AURA-1 chip design archive: docs, blog, research folders, tiny-gpu work, GDS viewer
7858a76 verified Olix (Olix Computing Ltd) — Research Notes
Compiled 2026-08-10
Company overview
- London chip startup, founded March 2024 by James Dacombe (b. 2000, Harrogate; second-time founder — also founded CoMind, non-invasive brain-monitoring neurotech).
- Building rack-scale AI inference systems, designed end to end: chips, lasers, and the network connecting them.
- Two contrarian bets:
- No HBM — on-chip SRAM instead (closer in spirit to Groq/Cerebras).
- No copper interconnect — a novel photonic interconnect links the chips.
- Also skips advanced packaging (no CoWoS-style interposers) — deliberately avoids the parts of the supply chain with multi-year queues.
- First chip: DX-1 (X-1 platform) — a decode accelerator for the output-generation stage of LLM inference. Claim: >10,000 tokens/sec/user on ~100B-parameter models, better throughput/watt than general-purpose GPUs. Each chip in the platform is specialized for one stage of the inference pipeline.
- First customer access targeted H2 2027 (pre-silicon as of Aug 2026).
Funding
| Round | Date | Amount | Valuation |
|---|---|---|---|
| Series A | Feb 2026 | $220M | ~$1B |
| Series B | Aug 2026 | $312M | $3.3B |
- Series B is the largest chip funding round in European history (~$530M total raised).
- Backers: Arm, Hudson River Trading, Fundomo, Reed Hastings (Netflix co-founder), UK government's Sovereign AI venture fund.
- Key people: Prof. Nick McKeown (Stanford, ex-Intel SVP) joined the board; Matt Briers (ex-Wise CFO) is CFO.
Patents
No company-published list exists. Google Patents (applicant = "Olix Computing") shows two published applications, both PCT/GB filings, same inventor, same priority date — filed ~5 months after founding:
| Publication | Title | Priority | Filed | Published | Inventor |
|---|---|---|---|---|---|
| WO2026033226A1 | Method and system for optical computing | 2024-08-08 | 2025-08-08 | 2026-02-12 | Dawid Borycki |
| WO2026033227A1 | Method and system for optical computing | 2024-08-08 | 2025-08-08 | 2026-02-12 | Dawid Borycki |
- WO2026033226A1 = PCT/GB2025/051752. Classifications: G02F1/21, G02F1/225, G02F1/212 (Mach-Zehnder-type interference optics), G02B6/28 (optical coupling/bus waveguides).
- Substance (226): an optical data combiner using frequency-swept coherent light — light split into two paths, each phase-modulated by an input data stream, one path delayed, then recombined; the interference pattern encodes the combination and can implement logic gates and arithmetic (add/subtract/multiply/divide). Notably this is genuine optical computing, not just interconnect.
- Both have GB counterparts; legal status pending.
- Caveat: applications publish 18 months after filing, so anything filed after ~Feb 2025 is still invisible. Re-check Google Patents for "olix computing" every few months.
- Local copies:
WO2026033226A1.pdf/WO2026033227A1.pdf(full text incl. description + claims, rendered from Google Patents pages; the.htmlfiles are the originals). Official WO pamphlet PDFs were not yet available on Google Patents / WIPO as of this date.
Comparison vs Lightmatter
| Olix | Lightmatter | |
|---|---|---|
| Founded | 2024, London | 2017, MIT spinout (Boston) |
| Valuation | $3.3B (Aug 2026) | $4.4B (Series D, Oct 2024) |
| Total raised | ~$530M | ~$850M |
| Stage | Pre-silicon; customers H2 2027 | Passage interconnect (32–64 Tbps CPO) shipping 2026; Synopsys/Cadence integrations; in Nvidia NVLink Fusion ecosystem |
| Model | Vertically integrated inference systems — competes with Nvidia | Sells photonic interconnect into the GPU ecosystem — partners with Nvidia |
| Photonics role | Interconnect fabric (compute is electronic/SRAM) | Interconnect (pivoted away from photonic compute / Envise) |
Similar valuation league, very different risk: Lightmatter is ~FEED-complete with vendors signed; Olix is a well-funded conceptual design package. Market prices them similarly because Olix's addressable prize (whole systems) is larger.
Why they moved so fast
- Narrow problem — decode-only accelerator, not a general GPU; small design space per chip.
- Design choices dodge slow supply chains — no HBM, no advanced packaging; novelty concentrated in the interconnect, individual dies comparatively conventional.
- Capital + talent compress schedules — $530M in 18 months, veteran hires, second-time founder; peak inference demand + UK sovereign-AI push meant money arrived fast.
- Calibration — the fastest-moving thing is capital formation; performance figures are projections, no shipped silicon. Cf. Graphcore at the same stage.
Sources
- https://olix.com/news/company-raises-series-b
- https://www.thestack.technology/olix-fund-raise-british-ai-chips/
- https://www.datacenterdynamics.com/en/news/chip-startup-olix-raises-312m-at-33bn-valuation-backed-by-uk-govt-sovereign-ai-venture-fund/
- https://rcrtech.com/semiconductor-news/olix-nvidia-by-ditching-hbm-and-copper/
- https://techfundingnews.com/olix-220m-unicorn-photonic-ai-chips-inference/
- https://businesscloud.co.uk/news/james-dacombe-profile-meet-the-uks-25-year-old-unicorn-boss/
- https://thenextweb.com/news/olix-312m-series-b-3-3bn-uk-optical-ai-chip-nvidia-rival
- https://patents.google.com/patent/WO2026033226A1/en
- https://patents.google.com/patent/WO2026033227A1/en
- https://lightmatter.co/press-release/lightmatter-raises-400m-series-d-quadruples-valuation-to-4-4b-as-photonics-leader-for-next-gen-ai-data-centers/