Tenstorrent

Phase A — Understand the business

Company Overview

What it is. Tenstorrent is a Toronto-founded (2016), Jim-Keller-led AI-compute company that sells AI accelerators and licenses the underlying silicon IP — a deliberate two-engine model. Founded by Ljubisa Bajic, Milos Trajkovic and Ivan Hamer; Keller joined as President/CTO in Jan 2021 and became CEO in Jan 2023. Headcount ~1,000 after a Nov-2025 reduction.

How it makes money — three stacked layers on one IP base (Tensix AI cores + RISC-V CPUs + chiplet interconnect + open-source compiler):

1. Cards / workstations (transactional, self-serve): Blackhole PCIe cards — p100a $999 (28 GB GDDR6), p150a $1,399 (120 Tensix cores, 32 GB GDDR6, 300 W), QuietBox 4-card workstation $11,999.
2. Systems (enterprise, sales-led): Galaxy Blackhole server $110,000 (32 Blackhole chips, 23 PFLOPS dense FP8, 1 TB GDDR6, 16 TB/s, 56×800G Ethernet); Galaxy Supercluster $440,000 (4 systems), scaling to 144 nodes / >4,000 chips.
3. IP licensing (high-touch B2B — the profit engine): licenses the RTL source code for the TT-Ascalon RISC-V CPU and Tensix AI cores to OEMs, automakers, foundries and sovereign programs — customers own and customize the silicon. Crucially, this is source-code licensing, unlike Arm's black-box model. Most of Tenstorrent's bookings to date come from IP deals, not hardware.

Key products. Hardware generations Grayskull → Wormhole → Blackhole (current, GA 28 Apr 2026). IP: TT-Ascalon (RVA23-compliant, 64-bit out-of-order superscalar, 256-bit vector), TT-Ascalon S (compute-dense core for agentic AI, launched at TT-Deploy JP, Jun 2026), Tensix Neo AI cores, with Babylon (next-gen Ascalon, ~18-month cadence, higher IPC) in development. Software: TT-Metalium (MIT-licensed, open-source, "abstraction-level comparable to CUDA/OpenCL"), TT-Buda/TT-Forge; ~70 models running on Blackhole-gen hardware.

Customers/suppliers/competitors. Named licensees/partners: Samsung (foundry + IP — chiplets in production at Taylor, TX on SF4X 4 nm), LG Electronics (RISC-V AI in smart TVs/appliances), Hyundai Motor Group (automotive AI — also strategic investor), Rapidus (Japan 2 nm IP co-development), plus Japan's TT-Deploy JP ecosystem (Preferred Networks, Socionext, Turing, ai&). Suppliers: Samsung Foundry + GlobalFoundries (fabrication), GDDR6 memory vendors. Competitors below (Lens 3/12/13): Nvidia (CUDA), AMD, Arm/SiFive (IP), Groq/Cerebras/SambaNova/d-Matrix/Etched (accelerator startups), Qualcomm (both rival and suitor).

Contract structure. IP licensing = upfront license fee + per-chip/per-unit royalties (recurring, high-margin once shipped); hardware = one-time. Reported ~$150M in customer contracts — lumpy, milestone-driven, and concentrated in a handful of strategic names. This is the model's central tension: royalty streams are the prize, but they lag design-in by years.

Supply Chain

Fabless designer sitting mid-chain; names along the flow:

Upstream inputs → Tenstorrent → end customer:

• EDA/IP: Standard RISC-V ISA (open, royalty-free — the strategic core; removes Arm/x86 licensing dependency), internal RTL, third-party EDA (Synopsys/Cadence implied).
• Foundry (chokepoint): Samsung Foundry (SF4X 4 nm at Taylor, TX; investor + partner) and GlobalFoundries (mature nodes); Rapidus for future 2 nm in Japan. Diversified away from TSMC — a deliberate hedge, and a differentiator vs Nvidia/AMD's TSMC-CoWoS dependence.
• Memory: GDDR6 (not HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips.) — a cost/availability choice that dodges the HBM + CoWoS advanced-packaging bottleneck binding Nvidia/AMD. Trade-off: lower bandwidth ceiling than HBM3e parts.
• Interconnect: standard Ethernet (800G) for scale-out instead of proprietary NVLink/InfiniBand — the "use open standards" bet, again reducing single-vendor dependence.
• Downstream: cards (self-serve/Cirrascale cloud), Galaxy systems (enterprise), and IP licensees who take the design into their own fabs (Samsung, LG, Hyundai, Rapidus, sovereign programs in Japan/Korea/India/UAE).

