Qualcomm's $10B Tenstorrent Bid: What RISC-V AI Cards Mean for Home Labs in 2026

This article was originally published on runaihome.com

TL;DR: Qualcomm is reportedly in talks to buy Jim Keller's Tenstorrent for up to $10B, and the cards at the center of it — the Blackhole p150a (32GB, $1,399) and p100a (28GB, $999) — are buyable today. They give you more VRAM than a used RTX 3090 at lower power, but roughly half the memory bandwidth and a much younger software stack. For pure local LLM inference in 2026, a used 3090 still wins on tokens/sec and ecosystem.

	Blackhole p150a	Blackhole p100a	Used RTX 3090
Memory	32GB GDDR6	28GB GDDR6	24GB GDDR6X
Bandwidth	512 GB/s	448 GB/s	936 GB/s
Price	$1,399	$999	~$1,070 (used)
Power (TBP)	300W	300W	350W
Software	TT-Metalium (Apache 2.0), young	Same	CUDA, mature
Best for	RISC-V/open-stack devs, big-model headroom	Cheapest 28GB inference card	Fastest tok/s for the money

Honest take: Buy a Tenstorrent card if you want to develop on an open RISC-V stack or you need 28–32GB at 300W and you enjoy being early. If you just want a local LLM that runs fast tonight, a used RTX 3090 is still the better buy — and don't buy speculatively hoping the Qualcomm deal helps you. Acquisitions disrupt roadmaps before they improve them.

What Qualcomm is actually trying to buy

On June 15, 2026, The Information reported — and Reuters, Tom's Hardware, and The Register picked up — that Qualcomm is in advanced talks to acquire Tenstorrent at a valuation between $8 billion and $10 billion. The deal is said to be cash and stock with performance-based adjustments still under negotiation, and nothing is final. QCOM stock jumped over 4% on the news.

Tenstorrent is the AI-chip startup founded in Canada in 2016 and run by Jim Keller — the silicon architect behind Apple's A-series, AMD's Zen, and Tesla's FSD chip. What makes Tenstorrent different from every other AI-accelerator name you've seen acquired is the architecture: it's built on open RISC-V CPU IP paired with proprietary Tensix AI cores, and the entire software stack is open source. For a phone-SoC company like Qualcomm watching smartphone growth flatten, buying a RISC-V inference architecture is a shortcut into the part of AI spend that's growing fastest — inference, not training.

That's the boardroom story. The reason it matters on this site is narrower: the hardware Tenstorrent sells is consumer-accessible. You can put a PCIe card in a cart on tenstorrent.com right now. So the real question for a home lab isn't "will the deal close" — it's "does this hardware do anything a used 3090 or a 5060 Ti doesn't, and should I wait to find out?"

The cards you can actually buy

Tenstorrent sells two generations of PCIe inference cards. The current Blackhole generation is the headline; the older Wormhole generation is what most of the software has actually been tuned against.

Card	Tensix cores	SRAM	Memory	Bandwidth	TBP	Price
Blackhole p100a	120	180MB	28GB GDDR6	448 GB/s	300W	$999
Blackhole p150a	140	210MB	32GB GDDR6	512 GB/s	300W	$1,399
Wormhole n150d	72	108MB	12GB GDDR6	288 GB/s	160W	$1,099
Wormhole n300d	128 (2 ASIC)	192MB	24GB GDDR6	576 GB/s	$1,449	$1,449

The Blackhole p150a is the interesting one. It carries 32GB of GDDR6 — the same capacity as an RTX 5090, more than any 3090/4090 — for $1,399, draws 300W, and adds four QSFP-DD 800G ports for clustering cards together. The p100a undercuts it at $999 for 28GB. On paper, Tenstorrent claims the p150a matches an RTX 4090 in FP8 and BF16 TFLOPS at lower power (300W vs the 4090's 450W).

