After Rejecting Intel and Qualcomm, Tenstorrent Proves Its Independent Path Is Viable with $1.5 Billion Funding

On October 19, 2026, Tenstorrent announced the completion of a $1.5 billion Series E funding round at a post-money valuation of $6.2 billion. Before that, both Intel and Qualcomm had made acquisition offers for the company, and both were rejected.

According to The Globe and Mail, the acquisition offers from both giants valued the company at more than $5 billion, with Qualcomm's bid reportedly reaching as high as $8 billion to $10 billion. Tenstorrent CEO Jim Keller ultimately chose to remain independent and raise funding himself.

Valuation Doubles in Two Years: What Is the Logic?

When Tenstorrent completed its Series D round in late 2024, its post-money valuation was approximately $2.6 billion. Less than two years later, the Series E valuation jumped to $6.2 billion, an increase of more than 130%. Driving this jump were inference performance data from the Galaxy Blackhole server on the DeepSeek-R1-0528 671B model in May 2026: in a cluster of 16 Galaxy servers, decode speed exceeded 350 tokens/sec/user, time to first token was under 4 seconds at a 100K context, batch size supported 8 to 64, and maximum context supported was 128K. According to WCCFTech, Tenstorrent reported an inference cost of approximately $6 per million tokens, while the corresponding cost for NVIDIA's GB300 NVL72 rack was approximately $30—a difference of about five times.

In an interview, Jim Keller explicitly stated that the DeepSeek demonstration in May was the key turning point in his decision to reject the acquisition offers and instead raise funding independently.

Who Is Betting on This Path

This round was led by Trace Capital and Innovation X, with participation from ARK Invest, Fidelity Management & Research Company, Morgan Stanley, T. Rowe Price, Tether Investments, and existing investors Samsung and LG Electronics.

ARK Invest is a representative fund for the technology-disruption narrative; Fidelity and Morgan Stanley are traditional large institutional asset managers; T. Rowe Price prefers long-cycle value targets. The simultaneous appearance of these three very different types of investors on the same cap table shows that Tenstorrent's investment story has been categorized by capital with different risk appetites under frameworks such as technology disruption, infrastructure substitution, and sovereign compute.

Samsung and LG appearing among the existing investors carries a more direct strategic meaning: both companies are licensees of Tenstorrent IP, using its TT-Ascalon RISC-V CPU cores and Tensix AI cores to develop their own chips. This means Tenstorrent's business model also exports intellectual property for chip architecture.

Jim Keller Is Betting Not on a Single Chip but on a System

To understand Tenstorrent, one cannot avoid the name Jim Keller. He led the design of Apple's A4/A5 processors, AMD's Zen microarchitecture, and Tesla's autonomous-driving chips, making him one of the few architects in the contemporary semiconductor industry to have left landmark work at multiple top companies. He officially became Tenstorrent's CEO in 2023, and that same year the company moved its headquarters from Toronto to Austin, Texas.

In the official announcement, Keller summarized the company's core strategy: "We built a unified architecture and sell it in various form factors." From a single accelerator card and workstation to racks, superclusters, and the IP cores licensed to Samsung and LG, all use the same underlying design philosophy. This differs structurally from NVIDIA's path—NVIDIA's ecosystem moat lies in the CUDA software stack and accompanying system integration, while Tenstorrent is betting on an open-source software stack plus a licensable IP architecture.

George Bousis, a partner at Protagonist, specifically noted in the funding announcement: "Very few companies can design both chips and the software that runs on them; Tenstorrent has done it."

Inference Is the Real Battleground

From 2024 to 2025, the market narrative around AI chips centered on training compute. As mainstream frontier models gradually mature, the weight of inference costs in actual business operations has risen rapidly. Training happens once; inference runs forever.

This is the structural reason Tenstorrent chose to enter the inference market. On the training side, NVIDIA's H100/H200 ecosystem advantage is almost impossible to shake in the short term; on the inference side, however, the requirements for latency, throughput, and unit cost are completely different, opening up new competitive space. The technical design of Galaxy Blackhole—32 Blackhole chips, 6.2 GB of on-chip SRAM plus 2.9 PB/s bandwidth, and 1 TB of DRAM with 16 TB/s bandwidth—is a trade-off explicitly aimed at high-volume inference scenarios. The Galaxy Blackhole server, starting at $110,000, targets enterprises and sovereign customers capable of building their own compute infrastructure. The funds will be used for large-scale production of Galaxy servers and global commercial expansion.

The Real Meaning of the Independent Path

Tenstorrent's rejection of Intel and Qualcomm's acquisition offers deserves separate analysis from an industrial-structure perspective. Intel has performed poorly in AI accelerator chips in recent years, and the Gaudi series has a rather limited market share; Qualcomm has a strong position in edge AI but has likewise yet to achieve scale in the data-center inference market. Behind both companies' bids was the use of M&A to make up for shortcomings in their own AI chip strategies.

For Tenstorrent, being acquired would mean that both its IP licensing strategy and its open-source software path would be constrained by the new parent company's interests. As IP licensees, Samsung and LG would not want to see this architecture monopolized by Intel or Qualcomm. Independent operation is a precondition for Tenstorrent to maintain the credibility of its IP licensing business model.

The cost of the independent path is also clear: the company must bear alone the risks of production ramp-up, building a global sales organization, and direct competition with NVIDIA's CUDA ecosystem. Whether the $1.5 billion funding can carry it through this phase depends on whether Galaxy servers can complete the leap from "technical validation" to "volume shipments" in 2026-2027. A 5x TCO advantage is a powerful number, but history is full of cautionary tales about the engineering complexity between a hardware demonstration and large-scale delivery.

What This Funding Round Shows

Diversification of AI compute is no longer just a slogan. ARK, Fidelity, and Morgan Stanley betting simultaneously on the same non-NVIDIA chip company shows that institutional capital has formed a basic judgment: the supplier landscape for AI inference infrastructure will not consist of only one company.

Tenstorrent's position in this contest is relatively unique: it is neither a direct competitor to NVIDIA (it lacks the capital scale to build a full-stack ecosystem) nor an edge AI chip vendor (it does not pursue the consumer end). Instead, it is trying to establish a different set of rules at the specific level of enterprise inference and sovereign compute through an open architecture and IP licensing. Rejecting acquisition by Intel and Qualcomm is essentially about defending the integrity of those rules.

Whether the $6.2 billion valuation is reasonable still needs to be verified by product shipment data. But this funding round itself has already accomplished one thing: it has turned "there are alternatives to NVIDIA" from a technical discussion into a commercial proposition endorsed by mainstream institutions. For the entire AI compute supply chain, this is a structural signal.