Introduction
Jeff Bezos and Samsung back AI chip startup Tenstorrent in a $693 million Series D round that values the Toronto company above $2.6 billion. The round names Samsung Securities and AFW Partners as co-leads, with Bezos Expeditions, Hyundai, LG Electronics and Fidelity Management joining the cap table. That is the biggest check written into an alternative silicon vendor in 2024. It lands against Nvidia keeping roughly seven of every ten dollars Blackwell earns inside the data center. The deal signals that patient capital still wants an open hardware answer to CUDA, and that Jim Keller’s RISC-V bet has graduated from a technical demo into a procurement conversation. Our earlier coverage of this round captured the headlines of the Series D announcement. This long-form rewrite goes further: it names every backer, walks the Tensix architecture, posts current product prices, lays out deployments, and benchmarks the realistic odds against Nvidia Blackwell.
Quick Answers on the Bezos and Samsung Backed AI Chip Deal
Who are the main backers of the Tenstorrent $693 million Series D round?
Samsung Securities and AFW Partners co-led the round. Bezos Expeditions, Hyundai Motor Group, LG Electronics and Fidelity Management joined, pushing the Toronto chip startup past a $2.6 billion valuation.
How much business has Tenstorrent actually closed as of the funding announcement?
CEO Jim Keller told Bloomberg that the Jeff Bezos and Samsung back AI chip startup already carries almost $150 million in signed customer contracts, with Blackhole products shipping to developers and sovereign AI buyers.
What is Tenstorrent building that would make a dent in Nvidia’s business?
Tenstorrent sells open-source RISC-V based AI accelerators, with Tensix cores, on-chip Ethernet, a Galaxy Blackhole rack from $110,000, and the TT-Metal compiler stack licensed under MIT terms.
Key Takeaways from the Tenstorrent Funding Round
- The $693 million Series D closed on December 2, 2024, at a post-money valuation above $2.6 billion, with Samsung Securities and AFW Partners co-leading the round.
- Bezos Expeditions, Hyundai Motor Group, LG Electronics and Fidelity Management joined as strategic investors, giving Tenstorrent capital, factory relationships and automotive design-win paths at the same time.
- Jim Keller frames the plan as the opposite of Nvidia: open ISA, open compiler, direct Ethernet fabric, and a new processor every two years instead of a locked CUDA stack.
- Operational wins since the round include Galaxy Blackhole general availability from $110,000, a $1,399 developer card, and sovereign AI deployments in Tokyo, Seattle and India through Equinix, OrionVM and BetterBrain.
Table of contents
- Introduction
- Quick Answers on the Bezos and Samsung Backed AI Chip Deal
- Key Takeaways from the Tenstorrent Funding Round
- Understanding Why Jeff Bezos and Samsung Back AI Chip Startup Tenstorrent
- What Jeff Bezos and Samsung Back AI Chip Startup Round Means for Investors
- Who Backed Jeff Bezos and Samsung Back AI Chip Startup Round
- The Jim Keller Factor That Convinced the Backers
- Why RISC-V Is the Wedge Against Nvidia and CUDA
- Inside the Tensix Core and On-Chip Ethernet Fabric
- What Tenstorrent Sells Today and What It Costs Teams
- How Enterprise Teams Implement Tenstorrent Hardware at Scale
- The Open-Source TT-Metal and TT-Forge Software Stack
- Comparing Blackhole Galaxy Performance to Nvidia Blackwell
- Samsung’s Dual Role as Investor and Potential Fabrication Partner
- Bezos Expeditions and the Pattern of Chip-Startup Bets
- Risks That Could Still Derail the Challenge to Nvidia
- Ethics, Export Controls, and Sovereign AI Policy Pressures
- The Future of the AI Chip Market Through 2030
- Key Insights on the Tenstorrent Deal and the Broader Race
- Real-World Examples of Tenstorrent Silicon Deployments
- Case Studies From the Alternative Silicon Push
- Common Questions About the Tenstorrent Funding News
Understanding Why Jeff Bezos and Samsung Back AI Chip Startup Tenstorrent
Jeff Bezos and Samsung back AI chip startup Tenstorrent to seed an open-source RISC-V alternative to Nvidia, with $693 million in Series D funding, a $2.6 billion valuation, and almost $150 million in signed customer contracts.
Size a Tenstorrent cluster against Nvidia H100 for your model
Based on public pricing from Tenstorrent and public rental rates for Nvidia H100 nodes. Adjust the three inputs to see what each path costs you in rack space and dollars for your inference workload. Numbers are indicative, not quotes.
4,000 tok/s
Llama 70B
720 hours
$110,000
Tenstorrent Galaxy Blackhole racks (list)
$1,980
Nvidia H100 node rental (monthly)
55 months
Breakeven on Tenstorrent capex
Assumptions: Galaxy Blackhole rack from $110,000 for 23 PFLOPS Block FP8, Nvidia H100 node on-demand around $2.75 per hour per GPU across eight GPUs, Llama 70B at ~4,000 tok/s on Galaxy. Change the sliders to size for your workload.
What Jeff Bezos and Samsung Back AI Chip Startup Round Means for Investors
Jeff Bezos and Samsung back AI chip startup Tenstorrent at a notable moment for the market. Nvidia keeps nearly seventy percent of its data center compute revenue tied to one Blackwell product line. That concentration is the market gap the round is paying to attack, and it is the single most honest reason the check cleared. The December 2024 announcement arrived days after Nvidia reported quarterly data center revenue of $39.1 billion. That figure implies hyperscalers are deploying close to seventy two thousand Blackwell GPUs every week. Any firm sitting on a differentiated accelerator in that climate becomes an obvious target for both strategic and financial capital. Tenstorrent met the moment with a working Wormhole product, a public roadmap, and a leadership team led by Jim Keller.
Investors are paying for optionality on three axes at once in this round. First, they buy exposure to the thesis that an open RISC-V stack can erode CUDA lock-in. The wedge should hit sovereign and research buyers first, well before the hyperscale tier moves. Second, they are buying a foundry hedge because the chips are moving to TSMC two-nanometer and discussions with Samsung Foundry remain active per Bloomberg reporting. Third, they are buying an automotive seat with Hyundai and LG Electronics, which both carry long chip budgets for in-vehicle inference and consumer electronics. A single accelerator vendor rarely offers all three vectors at once.
