Intel and Qualcomm both reportedly came shopping for Tenstorrent this year, and on Monday Jim Keller answered with $1.5 billion of somebody else’s money. Tenstorrent announced on October 19 that it closed a Series E at a $6.2 billion post-money valuation, led by Trace Capital and Innovation X, and says the cash goes into volume manufacturing of its Galaxy servers, global commercial expansion, and the RISC-V CPU and AI cores it licenses to other chipmakers. When I wrote up Tenstorrent’s Developer Day in May 2025, Keller’s team had raised roughly $1 billion in total and was selling $999 cards next to fully open repos, and this single round is half again the lifetime total I cited then.
Since Monday, Keller explained the decision to stay independent on Bloomberg Television, Japan’s research funding agency picked Tenstorrent and Preferred Networks for a five-year chip project on Wednesday, and on Thursday Tenstorrent pushed the beta of its Open Chiplet Atlas Harness to GitHub under Apache 2.0. Those three moves show where the money goes and who Tenstorrent wants building on it.
Intel and Qualcomm Circled, Keller Raised Anyway
Axios reported that the startup “was courted earlier this year by Intel and Qualcomm, but is opting to remain independent. At least for now.” Tenstorrent declined to comment to Axios on “market speculation,” so the detail on those talks comes from reporters and unnamed sources. Bloomberg first reported early takeover conversations with both chipmakers in May, and in June Reuters, citing The Information, said Qualcomm was discussing a deal worth $8 billion to $10 billion. TechCrunch’s Series E story puts the earlier interest at valuations of as much as $5 billion, and Tenstorrent has not confirmed any of those figures.
Keller offered his own reason on Bloomberg Television. According to TechCrunch’s account of that interview, he said Tenstorrent chose to stay independent and raise more capital after its Galaxy servers showed they could run DeepSeek models. Reuters, via CNA, reports that the December 2024 round raised more than $693 million at a $2 billion pre-money valuation, so the company’s price more than doubled in under two years while the reported buyout numbers spread from $5 billion to $10 billion around it.
Galaxy Is Where the $1.5 Billion Lands
The money is pointed at one very heavy box. Galaxy Blackhole is a 6U, air-cooled server with 32 Blackhole chips, 1 TB of GDDR6 at 16 TB/s, and a rated 23 PFLOPS of Block FP8, according to the Galaxy product page. It lists at $160,000, averages 8 to 10 kW with a 12 kW max, and weighs 262 pounds, so plan the rack power and the lift crew before the purchase order. Four of them make a $640,000 supercluster, and Tenstorrent quotes builds of 4 to 36 or more Galaxy systems stitched together over plain Ethernet.
Named deployments are starting to show up on Tenstorrent’s newsroom. Smallest.ai said on October 2 that its Lightning V2 text-to-speech model runs natively on Galaxy Blackhole for on-prem voice agents, and one of its inference engineers credits the network-on-chip and larger SRAM for “4x gains” against comparable GPU infrastructure, without describing the benchmark behind it. In September, Japan’s ai& launched JapanFold, a drug discovery service that runs open structural biology models such as Boltz-2 and OpenFold3 on Galaxy superclusters inside Japan.
Tenstorrent also claims a supercluster renders 81 frames of 720p video in 2.4 seconds, and Protagonist’s George Bousis said its “video generation results beat industry benchmarks ahead of our expectations,” without naming the benchmark. Our team at TechForward has not run a Galaxy, and every performance figure in this section belongs to Tenstorrent or its customers.
The $999 Card Still Ships in Two Weeks
This is the part of the Developer Day story I worried a rack business would bury, and it hasn’t. The Blackhole p100a still lists at $999 with 120 Tensix cores and 28 GB of GDDR6, the p150a adds four 800G QSFP-DD ports and 32 GB for $1,399, and the cards page shows both in stock and shipping in two weeks. The $11,999 Blackhole QuietBox I listed last May now shows as sold out, and its successor, the liquid-cooled QuietBox 2, lists at $9,999 with four Blackhole processors on two p300c cards and runs open-weight models up to 120B parameters from a wall outlet, according to Tenstorrent.
