Jim Keller's Tenstorrent Bet: Can Open-Source Chips Topple NVIDIA's Empire?

> Note: The original video title misspelled the company as "TenStorent"—the correct name is Tenstorrent. Jim Keller's resume speaks for itself: AMD Zen, Apple A4/A5, Tesla FSD, Intel Xe. Now the 64-year-old chip legend is making what may be his final big bet: using open-source RISC-V architecture to seize the AI inference table from NVIDIA. > > Meanwhile, players have already been cleared off the table. In December 2025, NVIDIA reportedly spent ~$20 billion to acquire Groq's LPU technology and team—the largest deal in its history. Will Tenstorrent be next, or does it hold something NVIDIA cannot buy?

1. Who Is Jim Keller: Four Era-Defining Chips

Keller's地位 in chips is often compared to having Hinton + LeCun + Bengio combined in AI. Nearly every generation-shifting chip of the past 20 years carries his signature:

| Period | Chip | Company | Impact | |:---:|:---|:---|:---| | 1998-1999 | AMD K7/K8 | AMD | First real threat to Intel in x86 | | 2008-2012 | Apple A4/A5 | Apple | Foundation of iPhone/iPad, ARM mobile era | | 2012-2015 | AMD Zen | AMD | Brought AMD back from the brink; still its cash cow | | 2015-2018 | Tesla FSD | Tesla | First automotive-grade AI inference chip | | 2018-2020 | Intel Xe | Intel | Attempt to rescue Intel's process troubles; failed | | 2021-now | Tenstorrent | Tenstorrent | RISC-V + open AI chips challenging NVIDIA |

Keller's philosophy: don't optimize what others have optimized—redefine the problem. But this time the incumbent is NVIDIA: a $3+ trillion company controlling chip, software, and ecosystem end-to-end.

2. Tenstorrent's Technical Approach: Neither GPU Nor TPU

2.1 Tensix Cores: RISC-V + Matrix Engine Hybrid

• Each Tensix core packs 5 RISC-V processors + vector/matrix units + local SRAM
• No hardware cache hierarchy (L1/L2/L3)—data movement is explicitly software-controlled
• Chips interconnect via standard Ethernet, no NVLink or external switches needed
• Mesh topology with point-to-point core communication

| Dimension | NVIDIA GPU | Tenstorrent Tensix | |:---|:---|:---| | Control | SIMT | 5 independent RISC-V cores | | Memory | Hardware-managed caches | Explicit SRAM management, no caches | | Interconnect | NVLink/InfiniBand + switches | On-chip Ethernet, direct torus | | Software | CUDA (proprietary) | TT-Forge/TT-Metalium (open source) | | Scaling | Dedicated interconnect hardware | Standard Ethernet |

The bet: AI dataflow patterns are predictable and don't need general-purpose GPU flexibility. Skip transistor spend on caches, branch prediction, and out-of-order execution; spend it on matrix units instead.

2.2 Fully Open-Source Stack

• TT-Forge: open AI compiler (PyTorch/JAX/ONNX), public beta
• TT-Metalium: low-level SDK for kernel writing, MIT license
• TT-LLK: low-level kernel software
• RISC-V ISA: royalty-free open instruction set

Tenstorrent wants to repeat Linux's victory over Unix: use open-source ecosystem stickiness to counter a proprietary ecosystem's short-term advantage.

2.3 Chip Roadmap

| Generation | Time | Notes | |:---|:---|:---| | Grayskull | 2020 | Proof of concept | | Wormhole | 2022-2023 | Better interconnect, training + inference | | Blackhole | 2024-2026 | Current flagship: 480 Tensix cores, 2,654 TFLOPS (BlockFP8) | | Grendel | 2026+ | Chiplet architecture; CPU and AI compute as separate tiles |

Grendel builds on the Open Chiplet Atlas standard—a "USB for chips" enabling cross-vendor chiplet interop.

3. Performance Claims vs. Reality

3.1 Official Claims (TT-Deploy launch, May 2026)

• Galaxy Blackhole server: DeepSeek R1 inference at 350+ tokens/second
• TCO: claimed 5x lower than NVIDIA GB300
• "Blitz Mode": generative AI video claimed 10x faster than current GPUs
• Keller: "We are committed to crushing everybody at everything"

3.2 Reality Check

1. What scenario is 350 tokens/s? Single vs. multi-user? Batch size? Quantization? Prefill/decode mix? Conditions were not fully disclosed. 2. How is 5x TCO computed? TCO includes hardware, power, cooling, migration costs, and retraining—migration from CUDA to TT-Forge may be omitted. 3. 10x video speed vs. what? The comparison baseline is unclear. 4. Software maturity: TT-Forge is in public beta; "runs" and "runs well" for production are separated by months or years of engineering. CUDA took ~5 years to mature; AMD's ROCm has chased for a decade.

