Semiconductors
PrivateThe cleanest pure-play on the "inference is a distinct, non-GPU category" thesis Nvidia itself just validated by paying $20B for Groq — but Corsair's SRAM-centric bet is a working solution for ≤100B dense models in the exact quarter the frontier pivoted to memory-hungry reasoning/agentic workloads, so the whole equity rides on Raptor/3DIMC landing on time in 2026. WATCHING, not yet ownable (private; no tradeable security).
Research
The verdict
The cleanest pure-play on the "inference is a distinct, non-GPU category" thesis Nvidia itself just validated by paying $20B for Groq — but Corsair's SRAM-centric bet is a working solution for ≤100B dense models in the exact quarter the frontier pivoted to memory-hungry reasoning/agentic workloads, so the whole equity rides on Raptor/3DIMC landing on time in 2026. WATCHING, not yet ownable (private; no tradeable security).
d-Matrix is a Santa Clara AI-inference-silicon startup, founded 2019 by Sid Sheth (CEO) and Sudeep Bhoja (CTO) — both ex-Inphi, the DSP/interconnect company Marvell bought for ~$10B. The company builds a full-stack datacenter inference platform (explicitly not training), sold as PCIe cards + networking + software + a rack reference design, not as a cloud service.
What it actually sells (the product stack):
Business model. Hardware unit sales (cards + NICs + racks via OEMs like Supermicro) plus a software/services attach that management wants to grow to a ~$100M ARR pre-IPO milestone. The core pitch to buyers: at rack level, Corsair (increasingly paired with GPUs) delivers ~10× faster token generation, ~3× lower cost, and 3–5× better energy efficiency than GPU-only setups for latency-sensitive workloads.
Customers / channel. Named segments only — "hyperscalers, neoclouds, and frontier AI labs" as priority, plus sovereign/regional clouds and telco/edge. No individual customer names are public (n/a — private, not disclosed). Microsoft is a strategic investor (M12) but has not been disclosed as a Corsair buyer. Channel-to-market is OEM/system-integrator led (Supermicro), with named tech partners Arista + Broadcom (networking) and manufacturing partners TSMC + Alchip.
Suppliers. TSMC (N6 wafers), Alchip Technologies (ASIC design/backend + the 3D-DRAM co-development), LPDDR5X memory suppliers (Samsung Ventures is an investor — a possible supply tell but unconfirmed as a supplier), Broadcom/Arista silicon for the rack fabric.
Upstream inputs → d-Matrix → end customer, named at every hop (web-only; a private co discloses less than a 10-K would, but the partner set is unusually well-named for a startup):
Chokepoints: (1) TSMC single-source foundry; (2) Alchip on the 3D-DRAM roadmap; (3) advanced-packaging capacity (chiplet + eventual 3D stack) is itself supply-constrained industry-wide. Not exposed to the HBM allocation war — a real, differentiated supply-chain advantage vs GPU vendors, and one of the more defensible lines in the whole thesis.
The real, durable edges:
The moat that is NOT yet real — and it is the whole ballgame:
Bargaining power. Weak-to-neutral today. As a pre-scale buyer of TSMC N6 capacity and packaging, d-Matrix has little leverage on suppliers; against customers it is the unproven insurgent, so it must give value (lower price, engineering support) to win designs. That flips only if a hyperscaler standardizes on it.
n/a — private, not disclosed. No segment P&L exists (no segments.csv rows on the research layer; company files nothing). Directionally, revenue is ~100% datacenter inference hardware today, with a software/services line management wants to grow toward ~$100M ARR pre-IPO. Geographically the disclosed push is North America core + EMEA and APAC expansion for sovereign/energy-sensitive buyers, with engineering/support hubs described as spanning multiple continents (India/Australia referenced in third-party profiles, not company-confirmed). Any hard segment split would be fabrication — flagged as not-sourced.
The round history is the scoreboard for a private:
| Round | Date | Amount | Lead / notable | Post-money valuation |
|---|---|---|---|---|
| Seed + early | 2019–2021 | ~undisclosed | Playground Global (incubated) | n/a — not disclosed |
| Series A | Apr 2022 | $44M | Playground Global, M12 (Microsoft) | n/a — not disclosed |
| Series B | Sep 2023 | $110M | Temasek (+ M12, Playground, Nautilus, Entrada, Samsung Ventures, Mirae, Ericsson…) | not disclosed; "heavily oversubscribed" |
| Series C | Nov 2025 | $275M | BullhoundCapital, Triatomic, Temasek co-lead; new: QIA, EDBI; follow-on M12, Nautilus, Industry Ventures, Mirae | $2.0B |
Total raised: ~$450M over 3 priced rounds, from ~27–28 investors, at a $2B post as of Nov 2025. The Series B→C step-up cannot be computed as a clean multiple (Series B post undisclosed → n/a), but the $44M→$110M→$275M cadence with escalating sovereign-fund participation (Temasek repeat + QIA + EDBI) is the classic pre-IPO capital-stacking signature.
