Open Silicon’s $6.2 Billion Moment: How Tenstorrent’s Galaxy Push Challenges Nvidia’s Inference Monopoly

On October 19, 2026, Tenstorrent closed a $1.5 billion Series E that valued the Santa Clara-based AI chip company at $6.2 billion post-money. The round, led by Trace Capital and Innovation X and joined by ARK Invest, Fidelity, Morgan Stanley, T. Rowe Price, SanDisk, and others, more than doubled the firm’s valuation from the roughly $2.6 billion mark set after its late-2024 financing. CEO Jim Keller, the microprocessor architect behind AMD’s Zen, Apple’s A4/A5, and Tesla’s Full Self-Driving chip, framed the capital as fuel for volume manufacturing of Galaxy servers and expansion of the company’s licensable RISC-V AI and CPU intellectual property.

This is not merely another mega-round in a crowded field of Nvidia challengers. It is a deliberate bet that open architectures, Ethernet-based scale-out, and software stacks that developers can inspect and modify will capture a durable share of inference workloads—chatbots, coding agents, video generation, and multimodal systems—where cost, customizability, and supply-chain optionality matter as much as peak FLOPS. The timing coincides with visible strains in the broader AI infrastructure build-out: project delays, grid constraints, and questions about whether frontier-lab revenues will cover the full cost of proprietary accelerator fleets.

What the Round Actually Funds

Tenstorrent has already begun commercial shipments of its Galaxy platforms. The Blackhole generation packs 32 RISC-V-based accelerators into a 6U chassis, delivering roughly 23 petaflops of FP8 compute, 1 TB of GDDR6, and 16 TB/s of memory bandwidth, interconnected by a dense Ethernet mesh rather than proprietary NVLink-style fabrics. Earlier Wormhole systems offered lower density but the same architectural thesis: one design point that scales from a single card or workstation to racks and superclusters. The new capital is earmarked for volume manufacturing of these servers, next-generation silicon, and continued licensing of TT-Ascalon RISC-V CPUs and Tensix AI cores to partners including Samsung and LG.

Keller has emphasized that the company sells the same architecture at every size and that open-weight models run effectively on the systems. Demonstrations involving DeepSeek-class models were cited as evidence that the hardware could serve competitive open models without requiring customers to lock into a single vendor’s software stack. The funding also supports expansion of the open-source TT-Metalium SDK and inference server tooling that aims to give engineers lower-level access than typical CUDA-centric environments.

The investor mix is revealing. Participation from ARK, Fidelity, Morgan Stanley, and T. Rowe Price signals that large asset managers see a multi-year inference market large enough to support multiple hardware winners, not a permanent Nvidia monopoly. Earlier acquisition interest from Intel and Qualcomm at valuations up to $5 billion, which Tenstorrent ultimately declined, further underscores that strategic buyers recognized the IP and the team even before this round.

Why Open RISC-V Inference Matters Now

Nvidia’s CUDA ecosystem and co-designed systems (Grace, Blackwell, and the emerging Vera CPU for agentic workloads) deliver unmatched performance density for training and many inference scenarios. Yet the economics of running open-weight models at scale are different. Customers increasingly want the ability to fine-tune, serve, and modify models without negotiating custom software contracts or waiting for vendor support. RISC-V’s royalty-free instruction set removes one licensing layer; Ethernet-based interconnects allow standard networking gear instead of proprietary switches; and an open SDK lowers the barrier for in-house optimization.

This aligns with broader industry pressures documented in recent analysis. TSMC’s record quarters and the concentration of leading-edge capacity have made advanced wafers and advanced packaging scarce and geopolitically sensitive, as detailed in Wafer Economics in the Age of Agents. Memory density advances from Chinese makers such as CXMT have already introduced volatility into the AI chip cycle, covered in The 4F² Shockwave. At the same time, the physical limits of power and grid interconnection have delayed or blocked multiple large campuses, including Microsoft’s withdrawal from the Nscale UK project analyzed in The Energy Wall No Chip Can Bypass.