Keller’s line on Monday was “We built one architecture and sell it at every size,” the same pitch he made on stage last year when he said “Everybody who focuses on just one thing is going to fail.” The open software is what makes that line testable. The stack is public on GitHub from the TT-Forge compiler down to TT-Metalium kernels, the nightly development tags I pointed to last year are still landing, and tt-metal cut a v0.81.0 release candidate on Wednesday. Tenstorrent says Tensix keeps one programming model from a single chip to a rack, so a kernel you tune on a $999 card targets the same cores that ship inside Galaxy, and the Galaxy page says 90% of Hugging Face models “just work,” a number worth checking against your own model list.
Three Days Later, the Chiplet Blueprints Went Open Source
Thursday’s GitHub drop is the Open Chiplet Atlas Harness, the reference design behind the chiplet interoperability standard Tenstorrent launched a year ago, released as an open source beta. The tt-oca-harness repo ships full RTL for the plumbing every chiplet needs, including a security processor that acts as the root of trust, a system management controller for boot, clocks, power and telemetry, cross-chiplet debug and test ports, and an AXI-over-UCIe bridge for die-to-die traffic, plus a SystemC virtual platform so firmware teams can start before silicon. BOS Semiconductors contributed the original AXI-over-UCIe RTL, lowRISC helped open source it, and the full release is targeted for January 2027.
“The OCA architecture gives chiplet designers the freedom to build without depending on a single vendor,” said Wei-Han Lien, Tenstorrent’s chief CPU architect. A day earlier, Preferred Networks and Tenstorrent announced that Japan’s NEDO selected their joint proposal for a five-year research program on AI processors built with photonics-electronics integration, where PFN develops AI chiplets with 3D-stacked memory and Tenstorrent develops open-architecture CPU chiplets that support confidential computing.
Both moves feed the third use of funds. Tenstorrent licenses its TT-Ascalon RISC-V CPU and Tensix AI cores to chipmakers, names Samsung and LG as licensees, and lists Samsung, LG Electronics, Hyundai Motor Group and Bezos Expeditions among its backers. If you design silicon, the harness is RTL you can clone today, and the licensed cores are the commercial product sitting on top of it.
Watch-Points Before You Rack a Galaxy
The Developer Day piece ended with execution risk, and most of that list survived a $6.2 billion valuation, so four items deserve a look before anyone signs a Galaxy purchase order.
- The Form D question. A Form D that Tenstorrent Holdings filed on September 9 reports $1.29 billion in equity sold to 134 investors, with a first sale on August 26. Tenstorrent has not said whether that filing is part of this Series E, so treat the two figures separately until it does.
- Revenue stays private. That same filing answers “Decline to Disclose” on revenue, so nobody outside the company can set a $6.2 billion price against sales.
- Vendor benchmarks. The video, voice and model-coverage claims need independent runs, and the software polish gap with CUDA I flagged last year is still the thing to watch in tt-metal release notes.
- “At least for now.” Axios chose those words, and they leave the buyout door open.
The OCA Harness is also a beta, built for early adopters and contributors, so chiplet teams should treat the January 2027 full release as the date that matters for tape-out plans.
Same Keys, Much Bigger Garage
Last May I wrote that Tenstorrent handed the community a set of keys and a $999 entry pass, and both made it through the raise. The p100a card still sells for $999, the stack is still open on GitHub, and most of the new money goes into building Galaxy servers in volume. Three days after the announcement, the company also put the RTL for its chiplet harness on GitHub. Axios left the door open on a sale with “at least for now,” so the cheap way to find out whether the software is ready for your workloads is a $999 card running your own models, well before anyone prices a $160,000 Galaxy.
Read Tenstorrent’s Series E announcement and the Axios report on the buyout interest, then price a rack on the Galaxy page or start small with a $999 Blackhole card. Clone tt-metal, browse the OCA Harness docs if you design chips, and look at the licensed IP. If you are new to the story, our Developer Day deep dive covers how the open stack got here, and our GPT-OSS piece covers the open-weight models these boxes are built to serve.