3.3 GB300 vs. Galaxy Blackhole

| Dimension | NVIDIA GB300 | Tenstorrent Galaxy Blackhole | |:---|:---|:---| | Architecture | Blackwell GPU + Grace CPU | RISC-V + Tensix | | Memory | HBM3e | GDDR6 / on-chip SRAM | | Software | CUDA (20 years) | TT-Forge (beta) | | Cloud availability | All major clouds | Koyeb and partners | | Model support | Nearly all AI models | Compile-dependent, limited | | Production validation | Global deployment | Small scale; Japan/Korea sovereign projects |

Bottom line: Tenstorrent may win on specific workloads (particular LLM inference, batch sizes), but generality and ecosystem maturity lag NVIDIA by orders of magnitude—realistic for a 5-year-old startup.

4. The $20B Story: Groq as Cautionary Tale

Groq's LPU (by Google TPU father Jonathan Ross) is radical: no caches, no branch prediction, no out-of-order execution—compiler-controlled deterministic execution with on-chip SRAM replacing HBM. In early 2024 it ran Llama 2 70B at 241 tokens/second, 2x+ GPUs of the time.

On December 24, 2025, NVIDIA announced acquiring Groq's core technology and team (including Ross and Sunny Madra) for ~$20 billion—nearly 3x the Mellanox deal. Motivations:

1. GPUs optimize compute-heavy prefill; decode is memory-bandwidth-bound, where Groq's SRAM architecture excels 2. Buying a competitive threat is cheaper than losing inference market share 3. Acquiring the TPU design school and its talent

Warning for Tenstorrent: Groq's exit proves novel inference architectures have value—but the ceiling of that value may be NVIDIA's acquisition price. Keller's counter is open source: NVIDIA can buy a company, but not an open ecosystem. TT-Forge is MIT-licensed; RISC-V is an open ISA. This is the Linux-vs-Microsoft playbook.

5. Business Model: Three Legs

Hardware sales: TT-QuietBox 2 ($9,999 workstation, 4x Blackhole, runs 120B-parameter models locally), TT-LoudBox servers, PCIe cards, Galaxy Blackhole rack-scale systems.

IP licensing (the ARM play): Ascalon RISC-V CPU IP licensed to LG, Hyundai, Japan's LSTC; Neo AI core IP to sovereign AI projects; Open Chiplet Atlas interconnect standard. Most bookings come from IP deals—selling designs, not chips.

Sovereign AI: Japan, Korea, Canada and others building sovereign AI compute on Tenstorrent—free from US export controls and NVIDIA supply chains.

6. Funding

| Time | Round | Amount | Valuation | Investors | |:---|:---|:---|:---|:---| | Dec 2024 | Series D | $1B+ (oversubscribed) | $2.6B | Samsung, Bezos Expeditions, Fidelity, Hyundai, LG, XTX Markets, Baillie Gifford | | Nov 2025 | Target | - | $3.2B | - |

$2.6B vs. NVIDIA's $3 trillion is a 1000x gap—but Tenstorrent need not beat NVIDIA outright. It needs an inference beachhead or ARM-style IP success. Key question: can its cash burn until the software ecosystem matures?

7. The Broader Open-Chip Movement

| Company | Approach | Status | Valuation/Market cap | |:---|:---|:---|:---:| | Cerebras | Wafer-scale (WSE) | IPO May 2026 (Nasdaq $CBRS) | ~$56B | | Etched | Transformer ASIC (Sohu) | Not shipping | ~$5B | | Groq | SRAM-based LPU | Acquired by NVIDIA Dec 2025 | $20B | | Tenstorrent | RISC-V + open ecosystem | Shipping + IP licensing | ~$3.2B | | AMD | ROCm + GPU | In production | ~$350B | | Google | TPU v7 (Ironwood) | In production | Internal | | Amazon | Trainium/Inferentia | In production | Internal | | Microsoft | Maia/Cobalt | In production | Internal |

Trends: the inference market is fragmenting from NVIDIA monopoly toward multipolar competition; hyperscalers build in-house chips; independents get acquired, IPO, or go open-source; valuation logic has shifted from "can you build a chip" to "can you build an ecosystem."

8. Conclusion: Keller's Final Big Bet

Can Tenstorrent topple NVIDIA? Honest answer: not short-term; possibly long-term.

Short-term (1-3 years): immature software, CUDA lock-in, workload-specific advantages only, huge valuation gap.

Long-term (5-10 years): RISC-V datacenter share rising (projected 25% by 2026); open-ecosystem stickiness once formed is hard to displace; sovereign AI demand; IP licensing may scale better than chip sales; and if inference overtakes training as the bigger market, Tenstorrent's inference-optimized architecture gains.

Keller isn't betting Tenstorrent defeats NVIDIA—he's betting the AI compute market structuralizes and fragments, with multiple specialized architectures coexisting. Groq got a $20B exit ticket; Keller wants Tenstorrent to become the next ARM—not acquired, but infrastructure no chip can bypass.

References:

• Tenstorrent: https://tenstorrent.com
• TT-Forge GitHub: https://github.com/tenstorrent
• "NVIDIA's $20B Groq Deal" - Yahoo Finance, 2025-12-26
• "Tenstorrent Vows to 'Crush Everyone'" - WCCFtech, 2026-05-02
• "RISC-V in 2026: 25% Market Share" - AEStech, 2026-05-03
• Cerebras IPO: Nasdaq $CBRS, May 2026, ~$56B valuation
• "Nvidia to license AI chip challenger Groq's tech" - TechCrunch, 2025-12-24