Etched AI ($21B), Fractile and DensityAI vs Nvidia for LLM inference: products, claims vs evidence, funding, and the public exposure via $NVDA, $MU, $TSM.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
Etched AI is a $21B private chip startup selling rack-scale inference clusters. Its first rack runs at Jane Street, and it claims more than 80% of peak FLOPs at under half the voltage of a normal AI chip. Fractile (UK) and DensityAI (ex-Tesla Dojo) are betting on DRAM bandwidth instead, and neither has shipped silicon. As of October 30, 2026, none of the three has a single independent benchmark, and Nvidia ($NVDA) already sells a disaggregated inference rack built on Groq's technology.
👉 Track the public side of this fight: the Custom Silicon basket holds Nvidia, AMD, Cerebras, Broadcom, Marvell and the ASIC design houses, and the Memory Supercycle basket holds the DRAM and HBM makers every one of these chips depends on.
Search "Fractile vs Density chips" and Google's AI Overview tells you Fractile does in-memory compute, while "Density AI chips (e.g., Etched)" hard-wire transformer math into silicon, hit 80%+ of peak FLOPs and run math blocks at under half the standard voltage.
Most of that is wrong, and it's wrong in an instructive way. The overview merged three companies into two:
Here is each claim checked against the record:
| AI Overview claim | Who actually said it | Verdict |
|---|---|---|
| Fractile does "in-memory compute" | Fractile's 2022–2024 SRAM design | Outdated. Fractile has since moved to logic plus DRAM |
| Fractile partners on custom high-bandwidth DRAM | CEO Walter Goodwin on No Priors; WSJ | Accurate as a plan. Partners and specs are undisclosed |
| Fractile serves thousands of tok/s at ultra-long context | Fractile's targets | False as stated. Nothing ships before 2027 |
| "Density AI chips (e.g., Etched)" | Two unrelated companies | Conflation |
| Transformer/MoE math hardened into silicon | Etched's 2024 Sohu pitch | Partly outdated. Etched now claims MoE and Mamba support |
| 80%+ peak FLOPs utilization | Etched | Unverified vendor claim |
| Under half the standard voltage | Etched | Unverified vendor claim |
Etched's product changed a lot between 2024 and 2026. Sohu, announced in 2024, was a single ASIC that hard-coded the dense transformer graph, built on TSMC's 4nm-class N4P process with 144 GB of HBM3E per chip (TechTimes).
The 2026 product is a rack-scale "frontier inference cluster" with two parts (etched.com):
COO Robert Wachen describes it that way himself: low voltage for prefill, shared memory for decode, and full systems rather than chips (TechCrunch). Etched dropped the "transformer-only" positioning and now says its systems run DeepSeek, Qwen, Mamba and Llama.
Every Etched performance number comes from Etched. The famous 2024 claim was that an 8-chip Sohu server does more than 500,000 tokens/sec on Llama-70B and "replaces 160 H100s." That was extrapolated from MLPerf H100 results at 2,048 input and 128 output tokens with large batches (SDxCentral). An independent breakdown by Peak FLOPS found the ~20x advantage plausible at high batch and short context, but falling to about 1.4x at batch size 1, where decode is bound by HBM bandwidth (Peak FLOPS). Those numbers describe a design Etched has since reworked anyway.
The only outside signal is Jane Street. It tested the chip, said it was "pleased with the early results" and runs the first rack in its own data center (Etched). Jane Street is also Etched's first customer and the lead investor in its latest round, so the endorsement is real but not arm's-length.
The funding moved much faster than the evidence:
| Round | Date | Amount | Valuation | Lead |
|---|---|---|---|---|
| Series A | June 2024 | $120M | n/a | Primary |
| Growth round | December 2025 | $500M | $5B | Stripes |
| Series C | July 23, 2026 | $300M | $10.3B | Sequoia, with SK Hynix |
| Series D | August 18, 2026 | $700M | $21B | Jane Street |
In October, Etched was reportedly fielding offers at $40B to $50B (TechCrunch). It claims more than $1B of signed contracts, which are purchase orders rather than revenue. Its software stack is proprietary, with no CUDA, vLLM or TensorRT compatibility, and first shipments slipped about 18 months, from a late-2024 target to summer 2026.
Fractile, founded in London in 2022 by Oxford PhD Walter Goodwin, started out building an SRAM in-memory-compute chip in the same family as Groq and Cerebras (eeNews Europe).
Then it changed course. On the No Priors podcast in October 2026, Goodwin said SRAM would not scale as context windows grew. In a "skunk works" effort with memory vendors and foundries, the company moved to a logic chip with high-bandwidth access to cheaper DRAM, on the argument that data-center economics come down to cost per gigabyte of memory (No Priors transcript). The WSJ, as summarized by SiliconANGLE, says the product uses neither conventional HBM nor on-chip SRAM (SiliconANGLE).