The catch is in the column most buyers skip: memory bandwidth. Decode speed — the tokens-per-second you actually watch stream — is bound by how fast the chip can read the model's weights out of memory, not by peak TFLOPS. We've made this point in our NPU vs GPU breakdown and it applies here exactly. The p150a's 512 GB/s is barely over half a used RTX 3090's 936 GB/s, and a third of an RTX 5090's ~1,792 GB/s. More capacity, slower reads.

The bandwidth math, in tokens

Here's why that gap matters. A 7B model in Q4_K_M is roughly 4.5GB of weights. To generate one token, the chip reads the active weights once. Crudely, decode throughput scales with bandwidth ÷ bytes-read-per-token, so the ceiling is set by GB/s.

A used RTX 3090 at 936 GB/s does roughly 95 tok/s on a 7B Q4 model — a number we've measured repeatedly across the site. Scale that by bandwidth and the Blackhole p150a's theoretical ceiling lands near 50–55 tok/s on the same model, before you account for software efficiency. And software efficiency is the second tax: early community benchmarks suggest Blackhole reaches 40–60% of its theoretical TFLOPS on real LLM workloads, versus 60–80% for NVIDIA cards with mature kernels. Stack the two together and the practical decode speed on a single Blackhole card for everyday 7B–14B models is well under what a 3090 delivers.

Where Tenstorrent's numbers look strong is aggregate, batched throughput — the data-center metric, not the single-user one:

System	Model	Throughput	Notes
TT-QuietBox 2 (4× Blackhole)	Llama 3.1 70B	476.5 tok/s	Aggregate, vendor figure
Wormhole Galaxy (32 ASIC)	Llama 70B, batch 32	~4,000–5,000 tok/s	Vendor, not independently verified
8× H100 SXM5 (vLLM)	Llama 70B, batch 32	~2,500–3,500 tok/s	For comparison

Read those carefully. The 476.5 tok/s on the TT-QuietBox 2 — four Blackhole ASICs, 480 Tensix cores, 2,654 TFLOPS BlockFP8, 128GB GDDR6, starting at $9,999 — is aggregate throughput across concurrent requests, not what one person watching one chat session feels. The Galaxy numbers come from Tenstorrent's own controlled runs on TT-Metal, not independent third-party audits, and a single user pulling one stream off any of these boxes sees a fraction of the aggregate. For a home lab running one or two sessions at a time, batched throughput is the wrong yardstick.

The software is the real story (and the real risk)

The thing that genuinely sets Tenstorrent apart isn't the silicon — it's that the entire stack is Apache 2.0 open source. TT-Metalium (the low-level kernel programming model) and TTNN (the operator library) are both Apache 2.0. There's a Tenstorrent-maintained fork of vLLM, and tt-inference-server wraps it in an OpenAI-compatible API so you can point existing tooling at it. If you've been frustrated by CUDA's black-box nature, this is the most open serious AI accelerator you can buy. It's the same "no-CUDA-required" pitch we examined with Intel Arc and AMD ROCm, but taken further — the kernels are yours to read and rewrite.

The flip side is maturity. As of early 2026, most verified model support and documentation targets the Wormhole n150/n300 cards; Blackhole software is earlier in its cycle. There's an experimental llama.cpp discussion thread for Grayskull/Wormhole, but it's community work, not a polished path. CUDA, by contrast, runs every model on day one with Ollama, LM Studio, llama.cpp, vLLM, and ComfyUI. With a Tenstorrent card you're buying into a roadmap, not a finished product — you'll spend time getting models running that would be a one-line ollama pull on NVIDIA. If you like building on an open stack (this is firmly FOSS territory), that's a feature. If you want to run a model tonight, it's friction.

RISC-V vs CUDA: why Qualcomm cares and you might not (yet)

The strategic logic is sound. Inference is becoming the dominant cost in AI, NVIDIA's CUDA moat is built around training, and a RISC-V architecture with an open stack is the kind of thing a hype