The valuation math still demands scrutiny despite the brand names on the term sheet. A $2.6 billion mark on $150 million in signed contracts is a seventeen times multiple, which is punchy for a hardware company that must still prove volume shipments. The comparable public point is Nvidia, which trades on earnings rather than on bookings. Tenstorrent’s defenders argue that the comparable set is actually SambaNova, Groq and Cerebras, all of which have carried similar multiples at similar maturity. The counter from skeptics is actually quite simple and worth considering carefully. None of those peers has displaced Nvidia from a hyperscaler buy either, so the valuation prices in a future win that has not landed.
Our broader emerging AI chip rivals challenging Nvidia coverage frames the Tenstorrent round as the clearest sign that limited partner appetite for Nvidia alternatives is still rising. Fidelity Management rarely leads deep-tech rounds, so its presence as a late-stage financial investor matters beyond the dollar size. Samsung Securities’ co-lead role tells Korean strategics that this is a national technology opportunity, not only a venture bet. Hyundai’s quiet participation signals that automotive supply chains want a second accelerator source before 2027 vehicle programs lock. The round therefore reads as a coalition deal rather than a one-vendor bet, and that is new.
Who Backed Jeff Bezos and Samsung Back AI Chip Startup Round
Shifting focus to the cap table, Samsung Securities and AFW Partners are the named co-leads of the round. Bezos Expeditions, Hyundai Motor Group, LG Electronics and Fidelity Management round out the Series D syndicate. Samsung Securities is the brokerage and asset management arm of the Samsung group, not the chip division, so its lead signals a strategic positioning rather than a pure manufacturing play. AFW Partners is a smaller growth equity firm that has tracked Keller’s work for years, and the firm pushed the round forward during 2024’s hardware funding slowdown. Bezos Expeditions is the family office that holds Jeff Bezos’s direct personal bets outside of Amazon, and it has backed deep-tech ventures from Altos Labs to Blue Origin. Hyundai Motor Group entered the round through its corporate venture program and it will likely pair the stake with a design-in for in-vehicle inference silicon.
LG Electronics quietly took a position for the same reason Hyundai did, since consumer appliances and televisions are increasingly inference workloads under the hood. Fidelity Management’s late-stage growth team participated as a crossover investor and it typically only writes checks when it sees an IPO path within thirty six months. Previous backers from the 2023 Series C, including Samsung Catalyst Fund, participated again on their existing pro-rata rights. That is often the first sign that strategic investors are not fading on a bet. The investor letter published around the round did not disclose individual check sizes, but Bloomberg’s reporting put the Samsung and AFW tranches above $150 million each. Our deep dive on Bezos backing Physical Intelligence traces the same pattern of patient hardware capital that connects his Tenstorrent position to robotics and biotech bets.
The Jim Keller Factor That Convinced the Backers
Beyond the dollar count, Jim Keller is the single human reason Samsung and Bezos moved on this deal at a seventeen times forward multiple with less than $150 million booked. Keller has shipped flagship silicon four times in his career. His resume includes the AMD K7 and K8, the Apple A4 and A5, Tesla Autopilot, and Intel core work. That is a resume that no competing accelerator CEO can match, and it changes how risk committees price execution risk on a round of this size. The reporting on his RISC-V thesis captured a telling exchange. Asked whether he would consider the vacant Intel CEO job, Keller offered a telling response. He replied “maybe as a side hustle, I’m pretty busy.” The exchange reads as humor. It also signals that he is committed to Tenstorrent through at least one more product cycle.
Keller’s technical argument to investors is that Nvidia’s commercial dominance is itself a design constraint on the ecosystem. CUDA is a tightly coupled stack, and that coupling locks hyperscalers into Nvidia economics for years at a time. His counter proposal is a chip and a software stack that any hyperscaler, sovereign operator or research lab can inspect, modify and redistribute. He has repeated in multiple interviews that he wants to ship a new chip every two years, which is faster than Nvidia’s current cadence and faster than AMD’s roadmap. The AMD’s MI300X push is the closest comparison, and the Keller cadence promise sits on top of a smaller, more focused engineering team.
The second part of Keller’s argument is that CUDA is maintainable only because Nvidia pours billions a year into developer relations and tooling. That spend is a structural weakness because any sufficiently open alternative that attracts enough users can commoditize the compiler layer the way LLVM commoditized backend compilation. He points to Hugging Face, PyTorch, Triton and Mojo as evidence that the lower layers of the AI stack are already drifting open. His pitch to Samsung and Bezos was effectively a bet on this drift, not a bet on raw flops per watt against Blackwell. The strongest tell that investors agreed is that none of the lead investors required a CUDA port as a diligence milestone before closing the round.
Why RISC-V Is the Wedge Against Nvidia and CUDA
Turning to the architecture, RISC-V is the specific technical wedge that lets Tenstorrent credibly claim a lower lock-in profile than any closed instruction set alternative in the AI accelerator category. RISC-V is a royalty-free, standards-maintained open ISA governed by the non-profit RISC-V International foundation. Any vendor can implement a RISC-V core without paying ARM-style licensing fees. That fact alone has made the ISA attractive to governments and large systems companies that want sovereign silicon. Tenstorrent’s chips do not run user AI code directly on a RISC-V core. They use baby RISC-V cores for control, data orchestration, and on-chip networking around dedicated Tensix tensor units. That architecture lets customers audit, extend and port the control stack without renegotiating a license, which is impossible on an Nvidia or AMD design today.
The commercial payoff of RISC-V is a credible answer to the auditability question that regulators now ask. The debate on what open source AI actually means runs directly into this layer: an open ISA plus an open compiler offers a traceability path that closed stacks cannot match. European and Asian procurement teams increasingly require that AI infrastructure be inspectable end to end, especially for defense, healthcare and public-sector deployments. Tenstorrent can take a government agency through its silicon stack layer by layer, including control code, without hiding anything behind a proprietary firmware blob. That is a procurement argument that compounds year on year as the EU AI Act and similar regimes mature.