The headline multiplier has shrunk as the design changed:
Goodwin didn't say which HBM chip the 25x is measured against, and it matters. Against a GB300 GPU's ~8 TB/s, 25x is about 200 TB/s per chip. Against Rubin's 19.2 TB/s it is about 480 TB/s. Both are above the 150 TB/s Nvidia quotes for the on-chip SRAM of one Groq 3 LPU (NVIDIA). That would be a remarkable result for DRAM, and no bandwidth, capacity, interface or packaging spec has been published to support it.
The commercial proof is one reported customer. Bloomberg reported an initial deal for Anthropic to buy about $250M of Fractile chips for 2027 delivery; neither company commented (The Next Web). Fractile then raised:
Fractile has about 150 staff across London, Bristol, San Francisco and Taipei. Goodwin says the platform will "ramp in the second half of next year," which means 2H 2027. The SRAM-to-DRAM pivot effectively restarted the product clock, so there is little room for a silicon respin before the Anthropic delivery.
DensityAI is the "Density" in the search query, and it is a real company, just not the one the AI Overview described. Ganesh Venkataramanan (who conceived Tesla's Dojo), Bill Chang and Ben Floering founded it in Mountain View. In August 2025 Tesla shut Dojo after about 20 engineers left to join the startup (remio.ai). In July 2026 DensityAI hired Tesla's former senior director of AI hardware design to lead packaging and systems (Electrek).
The Information reported in September 2026 that DensityAI was raising "hundreds of millions" at about $10B, led by a16z. It also reported an AWS purchase agreement that only applies if the chips hit performance targets; Amazon ($AMZN) declined to comment (Crypto Briefing).
What DensityAI has said about the product:
Architecturally DensityAI is closer to Fractile than to Etched. Both are memory-capacity-plus-bandwidth bets aimed at long-context inference. It made no voltage claims and no 80% FLOPs claim.
Three things make Nvidia's bar higher than the startups' pitch decks suggest.
Nvidia's installed base keeps getting faster in software. On the same GB300 NVL72 hardware, Nvidia's MLPerf DeepSeek-R1 Server result went from 2,907 to 8,064 tokens/sec per GPU (2.77x) between MLPerf v5.1 and v6.0 (NVIDIA). CoreWeave's own table shows 5,574 for the same test, so treat the top number as a best case (CoreWeave). Vera Rubin's first MLPerf preview shows roughly 1.6x to 2.5x GB300 per GPU (NVIDIA). Etched's 2024 baseline was the H100. A 2026 claim has to beat a GB300 that has more than doubled its own throughput since launch.
Nvidia adopted the startups' architecture. In December 2025 it licensed Groq's LPU technology and hired most of Groq's team, reportedly for about $20B (Reuters via Yahoo). By March 2026 it had shown the Groq 3 LPX rack: 256 SRAM-only LPUs with 128 GB of SRAM and 40 PB/s of bandwidth. Vera Rubin GPUs handle prefill and attention, the LPUs run the latency-critical expert layers, and Nvidia's Dynamo software splits the work (NVIDIA). That is the same prefill/decode split Etched built.
Artificial Analysis measured LPX at a median 3,431 output tokens/sec on Gemma 4 31B at 100K context, against 870 for the fastest public endpoint (NVIDIA). So "thousands of tokens per second at long context," the line Google gave Fractile, is something Nvidia has already shown, with third-party measurement, on a mid-size model. Fractile's opening is model size. An SRAM-only rack with 128 GB holds only about 128B parameters at FP8, while Fractile says it targets multi-trillion-parameter models. For context, Cerebras ($CBRS) already serves gpt-oss-120B at about 1,660–1,770 tokens/sec per user (Artificial Analysis), above Fractile's 1,200 tokens/sec target. That target only matters on much bigger models.
Nvidia's memory roadmap has cracks, and they keep challengers alive. Nvidia quietly cut Rubin's per-GPU HBM4 bandwidth from 22 to 19.2 TB/s (-13%) and NVLink 6 from 3.6 to 3.0 TB/s (TrendForce). That puts AMD's ($AMD) Helios ahead on published memory bandwidth. Rubin CPX, Nvidia's GDDR7 long-context prefill chip, is off the roadmap (Wccftech). Memory supply is the binding constraint, and that is the gap Fractile and DensityAI are built for. Etched's 80% FLOPs claim targets a different gap: GPU utilization on inference, which TechTimes puts at 30–40%. Even if Etched's number holds up, it helps most in compute-bound prefill and high-batch serving, which is where Nvidia's FP4 throughput is strongest.
