Tenstorrent Closes $1.5B Round as RISC-V Galaxy Servers Win Enterprise AI Customers

Tenstorrent Inc. announced Monday it has closed a $1.5 billion Series E financing round at a $6.2 billion post-money valuation, putting the Santa Clara AI chip company among the best-funded challengers to Nvidia's grip on AI computing infrastructure. The round was co-led by Trace Capital and Innovation X, with new institutional investors including Morgan Stanley, Fidelity, ARK Invest, and T. Rowe Price — a mix of growth equity and major asset managers whose participation signals the deal was underwritten by commercial evidence rather than early-stage promise. Tenstorrent's Galaxy Blackhole servers are already running production AI workloads at Virtu Financial Holdings, a high-frequency trading firm, and at data center operator Equinix.

The raise more than doubles Tenstorrent's last valuation. The company's December 2024 Series D — a $693 million round led by Samsung Securities and AFW Partners — was priced at $2.6 billion. In the intervening ten months, the company shipped its Galaxy Blackhole server into general availability, established a 120-system sovereign AI cluster in Japan, and turned away reported acquisition interest from Qualcomm at a rumored $8 to $10 billion. That last detail matters: the company could have taken a premium exit and instead chose to keep building.

RISC-V Architecture Bets Against the GPU Memory Hierarchy

Tenstorrent's hardware is built on the RISC-V open instruction set architecture — the same ISA that now underpins the majority of new AI accelerator designs entering production, as outlined in the RISC-V International 2025 Annual Report. The Blackhole chip, which forms the compute unit inside Galaxy servers, packs 768 RISC-V cores on a single die: 16 large SiFive Intelligence X280 cores handling general-purpose tasks and 752 smaller Tensix++ cores executing tensor math. A Galaxy Blackhole server packs 32 of these chips into a 6U chassis, delivering approximately 23.8 petaFLOPS of FP8 compute, starting at $110,000.

The architectural departure from GPU-based systems is not merely in the ISA. Nvidia's flagship inference hardware uses high-bandwidth memory — HBM3 or HBM3e — bonded close to the GPU die, with NVLink or InfiniBand connecting chips within and between servers. Blackhole uses GDDR6, which carries lower cost and lower bandwidth density than HBM, and routes the scale-out fabric through ten 400-gigabit-per-second Ethernet lanes etched directly onto the die. This eliminates the need for separate networking ASICs and, Tenstorrent argues, makes large inference deployments cheaper to build and operate. The architectural bet is that for long-running, high-batch-size inference on large mixture-of-experts models — which activate only a subset of parameters per token and therefore need high parallelism more than peak per-chip memory bandwidth — the economics of GDDR6 plus cheap Ethernet beat the economics of HBM plus NVLink.

Open Software Stack Separates Tenstorrent From Most AI Hardware Challengers

Hardware architecture alone does not displace an entrenched platform; the reason Nvidia's position is described as a moat rather than merely a market lead is that CUDA's fifteen-year head start in framework support, library coverage, and developer familiarity creates real switching costs for AI engineering teams. Tenstorrent's answer to this is an open-source software stack it has built alongside the silicon.

TT-Forge is a compiler that accepts PyTorch graphs and dispatches them to Blackhole hardware — the entry point for any team that currently builds on PyTorch without rewriting at the kernel level. Below it sits TT-Metalium, the core SDK for memory management, data movement, and lower-level kernel dispatch. TT-LLK provides the library of tensor math operations — matrix multiplications, activations, reductions — that feed into Metalium. All three components are open-source on GitHub. Protagonist co-founder and Managing Partner George Bousis described the combination as unusual: "Few companies design the chip and the software that runs on it." The claim is accurate as a general observation, though competitors like Groq have made similar architectural choices. The difference is that Tenstorrent has open-sourced its stack, which provides an independent verification path and reduces lock-in risk for buyers.

Read more: Microsoft Maia 300 AI chip challenges Nvidia with chiplet design and 7x throughput claim

Galaxy in Production at Virtu Financial and Japan's ai& Cluster

The production deployments disclosed alongside the funding round carry more evidential weight than typical launch benchmarks. Virtu Financial — a firm where trading outcomes are measured in microseconds and infrastructure decisions are made with quantifiable cost models — is running AI trading agents on Galaxy servers. Equinix has Galaxy hardware in its data center infrastructure. In Japan, ai&, a Tokyo-based AI infrastructure company, announced in June 2026 that it had deployed more than 120 Galaxy systems as part of a national sovereign AI initiative — making it the largest known non-GPU AI cluster outside a hyperscaler.

On that infrastructure, Tenstorrent claims its Galaxy cluster running DeepSeek R1-0528 671B — a 671-billion-parameter mixture-of-experts model — delivers more than 350 tokens per second per user across 16 servers at batch size 32, with a time-to-first-token under four seconds at 100,000 tokens of context. These figures are company-reported; no independent third-party benchmark reproduction has been published as of this writing.

Tenstorrent Claims 5x Inference Cost Advantage Over Nvidia GB300, but Benchmarks Are Company-Reported

The headline competitive claim is an estimated total cost of ownership of roughly $6 per million tokens on Galaxy hardware, against what Tenstorrent estimates as approximately $30 per million tokens on a comparable Nvidia GB300-based system — a 5x difference. This figure should be treated as a company claim rather than an independently established benchmark. The comparison methodology — which workloads, which batching conditions, which amortization assumptions — has not been published in a form that allows independent verification.

What can be said is that the architectural choices that reduce Blackhole's per-chip cost (GDDR6 versus HBM, on-die Ethernet versus NVLink ASICs) do create structural cost advantages that are real and measurable, even if the 5x ratio is a best-case number. Constellation Research analyst Holger Mueller framed the broader significance after the announcement: "Nvidia is not the only option anymore." That assessment tracks with the competitive picture — AMD's ROCm ecosystem and Groq's LPU architecture have also narrowed certain gaps — but Tenstorrent is the first to combine an open ISA, a fully open software stack, and institutional-scale production deployments in a single platform.

Read more: Amazon eyes $7 billion Decart deal to break Nvidia's grip on AI chip software

The capital raised will fund production ramping for Galaxy hardware, global workforce expansion, and R&D on the next stage of Tenstorrent's RISC-V roadmap, including the Ascalon CPU. Jim Keller, who designed semiconductor architectures at DEC, AMD, Apple, Tesla, and Intel before taking the CEO role at Tenstorrent in 2023, framed the company's thesis in a statement published Monday: "We've built one architecture and we sell it at every size. The future of compute is open, scalable and owned by the people building on it."

The more important milestone to watch than the funding announcement itself is whether independent benchmark organizations reproduce the DeepSeek inference numbers at scale — and whether Virtu Financial's AI trading infrastructure, one of the most demanding latency environments in enterprise software, keeps running on Blackhole silicon eighteen months from now.