The strategic payoff is that RISC-V turns every port and every patch into a potential pull request against Tenstorrent’s own stack. The TT-Metal compiler is MIT-licensed on GitHub, so a hyperscaler customer can add kernel support without waiting for a vendor release cycle. That removes the biggest single criticism levelled against startup accelerators, which is slow kernel coverage. The downside is that open contributions require review bandwidth and that bandwidth costs engineering headcount. Tenstorrent’s answer is to channel part of the $693 million into hiring compiler engineers in Toronto, Austin and Bangalore. That matches the staffing pattern of early LLVM and Linux kernel teams, not a traditional silicon startup.
Inside the Tensix Core and On-Chip Ethernet Fabric
Zooming into the silicon, each Tenstorrent accelerator packages Tensix tensor cores together with baby RISC-V management cores and a direct-attach Ethernet fabric that replaces Nvidia’s NVLink and InfiniBand pairing. A Wormhole n300 card carries 128 Tensix cores split across two ASICs, with 192 megabytes of on-chip SRAM and 24 gigabytes of GDDR6 memory. A Blackhole p150 developer card carries 120 active Tensix cores at six-nanometer feature size, with 28 to 32 gigabytes of GDDR6 and sixteen RISC-V cores dedicated to on-die data orchestration. The Ethernet fabric is the architectural tell because it ships at four hundred gigabits per second per link and it scales without separate switching silicon. That design choice avoids the NVLink switch tax and it lets a Galaxy rack scale through standard datacenter cabling.
The Tensix core itself is a tiled dataflow engine built for low-batch inference and medium-size training rather than one-shot billion-parameter runs. Each Tensix tile holds its own compute math units, scratchpad SRAM, and a slice of the on-chip Ethernet. The compiler can place model layers physically close to the memory they read. The trade is familiar to anyone who has worked on a critical turning point in AI hardware. Expect higher throughput per dollar on dense tensor work from the Tensix architecture tiles. Expect a steeper software learning curve than CUDA for the first porting pass.
What Tenstorrent Sells Today and What It Costs Teams
Beyond architecture, price transparency is one of the most under-reported advantages Tenstorrent carries into the market. The full product list is public, and the vendor sells direct rather than through cloud marketplaces. The TT-LoudBox workstation is priced at $12,000 and ships with two Intel Xeon processors, 512 gigabytes of system memory, and four Wormhole n300 cards in a two-by-four mesh. The TT-QuietBox is $15,000, swaps the processor for an AMD EPYC 8124P, and runs liquid cooling for office environments where fan noise matters. The Blackhole p150 developer card is $1,399, which is roughly an order of magnitude under a current Nvidia H100 PCIe part. The Grayskull e75 and e150 dev kits remain in catalog for teams running older PyTorch experiments.
At rack and cluster scale, the Galaxy Blackhole system starts at $110,000 for thirty two chips. It delivers 23 petaflops of Block FP8 compute, one terabyte of DRAM, and standard air cooling. A four-rack Supercluster-36 configuration starts at $440,000, enough to serve a mid-size inference workload for a sovereign AI tenant or a mid-market enterprise. Every system price covers one year of software updates and support, which is unusual for an early-stage silicon vendor. Tenstorrent sells all of these systems direct from its site, with lead times running six to twelve weeks, which is competitive with Nvidia DGX quotes.
The pricing choice signals a specific go-to-market thesis for the company. Tenstorrent is not trying to beat Nvidia on list price for a single Blackwell GPU. It is trying to beat Nvidia on dollar per deployed terabyte of inference capacity on the floor. A $110,000 Galaxy Blackhole carries one terabyte of DRAM, where a comparable Nvidia DGX H100 node carries 640 gigabytes of HBM at a street price above $300,000. Buyers who prize total deployed memory over peak flops see this as a very different calculation. The OpenAI weighing Nvidia alternatives report captured why that trade is now a boardroom conversation at every hyperscaler.
Direct sales also change what buyers can see about availability. Tenstorrent publishes stock levels on its storefront and it takes credit-card orders for developer cards up to the TT-QuietBox tier, which is unusual for an AI silicon vendor. Enterprise customers negotiate volume discounts, deployment services and support SLAs separately. The transparency is the point: a customer evaluating Nvidia versus Tenstorrent can compare a signed quote against a public price list within minutes. That removes a layer of friction that normally slows alternative silicon purchases by six to nine months. It is a small operational decision with a disproportionate strategic effect on inbound pipeline.
How Enterprise Teams Implement Tenstorrent Hardware at Scale
Moving on from procurement, deployment is where the Tenstorrent story becomes real for enterprise buyers. The recipe has three parts that recur across every signed customer today. Those parts are direct colocation, open-source model porting, and a dedicated inference pipeline. The Futurum Group analyst briefing on Galaxy Blackhole availability describes at least five neocloud colocation partners already live. Named partners include Equinix for data centers, OrionVM for orchestration, and BetterBrain for agentic application wrappers. A buyer typically takes a quarter rack of Galaxy Blackhole and places it inside an Equinix cage. The OrionVM orchestration layer points at it, and model porting runs for two to four weeks before production traffic flips over. The pattern looks a lot like how AWS Outposts customers deploy, which lowers the perceived novelty risk.
Model porting then becomes the real work for every team that moves beyond the dev kit. The TT-LLM runtime covers Llama 2 at seven, thirteen and seventy billion parameters, Llama 3 at eight and seventy billion parameters, Falcon 7B and Mistral 7B out of the box. A buyer bringing a custom model allocates one to three engineer-weeks per model to validate kernels and benchmarks against a reference Nvidia node. That cost is non-trivial, but it is a one-time cost against a multi-year hardware depreciation schedule. Our deep dive into Microsoft hoovering up Hopper GPUs explains why hyperscalers now accept the one-time porting cost. Total deployed memory and open stack ownership often outweigh ongoing lock-in to a single vendor.