| Etched | Fractile | DensityAI | Nvidia GB300 NVL72 | Nvidia Vera Rubin NVL72 | Nvidia Groq 3 LPX | |
|---|---|---|---|---|---|---|
| Core idea | LVI prefill chips plus pooled HBM+SRAM decode | Logic chip plus high-bandwidth DRAM | "SRAM-like bandwidth, DRAM-like capacity," 3D DRAM | 72 Blackwell Ultra GPUs, HBM3E | 72 Rubin GPUs, HBM4 | 256 SRAM LPUs beside Rubin |
| Key spec | A0 silicon on TSMC N4P; 2026 specs unpublished | None disclosed | None disclosed | 20 TB HBM3E, 576 TB/s | 20.7 TB HBM4, 1,400 TB/s | 128 GB SRAM, 40 PB/s |
| Headline claim | >80% peak FLOPs at under half voltage | 1,200 tok/s target; 25x HBM bandwidth | ≥90% utilization target | 8,064 tok/s/GPU DeepSeek-R1 (MLPerf v6.0) | ~1.6–2.5x GB300 (MLPerf preview) | 3,431 tok/s/user at 100K context |
| Independent evidence | Qualitative only (Jane Street, also an investor) | None | None | MLPerf, InferenceX | MLPerf preview | Artificial Analysis |
| Status, Oct 2026 | First rack at Jane Street | Pre-silicon; ramp 2H 2027 | No tape-out disclosed | Volume platform | Ramping | Early access |
| Valuation | $21B; offers at $40–50B | ~$6.5B+ | ~$10B (reported) | Public ($NVDA) | Public | Groq license ~$20B (reported) |
| Anchor customer | Jane Street | Anthropic (~$250M, reported) | AWS (conditional) | Hyperscalers, neoclouds | Hyperscalers | Not disclosed |
| Software | Proprietary, no CUDA/vLLM | Undisclosed | Undisclosed | CUDA, TensorRT-LLM, Dynamo | Same | Inside CUDA/Dynamo |
Nvidia system specs are from the GB300 and Vera Rubin product pages.
None of the three startups is listed, so there is no Etched stock to buy. The public exposure runs through their suppliers and through Nvidia, and memory is the common thread. Etched's decode tier is pooled HBM plus SRAM, Fractile's whole pitch is DRAM bandwidth, and DensityAI stacks DRAM. Nvidia, meanwhile, had to cut Rubin's bandwidth because of memory constraints.
*The Custom Silicon basket on Macroplane, equal-weighted, two-year view. It holds the listed chip designers and ASIC design houses competing for the same inference budgets.*
For Nvidia, the near-term threat from these three is smaller than their valuations suggest. None ships in volume before 2027, and Nvidia has already productized the disaggregated, SRAM-assisted decode design they are pitching. The Groq deal also shows the most likely exit for a startup whose chip complements GPUs rather than replacing them: a license or an acquihire.
Etched is a US startup that builds AI inference hardware. It started with Sohu, a transformer-only ASIC, and in 2026 shifted to rack-scale clusters that split inference into a low-voltage prefill tier and a shared HBM+SRAM decode tier. It has raised about $2B and was valued at $21B in August 2026.
No. Etched is private and has not announced an IPO. Its listed investor SK Hynix ($000660.KS) and its foundry TSMC ($TSM) are the closest public exposure.
No. DensityAI is a separate startup founded by former Tesla Dojo leaders. Google's AI Overview attached Etched's "low voltage, 80%+ FLOPs" claims to DensityAI's name, probably because coverage described Etched's design as raising "FLOP density."
Unknown. Fractile claims up to 25x the bandwidth of an HBM chip and targets 1,200 tokens/sec per user, but it has shipped no silicon of its current DRAM-based design and has published no benchmarks. Its platform is due to ramp in the second half of 2027.
Not before 2027, because none ships in volume until then. Nvidia has already built a prefill/decode split with its Groq 3 LPX rack, which was independently measured at about 3,400 tokens/sec per user at 100K context. The startups' best openings are very large models and memory cost per token.
No. This article is research and education based on public reporting as of October 30, 2026. Vendor performance claims are flagged as unverified. Do your own work before investing.
Etched is a US startup that builds AI inference hardware. It started with Sohu, a transformer-only ASIC, and in 2026 shifted to rack-scale clusters that split inference into a low-voltage prefill tier and a shared HBM+SRAM decode tier. It has raised about $2B and was valued at $21B in August 2026.
No. Etched is private and has not announced an IPO. Its listed investor SK Hynix ($000660.KS) and its foundry TSMC ($TSM) are the closest public exposure.
No. DensityAI is a separate startup founded by former Tesla Dojo leaders. Google's AI Overview attached Etched's "low voltage, 80%+ FLOPs" claims to DensityAI's name, probably because coverage described Etched's design as raising "FLOP density."
Unknown. Fractile claims up to 25x the bandwidth of an HBM chip and targets 1,200 tokens/sec per user, but it has shipped no silicon of its current DRAM-based design and has published no benchmarks. Its platform is due to ramp in the second half of 2027.
Not before 2027, because none ships in volume until then. Nvidia has already built a prefill/decode split with its Groq 3 LPX rack, which was independently measured at about 3,400 tokens/sec per user at 100K context. The startups' best openings are very large models and memory cost per token.
No. This article is research and education based on public reporting as of October 30, 2026. Vendor performance claims are flagged as unverified. Do your own work before investing.