Jim Keller designed AMD's K8 in 2003, when AMD was two bad quarters from irrelevance. Then the A4 and A5 at Apple. Then Zen at AMD again. Then Tesla's Full Self-Driving chip. Four architectures across four companies, each market-defining. He's the kind of architect whose involvement makes things real that weren't real before.

Since 2021 he's been running Tenstorrent. Building RISC-V AI accelerators. Three generations of silicon: Grayskull, Wormhole, and now Blackhole — a 6nm chip with 120 Tensix++ cores, 664 Block FP8 TFLOPS, and 400 Gbps Ethernet built in. Actual products. Available to order.

According to Reuters, confirmed June 16, Qualcomm wants to buy the company for between $8 and $10 billion. A year ago Tenstorrent was valued at $3.2 billion. Intel is also reportedly interested. Neither Qualcomm nor Tenstorrent has commented. Talks are ongoing with no guarantee a deal closes.

The question isn't whether the deal closes. It's what the bidding war tells you about where the AI chip race is going.

What RISC-V is and why it matters here

RISC-V is an open instruction set architecture. No company owns it. No royalties. Anyone can design silicon around it. That matters because the AI chip market is dominated by proprietary stacks: NVIDIA's CUDA ecosystem sits on top of NVIDIA hardware, ARM licenses its ISA for a fee. Running NVIDIA means being inside NVIDIA's software moat — which is exactly what has kept NVIDIA's gross margins near 78% as AI compute demand went vertical.

Tenstorrent built on RISC-V because they believe the long-term advantage is an open stack where you control everything from instruction set to compiler to runtime. Not as a philosophy statement. As a competitive strategy. If the software stack is open, a better chip doesn't need NVIDIA to cooperate.

Whether that bet is right is partly what Qualcomm's $8-10B bid is testing.

3×

→ Source: Reuters / The Register, June 2026

What Qualcomm is actually buying

Qualcomm has Snapdragon. Mobile silicon. Good at what mobile requires. They've been talking about a credible data center story for years without landing one. Tenstorrent gives them three things they can't easily build themselves:

1. A shipping AI accelerator with real customers and real silicon (Blackhole)
2. Jim Keller
3. A RISC-V head start measured in years, not months

Item 2 is the one I'd weight most. Keller's track record isn't just that he designs fast chips — it's that he designs chips that change competitive dynamics. AMD went from irrelevant to threatening Intel on server share, and a lot of that trace runs through Zen. Qualcomm is paying a premium for the certainty that kind of talent is inside, not outside.

Jim Keller's chip career

1. 2003

   AMD K8

   Rescued AMD from near-insolvency. Competed with Intel Pentium 4 on price/performance.

2. 2010

   Apple A4/A5

   Foundation of iPhone's first silicon advantage over Android.

3. 2012

   AMD Zen

   A second time. Zen architecture restored AMD as a serious server competitor to Intel.

4. 2018

   Tesla FSD chip

   In-house silicon for Full Self-Driving inference, replacing NVIDIA hardware in Tesla vehicles.

5. 2021

   Tenstorrent CEO

   Took over as CEO. Building RISC-V AI accelerators. Now the acquisition target.

Source spread

• The Register — Qualcomm circling Tenstorrent in $10B RISC-V power play — builder. Solid technical coverage of deal rationale and RISC-V implications. The Register cited Reuters as the originating source.
• Tom's Hardware — Qualcomm mulls taking over Jim Keller's Tenstorrent — builder. Valuation breakdown, deal structure (cash + stock + performance milestones).
• The Next Web — Intel and Qualcomm circle Tenstorrent — skeptic. Intel competitive interest and the no-guarantee framing.
• AI Unfiltered — Valuation tripled in one year — hype. Tracks the full Tenstorrent fundraising arc and contextualizes the premium being paid.

Pros & cons

What makes this acquisition compelling:

• Tenstorrent ships real silicon. Blackhole (6nm) is not vapor. Developer kits and workstations are available today. Customers are paying for it.
• RISC-V's open architecture is a structural long-term advantage if the software ecosystem matures. It's maturing — slower than anyone wants, faster than two years ago.
• Jim Keller's track record of producing real, market-shifting chip designs is unmatched in the semiconductor industry. This isn't a hype acquisition; the underlying technical depth is real.
• Intel competing for the same asset tells you the asset is credible. Intel doesn't go after vaporware at this price level.

What deserves skepticism:

• Qualcomm has tried the data center story before. They built Centriq, a competitive Arm-based server chip, in 2017. They cancelled it in 2018 after one large customer passed. Acquisitions don't repair a strategic commitment problem that runs through multiple leadership cycles.
• The deal structure includes performance-related milestone payments. That's appropriate for a technical roadmap this ambitious, but it's also Qualcomm signaling they aren't confident the full $10B is earned yet.
• RISC-V software is still years behind CUDA. The model that says "builders will migrate stacks for cost" has been right about every prior compute transition. It's been wrong about AI hardware specifically. CUDA lock-in is real and deep.
• Jim Keller under Qualcomm's corporate structure. He's not known for thriving inside large company bureaucracies — he's known for thriving in environments where he controls the chip roadmap. Acquisitions are notoriously bad at retaining this kind of talent. The risk isn't that Tenstorrent's chips are bad. The risk is that Keller leaves in year two.

What builders need to know

• Tenstorrent's Blackhole chips are available now as developer kits and workstations. If you're evaluating NVIDIA alternatives for inference, this is the most credible hardware available outside the hyperscaler custom silicon space.
• CUDA migration is not a weekend project. Tenstorrent's compiler and runtime toolchain is improving but if you're deeply embedded in CUDA, switching costs are real. Evaluate the software maturity before the hardware specs.
• If Qualcomm closes this, expect 2027+ for any Qualcomm-branded AI inference products. These timelines are long. Don't hold product roadmaps.
• Intel competing signals the AI chip battle is moving to the silicon layer, not just the model layer. Builders who want long-term optionality on hardware vendors — and who don't want to be permanently inside NVIDIA's software moat — should be paying attention to where this settles.
• The Keller premium is baked into the $10B. Anything Tenstorrent produces in the next two years assumes Keller stays post-acquisition. Watch post-close signals on retention.

Further reading

• The Register — Qualcomm said to be circling Tenstorrent in $10B RISC-V power play — the most detailed technical analysis of deal rationale
• Tom's Hardware — Qualcomm mulls taking over Jim Keller's Tenstorrent — valuation and deal structure breakdown
• The Next Web — Intel and Qualcomm circle Tenstorrent — the competitive bidding angle
• Tenstorrent — Blackhole hardware page — the actual product: specs, availability, workstation configs
• AI Unfiltered — Valuation tripled in one year — full fundraising context for the $8–10B premium