Single-source / chokepoint read: Tenstorrent's supply chain is structurally less exposed to the two industry chokepoints (CoWoS advanced packaging, HBM allocation) than the GPU incumbents — by design. Its dependency risk is instead foundry execution at Samsung/Rapidus (Samsung's yield history is mixed; Rapidus 2 nm is unproven) and export-control exposure on IP flows to China (Lens 10). Names present; chokepoints marked.

Competitive Advantages (moats)

The thesis in one line: Tenstorrent is the most complete open-stack counter-position to the three reinforcing moats that lock in the incumbents — CUDA (software), Arm/x86 (ISA), NVLink/InfiniBand (interconnect).

• The contrarian design axiom — Keller: "Let's just make a list of what Nvidia does, and we'll do the opposite". Open ISA (RISC-V), open MIT-licensed compiler (TT-Metalium), explicit data movement instead of a hardware cache hierarchy, Ethernet instead of NVLink, GDDR6 instead of HBM, and RTL source-licensing instead of black-box IP.
• Durable moat candidates:
• Licensing model + "silicon sovereignty." The genuinely differentiated asset. Selling RTL so a sovereign/OEM can own its silicon rides the strongest secular tailwind in compute (sovereign AI, supply-chain de-risking, anti-Arm-lock-in). This is a direct assault on Arm Neoverse's terms.
• Jim Keller as the moat. Recruiting/credibility flywheel — the one asset a well-funded rival cannot copy (Lens 9).
• Open-source community as a distribution wedge (the developer pivot) — auditable, no proprietary microcode/driver; attractive to sovereign buyers who distrust a US black box.
• Where the moat is thin (be honest): Software maturity is not a moat yet — it is the liability. TT-Metalium has no Nsight-equivalent profiler; real-world utilization runs 40–60% of theoretical TFLOPS vs 60–80% for mature Nvidia stacks. CUDA's 20-year ecosystem (cuDNN, NCCL, TensorRT) is the incumbent moat, and it is not closing on Tenstorrent's current timeline.
• Bargaining power. Weak today over customers (challenger selling into a Nvidia-default market; must discount hard on TCO — claims $6 vs ~$30 token cost, 5× TCO vs GB300 ). Improving over suppliers by design (foundry-diversified, no HBM/CoWoS hostage). Net: bargaining power is a future asset contingent on software + design-win conversion, not a present one.

Segments (revenue by product & geography)

our figures is empty — no research-layer segment data exists. Directional, web-only, unaudited:

• By product line: No official split disclosed. Management + trade press say IP licensing is the majority of bookings to date, with hardware (cards/Galaxy) still ramping post-Blackhole-GA (Apr 2026). The Nov-2025 pivot toward individual developers (away from enterprise sales) is a deliberate near-term de-emphasis of hardware enterprise revenue to build ecosystem.
• By geography: Heavily Asia-weighted demand — Korea (Samsung, LG, Hyundai), Japan (Rapidus, Preferred Networks, Socionext, Turing, sovereign program), plus disclosed datacenter deployments in India, UAE and China. R&D footprint: Toronto (HQ), US (Austin/Santa Clara/Bay Area/Seattle), Japan (new CPU/R&D team), India, and a China push via CoreLab/Allen Wu.
• Trend & cause: The center of gravity is licensing + sovereign/Asian OEM, accelerating as silicon-sovereignty demand rises; US enterprise hardware sales are the decelerating/de-prioritized leg (developer pivot + layoffs). n/a — segment economics not disclosed for margins by line.

---

Phase B — Measure performance (+private: funding, traction, cap table)

Funding & Valuation Trajectory (+private swap)

The valuation curve is the single clearest signal here — a ~3.2× mark-up in ~18 months, then a ~2.5–3× M&A re-rate on top:

*Source labeling of the 2021 round varies (Sacra: "Series B"; others: "Series C"). Total raised >$1.18B across ~10 rounds.

