What Happened
Qualcomm, the company whose Snapdragon processors power a large share of the world's smartphones, is in talks to acquire Tenstorrent, the AI-chip startup led by veteran architect Jim Keller, in a deal that financial reporting puts at between $8 billion and $10 billion. The discussions were first reported by The Information and quickly picked up across the trade press, though neither company has announced an agreement.
Tenstorrent is not a typical acquisition target. It designs AI accelerators built around its own Tensix cores and the open RISC-V instruction set, and it sells both finished cards and intellectual-property licenses to customers who want to design their own silicon. That combination — hardware, an open software stack, and licensable IP — is precisely the bundle Qualcomm would struggle to assemble on its own timeline.
The reported price marks a steep climb. In December 2024, Tenstorrent closed a Series D of more than $693 million at a $2 billion pre-money valuation, with backers including Samsung Securities, AFW Partners, LG, Hyundai Motor Group, Fidelity, Baillie Gifford and Jeff Bezos's Bezos Expeditions. An $8–10 billion figure implies roughly a four- to five-fold step up in under eighteen months — a measure of how much the value of AI silicon talent has moved.
At the center of the deal is Keller himself. He helped shape AMD's Zen architecture, Apple's early A-series chips and Tesla's self-driving hardware, and he became Tenstorrent's chief executive in 2023 after joining as chief technology officer. For Qualcomm, acquiring the team may matter as much as acquiring the products: in advanced silicon, a handful of senior architects can be worth as much as a portfolio of patents.
Why It Matters
For Qualcomm, the logic is about escaping the smartphone. The company has spent years trying to reduce its dependence on mobile, pushing into automotive, Windows-on-Arm laptops, extended reality and, most recently, the data center. Tenstorrent would accelerate the hardest of those moves.
Qualcomm has already shown its hand with AI200 and AI250, rack-scale inference systems aimed at data centers rather than single cards. The company says an AI200 rack carries 43 terabytes of memory and that one card can hold a 350-billion-parameter model, with commercial deployment targeted for 2026 and the AI250 following in 2027. What it lacks is a deep bench of AI-architecture engineers and a credible open-software story — exactly Tenstorrent's strengths.
The bigger backdrop is the shift in where AI money is spent. As models move from training to everyday inference, cost per token and efficiency per watt are becoming the metrics that decide infrastructure budgets. That is the ground on which challengers hope to chip away at Nvidia, whose GPUs and CUDA software still dominate. Tenstorrent's pitch — open, programmable, RISC-V-based accelerators tuned for inference — is aimed squarely at what one report called Nvidia's blind spot.
The company's flagship card, Blackhole, shows the approach. The p150a pairs 120 Tensix cores with 16 large RISC-V cores, 32 GB of GDDR6, 180 MB of on-chip SRAM and four 800G ports for lashing cards together. Rather than copy the GPU model exactly, it offloads data orchestration onto general-purpose cores — a design meant to ease the small-batch inference bottlenecks that matter most in production.
Reaction
Investors responded quickly. Qualcomm shares climbed on the reports, a sign that Wall Street reads the move as a serious bid for relevance in AI compute rather than a defensive flourish. The reaction reflects how hungry the market has become for a credible third option beyond Nvidia and AMD.
The deal also carries a defensive edge. Tenstorrent has become one of the most courted startups in the sector, and Intel has been mentioned as another potential suitor. If Qualcomm does not move, a rival might — and a competitor holding Keller's team and a maturing RISC-V accelerator line would be an uncomfortable outcome for any chipmaker trying to break into inference.
Not everyone is convinced the numbers add up. Skeptics note that an $8–10 billion price looks rich against Tenstorrent's current revenue, which remains modest relative to the valuation. The bull case rests on future capacity rather than today's sales: the team, the IP, the RISC-V roadmap and the chance to compress years of in-house development into a single transaction.
What's Next
The most immediate marker is timing. Qualcomm's Investor Day falls on June 24, and the company is widely expected to use it to lay out its data-center roadmap — the natural venue to confirm a deal of this size, or at least to signal how serious its ambitions are even if talks continue.
Plenty could still go wrong. Both companies have declined to comment, the price could shift, and terms might include milestone payments tied to product delivery. A transaction this large would also draw scrutiny from investors and, very likely, regulators weighing its effect on an AI-chip market already under close watch.
Integration is the quieter risk. Folding a fast-moving startup into a large, process-driven corporation can blunt the very speed that made it attractive. And competing against entrenched software stacks — CUDA, ROCm, and the in-house toolchains of the hyperscalers — takes years of sustained investment, as Tom's Hardware and others have noted while tracking Tenstorrent's earlier Wormhole cards.
Closing Thoughts
Whatever the outcome, the talks say something about the direction of the industry. The AI-chip race is no longer a simple duel between Nvidia and AMD; a layer of challengers is forming around inference, memory efficiency and more open, customizable architectures.
RISC-V sits at the heart of that story. Where Arm and x86 are licensed or proprietary, RISC-V is open and customizable, and Tenstorrent has made the marriage of AI acceleration and RISC-V central to its identity. As governments and large companies push for more control over their technology supply chains, that openness looks less like a niche preference and more like a strategic asset.
For Qualcomm, the prize is not the most powerful accelerator on a spec sheet. It is the ability to deliver better cost per token, higher efficiency per watt and more freedom to customize — the qualities that increasingly decide who wins inference workloads. Tenstorrent offers a fast, if expensive, way to argue that the company's future does not have to rest on smartphones. Whether the deal closes or not, the message is already clear: the competition for the economics of AI has moved well beyond raw training horsepower.
한글 요약
퀄컴이 짐 켈러가 이끄는 AI 반도체 스타트업 텐스토렌트를 80억~100억 달러에 인수하기 위해 협상 중이라고 복수의 외신이 전했다. 텐스토렌트는 자체 Tensix 코어와 개방형 RISC-V 아키텍처를 기반으로 한 AI 가속기와 IP 라이선스를 제공하는 기업으로, 2024년 12월 시리즈 D에서 20억 달러의 기업가치를 평가받았다. 이번에 거론되는 금액은 1년 반 만에 4~5배 뛴 수준으로, AI 실리콘 인재와 설계 역량의 몸값이 얼마나 올랐는지를 보여준다.
퀄컴은 스마트폰 의존도를 낮추고 데이터센터 추론(inference) 시장에 진입하기 위해 AI200·AI250 랙 스케일 시스템을 준비해 왔으나, AI 아키텍처 인재와 개방형 소프트웨어 역량은 부족했다. 텐스토렌트 인수는 이 공백을 빠르게 메우는 동시에, 엔비디아·AMD 양강 구도에 맞설 'RISC-V 기반 추론 가속기'라는 차별화 카드를 안겨준다. 6월 24일 열리는 퀄컴 인베스터 데이가 발표의 분수령이 될 전망이며, 인텔도 잠재 인수 후보로 거론된다.
다만 현재 매출 대비 가격이 높다는 회의론과 규제 심사, 대기업-스타트업 통합 리스크가 변수로 남아 있고 협상이 무산될 가능성도 있다. 그럼에도 이번 사례는 AI 칩 경쟁의 무게중심이 '학습 성능'에서 토큰당 비용·전력 효율·개방형 설계로 옮겨가고 있음을 분명히 보여준다. (출처: The Information, Tech Times, Tom's Hardware 등)