Jim Keller Says Tenstorrent Is Ready to Beat Cerebras, NVIDIA, and Others in AI Performance and Cost
Tenstorrent CEO Jim Keller is not losing sleep over Cerebras and its recent push into the public markets. In fact, he appears to welcome the competition. Keller says Cerebras’ IPO and rising valuation are “helpful” for the AI hardware market, but he also made it clear that Tenstorrent believes it can beat Cerebras “on everything.”
That confidence comes as Tenstorrent continues to position itself as a serious challenger in the fast-growing AI accelerator market, where companies are racing to deliver better performance, lower power use, and dramatically improved total cost of ownership.
Tenstorrent’s BlackHole Galaxy server is at the center of that strategy. The company recently introduced the system as a high-performance AI server designed to compete directly with some of the most expensive AI platforms in the industry. Tenstorrent claims its approach can deliver major cost advantages, including significantly better TCO compared with competing NVIDIA-based systems.
Jim Keller welcomes the Cerebras challenge
Cerebras recently attracted attention with performance claims for its WSE-CS3 Wafer Scale Engine hardware running the Kimi K2.6 1T model. The system is said to deliver close to 1,000 tokens per second on that large AI model, a figure that puts it firmly in the spotlight for high-end inference workloads.
Keller, however, does not seem impressed enough to back down. He said Tenstorrent can exceed that level of performance using large-scale BlackHole server deployments while doing so at a much lower total system cost.
“The Cerebras IPO and subsequent valuation was helpful, especially as we’re going to beat them on everything,” Keller said. “Challenge accepted!”
That statement captures Tenstorrent’s broader message: the company is not simply trying to match existing AI hardware leaders. It wants to undercut them on price, compete on performance, and scale into deployments large enough to appeal to major cloud providers, enterprise AI customers, and data center operators.
Tenstorrent’s answer to the KV cache problem
One of the biggest technical challenges in modern AI inference is handling the KV cache, which stores attention data during large language model generation. As models get larger and context windows expand, KV cache management becomes increasingly important for speed, memory use, and overall efficiency.
Keller said Tenstorrent’s design keeps the KV cache in the DRAM on the same chips handling decode workloads. From his perspective, that makes the issue much simpler.
“I am often asked how we handle the KV cache,” Keller said. “It’s in the DRAM on the same chips as the decode, we don’t even think about it. We’re really good at that.”
Tenstorrent’s Tensor chips can store KV cache locally in SRAM when possible. If more capacity is needed, data can be streamed to DRAM. While that can introduce some performance penalty, Keller argues that Tenstorrent’s architecture remains highly competitive because it has direct access to memory in a way that supports practical scaling.
This is an important point in the AI chip race. Many companies can demonstrate strong peak performance, but real-world AI inference depends heavily on memory architecture, bandwidth, latency, software support, and deployment cost. Tenstorrent is betting that its design choices will offer customers a more balanced and affordable path.
Possible Intel or Qualcomm deal could center on RISC-V
Keller also addressed speculation around Tenstorrent’s relationship with major chip companies, including Intel and Qualcomm. While he did not confirm any acquisition-related talk, he did say he has met with the CEOs of both companies and hopes to secure a major deal.
The focus appears to be Tenstorrent’s RISC-V CPU intellectual property.
“I’m hoping to get a big deal out of one of those guys, because our RISC-V CPU IP is great,” Keller said.
RISC-V has become one of the most closely watched CPU architectures in the semiconductor industry because it offers an open and flexible alternative to traditional proprietary instruction sets. Tenstorrent has invested heavily in RISC-V, and a major partnership with a company like Intel or Qualcomm could significantly expand its reach.
Keller also revealed that a hyperscale company has shown interest in Tenstorrent’s AI IP for a smaller chip. That suggests Tenstorrent’s technology may not be limited to massive AI servers. Its designs could also be adapted for custom silicon, edge AI, specialized accelerators, or compact data center products.
Tenstorrent is preparing for a potential IPO
Tenstorrent itself is also aiming toward an IPO. Keller said investors are very interested in the idea, while the company continues working to expand its global presence and strengthen its supply chain.
Demand for Tenstorrent’s AI hardware appears to be rising quickly. The company has reportedly received strong orders for Galaxy systems, including a 96-Galaxy pod order from an offshore customer. That deployment would include 3,072 BlackHole chips, showing that customers are already looking at Tenstorrent for large-scale AI infrastructure.
The timing could work in Tenstorrent’s favor. Demand for AI hardware continues to outstrip supply, and many customers are searching for alternatives to expensive and difficult-to-source systems from established market leaders.
Tenstorrent says its pricing is one of its biggest advantages. According to the company’s positioning, an AI hardware order that might cost around $100 million with NVIDIA could cost closer to $20 million through Tenstorrent, depending on the configuration and workload requirements.
That kind of price difference is exactly why Tenstorrent is getting attention. AI companies, cloud providers, and enterprises are under pressure to expand compute capacity while controlling infrastructure costs. If Tenstorrent can deliver competitive performance at a fraction of the cost, it could become one of the most disruptive players in AI hardware.
BlackHole Galaxy demand is already building
Tenstorrent is currently working on around 1,000 Galaxy servers, and Keller indicated that roughly half have already been sold. That is a strong signal for a company still building its place in a market dominated by larger names.
The broader AI accelerator market is moving fast, but it is also becoming more cost-conscious. Companies no longer want performance alone. They want better efficiency, lower hardware costs, easier scaling, and a reliable supply chain. Tenstorrent is trying to deliver all of those at once.
Jim Keller’s message is clear: Tenstorrent is not intimidated by Cerebras, NVIDIA, or any other AI hardware competitor. With BlackHole Galaxy servers, RISC-V CPU IP, growing customer demand, and possible partnerships with major chipmakers, Tenstorrent is presenting itself as a serious force in the next phase of AI computing.
The competition in AI chips is only getting more intense, but Keller seems ready for it. As he put it, “Challenge accepted.”