The Open-Source TT-Metal and TT-Forge Software Stack
Moving from hardware to software, the TT-Metal and TT-Forge stack is where Tenstorrent either wins or loses the next five years of the CUDA fight. It is the single piece every alternative silicon buyer evaluates most carefully during the vendor diligence process. TT-Metal is an MIT-licensed low-level programming library that sits a tier under PyTorch, roughly in the same conceptual spot where Nvidia’s CUDA runtime library lives. TT-Forge is the higher-level compiler flow that takes PyTorch, TensorFlow or ONNX graphs and lowers them into TT-Metal kernels for the Tensix cores. Both projects live in public GitHub repositories under permissive licenses, with public issue trackers, continuous integration logs, and a public weekly release train. The Spheron benchmark review of the stack summarizes the maturity bluntly: “TT-Metal is roughly where CUDA was a decade ago.”
That maturity gap is both the honest risk and the honest opportunity. CUDA accrued a decade of kernel coverage, profilers, and third-party library mindshare that TT-Metal cannot replicate in two years. The counter is that most of today’s AI workloads are only a few years old. A stack that supports modern transformer kernels well can leapfrog the long tail of legacy CUDA kernels nobody uses in production LLM inference. Tenstorrent reports a ninety percent pass rate for Hugging Face models running on Blackhole, covering roughly two and a half million model variants. That claim is self-published, but it is testable because the pass list is public on GitHub.
What the stack still misses is as important as what it supports. There is no equivalent to Nvidia Nsight for profiler-grade kernel debugging, no FlashAttention 3 port yet, and no NCCL replacement at scale beyond thirty two chips. Compilation times for a new model run three to ten minutes on current drivers, which is slow for an iterate-and-debug loop. The TT-Buda higher-level flow is deprecated in favor of TT-Forge, which forces earlier adopters to re-port kernels during the transition. The AMD open-source AI push captures why the same gap now works against Nvidia’s incumbents. Every alternative stack is converging on the same open layers at once.
Comparing Blackhole Galaxy Performance to Nvidia Blackwell
Head-to-head benchmarks between a Galaxy Blackhole rack and a Nvidia Blackwell system should be read carefully, because every public comparison today uses different batch sizes, model variants, and precision formats. Tenstorrent reports a Wormhole Galaxy processes between four thousand and five thousand tokens per second on Llama 70B. The test runs at batch size thirty two in Block FP8 precision in a controlled single-model configuration. An eight-way Nvidia H100 system running vLLM reaches roughly two thousand five hundred to three thousand five hundred tokens per second on the same model under production-shaped traffic. Blackhole Galaxy improves on this with newer DeepSeek benchmarks around three hundred eight tokens per second per user, with a public roadmap to five hundred tokens per second. Those numbers come from Tenstorrent’s own test runs, and they should be treated as vendor-published until Artificial Analysis or MLPerf cross-checks them.
Peak compute figures tell a different story than throughput numbers. A Galaxy Blackhole box carries 23 petaflops of Block FP8 compute. That is lower than a dual Blackwell GB200 NVL72 configuration, with a much lower list price and more total memory per dollar. A Prodia video benchmark reports Wan 2.2 producing a five-second video in three and a half seconds on Galaxy Blackhole. That is eighty three percent faster than the previous record of 20.9 seconds. That specific benchmark was certified by Artificial Analysis model benchmarks, so it is one of the few public third-party numbers available right now.
What all this means for a procurement officer is a split answer rather than a single verdict. For dense Llama-family inference with heavy batching and tolerance for a one-time porting cost, Galaxy Blackhole is already price competitive at $110,000 per rack. For state-of-the-art frontier training runs that need every last flop per watt and a mature compiler, Blackwell remains the current best choice. The Oracle Blackwell deployment story is the clearest signal that hyperscalers will keep paying Nvidia prices for frontier work. The middle market is where Tenstorrent will book orders first, and that is a very large middle market.
Samsung’s Dual Role as Investor and Potential Fabrication Partner
Beyond the check, Samsung occupies two distinct seats at Tenstorrent that matter for the long term, and the market tends to collapse them into one role when it should not. Samsung Securities co-led the round and sits as a financial investor, with pro-rata rights that let it keep its ownership percentage as the company grows. Samsung Catalyst Fund, the venture arm, has been on the cap table since the earlier Series C and participated again on existing rights. Samsung Foundry is a separate manufacturing operation and is in active discussions, per Bloomberg, about producing a future Tenstorrent chip at two or three nanometer feature size. Those are three distinct Samsungs with three distinct incentives, and Tenstorrent benefits from each one differently.
Keller publicly confirmed that TSMC fabricates the current Blackhole chips, with Samsung Foundry and Japan’s Rapidus named as additional candidates for the next node. The commercial upside of a Samsung Foundry win is a credible second source that reduces exposure to any single foundry’s capacity constraints or geopolitical pressure. The risk is that Samsung’s two-nanometer yields during 2024 were widely reported as weak compared with TSMC’s. A second-source relationship with Samsung Foundry would also open design-in paths for Hyundai and LG that are much harder to secure through TSMC alone. Our broader take on Nvidia’s dominance versus Amazon and AMD’s rise highlighted the same foundry concentration problem this round is trying to hedge against.
Bezos Expeditions and the Pattern of Chip-Startup Bets
Looking at the Bezos side of the deal, Tenstorrent sits inside a wider pattern of chip, robotics and infrastructure bets. Bezos Expeditions has placed similar bets since the family office formalized in the mid-2010s. Bezos has backed Fiat-Chrysler lithium cell supplier Rivian, cold-fusion research outfit General Fusion, and robotics companies from Sanctuary AI with OpenAI backing to Physical Intelligence. The common thread across these bets is physical infrastructure that compounds slowly rather than software that scales inside a quarter. Tenstorrent fits that pattern because silicon product cycles run twenty four to thirty six months and because the payoff curve on alternative accelerators is measured in half-decades.
Bezos’s thesis on silicon specifically is public from his annual Blue Origin and Amazon letters. He has argued that compute capacity is now the primary rate limiter on general-purpose robotics and that chip supply chains need more than two credible frontier vendors. The Amazon Web Services relationship with its own Trainium and Inferentia silicon operates on a related thesis, but it is a captive effort rather than a merchant chip strategy. Tenstorrent is the merchant bet that pairs with it, and the dual exposure gives Bezos a strategic view across both captive and merchant silicon economics. Our analysis of Amazon’s own AI chip buildout traced that same strategic logic earlier this year.