The M&A overlay (the live catalyst):

• 18 May 2026 — Bloomberg: early takeover interest from both Intel and Qualcomm.
• 15 Jun 2026 — Reuters/The Information: Qualcomm in talks to acquire for $8–10B.
• 30 Jun 2026 — Keller publicly denies active Qualcomm acquisition talks; says a strategic deal / JV is more likely than acquisition by a GPU company, and "investors are very hot on IPO".

Burn signal: ~$1.18B raised, ~1,000 staff, multi-node tape-outs, a Nov-2025 7.5% layoff — burn is high and the developer pivot + cuts read as runway/efficiency management ahead of the next raise or exit. Conflict surfaced honestly: the $8–10B M&A chatter and Keller's denial coexist — treat the M&A number as an interest signal, not a struck price.

Founder & Ecosystem Signal (+private: founder interviews replace earnings calls)

No earnings calls exist. Reading Keller's public posture across 2025–26 interviews:

• Consistent narrative, escalating confidence. "Do the opposite of Nvidia," "AI still obeys the old laws of compute," and — post-Cerebras-IPO — "set to beat them on everything." Tone has shifted from underdog to challenger-with-momentum.
• What he keeps saying: open stack, RISC-V sovereignty, licensing > box-selling ("enterprise sales will ultimately be more profitable"), IPO ambition, foundry/geographic diversification.
• What changed: explicit acknowledgment of software immaturity and a strategic pivot to developers first; open confirmation he's met Intel & Qualcomm CEOs "hoping to get a big deal" — while denying a sale. Reads as a founder optimizing for either a premium strategic partnership or independence-to-IPO, keeping both doors open.
• Signal quality: high credibility on silicon roadmap; promotional on TCO/perf claims that lack independent production-scale verification.

Cap Table & Comps (secondary/M&A marks) (+private swap)

Cap-table quality — a strong IPO-proximity syndicate:

• Crossover funds (the IPO tell): Fidelity (lead of the latest round — a classic pre-IPO crossover signal), Baillie Gifford — both are late-stage/public-market investors whose entry historically precedes an IPO by 12–24 months.
• Strategics: Samsung (Securities + Catalyst Fund + foundry), Hyundai Motor Group, LG Electronics, XTX Markets — customer-investors that also de-risk demand.
• Marquee/other: Bezos Expeditions, Eclipse Ventures, Real Ventures, AFW Partners, Corner Capital, MESH, Export Development Canada, Healthcare of Ontario Pension Plan.
• Read: a syndicate built for an exit — strategic buyers who could acquire, plus crossover funds who underwrite an IPO. This is why the M&A interest is credible.

Comps — by model/mechanism, not P/E (private accelerator + IP-licensing peers). Multiples that aren't sourced are marked n/a:

Implied revenue multiple (heavily caveated ): on ~$150M bookings (not recognized revenue, unaudited), the ~$3.2B mark is ~21× and the $8–10B M&A range is ~53–67×. GetLatka's "$501.6M ARR" figure is almost certainly auto-scraped and unreliable — do not use; a ~$320–370M 2025 run-rate appears in one low-quality source and conflicts with the ~$150M contracts figure — surfaced, not reconciled. Bottom line: the valuation prices a future royalty franchise, not today's P&L.

Catalysts (funding & product events that re-rated the name) (+private swap)

Events that moved perception/valuation over the cycle:

• Jan 2023 — Keller becomes CEO → credibility re-rate.
• Aug 2023 — Hyundai/Samsung $100M strategic up-round → demand validation.
• Dec 2024 — $693M Series D at $2.6B → scale capital.
• 2025 — Ascalon productized as licensable IP; Rapidus 2 nm partnership; developer pivot + 7.5% layoff (Nov); former Arm-China CEO Allen Wu joins the China push (Dec) → strategy sharpens, execution risk flagged.
• Nov 2025 — reported ~$800M Fidelity round at ~$3.2B → crossover mark.
• 28 Apr 2026 — Blackhole GA + Galaxy servers ship → first real hardware commercialization.
• 14 May 2026 — Cerebras IPO (~$40B) → re-rates the entire non-GPU accelerator class, Tenstorrent included.
• 18 May → 15 Jun 2026 — Intel/Qualcomm takeover interest → Qualcomm $8–10B talks → the dominant catalyst; 30 Jun Keller denial.
• Jun 2026 — TT-Ascalon S launch + Japan (TT-Deploy JP) expansion.

