· investment-strategies · 6 min read
What Is Etched’s Valuation in 2026? $21B After $700M
Etched closed $700M at $21B on August 18, 2026, led by Jane Street after the trading firm tested and installed a rack. That doubles July’s $10.3B Series C in a month. Revenue was not disclosed.
Etched is worth $21 billion after a $700 million round on August 18, 2026, led by Jane Street. That is one month after a $300 million Series C at $10.3 billion. Jane Street is also the first named customer — a rack is already in its datacenter.
Unexpected truth: the lead is not a Sand Hill fund writing a thesis check. It is a trading firm that tested the silicon, bought a rack, then led. Revenue was not disclosed.
This page is for people deciding whether Etched (or custom inference silicon) is worth another hour. Last verified August 19, 2026. Fact vs our read is labeled below.
Five-minute decision
| If you need… | Verdict |
|---|---|
| What happened | Funded. $700M @ $21B, Jane Street lead, Aug 18, 2026. First rack shipped to the lead. |
| What Etched is | Rack-scale inference clusters (prefill + decode silicon), not a GPU neocloud. |
| Whether it is “shipping” | One named rack. Jane Street. Order-book conversion beyond that is still company language. |
| Whether to diligence | Yes, if you underwrite custom inference vs NVIDIA. No, if you need disclosed revenue before a first call. |
Investigate further when: you believe a buyer-lead that already runs the hardware is a stronger signal than another software round at a similar mark.
Wait or pass when: your bar is public $/token benchmarks, or you think Groq’s NVIDIA-architecture cloud already occupies the inference spend you care about.
What happened
| Field | Detail |
|---|---|
| Company | Etched (etched.com); San Jose / Silicon Valley |
| Round | $700M · $21B valuation (company, Aug 18) |
| Lead | Jane Street (customer + investor; no /fund/ page here) |
| Directory funds named | Kleiner Perkins, Sequoia, a16z, Tiger Global, Bain Capital Ventures, Neo |
| Other names in the post | Peter Thiel, Stripes, Primary, Positive Sum, Diffusion, Argo, Blackstone |
| Traction disclosed | First rack shipped to Jane Street; Jane Street quote on early results |
| Prior | $300M Series C @ $10.3B, July 23, 2026, Sequoia-led |
| Use of proceeds | Company: accelerate production; ramp toward “Gigawatt-scale” |
The company post does not label this round Series D. TechCrunch and Dealroom call it Series D. Treat $700M / $21B as company fact and Series D as press taxonomy.
What Etched actually does
Product (company): full frontier inference clusters — systems, not loose ASICs. Co-founder Robert Wachen told TechCrunch the stack still splits inference into prefill (understand the prompt) and decode (emit tokens).
- Low-voltage prefill silicon → denser transistors, less heat.
- Cluster-scale memory for decode → many chips share a fast pool.
Early Etched marketing implied one model etched into silicon. The company now says systems run any frontier model. That is a product claim, not a public benchmark pack.
Job: cheaper, faster token serving than general-purpose GPU fleets for production inference.
How big is Etched actually?
| Question | Disclosed? | What we have |
|---|---|---|
| Valuation | Yes | $21B (company, Aug 18) |
| Round size | Yes | $700M |
| Prior mark | Yes | $10.3B (July 23 Series C) |
| First customer | Yes | Jane Street (named rack) |
| Revenue / margins | No | Not in the Aug 18 post |
| Remaining order book | Prior claim | ~$1B orders booked as of the July vintage — not restated as a new audited figure here |
| Total raised | Company wording | HPCwire reprint of the announcement cites $1.9B cumulative — do not rebuild the cap table from headlines |
| Headcount / fab status | Prior | July coverage: ~400 people; TSMC silicon; Milpitas facility. Not restated as new numbers on Aug 18 |
One named deployed rack is the new fact. The $21 billion mark is the price of that fact plus the July order-book story.
Worked numbers (indicative)
- Setup: Dec 2025 — about $5B. July 23 — $300M at $10.3B.
- Move: Aug 18 — $700M at $21B, Jane Street as customer-lead.
- Punch: Valuation doubled in 26 days. If $700M is new primary at $21B post, implied dilution is about 3.3% ($700M / $21B). That is arithmetic on the headlines, not a disclosed option pool or preference stack.
Same-week context (not Etched’s P&L): Groq printed $3.5B on August 17 for a neocloud after NVIDIA’s licensing deal. These are not substitutes.
Etched vs Groq — what differs
| Etched | Groq | |
|---|---|---|
| Latest round | $700M @ $21B, Aug 18 | $350M @ $3.5B, Aug 17 |
| Lead | Jane Street (customer) | Disruptive; planned NVIDIA |
| Product | Custom inference clusters | NVIDIA-architecture cloud capacity |
| What “shipping” means | One named rack (Jane Street) | 13 data centers; 6M+ developers (company) |
| Prior mark | $10.3B (July) | ~$6.9B chip-company story, then a reset |
| Main risk | Custom silicon vs CUDA gravity | Neocloud commodity vs CoreWeave/hyperscalers |
Choose Etched when you are underwriting non-GPU inference hardware. Choose Groq when you are underwriting megawatts on NVIDIA architecture. Do not mash the two $B marks.
Etched vs NVIDIA: NVIDIA is the default software and systems vendor. Etched is betting prefill/decode co-design beats general GPUs on TCO for some serving loads. No public bake-off is in the Aug 18 post.
Why investors likely wrote $700M
Known evidence (company + Jane Street quote)
- Named first customer that is also the lead.
- Jane Street: tested the chip, “pleased with the early results,” rack running in its datacenter.
- Returning Sequoia and a16z from the July C.
- New directory names on this close: Kleiner Perkins (prominent in the post), plus Tiger, Bain Capital Ventures, Neo.
- Use of proceeds aimed at production, not a first tape-out.
Our interpretation (not an IC memo)
| Bet | Why it could justify $21B |
|---|---|
| Buyer-lead | A latency-sensitive trading firm funded the factory after using the part |
| Speed of marks | $5B → $10.3B → $21B in eight months is category-FOMO priced in |
| Systems, not chips | Full racks are harder to copy than a slide of an ASIC |
| CUDA alternative | Some LPs still want a non-NVIDIA inference outcome |
What we cannot claim: that Jane Street is a large fraction of revenue, that the July ~$1B order book has converted, or that $21B is a revenue multiple.
What has to be true for Etched to win
- Jane Street’s rack becomes a repeatable sales motion, not a one-off lab.
- Software for non-transformer and MoE models stays good enough vs CUDA.
- TSMC / memory supply actually ramps toward the company’s gigawatt language.
- Hyperscalers do not absorb the same workloads on captive ASICs.
- $700M buys delivered racks, not just a higher paper mark.
What could break the thesis
- General GPUs keep winning on compiler and ecosystem even at worse $/token.
- The first customer stays the only named customer for too long.
- Model architectures move faster than Etched’s tape-out cycle.
- “Series D” press plus $1.9B cumulative language hides a messy recap.
When not to use this page
- Do not treat $21B as a July figure. July was $10.3B.
- Do not treat Jane Street as a conventional venture fund in our directory.
- Do not use Etched’s mark as Groq’s mark, or the reverse.
- Do not invent revenue from the order-book claim in the July article.
Where to go next
Read the July memo if you need the silicon story: Etched $300M Series C at $10.3B. Then Groq’s $3.5B reset if you are mapping inference cloud. Entity: Etched. Month tape: August 2026 VC news.