In that environment, a hardware platform that can be licensed, customized, and deployed with more conventional networking and power envelopes becomes attractive to sovereign clouds, enterprise inference platforms, and companies that prefer to own rather than rent their compute. Tenstorrent’s decision to stay independent and scale Galaxy production rather than sell the company is a direct expression of that thesis.

Competitive Context and Real Limits

Tenstorrent is not alone. Other startups and hyperscaler custom silicon programs (Google TPUs, Amazon Trainium, Microsoft Maia, Meta’s efforts) are also targeting inference efficiency. Nvidia continues to expand its own agentic CPU story with Vera, emphasizing high single-thread performance under full-socket load and monolithic die design for predictable latency. The difference is philosophical as much as technical: Nvidia optimizes the tightly integrated AI factory; Tenstorrent optimizes for open, multi-vendor, multi-size deployment.

Performance claims require independent verification. Galaxy systems have been shown running models such as Llama-class and DeepSeek variants, and video-generation workloads are part of the target set, but sustained tokens-per-second, power efficiency, and total cost of ownership at production scale remain to be proven across diverse customer environments. Software maturity—compiler quality, kernel coverage, and ecosystem tooling—will determine whether developers stay or return to CUDA for production reliability. Licensing revenue from Samsung, LG, and others provides a second path to returns even if server volumes ramp more slowly than hoped.

Local and edge inference trends also interact with this story. Meta’s release of large open models engineered to run on high-end Apple Silicon, examined in The Quiet Migration, shows that not every token needs a distant data center. Tenstorrent’s workstation and card offerings sit in the same continuum: offering developers a path from desktop experimentation to rack-scale serving without changing architectures. Google’s orbital TPU tests, covered in A Kilowatt in Orbit, further illustrate that the industry is exploring extreme form factors; open, licensable IP may prove more adaptable to non-standard environments than closed systems.

What to Watch

Three near-term signals will clarify whether the $6.2 billion valuation is justified. First, shipment and utilization metrics for Galaxy systems already deployed at customers such as Virtu Financial and Equinix. Second, the breadth and performance of the open software stack relative to production needs for coding agents and video generation. Third, additional design wins for the licensed CPU and AI-core IP beyond the existing Samsung and LG relationships.

Longer term, the outcome hinges on whether inference demand fragments across multiple architectures or reconsolidates around the highest-performance proprietary systems. Power constraints, memory pricing, and geopolitical supply-chain risk all favor optionality. If open-weight models continue to close the capability gap with closed frontier systems, the value of flexible, lower-cost serving hardware rises. Conversely, if agentic workloads prove unusually sensitive to low-latency single-thread CPU performance and tightly coupled GPU-CPU fabrics, Nvidia’s integrated approach retains the advantage.

For enterprises and sovereign operators, the practical question is simpler: can they achieve acceptable latency and cost on open hardware for the majority of their inference traffic while reserving premium proprietary capacity for the hardest residual workloads? Tenstorrent’s round supplies the balance sheet to test that proposition at greater scale.

Conclusion

Tenstorrent’s $1.5 billion raise at a $6.2 billion valuation is best understood as a capital-markets endorsement of open, RISC-V-based inference hardware at a moment when proprietary AI infrastructure is encountering real-world frictions of power, packaging, and cost. Jim Keller’s track record gives the company credibility; Galaxy’s commercial availability gives it a product to sell; and the licensing model gives it a second revenue path. Whether that combination displaces a meaningful share of Nvidia-centric inference remains an empirical question that the next 12–18 months of deployments will answer.

In the meantime, the round reinforces a broader pattern: the AI hardware market is large enough, and the constraints on any single supplier severe enough, that well-funded alternatives with differentiated architectures can attract both strategic and financial capital. Open silicon is no longer a research curiosity. It is a funded commercial strategy.

Official sources: Tenstorrent and Reuters coverage of the funding.