Pattern: this name re-rates on (1) strategic-investor validation, (2) IP/licensing milestones, and (3) M&A/comparable exits — far more than on raw benchmark wins. The market is pricing optionality and scarcity (the last independent open/RISC-V challenger), not shipped units.

Traction & unit economics (+private add-on). Bookings ~$150M (majority IP); ~70 models on Blackhole; Cirrascale cloud availability (May'26); marquee design-ins (Samsung/LG/Hyundai/Rapidus). Unit economics claimed strong on TCO ($6 vs ~$30 token cost; 5× vs GB300) but not independently verified at production scale — the single most important unproven number in the file.

---

Phase C — Judge people & books

Management (founder archetype)

• Jim Keller (CEO) — arguably the strongest single-operator asset in the challenger field. Track record, quantified: lead architect AMD K7/K8 (Athlon 64), Apple A4/A5 (the chips that launched the iPhone/iPad silicon era), AMD Zen (the architecture that resurrected AMD vs Intel), Tesla Autopilot HW, Intel SVP (2018–2020). He is the rare architect who has shipped category-defining silicon at four different companies. Joined Tenstorrent 2021, CEO since Jan 2023.
• Keith Witek (COO) — ex-Google, ex-SiFive (directly relevant: RISC-V IP-licensing GTM), ex-Tesla, ex-AMD; long history with Keller. The SiFive pedigree matters — it's the exact licensing motion Tenstorrent is scaling.
• Founders Bajic/Trajkovic/Hamer built the original Tensix architecture; Bajic stepped back from the CEO seat when Keller took over (current involvement not verified — n/a).
• Skin in the game / insider ownership: not disclosed (private) — n/a — not disclosed.
• Capital-allocation history: disciplined-under-pressure — foundry/geographic diversification, an open-source developer wedge, and a 7.5% layoff + enterprise-to-developer pivot that reads as rational focus, not distress. The risk is the opposite of empire-building: whether a research-led org can build the boring enterprise software/support motion that converts design-ins to royalties.
• Founder archetype: visionary technical founder-operator (Keller), paired with a licensing-GTM operator (Witek). Ideal for a deep-tech IP/roadmap company; the open question is commercial/enterprise execution, historically not Keller's core strength.
• Red flags: promotional TCO/benchmark claims; a firmware core-count downgrade (Blackhole p150 cut 140→120 Tensix cores post-sale, ~1–2% perf drop) — minor, but a QA/marketing-integrity ding; mixed Glassdoor sentiment around the layoffs.

Forensic Red Flags + Regulatory

Accounting/forensics: As a private company with no audited public financials, standard forensic ratio work (accruals, receivables-vs-revenue, SBC-flattered non-GAAP) cannot be performed — no statements exist. This opacity is itself the primary "red flag" for anyone marking the position: the ~$150M bookings, run-rate and TCO claims are management-sourced and unverifiable. Booking-quality risk is real: IP-license revenue is lumpy and milestone-driven, and "contracts" ≠ recognized recurring royalties.

Regulatory findings (per regulatory/regulatory-findings.md, generated 2026-07-10):

• SEC (EDGAR LR + AAER): total_sec_findings: 0 — Tenstorrent has no CIK, is not an SEC registrant, and no enforcement search is possible. No litigation releases or AAERs.
• Non-SEC / web search ("Tenstorrent" (FTC OR DOJ OR export controls OR sanctions) enforcement): no enforcement actions, fines, or consent decrees found. The material regulatory exposure is forward-looking, not historical:
• Export-control / geopolitics (the key risk). Tenstorrent is actively pushing RISC-V into China via CoreLab and former Arm-China CEO Allen Wu (Dec 2025), and discloses datacenter customers deploying in China, UAE, India, South Korea. US–China IP-flow and advanced-compute export controls could restrict licensees, packaging, or customers. Tenstorrent has filed a formal response to the US "American AI Exports Program" RFI — i.e. it is engaging policymakers directly, which cuts both ways (proactive, but signals the exposure is live).
• RISC-V's appeal as a sovereignty hedge (no US/UK license dependency) is precisely what makes its China footprint a policy target.
• Verdict: No material historical regulatory/legal findings — verified via SEC EDGAR EFTS (LR, AAER, 0 findings), web enforcement search, and the absence of any public 10-KForm 10-KA company’s audited annual report to the US regulator. The most complete thing it publishes. Item 3 (no filer), as of 2026-07-10. Forward export-control risk is the standout, and it is structural.

