Intel and Qualcomm are reported in talks for Tenstorrent worth up to $10 billion

Real-time voice AI got a price tag this month, and it is low enough to move money. On October 2, 2026, Smallest.ai, a voice-AI research lab, put its Lightning V2 speech model on Tenstorrent'sPVT~$2.6B — Tenstorrent, private, latest valuation ~$2.6B Galaxy Blackhole servers, on usage-based pricing with no upfront commitment ¹. The engineering paper behind the launch↗ claims the same workload runs for roughly a quarter of the accelerator cost of an Nvidia▼−1.59% — Nvidia, down 1.59 percent today L40S ².

The claim is vendor-published and unverified, which is exactly why it matters. Intel▼−6.29% — Intel, down 6.29 percent today and Qualcomm▼−3.81% — Qualcomm, down 3.81 percent today are reported to be bidding as much as $10 billion↗ for Tenstorrent, a price any bidder can now test before signing ³⁴.

The price break had to be engineered into the model

Tenstorrent's advantage starts in arithmetic. Its Blackhole chips run BlockFloat8, an eight-bit number format, natively on silicon that lists near $1,000 a device, while the paper's authors note that GPUs with native support for the format sit near $40,000 ⁵. Speech is the hard case: text-to-speech models generate continuous waveforms, and small numerical errors compound across denoising steps into audible artifacts ².

Smallest.ai rebuilt Lightning V2 so that more than 80% of the model runs in that eight-bit format, using the chip's internal memory network to cut data movement, and the audio held: a 0.071 perceptual-score gap against the L40S, which the authors call minor variation ²⁵.

The fleet-scale discount is partly projection

At fleet scale, 550 simultaneous five-second requests, the authors price the workload at about $100,000 of L40S cards or about $27,000 of Tenstorrent P100s ⁵. The P100's latency figure, though, is carried over from measured P150 results↗, so the cheaper fleet number rests partly on unmeasured latency ⁵. Layers too numerically sensitive to quantize, and a compiler not yet fully tuned, mark where measurement ends and projection begins ⁵.

Japan repeated the pattern on language models

Japan supplied a second data point the same day. Fixstars, a Tokyo performance-engineering firm, and Macnica, a Tokyo-listed chip distributor (TSE: 3132), tuned the open-weight gpt-oss-120b model on a Tenstorrent workstation↗ ⁶⁷. A key matrix-multiply step ran 3.68 times faster, and per-token generation time fell from 144.85 to 99.49 milliseconds, a 31 percent cut ⁶. Porting models onto this silicon is already a billable service Macnica sells ⁷.

The bids price a token nobody has independently measured

Tenstorrent reached general availability with Galaxy Blackhole on April 28, 2026, and raised a $1.29 billion Series E that August↗ ⁴⁸. Bloomberg reported Intel and Qualcomm in early talks at $5 billion or more, and The Information, since confirmed by Reuters, put a Qualcomm deal at $8 to $10 billion; no deal has closed ³⁴. Nvidia's AI silicon generates margins above 70 percent, and a cost break that survives verification is the arithmetic that would compress them ⁴.

Verification is the open link. Comprehensive independent benchmarks of Tenstorrent on production workloads do not yet exist, and Bernstein analyst Stacy Rasgon warns that a buyer may be acquiring Tenstorrent chief executive Jim Keller, an architect with a record of leaving big companies within a few years ⁴. Any bidder can now run the workload and read the invoice before signing, and that invoice will settle the argument.