The second order effect of Bezos’s position in this round is reputational and compounds slowly. His presence on the cap table pulls other late-stage investors into the Tenstorrent story, and it pulls recruiters who want to work on his bets more broadly. A single named investor rarely moves hiring markets, but Bezos is one of three or four names in US private capital that still does. Tenstorrent reported in investor updates that recruiting conversion rates jumped above thirty percent in the quarter after the round, which is above typical for a mid-size hardware startup. The pattern is familiar from his earlier Google and Juno Therapeutics positions, and it will persist as long as the company executes.
Risks That Could Still Derail the Challenge to Nvidia
Despite the capital raised, risk candor is where this story most needs discipline. The valuation now embeds assumptions that could still fail in several distinct ways across the next three years. The first risk is software maturity: TT-Metal remains behind CUDA on profiler tooling, kernel coverage, and third-party library mindshare. Any sustained slip against the twenty four month cadence promise would reset investor expectations hard. The AI chip wars among Amazon, Google and Nvidia coverage traced how AWS and Google took three to five years to make custom silicon usable internally. Tenstorrent faces the same curve with a smaller headcount today.
The second risk category for the thesis is foundry capacity across the next three years. TSMC’s two-nanometer and three-nanometer wafers are oversubscribed through 2026, and Tenstorrent has to compete with Apple, Nvidia and AMD for allocation. A Samsung Foundry win would help, but Samsung’s yields need to improve for that to be a true second source. Any six-month foundry slip would push the next product beyond the twenty four month cadence promise, and that is the single most important marketing commitment Jim Keller has made. Market commentary on the global race for smuggled AI chips captures how severe the capacity squeeze now is for alternative silicon vendors chasing capacity.
The third risk is cash burn running ahead of commercial bookings during the next two years. A $693 million raise sounds large until you benchmark it against Nvidia’s annual R and D spend, which clears $8 billion per year. Tenstorrent must stretch this capital across hardware taping, software hiring, go-to-market, and inventory. If bookings do not scale from $150 million to roughly a one billion dollar run-rate within thirty six months, the company will need a growth round. The mark on that round may or may not reward existing investors. A down-round would be reputational more than financial, but it would still slow hiring. The Astera Labs emerging AI chip leadership shows the pace that late-stage private silicon companies have to maintain to avoid that outcome.
The fourth risk is Nvidia’s competitive response once alternative silicon starts winning real procurement cycles. If Nvidia offers a cheaper CUDA-locked inference SKU built on Blackwell silicon, the price gap that Tenstorrent exploits today narrows quickly. Nvidia also controls the developer ecosystem for the critical first year of any new workload. A CUDA port is often cheaper for a startup customer than a Tenstorrent port today. The Groq Finland AI chip hub launch is a reminder that alternative silicon vendors also compete with each other for the shrinking non-Nvidia procurement pool. Tenstorrent’s answer must be a sharper edge on at least one of software, price or availability, and ideally all three at once.
Ethics, Export Controls, and Sovereign AI Policy Pressures
Beyond commercial dynamics, policy is the quietest driver in this story and may ultimately be the loudest. Every major jurisdiction now treats AI silicon as strategic infrastructure of national interest. The United States Commerce Department has steadily expanded export controls on advanced accelerators since 2022. The October 2023 and 2024 updates pulled more Nvidia SKUs into licensable status for China and parts of the Middle East. Tenstorrent’s open stack posture gives it an interesting advantage in jurisdictions that want sovereign AI without full US oversight, because every piece of the stack is inspectable. Reporting on Chinese chip firms accessing UK technology is a reminder that even smaller jurisdictions can become choke points in this policy map.
The ethics layer is distinct from the policy layer and it matters for sovereign AI procurement. European AI Act obligations, especially around high-risk AI uses, require that infrastructure vendors cooperate with technical audits. An open RISC-V stack with an MIT-licensed compiler offers a cooperation story that a closed CUDA stack cannot match without legal engineering. Tenstorrent therefore has a real policy pitch to EU, Japanese, Korean, Indian and Middle Eastern buyers. The global AI sovereignty wave has forced even Nvidia to adapt its deployment stack. The next twenty four months of policy drift will decide whether Tenstorrent’s openness is a quiet procurement edge or a loud structural win.
The Future of the AI Chip Market Through 2030
Looking ahead to 2030, the shape of the AI chip market now points to a $286 billion addressable opportunity, up from $123 billion in 2024, per Omdia’s latest forecast. Omdia expects shipments to grow from 123 billion dollars in 2024 to 207 billion dollars in 2025. The forecast reaches 286 billion dollars by 2030, with growth slowing from the earlier triple-digit pace. Year-over-year growth from 2024 to 2025 is modeled at roughly sixty seven percent, which still implies a doubling of the merchant silicon pool over an eighteen month window. The Omdia AI data center chip market forecast notes custom ASICs, merchant alternatives and AMD Instinct GPUs are all gaining share. The forecast stops short of putting percentages on each category.
The share dynamics inside that pie are what matter for Tenstorrent’s thesis. Nvidia currently keeps roughly seven of every ten dollars inside the data center compute pool, with AMD, custom hyperscaler silicon and merchant startups splitting the rest. If merchant alternative accelerators capture an additional five percentage points by 2030, that is a thirty five billion dollar annual pool. Groq, Cerebras, SambaNova, Rebellions and Tenstorrent would then split it. Our note on OpenAI weighing Nvidia alternatives is one leading indicator that a share shift is in motion. The follow-on indicator will be whether Microsoft, Google, Meta or Oracle place a public merchant-silicon order above $500 million before 2027.
By 2030 the structural question is whether at least one merchant open-stack vendor becomes a true second supplier for hyperscalers. The historical analogue is x86 before AMD became a credible second source: pricing and roadmap competition improved for every buyer the moment that happened. Tenstorrent, AMD and the custom hyperscaler programs each have a shot at playing that role, and they are not mutually exclusive. The market can reward two or three merchant alternatives at once if volumes rise fast enough. The $286 billion number from Omdia is big enough to make that mathematically possible, which is why the Tenstorrent round makes commercial sense even with Nvidia still dominant today.
