semiconductorsGraveyard archive2014–2025

Esperanto Technologies

Esperanto Technologies spent a decade trying to prove that a thousand tiny RISC-V cores could beat Nvidia's brute-force approach to AI inference — then lost its engineers one by one to rivals offering four times the pay, and quietly wound down in mid-2025.

Years active

2014–2025

Total raised

$63M

Peak headcount

~140

The bet

Dave Ditzel had already changed computing once. In 1980 he co-authored "The Case for the Reduced Instruction Set Computer" with David Patterson — the paper that seeded the RISC revolution. He went on to found Transmeta, whose power-sipping x86-compatible chips drew a young Linus Torvalds as an engineer. By 2014, Ditzel was ready to place another long-odds bet: that the open RISC-V instruction set could be the backbone of the AI-chip era.

Founded in Mountain View, Esperanto operated quietly before emerging at a RISC-V Workshop with a contrarian thesis. Instead of chasing raw peak throughput with a few fat cores, Ditzel proposed tiling more than a thousand lean, energy-efficient RISC-V cores across a single die and letting the parallel nature of machine-learning inference do the rest. Do more per watt, not merely more per clock — the philosophy he had chased his whole career.

The build

Investors bought the thesis. In November 2018 Esperanto closed a $58M Series B, bringing total funding to roughly $63M, with backers including Western Digital. The money funded a serious silicon program: a 7nm system-on-chip, the ET-SoC-1, integrating 1,092 RISC-V cores (1,088 small "ET-Minion" inference engines plus four larger "ET-Maxion" cores) designed to stay under ~20 watts while delivering 100+ TOPS of inference.

The chip taped out at TSMC and was revealed at Hot Chips in August 2021 — real silicon, not a slide deck, and the largest RISC-V core count put on a single commercial chip at the time. By April 2022 Samsung SDS was running a formal evaluation and publicly praised the chip's near-linear scaling across cluster configurations. For a moment, the watts-per-inference pitch looked like it might land with the hyperscalers quietly uneasy about Nvidia's pricing power.

The wall

Evaluation is not revenue. The launch of ChatGPT in late 2022 reoriented every enterprise AI budget toward Nvidia's H100s and the CUDA ecosystem — a decade-deep software moat that Esperanto's stack could not escape. The market rewarded raw throughput and ecosystem compatibility over the energy efficiency that was Esperanto's entire differentiator.

The talent problem compounded it. CEO Art Swift later put it bluntly: deep-pocketed competitors raided the company with packages "two, three, even four times higher" than a startup with $63M to its name could match. Engineers holding the irreplaceable microarchitecture knowledge left steadily. A late pivot — a Generative AI Appliance announced in 2023 to run open LLMs on ET-SoC-1 clusters — was credible but entered an already-crowded field against Nvidia cards that got more efficient every generation.

The wind-down

By mid-2025 the attrition crossed a point of no return. Esperanto wound down its silicon business, cutting roughly 90% of its Mountain View staff — from a peak near 140 — and closing its European subsidiaries in Spain and Serbia. Swift and a skeleton crew stayed to find a buyer or IP licensee.

In November 2025, Ainekko, an open-source AI-hardware startup, acquired Esperanto's full IP portfolio — RTL, tooling, frameworks, reference designs — and committed to open-sourcing the many-core RISC-V architecture on GitHub. The ET-SoC-1 design would live on, just not as a commercial product. Swift noted that by the time of the closure, about 95% of former employees had already found new roles. A decade-long bet on a more elegant way to do inference ended not with a failed chip, but with a company that simply could not out-pay the giants it had hoped to out-engineer.

What worked, what broke

What worked

Taped out the ET-SoC-1 at TSMC in 7nm — 1,092 RISC-V cores, the most ever integrated on a single commercial chip at its 2021 debut
Demonstrated near-linear performance scaling in a Samsung SDS evaluation, earning rare public praise from a hyperscaler-tier customer
Validated energy-efficient ML inference well under 20W per chip on real silicon, proving out the low-power many-core thesis
Built a production-grade software toolchain complete enough that Ainekko acquired and open-sourced it as a full platform in 2025
Sustained an 11-year silicon program across two AI eras on only ~$63M — lean by semiconductor-startup standards

What broke

Talent hemorrhage: rivals poached engineers at 2–4x Esperanto's pay ceiling, eroding irreplaceable microarchitecture expertise
The generative-AI boom rewarded raw throughput and CUDA compatibility over watts-per-inference, negating Esperanto's core differentiator
Undercapitalized: ~$63M total could not fund both competitive salaries and the multi-year silicon iteration needed to reach scale customers
Evaluation-to-revenue gap: traction gated by evaluation programs rather than committed purchase orders, just as speed mattered most
Nvidia's CUDA moat made switching costs prohibitive for enterprises already running GPU inference pipelines

Sources

Obituary authored Jun 4, 2026 via Sonnet 4.5 + web_search.