Burn signal (inferred, not disclosed): 283 employees as of Apr 30 2026, a leading-edge TSMC N6 tapeout, a second 4nm tapeout in flight (Raptor), plus a cash acquisition of GigaIO's datacenter business — this is a heavy-burn hardware company. $275M raised Nov 2025 against a from-scratch silicon + software + fabric buildout implies a runway measured in a small number of years, not many; expect another raise before or into an IPO window (see Lens 11). All burn figures ``; none disclosed.
The single most important recent "print" isn't a funding round — it's Corsair entering full production on 2026-06-09 with volume shipments to priority customers slated for summer 2026. For a hardware startup, "in volume production" is the equivalent of the first real revenue quarter.
No earnings calls exist; the tape is founder interviews + launch cadence. The arc over ~18 months:
Sentiment read: consistently high-conviction and increasingly validated by the field — but the substance has migrated from "replace the GPU" → "solve memory" → "own the rack" → "complement the GPU." That migration is either admirable adaptability or a moving goalpost; a skeptic reads it as the pure-play SRAM thesis quietly narrowing as reasoning workloads scaled.
Syndicate quality — the IPO-proximity tell. d-Matrix's cap table is stacked with exactly the investor types that precede an IPO:
Secondary marks: n/a — not publicly disclosed.Mechanism/peer comps (the relevant comparison for a private chip co is architecture + valuation vs the cohort, not P/E):
| Company | Approach | Status / valuation | Read-through |
|---|---|---|---|
| d-Matrix | DIMC (compute-in-SRAM) + LPDDR5X, chiplet, moving to 3D-DRAM | Private, $2.0B (Nov 2025) | The subject |
| Groq | Deterministic LPU, SRAM-heavy | Nvidia licensed tech ~$20B, acqui-hired founder Dec 2025; cloud stays independent | Direct validation of the low-latency-inference category; also removes the loudest competitor |
| Cerebras | Wafer-scale | IPO'd May 2026, ~$80B intraday peak | The comp that shows the public window is open and can be enormous |
| SambaNova | Reconfigurable dataflow (RDU) | Once $5B (2021); Intel signed non-binding term sheet ~$1.6B, 2026 | The cautionary comp — private AI-chip marks can round-trip down ~68% |
| Etched | Transformer-only ASIC | Out of stealth 2026-06-30, >$1B customer contracts, ~$800M raised | The disruptor threat — even more specialized than d-Matrix |
| Tenstorrent / MatX / Positron / Rebellions | Various inference ASIC/RISC-V | Private, repeatedly financed | The category is crowded and well-capitalized |
| Nvidia | GPU + CUDA + NVLink + (now) Groq LPU IP | Incumbent, ~$3–4T-class | The gravity well every line must escape |
Valuation multiples (EV/Sales, P/E): n/a (private; no disclosed revenue). The only honest valuation anchor is the $2B Nov-2025 post and the cohort read-throughs above. At $2B on an unaudited pre-scale revenue base, d-Matrix is priced as a credible-contender option, well below Cerebras's public mark and well above SambaNova's distressed Intel number — i.e. the market currently rates it a live, un-broken bet.
No ticker, so "what moves the mark" = the events that reprice a private. Over the company's life, the value-inflection events were: (1) Series A/B/C closes (each a step-up); (2) Corsair unveil, Nov 2024 (stealth → real product); (3) Hot Chips 3D-DRAM/Raptor reveal, Aug 2025 (answered the capacity-ceiling objection with a roadmap); (4) SquadRack + partner ecosystem, late 2025 (Arista/Broadcom/Supermicro credibility); (5) GigaIO acquisition, Apr 2026 (owns the fabric); (6) full production, Jun 2026 (execution proof); and — crucially — (7) an exogenous catalyst: Nvidia's ~$20B Groq deal, Dec 2025, which validated the entire non-GPU low-latency-inference category and, per multiple analysts, sets d-Matrix up for a materially higher next round. Pattern read: the mark reacts to category validation + execution milestones, and the single biggest future re-rate catalyst is a named-hyperscaler design win or a credible S-1.
n/a — not disclosed, but founder-led-since-inception is the archetype.n/a (pre-profit).For a private company there is no income statement / balance sheet / cash-flow statement to forensically dissect — so the standard revenue-recognition/SBC/goodwill checks are n/a — no audited financials filed. What can be assessed is disclosure quality and verification integrity, and here there is one substantive flag:
Regulatory findings (from regulatory/regulatory-findings.md, fetched 2026-07-06):
n/a — no 10-K exists."d-Matrix" (FTC OR DOJ OR settlement OR fine OR penalty) enforcement returned no material enforcement, litigation, or consent-decree hits as of 2026-07-06.This is the be-early payoff lens. d-Matrix is not currently in research/private-watch.json — it should be added (see Open items); on the readiness scale (1 early → 5 S-1/IPO-imminent), it sits at roughly 3.5 — late-stage, pre-IPO but not filed.