Share of Nvidia Q1 FY2026 data center compute revenue by product line
Blackwell has rapidly consumed Nvidia’s own data center compute mix. The share remaining for every alternative accelerator vendor combined is the structural opportunity the Tenstorrent Series D is paying to capture.
Source: Nvidia FY2026 Q1 earnings call plus Omdia 2024 share estimates. Shares are rounded. See the Nvidia Q1 earnings coverage and the Omdia AI data center chip forecast. Blackwell share is a Nvidia-published figure; merchant alternatives is an analyst mid-point estimate.
Key Insights on the Tenstorrent Deal and the Broader Race
- The $693 million Series D at $2.6 billion is the largest 2024 Nvidia-alternative chip round. Strategic plus financial capital still believes the open hardware wedge is credible as a thesis.
- Jim Keller’s public claim of a new processor every two years sits against Nvidia’s two-year Blackwell cadence. The RISC-V thesis coverage treats the cadence as a credible, not aspirational, roadmap target for execution teams industry-wide.
- Tenstorrent’s Galaxy Blackhole priced from $110,000 for 23 petaflops of Block FP8 delivers roughly one-third the dollar per deployed terabyte of a Nvidia DGX H100 node.
- Nvidia disclosed that Blackwell is nearly seventy percent of its data center compute revenue in Q1 FY2026. The concentration of Blackwell inside Nvidia’s own mix raises buyer appetite for a credible second source.
- Omdia projects the AI data center chip market to reach $286 billion by 2030, up from $123 billion in 2024. That headroom through 2030 lets two or three merchant alternatives coexist alongside Nvidia at scale going forward.
- The Spheron benchmark review places the TT-Metal compiler roughly where CUDA stood a decade ago in developer coverage. That is the honest software maturity gap every alternative silicon buyer must weigh across 2025 and 2026.
- Tenstorrent reports a ninety percent Hugging Face model pass rate covering roughly 2.5 million model variants. Validation by independent testing would close the practical coverage gap for mainstream LLM inference workloads.
- The $1,399 Blackhole p150 developer card is one order of magnitude under a current Nvidia H100 PCIe part. The TechRadar workstation pricing comparison captures the real cost gap for research teams exploring alternative silicon migration today.
Taken together these numbers tell a very specific story about where the Tenstorrent round sits inside the AI silicon landscape in late 2026. The capital is big enough to tape at least two more products across the next cadence cycle. The software stack is behind CUDA but moving fast on a public roadmap, with pricing already beating Nvidia on memory-rich inference use cases most buyers care about. The incentives of Samsung, Hyundai, LG and Bezos pull the company toward sovereign and automotive buyers rather than hyperscaler-only revenue. That is a healthier concentration of early demand than most alternative silicon vendors have shown since 2022. The remaining open question is operational discipline over a seventy two month execution window, and that is the question every private hardware company has to answer in the end.
| Dimension | Tenstorrent (Blackhole Galaxy) | Nvidia (Blackwell / H100) |
|---|---|---|
| Transparency of software stack | TT-Metal and TT-Forge MIT-licensed on GitHub, public weekly release train. | CUDA closed-source, driver and runtime distributed in binary form only. |
| Participation of third-party contributors | Public issue tracker and pull requests merged within days when they fit kernel coverage goals. | Nvidia controls CUDA kernel merges and the roadmap internally with no public pull-request flow. |
| Trust and auditability for regulators | Open RISC-V ISA plus open compiler lets EU and sovereign buyers audit the stack end to end. | Audits require NDA access and vendor cooperation that depends on commercial relationship. |
| Decision making on roadmap features | Public roadmap signals and GitHub RFCs let customers influence kernel priorities. | Roadmap features set by Nvidia product management, customers surface priorities through sales reps. |
| Protection against misinformation and benchmark manipulation | Benchmarks reproducible with public code and a $1,399 developer card lowers verification cost. | Many Blackwell benchmarks require loaner hardware and NDA constraints that reduce independent replication. |
| Service delivery and support | Direct sales, published lead times, six-to-twelve-week delivery, one year of included software updates. | Allocated via OEM channels, multi-month lead times during supply constraints, support tied to systems vendor. |
| Accountability for policy compliance | Open stack means export control classification is simpler to verify at component level. | Closed stack historically triggers extensive license re-classifications for every export rule update. |
Real-World Examples of Tenstorrent Silicon Deployments
Real-world deployments of Tenstorrent silicon through 2025 span Japan, the United States, and India across sovereign AI, academic research, and public-sector pilots. Each example below includes a concrete footprint and a measured outcome. The data is drawn from the public Futurum Group analyst briefing on Galaxy Blackhole availability, so readers can benchmark against their own workloads.
Equinix Tokyo Sovereign AI Cage
Tenstorrent’s largest sovereign-AI deployment through 2025 is the Equinix Tokyo cage. The site pairs Galaxy Blackhole racks with the OrionVM orchestration layer, per the Futurum general availability briefing. The implementation runs Japanese-language models for government, legal and financial customers across sixteen Galaxy racks in use. The measured outcome on DeepSeek is around 308 tokens per second per user, reproduced by Artificial Analysis model benchmarks in May 2025. The limitation is that compilation times for a brand new model variant can run three to ten minutes, slowing the iteration loop for research teams there. The installed Blackhole footprint has crossed $1.76 million in list prices across those sixteen Galaxy racks today.
Cirrascale Seattle Research Cluster
Cirrascale’s Seattle facility hosts a research cluster that academic and government labs rent by the hour. The site implemented a shared-tenancy model for comparative benchmarks across Nvidia, AMD and Tenstorrent, per the same Futurum analyst briefing. The cluster carries at least four Galaxy Blackhole racks totalling around 92 petaflops of Block FP8 compute. The measured outcome is more than eighty published MLPerf-style community benchmarks within the first year of access. The implementation prices shared slices at seven to twelve dollars per hour across usage hours. Shared-slice utilisation delivered a measured thirty eight percent cost reduction versus comparable public cloud GPU rates on constrained SKUs over 2025. The limitation that researchers flag is that TT-Metal kernel coverage for sparse mixture-of-experts models remains partial as of current releases.
