Jim Keller Kills the Qualcomm Rumor — Tenstorrent Is Building for an IPO, Not a Sale

When reports emerged in mid-June that Qualcomm was in advanced talks to acquire Tenstorrent for up to $10 billion, the AI chip world paid close attention. The strategic logic was obvious: Qualcomm needed an AI inference engine capable of challenging Nvidia, and Tenstorrent's Blackhole chip — built on the open RISC-V instruction set — looked exactly like that. Then, on June 30, CEO Jim Keller walked into a Tokyo media event and shut it all down.

"We are not in discussions with Qualcomm about an acquisition," Keller said. He emphasized that Tenstorrent is focused on building its own business — including IP licensing and high-performance scalable AI infrastructure work across Japan and other markets. He left the door open for "partnerships in strategic areas," but the message was unambiguous: Tenstorrent is not for sale.

What Qualcomm Was Actually After

The acquisition thesis made real sense. Qualcomm was reportedly circling Tenstorrent at a $8–10B valuation — a RISC-V power play designed to crack Nvidia's grip on the AI inference market. Tenstorrent's Blackhole chip, which reached general availability on April 28, 2026, is architecturally different from Nvidia's H100s and B200s. Rather than stacking GPU parallelism, each Tensix core bundles five RISC-V processors alongside dedicated matrix and vector engines connected through a mesh-network fabric. Keller has argued this design handles the real bottleneck of modern AI inference — memory bandwidth and interconnect efficiency — better than the GPU model can.

Qualcomm, for its part, has been spending aggressively to escape its smartphone-chip reputation. On June 24, the company announced a $14 billion bet on RISC-V and an open compiler stack for AI inference — a direct attempt to undercut Nvidia's proprietary CUDA ecosystem. Tenstorrent's architecture would have been a ready-made piece of that strategy.

Building for an IPO, Not an Exit

Keller has made no secret of where he wants to take Tenstorrent. In recent interviews, he's described wanting to challenge Cerebras, Nvidia, and the broader AI hardware market on the company's own timeline. The commercial momentum backs that up: Tenstorrent recently shipped its largest order to date — a 96-Galaxy cluster comprising 3,072 Blackhole chips destined for a customer outside the US. That's not a startup scrambling for liquidity. That's a company filling serious order books.

The denial also cooled Qualcomm's deal prospects publicly, with analysts noting the deal now looks considerably less likely. For Qualcomm, that means its $14B open RISC-V infrastructure bet has to stand on its own — without Tenstorrent's chip architecture as a shortcut.

What It Means for AI Inference

Whether or not the acquisition talks were real, the story confirms that the AI inference chip market is heating up fast enough to generate $10B rumors around a company that only shipped its flagship chip three months ago. Nvidia still dominates AI training with the Blackwell platform, but inference — where the majority of enterprise AI spending is now concentrated — is increasingly contested territory. Tenstorrent, Cerebras, AMD, and a highly motivated Qualcomm are all fighting for that market.

For developers and enterprises evaluating inference infrastructure, the takeaway is that meaningful competition to CUDA is becoming real. Tenstorrent's open RISC-V approach means customers adopting Blackhole hardware aren't locked into a proprietary instruction set. That's a genuinely different value proposition from every GPU vendor in the market — and one Jim Keller clearly wants to pursue as a public company, not as a division of Qualcomm.