
TT-Metalium is the low-level, open-source SDK that exposes Tenstorrent hardware directly. It gives developers explicit control over the RISC-V cores, the Network-on-Chip, and the matrix and vector engines inside each Tensix core. Metalium is conceptually similar in role to CUDA at a low level, but every layer of it is publicly available on GitHub [16]. The TTNN library is built on top of Metalium and provides PyTorch-style operator kernels.
TT-Forge is Tenstorrent's MLIR-based compiler stack. It accepts models from high-level frameworks such as PyTorch, JAX, and TensorFlow and compiles them down through MLIR dialects and TTNN kernels to run on Tenstorrent hardware. As of 2025, more than 800 model variants were tracked in continuous integration. TT-Forge is positioned as a general path for developers who do not want to write custom kernels [15].
TT-Buda was an earlier compiler stack that compiled models from several frameworks directly to Tenstorrent hardware. It has since been archived in favor of TT-Forge built on TT-Metalium, although the source remains on GitHub.
Tenstorrent also maintains a presence on Hugging Face, where it publishes pre-compiled model artifacts for its hardware, and the company runs a developer cloud that provides remote access to Wormhole and Blackhole systems.
Tenstorrent has raised money in a series of progressively larger rounds, with strategic participation from major Asian electronics conglomerates and prominent technology investors. The Series D round in particular was widely covered as one of the largest AI hardware financings of 2024 [2].
| Round | Date | Amount (USD) | Notable investors |
|---|---|---|---|
| Seed | 2016 | ~$5.9 M | Real Ventures, Eclipse Ventures |
| Earlier rounds | 2018-2022 | undisclosed (multi-round) | Eclipse, Fidelity, Moore Strategic Ventures, Real Ventures |
| Series C extension | August 2023 | $100 M | Hyundai Motor Group ($30 M), Kia ($20 M), Samsung Catalyst Fund |
| Series D | December 2, 2024 | $693 M+ | Samsung Securities, AFW Partners (co-leads); LG Electronics, Hyundai Motor Group, Bezos Expeditions, Fidelity Management & Research, Baillie Gifford, XTX Markets, Corner Capital, MESH, Export Development Canada, Healthcare of Ontario Pension Plan |
| Later round | Late 2025 | ~$800 M (reported) | Fidelity Management (lead), pushing valuation toward $3.2 B |
The Series D round closed at a pre-money valuation of approximately $2 billion and a post-money valuation of approximately $2.6 billion, was oversubscribed, and was advised by Wilson Sonsini Goodrich and Rosati [1][20]. Bezos Expeditions, the personal investment vehicle of Amazon founder Jeff Bezos, participated in the round and gave Tenstorrent additional visibility in the U.S. press as a potential challenger to NVIDIA [2].
In mid June 2026, multiple outlets including The Register reported, citing The Information, that Qualcomm was in advanced talks to acquire Tenstorrent in a deal valued at up to roughly $10 billion (with some reports placing the range at $8 billion to $10 billion), potentially structured as a mix of cash and stock [3]. The reports framed the move as a way for Qualcomm to obtain a RISC-V-based AI accelerator architecture aimed at the inference workloads that increasingly dominate AI infrastructure cost, accelerating its push into the data center. Both companies declined to comment on the reports, which described the talks as not yet concluded and subject to change. If completed at the reported figures, the deal would value Tenstorrent at roughly four times its December 2024 post-money valuation and would rank among the largest acquisitions in Qualcomm's history.
Tenstorrent's investor base substantially overlaps with its IP customer base. Several of the strategic relationships are unusual for a fabless silicon startup in that they combine licensing, joint chiplet development, and equity investment.
| Partner | Nature of relationship | Notes |
|---|---|---|
| LG Electronics | IP licensee, equity investor, Open Chiplet Atlas partner | Joint development of RISC-V, AI, and video codec chiplets for LG premium TVs, automotive products, and Tenstorrent data center products |
| Hyundai Motor Group | IP licensee, equity investor | $30 M direct investment in 2023; partnership to co-develop CPUs and AI co-processors for future Hyundai mobility vehicles and robots |
| Kia | Equity investor | $20 M direct investment in 2023 as part of the Hyundai-led tranche |
| Samsung (Catalyst Fund and Securities) | Lead investor in Series C extension and Series D | Strategic ties to Samsung Foundry for advanced-node manufacturing |
| Bezos Expeditions | Series D investor | Personal investment vehicle of Jeff Bezos |
| SiFive | Technology partner | SiFive X280 cores integrated into Blackhole alongside Tenstorrent's own RISC-V designs |
| Japan ecosystem (Rapidus and government partners) | RISC-V and chiplet collaboration | Tenstorrent RISC-V and chiplet technology selected as part of Japan's national AI infrastructure plan |
| Open Chiplet Atlas members | Standards coalition | LG, Axelera AI, AheadComputing, AIDC, BOS Semiconductors and others |
The $100 million Series C extension in August 2023 was led by Hyundai Motor Group and Samsung, with Hyundai contributing $30 million and Kia $20 million [17].