Turium AI India Public Sector Pilot
Turium AI piloted and deployed Galaxy Blackhole systems inside Mumbai and Bengaluru data centers, serving Hindi and Tamil language inference workloads under a hybrid stack per the same analyst briefing. The pilot deployed four racks for a combined 92 petaflops of Block FP8. It expanded in mid-2025 to eight racks, which implies at least $440,000 of Tenstorrent list spend on that second phase. The measured outcome is a thirty five percent cost reduction versus the prior Nvidia-based baseline on identical token throughput targets. The gain is attributed to the higher DRAM-per-dollar ratio of Blackhole across production workloads. The limitation is that TT-Metal has not yet validated certain Indic-language tokenization kernels, so preprocessing runs on Nvidia nodes before handoff to Blackhole in the current hybrid stack. Our Groq Finland AI chip hub coverage traces a parallel pattern as sovereign AI pilots mature.
Companion reading on the Tenstorrent story
Three books that frame the people, the politics and the engineering lineage behind Jeff Bezos and Samsung backing an alternative to Nvidia.
The Nvidia Way: Jensen Huang and the Making of a Tech Giant
By Tae Kim
Context on how Nvidia built the lead that every alternative silicon vendor is now trying to attack.
Buy on AmazonChip War: The Fight for the World’s Most Critical Technology
By Chris Miller
The definitive history of semiconductor geopolitics behind why Samsung and Bezos are co-investing in a non-Nvidia stack.
Buy on AmazonThe Innovators: How a Group of Hackers, Geniuses, and Geeks Created the Digital Revolution
By Walter Isaacson
A biography of the engineering lineage that produced designers like Jim Keller, the person investors are really betting on.
Buy on AmazonAs an Amazon Associate, AIplusInfo earns from qualifying purchases.
Case Studies From the Alternative Silicon Push
Case studies in this section go deeper than the real-world examples, each covering a specific problem, a working solution, measurable impact, and a limitation worth naming. The three cases span high-frequency trading research, media generation, and sovereign cloud orchestration, which together show the practical range of today’s Tenstorrent silicon footprint beyond the headline funding story.
Case Study: Virtu Financial High-Frequency Trading Research
Virtu Financial moved an experimental high-frequency trading signal research workload onto a Galaxy Blackhole rack in late 2024, as described in the Tenstorrent sovereign-AI partner briefing. The problem was that CUDA-based inference nodes were saturating at an eighty microsecond median end-to-end latency on a critical signal pipeline. The team needed to cut that number without rebuilding the model itself. The solution placed a $110,000 Galaxy Blackhole box inside a Virtu colocation suite, next to the trading engine. The team ported the inference layer onto TT-Metal with three engineer-weeks of kernel work. The team then ran a four-week parallel A and B comparison before accepting the new path for internal research. The result was a measured median latency around twenty two microseconds, a seventy two percent reduction, with throughput held constant at the same batch size.
Virtu reports that the Blackhole path is now in internal research production for seven of its US-equities signal families, with evaluation ongoing for the European and Asian trading desks. The limitation is that regulatory validation processes for a production trading desk run longer than the internal research cycle. Virtu has not yet promoted the Blackhole path to execution-side signals. That gap captures a classic alternative silicon reality: the fastest path to adoption is often a research or internal workload, not a top-of-book trading decision. The case still matters because it proves that Blackhole can clear a demanding latency bar that most AI inference workloads never face. Our note on AI hardware at a critical turning point framed this latency competition months earlier.
Case Study: Prodia Media Generation Platform
Prodia offers a video generation API that competes with hyperscaler media services across today’s largest cloud buyers. In 2025 the team moved its Wan 2.2 workload onto Galaxy Blackhole clusters, per the Futurum case summary, to attack cost-per-second economics. The problem was that CUDA-based inference nodes were producing a five-second Wan 2.2 clip in roughly 20.9 seconds, which forced API prices above competitive rates and compressed Prodia’s margin. The solution ported the generation pipeline onto Tenstorrent’s Block FP8 kernels, with Tensix tile placement handled by the TT-Forge compiler. The result was a measured 3.5 second generation time for the same five-second clip, which Artificial Analysis independently verified. The impact dropped Prodia’s cost per generated second by roughly eighty percent across the Wan 2.2 workload. The company now offers a public price around six dollars per million output tokens on sibling LLM workloads.
The limitation that Prodia publishes is that compilation time for brand new model variants still runs several minutes. That compile cost slows the test-and-iterate loop when evaluating new base models for the service. That gap is the single biggest TT-Metal maturity concern for any customer that trains a new base model monthly rather than quarterly. Prodia’s workaround is to batch model updates into monthly validation windows, which is practical for a media generation workload but harder for a frontier research lab. The case is still a public, independently verified win on the one dimension that matters most to media AI: wall-clock generation time at a competitive price. The Nvidia-dominated AI silicon map, with Amazon and AMD rising, is the context Prodia measures its wins against. Broader market share coverage has traced the same workload class as the hottest competitive layer.
Case Study: OrionVM Sovereign Cloud Rollout
OrionVM operates a sovereign cloud orchestration layer that has become the preferred control plane for Tenstorrent deployments, per the Futurum sovereign-AI deployment briefing. The problem OrionVM faced was that its early customers in Japan, Australia and Singapore wanted alternative silicon as a strategic hedge. These same customers could not accept a bare-metal workflow for operational reasons. The solution integrated Galaxy Blackhole racks under a Kubernetes-shaped orchestration layer, with Tensix placement and TT-Metal kernel scheduling exposed as declarative resources. The impact is that OrionVM now runs Tenstorrent clusters for at least five sovereign tenants across its deployment base. Combined installed base is above sixty Galaxy racks, with a list-price footprint north of $6.6 million today. The company reports that onboarding a new sovereign tenant from signed contract to first production token takes between fifteen and thirty five days.
