Tenstorrent, the AI chip company led by veteran processor designer Jim Keller, announced on October 19, 2026 that it closed a $1.5 billion Series E at a $6.2 billion post-money valuation. The number is large even by this year’s standards, but the more useful question for founders is what the round says about where money is flowing in AI infrastructure: away from pure model-building and toward the hardware and licensing layers underneath it.
What was raised, and from whom
According to the company’s release, Trace Capital and Innovation X led the round. The participant list is unusually broad, mixing public-market names such as Fidelity Management & Research, T. Rowe Price, ARK Invest and Morgan Stanley with strategic and trading-sector backers including Sandisk, Tether Investments, Virtu Financial and XTX Markets. Dealroom’s write-up puts the valuation at more than double the roughly $2.6 billion attached to the company’s roughly $700 million Series D in late 2024.
The lead investors are themselves a story. Per the release, Trace Capital was founded only in 2026 by Ben Mahdavi, a former co-founder of Quiet Capital, while Innovation X is a late-stage firm founded in 2016 with more than $4.5 billion under management as of the end of 2025. A brand-new fund co-leading a billion-dollar-plus round is a reminder of how much late-stage capital is chasing a small set of AI infrastructure names.
The product: one architecture, sold at every size
Tenstorrent builds AI processors and systems on the open RISC-V instruction set. It sells compute from single cards and workstations up to racks and multi-rack superclusters, and it licenses its TT-Ascalon CPU design and Tensix AI cores to other chip designers; the company names Samsung and LG among licensees. Keller’s own summary, as quoted in the coverage, is that the firm built one architecture and sells it at every size.
The company says the new capital will go toward deploying its Galaxy servers at scale, serving open-weight models for language and video generation, and advancing the IP it licenses. One analysis of the round reports that Galaxy systems began shipping in April 2026 and that a May company update described them as in volume production, including a 36-system supercluster operating as a single computer. Treat those as company-reported claims: the same analysis notes that Tenstorrent has not disclosed how many systems customers have received, what they cost to own and run, or how its software performs against Nvidia on like-for-like workloads.
A sale was on the table, and it chose capital instead
Dealroom, citing Bloomberg, reports that Tenstorrent held takeover talks with Intel and Qualcomm earlier this year but opted for fresh funding over a sale. A separate analysis points to an Axios account of the same courting and notes that the company declined to comment and has called such reports market speculation. The safe reading is that suitors were reportedly interested; no offer terms are public, and Tenstorrent has not confirmed any talks.
Keller has linked investor appetite to shipments, pointing to a May demonstration of Galaxy servers running DeepSeek models as a turning point. That sequencing matters. Investors appear willing to fund a hardware company again once there is physical product in customers’ racks, not just a roadmap.
The honest risks
Several outlets that covered the deal are careful to separate investor confidence from commercial proof. The Magica analysis lists the open questions plainly: whether Tenstorrent can move from first shipments to repeatable volume manufacturing, whether developers stay as models change faster than software can be tuned, and whether licensing agreements turn into recurring customers rather than isolated design wins. Independent benchmarks that disclose model, batch size, latency, throughput, power and total cost are still missing.
The competitive context is also crowded. Dealroom lists other 2026 AI-chip raises, including Cerebras at $1 billion and Positron at $875 million on a $5 billion valuation. Nvidia remains the incumbent every one of these companies is measured against, and Tenstorrent’s argument is that an open, customizable stack will be cheaper and easier to adapt. That is a thesis, not yet a result.
Smaller facts also deserve care. Sources differ on where the company is based: the release and most outlets put headquarters in Santa Clara, while at least one listing calls it Toronto-based. The company moved its headquarters from Toronto to Santa Clara in 2023 and keeps a large Canadian presence, which likely explains the confusion.
Why this round matters beyond chips
Three signals stand out. First, capital is concentrating in the layers where scarcity is physical: compute, memory and power. Second, open standards are being treated as an investable strategy rather than a charity project, with RISC-V as the clearest example. Third, strategic investors from storage and trading firms, who buy huge amounts of compute, are putting money in, which suggests customers want a second source of supply, not only a cheaper one.
For founders building on top of AI models, the practical consequence is indirect but real. If alternative hardware stacks mature, inference costs could become less tied to a single vendor’s pricing. Nothing in this announcement proves that will happen soon, and any price relief would depend on volume production and software maturity that remain unverified.
What This Means for Philippine Founders
Most Philippine startups will never buy a chip, but many will feel compute pricing through their cloud and API bills. A sensible response is to avoid hard-wiring a product to one provider’s proprietary tooling. Building on open-weight models and standard inference interfaces keeps the option open to move workloads if cheaper capacity appears, and it costs little to do early.
On the funding side, the round shows how selective late-stage investors are: money is abundant for infrastructure with shipped product and thin for everything else. We could not confirm any Tenstorrent presence or partnerships in the Philippines, so local founders should not assume access to this hardware. The more realistic takeaway is to track benchmark data as it emerges and to pitch investors with unit-economics evidence rather than roadmaps.
Share this article