Tenstorrent occupies a distinct niche among the post-2015 wave of AI chip startups challenging NVIDIA. Whereas competitors have generally built proprietary architectures and proprietary software stacks, Tenstorrent has bet on RISC-V and open-source software as a competitive wedge.
| Company | Founded | Headquarters | Architecture | Software stack | Primary focus | Reported peak valuation |
|---|---|---|---|---|---|---|
| Tenstorrent | 2016 | Santa Clara, CA / Toronto | Tensix cores plus RISC-V CPUs, open chiplet ecosystem | Open-source (TT-Metalium, TT-Forge) | Training and inference, IP licensing | ~$2.6 B (Dec 2024); ~$3.2 B reported late 2025; up to ~$10 B Qualcomm bid reported June 2026 |
| NVIDIA | 1993 | Santa Clara, CA | GPU streaming multiprocessors, NVLink, Grace Hopper | Proprietary (CUDA) | Training and inference, gaming, graphics | Public company, multi-trillion-dollar market cap |
| Cerebras | 2016 | Sunnyvale, CA | Wafer-scale single die (WSE) | Proprietary Cerebras SDK | Inference and training, especially LLM inference latency | ~$15 B+ targeted IPO range (2026) |
| Groq | 2016 | Mountain View, CA | 144-wide VLIW Tensor Streaming Processor (LPU) | Proprietary, ahead-of-time scheduled | Inference latency on LLMs | ~$6.9 B (Aug 2025 reports) |
| SambaNova | 2017 | Palo Alto, CA | Reconfigurable Dataflow Unit (RDU, CGRA) | Proprietary SambaFlow | Enterprise dataflow systems and agentic workflows | ~$5 B (2021) |
Groq, Cerebras, and SambaNova all rely on highly customized architectures and proprietary software. Tenstorrent's strategy is the opposite: standard RISC-V cores, an open chiplet specification, and an open-source compiler and runtime. Whether that openness is enough to attract developers away from NVIDIA's mature CUDA ecosystem is the central commercial question facing the company.
Grayskull and Wormhole were both fabricated at GlobalFoundries on a 12 nm process. Blackhole moved to TSMC's 6 nm node, and the upcoming Aegis chiplets are planned at Samsung Foundry on 4 nm and 3 nm. The simultaneous use of three different foundries gives Tenstorrent unusual flexibility in supply but also adds engineering complexity, since each migration requires re-validating Tensix and RISC-V designs against a new process design kit.
Tenstorrent has emphasized that its choice of GDDR6 over High Bandwidth Memory is deliberate. HBM supply is dominated by a small number of suppliers and is largely absorbed by NVIDIA, AMD, and the hyperscaler ASIC programs. By targeting GDDR6, Tenstorrent can avoid HBM allocation constraints and ship developer-priced cards at retail price points that are at least an order of magnitude below comparable HBM-based accelerators, although at lower memory bandwidth per watt.
As of mid 2026, Tenstorrent is operating along three distinct revenue streams: discrete AI accelerator cards and Galaxy systems, licensable IP blocks (Ascalon CPUs, Tensix AI cores, and OCA chiplets), and a developer cloud built on Wormhole and Blackhole hardware. The combination of low-cost discrete cards, open software, and high-end licensable IP is intended to build a developer base from the bottom while extracting royalty revenue from large electronics OEMs at the top.
The company's positioning relative to NVIDIA is explicitly contrarian. Where NVIDIA's Hopper and Blackwell architectures stack large amounts of HBM behind massive monolithic GPU dies and lock developers into CUDA, Tenstorrent ships modest-die GDDR6 accelerators connected by Ethernet, exposes the underlying RISC-V cores directly to developers, and publishes the entirety of its compiler and runtime as open source. Jim Keller has repeatedly framed this as the only viable long-term route to competition, on the grounds that any closed alternative will struggle to escape NVIDIA's gravitational pull.
The immediate commercial test will be Galaxy Blackhole adoption and the conversion of Tenstorrent's strategic investors (especially LG, Hyundai, and Samsung) into volume IP customers. The longer-term test is whether Tenstorrent's Ascalon CPUs and Tensix accelerators can credibly serve as a RISC-V-based foundation for high-performance data center compute outside of the company's own products, in the same way Arm's Neoverse line did for the Arm ecosystem during the 2020s. The reported June 2026 Qualcomm acquisition interest underscores how much strategic value the broader industry now places on a credible RISC-V AI alternative to NVIDIA [3].