---

Phase D — Project & stress-test

IPO / Exit Readiness & Path-to-Tradeable (+private swap)

Estimated readiness: 4/5 on the private-watch scale (4 = pre-IPO / secondary-active). Note: Tenstorrent is NOT currently in research/private-watch.json — it should be added (proposed entry in Open items / handoff below); I did not mutate that shared file in this unattended run.

Two live, competing paths to a tradeable mark:

• Path 1 — IPO. Crossover investors (Fidelity lead, Baillie Gifford) are in; Keller: "investors are very hot on IPO"; Cerebras's May-2026 IPO (~$40B) is the proof-of-exit for the class. No S-1 filed; no official timeline — 2027+ is plausible but speculative. Milestones that unlock an S-1: audited financials + recurring royalty revenue (not just bookings), software-maturity credibility, and a demonstrable enterprise motion.
• Path 2 — Strategic acquisition. Qualcomm $8–10B interest (Reuters, Jun'26) + prior Intel interest; Keller denies a sale but courts a "big deal." Strategic logic is strong: a buyer gets a licensable IP block + a credible non-Arm CPU roadmap + Keller's team in one transaction. Intel needs an anti-Nvidia repositioning post-Gaudi; Qualcomm wants data-center + a RISC-V hedge.

No our model logged (per --watchlist rule — and there's no EPS/binary to score; the resolvable event is IPO-or-acquisition, not an earnings line).

Path-to-tradeable read: the realistic near-term resolution is M&A at a premium (highest-probability catalyst on the tape right now), with an IPO as the independent-path alternative if Keller holds out and royalty revenue matures. Either way the value crystallizes on the IP/team, not standalone unit economics.

Bull vs Bear

Bull case. Tenstorrent is the only company assembling a full open-stack alternative to the CUDA/Arm/NVLink triple-moat, led by the most credible silicon architect alive, riding the strongest secular tailwind in compute (sovereign AI + silicon sovereignty + anti-Nvidia diversification). The RTL-licensing model is genuinely differentiated — sell the source, let Samsung/LG/Hyundai/Japan/India own their silicon — and licensing is where the durable, high-margin royalty economics live. Marquee design-ins are real; the cap table is exit-built; and $8–10B takeover interest plus a ~$40B Cerebras IPO validate both the scarcity value and the exit path. If even a fraction of the design-ins convert to volume royalties, a ~$3.2B mark is cheap versus the option value.

Bear case (2–3 permanent-impairment risks).

1. Software never catches CUDA on the timeline that matters. 40–60% real-world utilization, no mature profiler, ~70 models — the ecosystem gap is the whole game, and it compounds in Nvidia's favor. A great chip with an immature stack is a demo, not a franchise.
2. The bookings don't become royalties. ~$150M of lumpy IP contracts is not a recurring-revenue business yet; design-ins can stall, and the enterprise motion Tenstorrent just de-prioritized (developer pivot + layoffs) is exactly the muscle needed to convert them.
3. The comp set is a warning. Groq got absorbed (~$20B Nvidia license), SambaNova took a down-round to $2.2B (from $5.1B), and Cerebras — the "winner" — needed a public listing to crystallize value. The independent-standalone path is where accelerator startups go to be repriced.

Pre-mortem (18 months out, thesis broke): Software maturity slipped another year; two marquee design-ins pushed out; a Qualcomm/Intel deal failed to close (regulatory or price), removing the M&A floor; export controls clipped the China/sovereign pipeline; the next raise came as a flat-or-down round as the AI-hardware funding window narrowed — and Tenstorrent got repriced toward SambaNova's outcome, not Cerebras's.