Milestones that unlock an S-1, and where d-Matrix stands:
Estimated window: an S-1 is plausible in the ~12–24 month horizon (call it 2027–2028) if (a) Raptor ships and (b) at least one named large-scale deployment converts — but it is not imminent (no crossover round, no disclosed revenue scale, no filing signal). The nearer-term, more-likely event is a higher-priced private round (Series D at a multiple of the $2B mark), which multiple analysts flagged as probable in the wake of the Nvidia-Groq deal. A strategic acquisition is a real alternative exit — the SambaNova/Intel and Groq/Nvidia precedents show incumbents are buying inference IP; d-Matrix's owned fabric (post-GigaIO) + differentiated DIMC make it an attractive target. ``
No EPS forecast is logged (forecast.ts create skipped) — there is no share count, no revenue base, and no tradeable security to Brier-score. The tracked prediction that would matter is binary and better logged in /thesis if/when a position is contemplated: "d-Matrix ships Raptor (3DIMC) commercially before end-2026" and "d-Matrix files an S-1 or is acquired before end-2027."
Bull case. Inference is about to dwarf training — Jensen Huang himself frames eventual inference demand as "a billion times" training, with inference projected at 70–80% of AI compute by ~2035; the inference market is sized at ~$255B by 2030 (~17–19% CAGR). In that world there is room for a specialized, non-GPU, memory-efficient inference platform, and d-Matrix is the cleanest pure-play on it: a domain-perfect founding team (ex-Inphi, $10B outcome), a genuinely differentiated architecture that routes around the HBM bottleneck, a named blue-chip ecosystem (Microsoft, Arista, Broadcom, Supermicro, TSMC, Alchip), sovereign-fund capital runway (Temasek/QIA/EDBI), owned scale-up fabric (GigaIO), and — the kicker — Nvidia just paid ~$20B for Groq, validating the exact "low-latency inference is a distinct category" thesis d-Matrix is built on. If Raptor lands and one hyperscaler standardizes, the next mark is a large multiple of $2B, with an $80B Cerebras public comp overhead as the ceiling-raiser.
Bear case (2–3 permanent-impairment risks).
Pre-mortem (18 months out, thesis broke): It's early 2028. Raptor shipped a year late and its 3D-DRAM yields disappointed; reasoning models made SRAM-first inference a niche; the one hyperscaler pilot never converted to volume because re-porting off CUDA wasn't worth 3×; a Series D got done at a flat-to-down $2B or the company sold to a strategic for ~$3B — a fine outcome for founders, a mediocre one for the last-round sovereign funds. The "10×" claims aged badly because MLPerf numbers, when finally submitted, showed a real-but-narrower edge.
Multiples too high? n/a on any earnings multiple. On a category-option basis, $2B looks reasonable-to-cheap relative to the $80B Cerebras public comp and the inference TAM — but "cheap vs the euphoria comp" is exactly the trap SambaNova investors fell into at $5B.
Contrarian view (what the market is refusing to see): The consensus post-Groq-deal read is "inference-silicon is validated → d-Matrix's next round is up-and-to-the-right." The thing being under-weighted is that Nvidia validated the category by buying the leader and folding LPU IP into CUDA/NVLink — which makes the independent path harder, not easier. The most likely good outcome for d-Matrix may not be a triumphant IPO but a strategic acquisition into an incumbent that needs its DIMC + fabric — a fine but capped result the current "next unicorn round" framing isn't pricing.
Dismantling the bull case as a skeptical short would:
Best analog franchise on Earth, mid-cycle, fully priced — the FCF-inflection thesis is now consensus at ~40x forward and above Street targets; you're buying quality at a cyclical-optimism peak, with China share-loss the under-priced tail. WATCHING, not chasing.
The pure-play picks-and-shovels winner of AI-chip test, printing a vertical Q1'26 (+87%, $2.53 EPS) — but the stock fell ~14% on it because Q2 guidance steps DOWN sequentially and a ~54x P/E prices permanent acceleration; great business, demanding price, cyclical tape. NEUTRAL/WATCHING into the next print.
Best-in-class EDA franchise temporarily wearing an Ansys-debt-and-amortization disguise — the GAAP "collapse" is accounting, not the business; the real risk is paying ~35x forward for a name whose Design-IP leg is structurally cracked and whose synergy math doesn't pay until FY2028.