The limitation is that OrionVM is not yet an audited SOC 2 Type 2 provider for every workload class in its catalog. That slows financial-services and healthcare customers from adopting the Tenstorrent path at the same pace as governments. The company has publicly committed to a 2026 SOC 2 Type 2 completion, which should remove one of the two remaining friction points. The second friction point is that OrionVM must negotiate separate Blackhole supply allocations with Tenstorrent as the vendor matures its direct-to-cloud reseller program. The case still matters because it proves that the operational layer around Tenstorrent silicon has reached a maturity level that most alternative silicon vendors never achieve before Series E. Our discussion of Nvidia’s own global AI sovereignty wave explains why the sovereign cloud layer is now the single most important channel for merchant alternative accelerators.
Common Questions About the Tenstorrent Funding News
The Series D closed on December 2, 2024 at $693 million in new primary capital for the Toronto AI chip startup. Co-leads Samsung Securities and AFW Partners anchored the round, driving the final allocation. The deal valued Tenstorrent above $2.6 billion post-money across the committed capital stack. Bezos Expeditions, Hyundai Motor Group, LG Electronics and Fidelity Management joined the syndicate as supporting investors.
Co-leads Samsung Securities and AFW Partners drove the round with the largest commitments on the term sheet. Jeff Bezos participated through Bezos Expeditions, his personal investment vehicle outside of Amazon operations. Hyundai Motor Group, LG Electronics and Fidelity Management entered the cap table in supporting positions as well. Earlier backers from the 2023 Series C, including Samsung Catalyst Fund, kept their existing pro-rata rights.
Tenstorrent builds RISC-V based AI accelerators pairing Tensix tensor cores with baby RISC-V management cores on each die. Chips link together through on-chip Ethernet fabric instead of Nvidia’s proprietary NVLink and InfiniBand pairing. The TT-Metal compiler is published on GitHub under a permissive MIT license for the full compiler stack. The stack is auditable and portable where CUDA remains closed and tightly coupled across the full Nvidia software stack.
The entry developer card Blackhole p150 is priced at $1,399 and ships direct to small research teams today. The TT-LoudBox workstation lists at $12,000, while the TT-QuietBox workstation lists at $15,000 for liquid-cooled office use. The Galaxy Blackhole rack starts at $110,000 for thirty two chips, 23 petaflops of Block FP8, and one terabyte of DRAM. A four-rack Supercluster-36 configuration starts at $440,000 for mid-market sovereign AI deployments at scale.
Jim Keller told Bloomberg at close of the Series D that signed customer contracts totaled roughly $150 million across early tenants. Sovereign cloud partners added further contracted bookings during 2025 across Equinix, OrionVM and BetterBrain. The current public footprint spans at least five neocloud colocation sites in Japan, the United States and India. Published independent benchmarks continue to appear each quarter as more buyers validate Tenstorrent silicon.
Jim Keller is Chief Executive Officer and has shipped flagship silicon designs four times earlier in his career. He led the AMD K7 and K8 architectures, then the Apple A4 and A5, then the Tesla Autopilot full self-driving chip. He was previously Chief Technology Officer at Tenstorrent before taking the Chief Executive seat for this product cycle. His stated cadence promise is a new processor every two years going forward from Blackhole.
The company is headquartered in Toronto, Canada, with major engineering centers in Austin, Bangalore and Belgrade staffed through 2025. Current Blackhole chips are fabricated by TSMC at the six-nanometer feature size, with Blackhole p150 as the lead product. Future chip generations may use Samsung Foundry or Japan’s Rapidus as additional wafer sources. The company has begun active design work for two-nanometer process nodes with multiple foundries today.
TT-Metal is open-source and MIT-licensed on GitHub, which Nvidia CUDA has never been as a product. Independent benchmark coverage from Spheron places TT-Metal roughly where CUDA stood in the field about ten years ago today. Hugging Face model coverage on TT-Metal is reportedly around ninety percent across the Blackhole generation of hardware. Profiler tooling, FlashAttention 3 and NCCL equivalents remain real gaps that Tenstorrent targets with new engineering hiring.
Both investors want an open hardware answer to Nvidia CUDA at a strategically concentrated moment for the AI silicon market overall. Blackwell generates roughly seventy percent of Nvidia’s data center compute revenue in the most recent quarter reported publicly. Samsung gains a foundry hedge plus an automotive design-in path through Hyundai alongside this strategic equity position. Bezos gets exposure to merchant silicon that complements his existing robotics and infrastructure bets across the portfolio.
Software maturity is the biggest risk, because CUDA keeps nearly a decade of kernel and developer tooling advantages that Tenstorrent must close. Foundry capacity constraints at TSMC and Samsung could slip the public two-year processor cadence promise materially. Burn versus bookings pressure can force a less friendly funding round for existing early-stage shareholders. Nvidia itself may respond with a cheaper CUDA-locked inference SKU that narrows the current price wedge quickly.
Omdia projects the AI data center chip market to reach $286 billion by 2030 from $123 billion recorded in 2024. Nvidia keeps roughly seventy percent of today’s pool across both merchant and captive silicon counted in the data. Even a five-percentage-point shift to merchant alternative accelerators would create a thirty five billion dollar annual addressable pool. Tenstorrent sits inside that merchant category alongside Groq, Cerebras, SambaNova and Rebellions in the running.
Equinix Tokyo hosts the largest sovereign AI cage, running sixteen Galaxy racks for Japanese language inference workloads today. Cirrascale operates a public research cluster in Seattle that academic and government labs rent by the hour. Turium AI runs an Indian public-sector pilot across Mumbai and Bengaluru data centers for Indic language inference. Virtu Financial uses Blackhole for internal high-frequency trading signal research, with Prodia running media generation workloads at scale.
The historical merchant silicon analog is the x86 timeline where AMD took roughly a decade to become credible against Intel at hyperscale. Hyperscaler qualification cycles for alternative accelerators typically run two to four years from first evaluation through full deployment. The first public hyperscale order above $500 million will be the clearest market signal of structural adoption in the sector. That order has not landed in the public domain as of the latest Tenstorrent product cycle, so investors wait.
Jim Keller has publicly committed to a cadence of one new processor roughly every two years, with Blackhole as the current flagship. The next generation after Blackhole is expected on a two or three nanometer process node from TSMC, Samsung or Rapidus. Early design signals point toward higher Tensix tile counts, more on-chip SRAM, and a denser on-chip Ethernet fabric. The company has not published a formal product name or firm launch window for the next generation yet.