Are multiples too high? On today's revenue, unambiguously yes (~21–60× bookings). On option value (IP + team + M&A + sovereignty tailwind), defensible but fully priced — there is little margin of safety at $3.2B and none at $10B.

Contrarian view (what the market refuses to see): The consensus frames Tenstorrent as "the Nvidia challenger." It is more accurately an IP-licensing/CPU company (an Arm/SiFive challenger) that happens to sell accelerators — and that reframing is more bullish on the durable business (royalties, sovereignty) and more bearish on the hardware narrative the hype is priced on. The real prize is Ascalon/Babylon RISC-V royalties, not beating Nvidia on tokens/sec.

Devil's Advocate (short-seller)

Dismantling the bull case:

• Revenue concentration + quality. A handful of strategic names (Samsung/LG/Hyundai/Rapidus) who are also investors — related-party-adjacent demand that may not generalize to arm's-length enterprise buyers. If two design-ins slip, the "$150M bookings" story cracks.
• The moat is a liability in disguise. "Open" is a distribution wedge, not a moat — anyone can fork RISC-V and MIT-licensed software. Tenstorrent's differentiation reduces switching costs, the very thing that makes CUDA a moat. It may be building the commodity layer of AI compute.
• Most dangerous competitor bulls underestimate: Qualcomm — as a rival, not a buyer. Qualcomm's own $14B RISC-V + open-compiler + AI200/AI250 inference push (Jun'26) means the suitor is also building the exact thing Tenstorrent sells. If the deal doesn't close, Qualcomm competes — with far more capital and a shipping mobile/edge channel. Add SiFive ($3.65B, better-capitalized pure-play RISC-V IP) taking licensing deals, and Arm defending Neoverse.
• Capital-allocation/incentive flags: promotional benchmark/TCO claims without independent production verification; a firmware core-count downgrade after sale; investor-customers whose marks may not reflect arm's-length value; opacity that prevents any real diligence.
• What must hold for today's price: software maturity closes materially; ≥several design-ins convert to volume royalties; the funding/M&A window stays open; export controls don't clip the sovereign/China pipeline; and Keller stays.
• If growth disappoints 20–30%: a flat/down next round, the M&A floor becomes the only support, and the mark compresses toward the SambaNova precedent.
• Single scenario that permanently impairs: a failed Qualcomm/Intel deal + a stalled software roadmap + a narrowing AI-hardware funding window arriving together — the name loses its exit floor precisely when it needs the most capital, and re-rates down.

Management Questions (ordered by information value)

1. Of the ~$150M in customer contracts, what share is recognized recurring royalty vs one-time license/NRE, and what's the contracted royalty backlog that converts over the next 8 quarters?
2. What is the concrete plan and timeline to close the software-utilization gap (40–60% → parity), and what independent, production-scale benchmark will you stand behind?
3. Are you optimizing for IPO or acquisition — and what specifically would make you sell to Qualcomm/Intel vs stay independent?
4. What is current cash runway at present burn, and are the reported ~$800M Fidelity round proceeds in the bank (closed) or in-progress?
5. How do you convert investor-customer design-ins (Samsung/LG/Hyundai) into arm's-length, third-party licensing revenue?
6. How much of your pipeline is export-control-exposed (China/UAE), and what's the contingency if IP-flow restrictions tighten?
7. What is the royalty rate and per-unit economics of an Ascalon/Tensix license at volume — the number that defines the terminal business?
8. Why de-prioritize enterprise sales via the developer pivot when royalties depend on the enterprise/OEM motion?
9. On what nodes/foundries (Samsung SF4X, Rapidus 2 nm, GF) is the roadmap committed, and what's your yield/exposure if Samsung or Rapidus slips?
10. What is your defensible answer to "open = commoditized" — where are the switching costs in a RISC-V + MIT-licensed stack?
11. Post-Groq (absorbed) and SambaNova (down-round), why is Tenstorrent's independent path different?
12. What is headcount trajectory after the 7.5% cut, and are further reductions planned for H2 2026?
13. What did the firmware core-count downgrade reveal about yield/binning, and how do you rebuild benchmarking trust?
14. What are Babylon's committed IPC/perf targets and tape-out date, and who are the lead licensees?
15. What insider/founder ownership and lock-ups exist, and how are incentives aligned through an IPO or sale?

---